Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c (revision 356156) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu.c (revision 356157) @@ -1,2754 +1,2750 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2011, 2017 by Delphix. All rights reserved. * Copyright (c) 2019 Datto Inc. */ /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */ /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */ /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef _KERNEL #include #include #include #endif /* * Enable/disable nopwrite feature. */ int zfs_nopwrite_enabled = 1; SYSCTL_DECL(_vfs_zfs); SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); /* * Tunable to control percentage of dirtied L1 blocks from frees allowed into * one TXG. After this threshold is crossed, additional dirty blocks from frees * will wait until the next TXG. * A value of zero will disable this throttle. */ uint32_t zfs_per_txg_dirty_frees_percent = 5; SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN, &zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied indirect blocks from frees allowed in one txg"); /* * This can be used for testing, to ensure that certain actions happen * while in the middle of a remap (which might otherwise complete too * quickly). */ int zfs_object_remap_one_indirect_delay_ticks = 0; /* * Limit the amount we can prefetch with one call to this amount. This * helps to limit the amount of memory that can be used by prefetching. * Larger objects should be prefetched a bit at a time. */ uint64_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE; const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" }, { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" }, { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" }, { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" }, { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" }, { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" }, { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" }, { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" }, { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" }, { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" }, { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" }, { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" }, { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" } }; const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { { byteswap_uint8_array, "uint8" }, { byteswap_uint16_array, "uint16" }, { byteswap_uint32_array, "uint32" }, { byteswap_uint64_array, "uint64" }, { zap_byteswap, "zap" }, { dnode_buf_byteswap, "dnode" }, { dmu_objset_byteswap, "objset" }, { zfs_znode_byteswap, "znode" }, { zfs_oldacl_byteswap, "oldacl" }, { zfs_acl_byteswap, "acl" } }; int dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, void *tag, dmu_buf_t **dbp) { uint64_t blkid; dmu_buf_impl_t *db; blkid = dbuf_whichblock(dn, 0, offset); rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, blkid, tag); rw_exit(&dn->dn_struct_rwlock); if (db == NULL) { *dbp = NULL; return (SET_ERROR(EIO)); } *dbp = &db->db; return (0); } int dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, void *tag, dmu_buf_t **dbp) { dnode_t *dn; uint64_t blkid; dmu_buf_impl_t *db; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); blkid = dbuf_whichblock(dn, 0, offset); rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, blkid, tag); rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); if (db == NULL) { *dbp = NULL; return (SET_ERROR(EIO)); } *dbp = &db->db; return (err); } int dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, void *tag, dmu_buf_t **dbp, int flags) { int err; int db_flags = DB_RF_CANFAIL; if (flags & DMU_READ_NO_PREFETCH) db_flags |= DB_RF_NOPREFETCH; err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); if (err == 0) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); err = dbuf_read(db, NULL, db_flags); if (err != 0) { dbuf_rele(db, tag); *dbp = NULL; } } return (err); } int dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, void *tag, dmu_buf_t **dbp, int flags) { int err; int db_flags = DB_RF_CANFAIL; if (flags & DMU_READ_NO_PREFETCH) db_flags |= DB_RF_NOPREFETCH; err = dmu_buf_hold_noread(os, object, offset, tag, dbp); if (err == 0) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); err = dbuf_read(db, NULL, db_flags); if (err != 0) { dbuf_rele(db, tag); *dbp = NULL; } } return (err); } int dmu_bonus_max(void) { return (DN_OLD_MAX_BONUSLEN); } int dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int error; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (dn->dn_bonus != db) { error = SET_ERROR(EINVAL); } else if (newsize < 0 || newsize > db_fake->db_size) { error = SET_ERROR(EINVAL); } else { dnode_setbonuslen(dn, newsize, tx); error = 0; } DB_DNODE_EXIT(db); return (error); } int dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int error; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (!DMU_OT_IS_VALID(type)) { error = SET_ERROR(EINVAL); } else if (dn->dn_bonus != db) { error = SET_ERROR(EINVAL); } else { dnode_setbonus_type(dn, type, tx); error = 0; } DB_DNODE_EXIT(db); return (error); } dmu_object_type_t dmu_get_bonustype(dmu_buf_t *db_fake) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; dmu_object_type_t type; DB_DNODE_ENTER(db); dn = DB_DNODE(db); type = dn->dn_bonustype; DB_DNODE_EXIT(db); return (type); } int dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) { dnode_t *dn; int error; error = dnode_hold(os, object, FTAG, &dn); dbuf_rm_spill(dn, tx); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); dnode_rm_spill(dn, tx); rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); return (error); } /* * returns ENOENT, EIO, or 0. */ int dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) { dnode_t *dn; dmu_buf_impl_t *db; int error; error = dnode_hold(os, object, FTAG, &dn); if (error) return (error); rw_enter(&dn->dn_struct_rwlock, RW_READER); if (dn->dn_bonus == NULL) { rw_exit(&dn->dn_struct_rwlock); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); if (dn->dn_bonus == NULL) dbuf_create_bonus(dn); } db = dn->dn_bonus; /* as long as the bonus buf is held, the dnode will be held */ if (zfs_refcount_add(&db->db_holds, tag) == 1) { VERIFY(dnode_add_ref(dn, db)); atomic_inc_32(&dn->dn_dbufs_count); } /* * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's * hold and incrementing the dbuf count to ensure that dnode_move() sees * a dnode hold for every dbuf. */ rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); *dbp = &db->db; return (0); } /* * returns ENOENT, EIO, or 0. * * This interface will allocate a blank spill dbuf when a spill blk * doesn't already exist on the dnode. * * if you only want to find an already existing spill db, then * dmu_spill_hold_existing() should be used. */ int dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = NULL; int err; if ((flags & DB_RF_HAVESTRUCT) == 0) rw_enter(&dn->dn_struct_rwlock, RW_READER); db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_exit(&dn->dn_struct_rwlock); ASSERT(db != NULL); err = dbuf_read(db, NULL, flags); if (err == 0) *dbp = &db->db; else dbuf_rele(db, tag); return (err); } int dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { err = SET_ERROR(EINVAL); } else { rw_enter(&dn->dn_struct_rwlock, RW_READER); if (!dn->dn_have_spill) { err = SET_ERROR(ENOENT); } else { err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); } rw_exit(&dn->dn_struct_rwlock); } DB_DNODE_EXIT(db); return (err); } int dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); DB_DNODE_EXIT(db); return (err); } /* * Note: longer-term, we should modify all of the dmu_buf_*() interfaces * to take a held dnode rather than -- the lookup is wasteful, * and can induce severe lock contention when writing to several files * whose dnodes are in the same block. */ int dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) { dmu_buf_t **dbp; uint64_t blkid, nblks, i; uint32_t dbuf_flags; int err; zio_t *zio; ASSERT(length <= DMU_MAX_ACCESS); /* * Note: We directly notify the prefetch code of this read, so that * we can tell it about the multi-block read. dbuf_read() only knows * about the one block it is accessing. */ dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH; rw_enter(&dn->dn_struct_rwlock, RW_READER); if (dn->dn_datablkshift) { int blkshift = dn->dn_datablkshift; nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; } else { if (offset + length > dn->dn_datablksz) { zfs_panic_recover("zfs: accessing past end of object " "%llx/%llx (size=%u access=%llu+%llu)", (longlong_t)dn->dn_objset-> os_dsl_dataset->ds_object, (longlong_t)dn->dn_object, dn->dn_datablksz, (longlong_t)offset, (longlong_t)length); rw_exit(&dn->dn_struct_rwlock); return (SET_ERROR(EIO)); } nblks = 1; } dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); #if defined(_KERNEL) && defined(RACCT) if (racct_enable && !read) { PROC_LOCK(curproc); racct_add_force(curproc, RACCT_WRITEBPS, length); racct_add_force(curproc, RACCT_WRITEIOPS, nblks); PROC_UNLOCK(curproc); } #endif zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); blkid = dbuf_whichblock(dn, 0, offset); for (i = 0; i < nblks; i++) { dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); if (db == NULL) { rw_exit(&dn->dn_struct_rwlock); dmu_buf_rele_array(dbp, nblks, tag); zio_nowait(zio); return (SET_ERROR(EIO)); } /* initiate async i/o */ if (read) (void) dbuf_read(db, zio, dbuf_flags); #ifdef _KERNEL else curthread->td_ru.ru_oublock++; #endif dbp[i] = &db->db; } if ((flags & DMU_READ_NO_PREFETCH) == 0 && DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { dmu_zfetch(&dn->dn_zfetch, blkid, nblks, read && DNODE_IS_CACHEABLE(dn)); } rw_exit(&dn->dn_struct_rwlock); /* wait for async i/o */ err = zio_wait(zio); if (err) { dmu_buf_rele_array(dbp, nblks, tag); return (err); } /* wait for other io to complete */ if (read) { for (i = 0; i < nblks; i++) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; mutex_enter(&db->db_mtx); while (db->db_state == DB_READ || db->db_state == DB_FILL) cv_wait(&db->db_changed, &db->db_mtx); if (db->db_state == DB_UNCACHED) err = SET_ERROR(EIO); mutex_exit(&db->db_mtx); if (err) { dmu_buf_rele_array(dbp, nblks, tag); return (err); } } } *numbufsp = nblks; *dbpp = dbp; return (0); } static int dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, numbufsp, dbpp, DMU_READ_PREFETCH); dnode_rele(dn, FTAG); return (err); } int dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, uint64_t length, boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; int err; DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, numbufsp, dbpp, DMU_READ_PREFETCH); DB_DNODE_EXIT(db); return (err); } void dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) { int i; dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; if (numbufs == 0) return; for (i = 0; i < numbufs; i++) { if (dbp[i]) dbuf_rele(dbp[i], tag); } kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); } /* * Issue prefetch i/os for the given blocks. If level is greater than 0, the * indirect blocks prefeteched will be those that point to the blocks containing * the data starting at offset, and continuing to offset + len. * * Note that if the indirect blocks above the blocks being prefetched are not in * cache, they will be asychronously read in. */ void dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, uint64_t len, zio_priority_t pri) { dnode_t *dn; uint64_t blkid; int nblks, err; if (len == 0) { /* they're interested in the bonus buffer */ dn = DMU_META_DNODE(os); if (object == 0 || object >= DN_MAX_OBJECT) return; rw_enter(&dn->dn_struct_rwlock, RW_READER); blkid = dbuf_whichblock(dn, level, object * sizeof (dnode_phys_t)); dbuf_prefetch(dn, level, blkid, pri, 0); rw_exit(&dn->dn_struct_rwlock); return; } /* * See comment before the definition of dmu_prefetch_max. */ len = MIN(len, dmu_prefetch_max); /* * XXX - Note, if the dnode for the requested object is not * already cached, we will do a *synchronous* read in the * dnode_hold() call. The same is true for any indirects. */ err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return; rw_enter(&dn->dn_struct_rwlock, RW_READER); /* * offset + len - 1 is the last byte we want to prefetch for, and offset * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the * last block we want to prefetch, and dbuf_whichblock(dn, level, * offset) is the first. Then the number we need to prefetch is the * last - first + 1. */ if (level > 0 || dn->dn_datablkshift != 0) { nblks = dbuf_whichblock(dn, level, offset + len - 1) - dbuf_whichblock(dn, level, offset) + 1; } else { nblks = (offset < dn->dn_datablksz); } if (nblks != 0) { blkid = dbuf_whichblock(dn, level, offset); for (int i = 0; i < nblks; i++) dbuf_prefetch(dn, level, blkid + i, pri, 0); } rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); } /* * Get the next "chunk" of file data to free. We traverse the file from * the end so that the file gets shorter over time (if we crashes in the * middle, this will leave us in a better state). We find allocated file * data by simply searching the allocated level 1 indirects. * * On input, *start should be the first offset that does not need to be * freed (e.g. "offset + length"). On return, *start will be the first * offset that should be freed and l1blks is set to the number of level 1 * indirect blocks found within the chunk. */ static int get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) { uint64_t blks; uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); /* bytes of data covered by a level-1 indirect block */ uint64_t iblkrange = dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); ASSERT3U(minimum, <=, *start); /* * Check if we can free the entire range assuming that all of the * L1 blocks in this range have data. If we can, we use this * worst case value as an estimate so we can avoid having to look * at the object's actual data. */ uint64_t total_l1blks = (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / iblkrange; if (total_l1blks <= maxblks) { *l1blks = total_l1blks; *start = minimum; return (0); } ASSERT(ISP2(iblkrange)); for (blks = 0; *start > minimum && blks < maxblks; blks++) { int err; /* * dnode_next_offset(BACKWARDS) will find an allocated L1 * indirect block at or before the input offset. We must * decrement *start so that it is at the end of the region * to search. */ (*start)--; err = dnode_next_offset(dn, DNODE_FIND_BACKWARDS, start, 2, 1, 0); /* if there are no indirect blocks before start, we are done */ if (err == ESRCH) { *start = minimum; break; } else if (err != 0) { *l1blks = blks; return (err); } /* set start to the beginning of this L1 indirect */ *start = P2ALIGN(*start, iblkrange); } if (*start < minimum) *start = minimum; *l1blks = blks; return (0); } /* * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, * otherwise return false. * Used below in dmu_free_long_range_impl() to enable abort when unmounting */ /*ARGSUSED*/ static boolean_t dmu_objset_zfs_unmounting(objset_t *os) { #ifdef _KERNEL if (dmu_objset_type(os) == DMU_OST_ZFS) return (zfs_get_vfs_flag_unmounted(os)); #endif return (B_FALSE); } static int dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, uint64_t length) { uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; int err; uint64_t dirty_frees_threshold; dsl_pool_t *dp = dmu_objset_pool(os); if (offset >= object_size) return (0); if (zfs_per_txg_dirty_frees_percent <= 100) dirty_frees_threshold = zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; else dirty_frees_threshold = zfs_dirty_data_max / 20; if (length == DMU_OBJECT_END || offset + length > object_size) length = object_size - offset; while (length != 0) { uint64_t chunk_end, chunk_begin, chunk_len; uint64_t l1blks; dmu_tx_t *tx; if (dmu_objset_zfs_unmounting(dn->dn_objset)) return (SET_ERROR(EINTR)); chunk_end = chunk_begin = offset + length; /* move chunk_begin backwards to the beginning of this chunk */ err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); if (err) return (err); ASSERT3U(chunk_begin, >=, offset); ASSERT3U(chunk_begin, <=, chunk_end); chunk_len = chunk_end - chunk_begin; tx = dmu_tx_create(os); dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); /* * Mark this transaction as typically resulting in a net * reduction in space used. */ dmu_tx_mark_netfree(tx); err = dmu_tx_assign(tx, TXG_WAIT); if (err) { dmu_tx_abort(tx); return (err); } uint64_t txg = dmu_tx_get_txg(tx); mutex_enter(&dp->dp_lock); uint64_t long_free_dirty = dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; mutex_exit(&dp->dp_lock); /* * To avoid filling up a TXG with just frees, wait for * the next TXG to open before freeing more chunks if * we have reached the threshold of frees. */ if (dirty_frees_threshold != 0 && long_free_dirty >= dirty_frees_threshold) { dmu_tx_commit(tx); txg_wait_open(dp, 0); continue; } /* * In order to prevent unnecessary write throttling, for each * TXG, we track the cumulative size of L1 blocks being dirtied * in dnode_free_range() below. We compare this number to a * tunable threshold, past which we prevent new L1 dirty freeing * blocks from being added into the open TXG. See * dmu_free_long_range_impl() for details. The threshold * prevents write throttle activation due to dirty freeing L1 * blocks taking up a large percentage of zfs_dirty_data_max. */ mutex_enter(&dp->dp_lock); dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += l1blks << dn->dn_indblkshift; mutex_exit(&dp->dp_lock); DTRACE_PROBE3(free__long__range, uint64_t, long_free_dirty, uint64_t, chunk_len, uint64_t, txg); dnode_free_range(dn, chunk_begin, chunk_len, tx); dmu_tx_commit(tx); length -= chunk_len; } return (0); } int dmu_free_long_range(objset_t *os, uint64_t object, uint64_t offset, uint64_t length) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return (err); err = dmu_free_long_range_impl(os, dn, offset, length); /* * It is important to zero out the maxblkid when freeing the entire * file, so that (a) subsequent calls to dmu_free_long_range_impl() * will take the fast path, and (b) dnode_reallocate() can verify * that the entire file has been freed. */ if (err == 0 && offset == 0 && length == DMU_OBJECT_END) dn->dn_maxblkid = 0; dnode_rele(dn, FTAG); return (err); } int dmu_free_long_object(objset_t *os, uint64_t object) { dmu_tx_t *tx; int err; err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); if (err != 0) return (err); tx = dmu_tx_create(os); dmu_tx_hold_bonus(tx, object); dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); dmu_tx_mark_netfree(tx); err = dmu_tx_assign(tx, TXG_WAIT); if (err == 0) { err = dmu_object_free(os, object, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } return (err); } int dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, dmu_tx_t *tx) { dnode_t *dn; int err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); ASSERT(offset < UINT64_MAX); ASSERT(size == -1ULL || size <= UINT64_MAX - offset); dnode_free_range(dn, offset, size, tx); dnode_rele(dn, FTAG); return (0); } static int dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { dmu_buf_t **dbp; int numbufs, err = 0; /* * Deal with odd block sizes, where there can't be data past the first * block. If we ever do the tail block optimization, we will need to * handle that here as well. */ if (dn->dn_maxblkid == 0) { int newsz = offset > dn->dn_datablksz ? 0 : MIN(size, dn->dn_datablksz - offset); bzero((char *)buf + newsz, size - newsz); size = newsz; } while (size > 0) { uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); int i; /* * NB: we could do this block-at-a-time, but it's nice * to be reading in parallel. */ err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, TRUE, FTAG, &numbufs, &dbp, flags); if (err) break; for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); bcopy((char *)db->db_data + bufoff, buf, tocpy); offset += tocpy; size -= tocpy; buf = (char *)buf + tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); } return (err); } int dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) return (err); err = dmu_read_impl(dn, offset, size, buf, flags); dnode_rele(dn, FTAG); return (err); } int dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, uint32_t flags) { return (dmu_read_impl(dn, offset, size, buf, flags)); } static void dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { int i; for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); bcopy(buf, (char *)db->db_data + bufoff, tocpy); if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; buf = (char *)buf + tocpy; } } void dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; if (size == 0) return; VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp)); dmu_write_impl(dbp, numbufs, offset, size, buf, tx); dmu_buf_rele_array(dbp, numbufs, FTAG); } void dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, const void *buf, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; if (size == 0) return; VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); dmu_write_impl(dbp, numbufs, offset, size, buf, tx); dmu_buf_rele_array(dbp, numbufs, FTAG); } static int dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn, uint64_t last_removal_txg, uint64_t offset) { uint64_t l1blkid = dbuf_whichblock(dn, 1, offset); int err = 0; rw_enter(&dn->dn_struct_rwlock, RW_READER); dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG); ASSERT3P(dbuf, !=, NULL); /* * If the block hasn't been written yet, this default will ensure * we don't try to remap it. */ uint64_t birth = UINT64_MAX; ASSERT3U(last_removal_txg, !=, UINT64_MAX); if (dbuf->db_blkptr != NULL) birth = dbuf->db_blkptr->blk_birth; rw_exit(&dn->dn_struct_rwlock); /* * If this L1 was already written after the last removal, then we've * already tried to remap it. */ if (birth <= last_removal_txg && dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 && dbuf_can_remap(dbuf)) { dmu_tx_t *tx = dmu_tx_create(os); dmu_tx_hold_remap_l1indirect(tx, dn->dn_object); err = dmu_tx_assign(tx, TXG_WAIT); if (err == 0) { (void) dbuf_dirty(dbuf, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } } dbuf_rele(dbuf, FTAG); delay(zfs_object_remap_one_indirect_delay_ticks); return (err); } /* * Remap all blockpointers in the object, if possible, so that they reference * only concrete vdevs. * * To do this, iterate over the L0 blockpointers and remap any that reference * an indirect vdev. Note that we only examine L0 blockpointers; since we * cannot guarantee that we can remap all blockpointer anyways (due to split * blocks), we do not want to make the code unnecessarily complicated to * catch the unlikely case that there is an L1 block on an indirect vdev that * contains no indirect blockpointers. */ int dmu_object_remap_indirects(objset_t *os, uint64_t object, uint64_t last_removal_txg) { uint64_t offset, l1span; int err; dnode_t *dn; err = dnode_hold(os, object, FTAG, &dn); if (err != 0) { return (err); } if (dn->dn_nlevels <= 1) { if (issig(JUSTLOOKING) && issig(FORREAL)) { err = SET_ERROR(EINTR); } /* * If the dnode has no indirect blocks, we cannot dirty them. * We still want to remap the blkptr(s) in the dnode if * appropriate, so mark it as dirty. */ if (err == 0 && dnode_needs_remap(dn)) { dmu_tx_t *tx = dmu_tx_create(os); dmu_tx_hold_bonus(tx, dn->dn_object); if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) { dnode_setdirty(dn, tx); dmu_tx_commit(tx); } else { dmu_tx_abort(tx); } } dnode_rele(dn, FTAG); return (err); } offset = 0; l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT + dn->dn_datablkshift); /* * Find the next L1 indirect that is not a hole. */ while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) { if (issig(JUSTLOOKING) && issig(FORREAL)) { err = SET_ERROR(EINTR); break; } if ((err = dmu_object_remap_one_indirect(os, dn, last_removal_txg, offset)) != 0) { break; } offset += l1span; } dnode_rele(dn, FTAG); return (err); } void dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs, i; if (size == 0) return; VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp)); for (i = 0; i < numbufs; i++) { dmu_buf_t *db = dbp[i]; dmu_buf_will_not_fill(db, tx); } dmu_buf_rele_array(dbp, numbufs, FTAG); } void dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, void *data, uint8_t etype, uint8_t comp, int uncompressed_size, int compressed_size, int byteorder, dmu_tx_t *tx) { dmu_buf_t *db; ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); VERIFY0(dmu_buf_hold_noread(os, object, offset, FTAG, &db)); dmu_buf_write_embedded(db, data, (bp_embedded_type_t)etype, (enum zio_compress)comp, uncompressed_size, compressed_size, byteorder, tx); dmu_buf_rele(db, FTAG); } /* * DMU support for xuio */ kstat_t *xuio_ksp = NULL; int dmu_xuio_init(xuio_t *xuio, int nblk) { dmu_xuio_t *priv; uio_t *uio = &xuio->xu_uio; uio->uio_iovcnt = nblk; uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); priv->cnt = nblk; priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); priv->iovp = uio->uio_iov; XUIO_XUZC_PRIV(xuio) = priv; if (XUIO_XUZC_RW(xuio) == UIO_READ) XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); else XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); return (0); } void dmu_xuio_fini(xuio_t *xuio) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); int nblk = priv->cnt; kmem_free(priv->iovp, nblk * sizeof (iovec_t)); kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); kmem_free(priv, sizeof (dmu_xuio_t)); if (XUIO_XUZC_RW(xuio) == UIO_READ) XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); else XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); } /* * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } * and increase priv->next by 1. */ int dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) { struct iovec *iov; uio_t *uio = &xuio->xu_uio; dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); int i = priv->next++; ASSERT(i < priv->cnt); ASSERT(off + n <= arc_buf_lsize(abuf)); iov = uio->uio_iov + i; iov->iov_base = (char *)abuf->b_data + off; iov->iov_len = n; priv->bufs[i] = abuf; return (0); } int dmu_xuio_cnt(xuio_t *xuio) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); return (priv->cnt); } arc_buf_t * dmu_xuio_arcbuf(xuio_t *xuio, int i) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); ASSERT(i < priv->cnt); return (priv->bufs[i]); } void dmu_xuio_clear(xuio_t *xuio, int i) { dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); ASSERT(i < priv->cnt); priv->bufs[i] = NULL; } static void xuio_stat_init(void) { xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); if (xuio_ksp != NULL) { xuio_ksp->ks_data = &xuio_stats; kstat_install(xuio_ksp); } } static void xuio_stat_fini(void) { if (xuio_ksp != NULL) { kstat_delete(xuio_ksp); xuio_ksp = NULL; } } void xuio_stat_wbuf_copied(void) { XUIOSTAT_BUMP(xuiostat_wbuf_copied); } void xuio_stat_wbuf_nocopy(void) { XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); } #ifdef _KERNEL int dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) { dmu_buf_t **dbp; int numbufs, i, err; xuio_t *xuio = NULL; /* * NB: we could do this block-at-a-time, but it's nice * to be reading in parallel. */ err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, TRUE, FTAG, &numbufs, &dbp, 0); if (err) return (err); #ifdef UIO_XUIO if (uio->uio_extflg == UIO_XUIO) xuio = (xuio_t *)uio; #endif for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = uio->uio_loffset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); if (xuio) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; arc_buf_t *dbuf_abuf = dbi->db_buf; arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); if (!err) { uio->uio_resid -= tocpy; uio->uio_loffset += tocpy; } if (abuf == dbuf_abuf) XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); else XUIOSTAT_BUMP(xuiostat_rbuf_copied); } else { #ifdef illumos err = uiomove((char *)db->db_data + bufoff, tocpy, UIO_READ, uio); #else err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, uio); #endif } if (err) break; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } /* * Read 'size' bytes into the uio buffer. * From object zdb->db_object. * Starting at offset uio->uio_loffset. * * If the caller already has a dbuf in the target object * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), * because we don't have to find the dnode_t for the object. */ int dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; dnode_t *dn; int err; if (size == 0) return (0); DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_read_uio_dnode(dn, uio, size); DB_DNODE_EXIT(db); return (err); } /* * Read 'size' bytes into the uio buffer. * From the specified object * Starting at offset uio->uio_loffset. */ int dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) { dnode_t *dn; int err; if (size == 0) return (0); err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_read_uio_dnode(dn, uio, size); dnode_rele(dn, FTAG); return (err); } int dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs; int err = 0; int i; err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy; int bufoff; dmu_buf_t *db = dbp[i]; ASSERT(size > 0); bufoff = uio->uio_loffset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); #ifdef illumos /* * XXX uiomove could block forever (eg. nfs-backed * pages). There needs to be a uiolockdown() function * to lock the pages in memory, so that uiomove won't * block. */ err = uiomove((char *)db->db_data + bufoff, tocpy, UIO_WRITE, uio); #else err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, uio); #endif if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); if (err) break; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } /* * Write 'size' bytes from the uio buffer. * To object zdb->db_object. * Starting at offset uio->uio_loffset. * * If the caller already has a dbuf in the target object * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), * because we don't have to find the dnode_t for the object. */ int dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; dnode_t *dn; int err; if (size == 0) return (0); DB_DNODE_ENTER(db); dn = DB_DNODE(db); err = dmu_write_uio_dnode(dn, uio, size, tx); DB_DNODE_EXIT(db); return (err); } /* * Write 'size' bytes from the uio buffer. * To the specified object. * Starting at offset uio->uio_loffset. */ int dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, dmu_tx_t *tx) { dnode_t *dn; int err; if (size == 0) return (0); err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dmu_write_uio_dnode(dn, uio, size, tx); dnode_rele(dn, FTAG); return (err); } #ifdef illumos int dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, page_t *pp, dmu_tx_t *tx) { dmu_buf_t **dbp; int numbufs, i; int err; if (size == 0) return (0); err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy, copied, thiscpy; int bufoff; dmu_buf_t *db = dbp[i]; caddr_t va; ASSERT(size > 0); ASSERT3U(db->db_size, >=, PAGESIZE); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); for (copied = 0; copied < tocpy; copied += PAGESIZE) { ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); thiscpy = MIN(PAGESIZE, tocpy - copied); va = zfs_map_page(pp, S_READ); bcopy(va, (char *)db->db_data + bufoff, thiscpy); zfs_unmap_page(pp, va); pp = pp->p_next; bufoff += PAGESIZE; } if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } #else /* !illumos */ int dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, vm_page_t *ma, dmu_tx_t *tx) { dmu_buf_t **dbp; struct sf_buf *sf; int numbufs, i; int err; if (size == 0) return (0); err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, &numbufs, &dbp); if (err) return (err); for (i = 0; i < numbufs; i++) { int tocpy, copied, thiscpy; int bufoff; dmu_buf_t *db = dbp[i]; caddr_t va; ASSERT(size > 0); ASSERT3U(db->db_size, >=, PAGESIZE); bufoff = offset - db->db_offset; tocpy = (int)MIN(db->db_size - bufoff, size); ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); if (tocpy == db->db_size) dmu_buf_will_fill(db, tx); else dmu_buf_will_dirty(db, tx); for (copied = 0; copied < tocpy; copied += PAGESIZE) { ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); thiscpy = MIN(PAGESIZE, tocpy - copied); va = zfs_map_page(*ma, &sf); bcopy(va, (char *)db->db_data + bufoff, thiscpy); zfs_unmap_page(sf); ma += 1; bufoff += PAGESIZE; } if (tocpy == db->db_size) dmu_buf_fill_done(db, tx); offset += tocpy; size -= tocpy; } dmu_buf_rele_array(dbp, numbufs, FTAG); return (err); } int dmu_read_pages(objset_t *os, uint64_t object, vm_page_t *ma, int count, int *rbehind, int *rahead, int last_size) { struct sf_buf *sf; vm_object_t vmobj; vm_page_t m; dmu_buf_t **dbp; dmu_buf_t *db; caddr_t va; int numbufs, i; int bufoff, pgoff, tocpy; int mi, di; int err; ASSERT3U(ma[0]->pindex + count - 1, ==, ma[count - 1]->pindex); ASSERT(last_size <= PAGE_SIZE); err = dmu_buf_hold_array(os, object, IDX_TO_OFF(ma[0]->pindex), IDX_TO_OFF(count - 1) + last_size, TRUE, FTAG, &numbufs, &dbp); if (err != 0) return (err); #ifdef DEBUG IMPLY(last_size < PAGE_SIZE, *rahead == 0); if (dbp[0]->db_offset != 0 || numbufs > 1) { for (i = 0; i < numbufs; i++) { ASSERT(ISP2(dbp[i]->db_size)); ASSERT((dbp[i]->db_offset % dbp[i]->db_size) == 0); ASSERT3U(dbp[i]->db_size, ==, dbp[0]->db_size); } } #endif vmobj = ma[0]->object; zfs_vmobject_wlock(vmobj); db = dbp[0]; for (i = 0; i < *rbehind; i++) { m = vm_page_grab(vmobj, ma[0]->pindex - 1 - i, VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (m == NULL) break; if (!vm_page_none_valid(m)) { ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL); vm_page_sunbusy(m); break; } ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); ASSERT(db->db_size > PAGE_SIZE); bufoff = IDX_TO_OFF(m->pindex) % db->db_size; va = zfs_map_page(m, &sf); bcopy((char *)db->db_data + bufoff, va, PAGESIZE); zfs_unmap_page(sf); vm_page_valid(m); - vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); - vm_page_unlock(m); vm_page_sunbusy(m); } *rbehind = i; bufoff = IDX_TO_OFF(ma[0]->pindex) % db->db_size; pgoff = 0; for (mi = 0, di = 0; mi < count && di < numbufs; ) { if (pgoff == 0) { m = ma[mi]; if (m != bogus_page) { vm_page_assert_xbusied(m); ASSERT(vm_page_none_valid(m)); ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); va = zfs_map_page(m, &sf); } } if (bufoff == 0) db = dbp[di]; if (m != bogus_page) { ASSERT3U(IDX_TO_OFF(m->pindex) + pgoff, ==, db->db_offset + bufoff); } /* * We do not need to clamp the copy size by the file * size as the last block is zero-filled beyond the * end of file anyway. */ tocpy = MIN(db->db_size - bufoff, PAGESIZE - pgoff); if (m != bogus_page) bcopy((char *)db->db_data + bufoff, va + pgoff, tocpy); pgoff += tocpy; ASSERT(pgoff <= PAGESIZE); if (pgoff == PAGESIZE) { if (m != bogus_page) { zfs_unmap_page(sf); vm_page_valid(m); } ASSERT(mi < count); mi++; pgoff = 0; } bufoff += tocpy; ASSERT(bufoff <= db->db_size); if (bufoff == db->db_size) { ASSERT(di < numbufs); di++; bufoff = 0; } } #ifdef DEBUG /* * Three possibilities: * - last requested page ends at a buffer boundary and , thus, * all pages and buffers have been iterated; * - all requested pages are filled, but the last buffer * has not been exhausted; * the read-ahead is possible only in this case; * - all buffers have been read, but the last page has not been * fully filled; * this is only possible if the file has only a single buffer * with a size that is not a multiple of the page size. */ if (mi == count) { ASSERT(di >= numbufs - 1); IMPLY(*rahead != 0, di == numbufs - 1); IMPLY(*rahead != 0, bufoff != 0); ASSERT(pgoff == 0); } if (di == numbufs) { ASSERT(mi >= count - 1); ASSERT(*rahead == 0); IMPLY(pgoff == 0, mi == count); if (pgoff != 0) { ASSERT(mi == count - 1); ASSERT((dbp[0]->db_size & PAGE_MASK) != 0); } } #endif if (pgoff != 0) { ASSERT(m != bogus_page); bzero(va + pgoff, PAGESIZE - pgoff); zfs_unmap_page(sf); vm_page_valid(m); } for (i = 0; i < *rahead; i++) { m = vm_page_grab(vmobj, ma[count - 1]->pindex + 1 + i, VM_ALLOC_NORMAL | VM_ALLOC_NOWAIT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (m == NULL) break; if (!vm_page_none_valid(m)) { ASSERT3U(m->valid, ==, VM_PAGE_BITS_ALL); vm_page_sunbusy(m); break; } ASSERT(m->dirty == 0); ASSERT(!pmap_page_is_mapped(m)); ASSERT(db->db_size > PAGE_SIZE); bufoff = IDX_TO_OFF(m->pindex) % db->db_size; tocpy = MIN(db->db_size - bufoff, PAGESIZE); va = zfs_map_page(m, &sf); bcopy((char *)db->db_data + bufoff, va, tocpy); if (tocpy < PAGESIZE) { ASSERT(i == *rahead - 1); ASSERT((db->db_size & PAGE_MASK) != 0); bzero(va + tocpy, PAGESIZE - tocpy); } zfs_unmap_page(sf); vm_page_valid(m); - vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); - vm_page_unlock(m); vm_page_sunbusy(m); } *rahead = i; zfs_vmobject_wunlock(vmobj); dmu_buf_rele_array(dbp, numbufs, FTAG); return (0); } #endif /* illumos */ #endif /* _KERNEL */ /* * Allocate a loaned anonymous arc buffer. */ arc_buf_t * dmu_request_arcbuf(dmu_buf_t *handle, int size) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); } /* * Free a loaned arc buffer. */ void dmu_return_arcbuf(arc_buf_t *buf) { arc_return_buf(buf, FTAG); arc_buf_destroy(buf, FTAG); } /* * When possible directly assign passed loaned arc buffer to a dbuf. * If this is not possible copy the contents of passed arc buf via * dmu_write(). */ void dmu_assign_arcbuf_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, dmu_tx_t *tx) { dmu_buf_impl_t *db; uint32_t blksz = (uint32_t)arc_buf_lsize(buf); uint64_t blkid; rw_enter(&dn->dn_struct_rwlock, RW_READER); blkid = dbuf_whichblock(dn, 0, offset); VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); rw_exit(&dn->dn_struct_rwlock); /* * We can only assign if the offset is aligned, the arc buf is the * same size as the dbuf, and the dbuf is not metadata. */ if (offset == db->db.db_offset && blksz == db->db.db_size) { #ifdef _KERNEL curthread->td_ru.ru_oublock++; #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_force(curproc, RACCT_WRITEBPS, blksz); racct_add_force(curproc, RACCT_WRITEIOPS, 1); PROC_UNLOCK(curproc); } #endif /* RACCT */ #endif /* _KERNEL */ dbuf_assign_arcbuf(db, buf, tx); dbuf_rele(db, FTAG); } else { objset_t *os; uint64_t object; /* compressed bufs must always be assignable to their dbuf */ ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); os = dn->dn_objset; object = dn->dn_object; dbuf_rele(db, FTAG); dmu_write(os, object, offset, blksz, buf->b_data, tx); dmu_return_arcbuf(buf); XUIOSTAT_BUMP(xuiostat_wbuf_copied); } } void dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, dmu_tx_t *tx) { dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; DB_DNODE_ENTER(dbuf); dmu_assign_arcbuf_dnode(DB_DNODE(dbuf), offset, buf, tx); DB_DNODE_EXIT(dbuf); } typedef struct { dbuf_dirty_record_t *dsa_dr; dmu_sync_cb_t *dsa_done; zgd_t *dsa_zgd; dmu_tx_t *dsa_tx; } dmu_sync_arg_t; /* ARGSUSED */ static void dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) { dmu_sync_arg_t *dsa = varg; dmu_buf_t *db = dsa->dsa_zgd->zgd_db; blkptr_t *bp = zio->io_bp; if (zio->io_error == 0) { if (BP_IS_HOLE(bp)) { /* * A block of zeros may compress to a hole, but the * block size still needs to be known for replay. */ BP_SET_LSIZE(bp, db->db_size); } else if (!BP_IS_EMBEDDED(bp)) { ASSERT(BP_GET_LEVEL(bp) == 0); bp->blk_fill = 1; } } } static void dmu_sync_late_arrival_ready(zio_t *zio) { dmu_sync_ready(zio, NULL, zio->io_private); } /* ARGSUSED */ static void dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) { dmu_sync_arg_t *dsa = varg; dbuf_dirty_record_t *dr = dsa->dsa_dr; dmu_buf_impl_t *db = dr->dr_dbuf; zgd_t *zgd = dsa->dsa_zgd; /* * Record the vdev(s) backing this blkptr so they can be flushed after * the writes for the lwb have completed. */ if (zio->io_error == 0) { zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); } mutex_enter(&db->db_mtx); ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); if (zio->io_error == 0) { dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); if (dr->dt.dl.dr_nopwrite) { blkptr_t *bp = zio->io_bp; blkptr_t *bp_orig = &zio->io_bp_orig; uint8_t chksum = BP_GET_CHECKSUM(bp_orig); ASSERT(BP_EQUAL(bp, bp_orig)); VERIFY(BP_EQUAL(bp, db->db_blkptr)); ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); ASSERT(zio_checksum_table[chksum].ci_flags & ZCHECKSUM_FLAG_NOPWRITE); } dr->dt.dl.dr_overridden_by = *zio->io_bp; dr->dt.dl.dr_override_state = DR_OVERRIDDEN; dr->dt.dl.dr_copies = zio->io_prop.zp_copies; /* * Old style holes are filled with all zeros, whereas * new-style holes maintain their lsize, type, level, * and birth time (see zio_write_compress). While we * need to reset the BP_SET_LSIZE() call that happened * in dmu_sync_ready for old style holes, we do *not* * want to wipe out the information contained in new * style holes. Thus, only zero out the block pointer if * it's an old style hole. */ if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && dr->dt.dl.dr_overridden_by.blk_birth == 0) BP_ZERO(&dr->dt.dl.dr_overridden_by); } else { dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; } cv_broadcast(&db->db_changed); mutex_exit(&db->db_mtx); dsa->dsa_done(dsa->dsa_zgd, zio->io_error); kmem_free(dsa, sizeof (*dsa)); } static void dmu_sync_late_arrival_done(zio_t *zio) { blkptr_t *bp = zio->io_bp; dmu_sync_arg_t *dsa = zio->io_private; blkptr_t *bp_orig = &zio->io_bp_orig; zgd_t *zgd = dsa->dsa_zgd; if (zio->io_error == 0) { /* * Record the vdev(s) backing this blkptr so they can be * flushed after the writes for the lwb have completed. */ zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); if (!BP_IS_HOLE(bp)) { ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); ASSERT(zio->io_bp->blk_birth == zio->io_txg); ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); zio_free(zio->io_spa, zio->io_txg, zio->io_bp); } } dmu_tx_commit(dsa->dsa_tx); dsa->dsa_done(dsa->dsa_zgd, zio->io_error); abd_put(zio->io_abd); kmem_free(dsa, sizeof (*dsa)); } static int dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, zio_prop_t *zp, zbookmark_phys_t *zb) { dmu_sync_arg_t *dsa; dmu_tx_t *tx; tx = dmu_tx_create(os); dmu_tx_hold_space(tx, zgd->zgd_db->db_size); if (dmu_tx_assign(tx, TXG_WAIT) != 0) { dmu_tx_abort(tx); /* Make zl_get_data do txg_waited_synced() */ return (SET_ERROR(EIO)); } /* * In order to prevent the zgd's lwb from being free'd prior to * dmu_sync_late_arrival_done() being called, we have to ensure * the lwb's "max txg" takes this tx's txg into account. */ zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); dsa->dsa_dr = NULL; dsa->dsa_done = done; dsa->dsa_zgd = zgd; dsa->dsa_tx = tx; /* * Since we are currently syncing this txg, it's nontrivial to * determine what BP to nopwrite against, so we disable nopwrite. * * When syncing, the db_blkptr is initially the BP of the previous * txg. We can not nopwrite against it because it will be changed * (this is similar to the non-late-arrival case where the dbuf is * dirty in a future txg). * * Then dbuf_write_ready() sets bp_blkptr to the location we will write. * We can not nopwrite against it because although the BP will not * (typically) be changed, the data has not yet been persisted to this * location. * * Finally, when dbuf_write_done() is called, it is theoretically * possible to always nopwrite, because the data that was written in * this txg is the same data that we are trying to write. However we * would need to check that this dbuf is not dirty in any future * txg's (as we do in the normal dmu_sync() path). For simplicity, we * don't nopwrite in this case. */ zp->zp_nopwrite = B_FALSE; zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); return (0); } /* * Intent log support: sync the block associated with db to disk. * N.B. and XXX: the caller is responsible for making sure that the * data isn't changing while dmu_sync() is writing it. * * Return values: * * EEXIST: this txg has already been synced, so there's nothing to do. * The caller should not log the write. * * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. * The caller should not log the write. * * EALREADY: this block is already in the process of being synced. * The caller should track its progress (somehow). * * EIO: could not do the I/O. * The caller should do a txg_wait_synced(). * * 0: the I/O has been initiated. * The caller should log this blkptr in the done callback. * It is possible that the I/O will fail, in which case * the error will be reported to the done callback and * propagated to pio from zio_done(). */ int dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; objset_t *os = db->db_objset; dsl_dataset_t *ds = os->os_dsl_dataset; dbuf_dirty_record_t *dr; dmu_sync_arg_t *dsa; zbookmark_phys_t zb; zio_prop_t zp; dnode_t *dn; ASSERT(pio != NULL); ASSERT(txg != 0); SET_BOOKMARK(&zb, ds->ds_object, db->db.db_object, db->db_level, db->db_blkid); DB_DNODE_ENTER(db); dn = DB_DNODE(db); dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); DB_DNODE_EXIT(db); /* * If we're frozen (running ziltest), we always need to generate a bp. */ if (txg > spa_freeze_txg(os->os_spa)) return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); /* * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() * and us. If we determine that this txg is not yet syncing, * but it begins to sync a moment later, that's OK because the * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. */ mutex_enter(&db->db_mtx); if (txg <= spa_last_synced_txg(os->os_spa)) { /* * This txg has already synced. There's nothing to do. */ mutex_exit(&db->db_mtx); return (SET_ERROR(EEXIST)); } if (txg <= spa_syncing_txg(os->os_spa)) { /* * This txg is currently syncing, so we can't mess with * the dirty record anymore; just write a new log block. */ mutex_exit(&db->db_mtx); return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); } dr = db->db_last_dirty; while (dr && dr->dr_txg != txg) dr = dr->dr_next; if (dr == NULL) { /* * There's no dr for this dbuf, so it must have been freed. * There's no need to log writes to freed blocks, so we're done. */ mutex_exit(&db->db_mtx); return (SET_ERROR(ENOENT)); } ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); if (db->db_blkptr != NULL) { /* * We need to fill in zgd_bp with the current blkptr so that * the nopwrite code can check if we're writing the same * data that's already on disk. We can only nopwrite if we * are sure that after making the copy, db_blkptr will not * change until our i/o completes. We ensure this by * holding the db_mtx, and only allowing nopwrite if the * block is not already dirty (see below). This is verified * by dmu_sync_done(), which VERIFYs that the db_blkptr has * not changed. */ *zgd->zgd_bp = *db->db_blkptr; } /* * Assume the on-disk data is X, the current syncing data (in * txg - 1) is Y, and the current in-memory data is Z (currently * in dmu_sync). * * We usually want to perform a nopwrite if X and Z are the * same. However, if Y is different (i.e. the BP is going to * change before this write takes effect), then a nopwrite will * be incorrect - we would override with X, which could have * been freed when Y was written. * * (Note that this is not a concern when we are nop-writing from * syncing context, because X and Y must be identical, because * all previous txgs have been synced.) * * Therefore, we disable nopwrite if the current BP could change * before this TXG. There are two ways it could change: by * being dirty (dr_next is non-NULL), or by being freed * (dnode_block_freed()). This behavior is verified by * zio_done(), which VERIFYs that the override BP is identical * to the on-disk BP. */ DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) zp.zp_nopwrite = B_FALSE; DB_DNODE_EXIT(db); ASSERT(dr->dr_txg == txg); if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { /* * We have already issued a sync write for this buffer, * or this buffer has already been synced. It could not * have been dirtied since, or we would have cleared the state. */ mutex_exit(&db->db_mtx); return (SET_ERROR(EALREADY)); } ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; mutex_exit(&db->db_mtx); dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); dsa->dsa_dr = dr; dsa->dsa_done = done; dsa->dsa_zgd = zgd; dsa->dsa_tx = NULL; zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); return (0); } int dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, dmu_tx_t *tx) { dnode_t *dn; int err; err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); err = dnode_set_blksz(dn, size, ibs, tx); dnode_rele(dn, FTAG); return (err); } void dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, dmu_tx_t *tx) { dnode_t *dn; /* * Send streams include each object's checksum function. This * check ensures that the receiving system can understand the * checksum function transmitted. */ ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); VERIFY0(dnode_hold(os, object, FTAG, &dn)); ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); dn->dn_checksum = checksum; dnode_setdirty(dn, tx); dnode_rele(dn, FTAG); } void dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, dmu_tx_t *tx) { dnode_t *dn; /* * Send streams include each object's compression function. This * check ensures that the receiving system can understand the * compression function transmitted. */ ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); VERIFY0(dnode_hold(os, object, FTAG, &dn)); dn->dn_compress = compress; dnode_setdirty(dn, tx); dnode_rele(dn, FTAG); } int zfs_mdcomp_disable = 0; SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN, &zfs_mdcomp_disable, 0, "Disable metadata compression"); /* * When the "redundant_metadata" property is set to "most", only indirect * blocks of this level and higher will have an additional ditto block. */ int zfs_redundant_metadata_most_ditto_level = 2; void dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) { dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)); enum zio_checksum checksum = os->os_checksum; enum zio_compress compress = os->os_compress; enum zio_checksum dedup_checksum = os->os_dedup_checksum; boolean_t dedup = B_FALSE; boolean_t nopwrite = B_FALSE; boolean_t dedup_verify = os->os_dedup_verify; int copies = os->os_copies; /* * We maintain different write policies for each of the following * types of data: * 1. metadata * 2. preallocated blocks (i.e. level-0 blocks of a dump device) * 3. all other level 0 blocks */ if (ismd) { if (zfs_mdcomp_disable) { compress = ZIO_COMPRESS_EMPTY; } else { /* * XXX -- we should design a compression algorithm * that specializes in arrays of bps. */ compress = zio_compress_select(os->os_spa, ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); } /* * Metadata always gets checksummed. If the data * checksum is multi-bit correctable, and it's not a * ZBT-style checksum, then it's suitable for metadata * as well. Otherwise, the metadata checksum defaults * to fletcher4. */ if (!(zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_METADATA) || (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED)) checksum = ZIO_CHECKSUM_FLETCHER_4; if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_MOST && (level >= zfs_redundant_metadata_most_ditto_level || DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) copies++; } else if (wp & WP_NOFILL) { ASSERT(level == 0); /* * If we're writing preallocated blocks, we aren't actually * writing them so don't set any policy properties. These * blocks are currently only used by an external subsystem * outside of zfs (i.e. dump) and not written by the zio * pipeline. */ compress = ZIO_COMPRESS_OFF; checksum = ZIO_CHECKSUM_NOPARITY; } else { compress = zio_compress_select(os->os_spa, dn->dn_compress, compress); checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? zio_checksum_select(dn->dn_checksum, checksum) : dedup_checksum; /* * Determine dedup setting. If we are in dmu_sync(), * we won't actually dedup now because that's all * done in syncing context; but we do want to use the * dedup checkum. If the checksum is not strong * enough to ensure unique signatures, force * dedup_verify. */ if (dedup_checksum != ZIO_CHECKSUM_OFF) { dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; if (!(zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_DEDUP)) dedup_verify = B_TRUE; } /* * Enable nopwrite if we have secure enough checksum * algorithm (see comment in zio_nop_write) and * compression is enabled. We don't enable nopwrite if * dedup is enabled as the two features are mutually * exclusive. */ nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_NOPWRITE) && compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); } zp->zp_checksum = checksum; zp->zp_compress = compress; ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; zp->zp_level = level; zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); zp->zp_dedup = dedup; zp->zp_dedup_verify = dedup && dedup_verify; zp->zp_nopwrite = nopwrite; zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ? os->os_zpl_special_smallblock : 0; } int dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) { dnode_t *dn; int err; /* * Sync any current changes before * we go trundling through the block pointers. */ err = dmu_object_wait_synced(os, object); if (err) { return (err); } err = dnode_hold(os, object, FTAG, &dn); if (err) { return (err); } err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); dnode_rele(dn, FTAG); return (err); } /* * Given the ZFS object, if it contains any dirty nodes * this function flushes all dirty blocks to disk. This * ensures the DMU object info is updated. A more efficient * future version might just find the TXG with the maximum * ID and wait for that to be synced. */ int dmu_object_wait_synced(objset_t *os, uint64_t object) { dnode_t *dn; int error, i; error = dnode_hold(os, object, FTAG, &dn); if (error) { return (error); } for (i = 0; i < TXG_SIZE; i++) { if (list_link_active(&dn->dn_dirty_link[i])) { break; } } dnode_rele(dn, FTAG); if (i != TXG_SIZE) { txg_wait_synced(dmu_objset_pool(os), 0); } return (0); } void __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) { dnode_phys_t *dnp = dn->dn_phys; doi->doi_data_block_size = dn->dn_datablksz; doi->doi_metadata_block_size = dn->dn_indblkshift ? 1ULL << dn->dn_indblkshift : 0; doi->doi_type = dn->dn_type; doi->doi_bonus_type = dn->dn_bonustype; doi->doi_bonus_size = dn->dn_bonuslen; doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; doi->doi_indirection = dn->dn_nlevels; doi->doi_checksum = dn->dn_checksum; doi->doi_compress = dn->dn_compress; doi->doi_nblkptr = dn->dn_nblkptr; doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; doi->doi_fill_count = 0; for (int i = 0; i < dnp->dn_nblkptr; i++) doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); } void dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) { rw_enter(&dn->dn_struct_rwlock, RW_READER); mutex_enter(&dn->dn_mtx); __dmu_object_info_from_dnode(dn, doi); mutex_exit(&dn->dn_mtx); rw_exit(&dn->dn_struct_rwlock); } /* * Get information on a DMU object. * If doi is NULL, just indicates whether the object exists. */ int dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) { dnode_t *dn; int err = dnode_hold(os, object, FTAG, &dn); if (err) return (err); if (doi != NULL) dmu_object_info_from_dnode(dn, doi); dnode_rele(dn, FTAG); return (0); } /* * As above, but faster; can be used when you have a held dbuf in hand. */ void dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; DB_DNODE_ENTER(db); dmu_object_info_from_dnode(DB_DNODE(db), doi); DB_DNODE_EXIT(db); } /* * Faster still when you only care about the size. * This is specifically optimized for zfs_getattr(). */ void dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, u_longlong_t *nblk512) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); *blksize = dn->dn_datablksz; /* add in number of slots used for the dnode itself */ *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> SPA_MINBLOCKSHIFT) + dn->dn_num_slots; DB_DNODE_EXIT(db); } void dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); *dnsize = dn->dn_num_slots << DNODE_SHIFT; DB_DNODE_EXIT(db); } void byteswap_uint64_array(void *vbuf, size_t size) { uint64_t *buf = vbuf; size_t count = size >> 3; int i; ASSERT((size & 7) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_64(buf[i]); } void byteswap_uint32_array(void *vbuf, size_t size) { uint32_t *buf = vbuf; size_t count = size >> 2; int i; ASSERT((size & 3) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_32(buf[i]); } void byteswap_uint16_array(void *vbuf, size_t size) { uint16_t *buf = vbuf; size_t count = size >> 1; int i; ASSERT((size & 1) == 0); for (i = 0; i < count; i++) buf[i] = BSWAP_16(buf[i]); } /* ARGSUSED */ void byteswap_uint8_array(void *vbuf, size_t size) { } void dmu_init(void) { abd_init(); zfs_dbgmsg_init(); sa_cache_init(); xuio_stat_init(); dmu_objset_init(); dnode_init(); zfetch_init(); zio_compress_init(); l2arc_init(); arc_init(); dbuf_init(); } void dmu_fini(void) { arc_fini(); /* arc depends on l2arc, so arc must go first */ l2arc_fini(); zfetch_fini(); zio_compress_fini(); dbuf_fini(); dnode_fini(); dmu_objset_fini(); xuio_stat_fini(); sa_cache_fini(); zfs_dbgmsg_fini(); abd_fini(); } Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c (revision 356156) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c (revision 356157) @@ -1,6029 +1,6027 @@ /* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2012, 2015 by Delphix. All rights reserved. * Copyright (c) 2014 Integros [integros.com] * Copyright 2017 Nexenta Systems, Inc. */ /* Portions Copyright 2007 Jeremy Teo */ /* Portions Copyright 2010 Robert Milkowski */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Programming rules. * * Each vnode op performs some logical unit of work. To do this, the ZPL must * properly lock its in-core state, create a DMU transaction, do the work, * record this work in the intent log (ZIL), commit the DMU transaction, * and wait for the intent log to commit if it is a synchronous operation. * Moreover, the vnode ops must work in both normal and log replay context. * The ordering of events is important to avoid deadlocks and references * to freed memory. The example below illustrates the following Big Rules: * * (1) A check must be made in each zfs thread for a mounted file system. * This is done avoiding races using ZFS_ENTER(zfsvfs). * A ZFS_EXIT(zfsvfs) is needed before all returns. Any znodes * must be checked with ZFS_VERIFY_ZP(zp). Both of these macros * can return EIO from the calling function. * * (2) VN_RELE() should always be the last thing except for zil_commit() * (if necessary) and ZFS_EXIT(). This is for 3 reasons: * First, if it's the last reference, the vnode/znode * can be freed, so the zp may point to freed memory. Second, the last * reference will call zfs_zinactive(), which may induce a lot of work -- * pushing cached pages (which acquires range locks) and syncing out * cached atime changes. Third, zfs_zinactive() may require a new tx, * which could deadlock the system if you were already holding one. * If you must call VN_RELE() within a tx then use VN_RELE_ASYNC(). * * (3) All range locks must be grabbed before calling dmu_tx_assign(), * as they can span dmu_tx_assign() calls. * * (4) If ZPL locks are held, pass TXG_NOWAIT as the second argument to * dmu_tx_assign(). This is critical because we don't want to block * while holding locks. * * If no ZPL locks are held (aside from ZFS_ENTER()), use TXG_WAIT. This * reduces lock contention and CPU usage when we must wait (note that if * throughput is constrained by the storage, nearly every transaction * must wait). * * Note, in particular, that if a lock is sometimes acquired before * the tx assigns, and sometimes after (e.g. z_lock), then failing * to use a non-blocking assign can deadlock the system. The scenario: * * Thread A has grabbed a lock before calling dmu_tx_assign(). * Thread B is in an already-assigned tx, and blocks for this lock. * Thread A calls dmu_tx_assign(TXG_WAIT) and blocks in txg_wait_open() * forever, because the previous txg can't quiesce until B's tx commits. * * If dmu_tx_assign() returns ERESTART and zfsvfs->z_assign is TXG_NOWAIT, * then drop all locks, call dmu_tx_wait(), and try again. On subsequent * calls to dmu_tx_assign(), pass TXG_NOTHROTTLE in addition to TXG_NOWAIT, * to indicate that this operation has already called dmu_tx_wait(). * This will ensure that we don't retry forever, waiting a short bit * each time. * * (5) If the operation succeeded, generate the intent log entry for it * before dropping locks. This ensures that the ordering of events * in the intent log matches the order in which they actually occurred. * During ZIL replay the zfs_log_* functions will update the sequence * number to indicate the zil transaction has replayed. * * (6) At the end of each vnode op, the DMU tx must always commit, * regardless of whether there were any errors. * * (7) After dropping all locks, invoke zil_commit(zilog, foid) * to ensure that synchronous semantics are provided when necessary. * * In general, this is how things should be ordered in each vnode op: * * ZFS_ENTER(zfsvfs); // exit if unmounted * top: * zfs_dirent_lookup(&dl, ...) // lock directory entry (may VN_HOLD()) * rw_enter(...); // grab any other locks you need * tx = dmu_tx_create(...); // get DMU tx * dmu_tx_hold_*(); // hold each object you might modify * error = dmu_tx_assign(tx, (waited ? TXG_NOTHROTTLE : 0) | TXG_NOWAIT); * if (error) { * rw_exit(...); // drop locks * zfs_dirent_unlock(dl); // unlock directory entry * VN_RELE(...); // release held vnodes * if (error == ERESTART) { * waited = B_TRUE; * dmu_tx_wait(tx); * dmu_tx_abort(tx); * goto top; * } * dmu_tx_abort(tx); // abort DMU tx * ZFS_EXIT(zfsvfs); // finished in zfs * return (error); // really out of space * } * error = do_real_work(); // do whatever this VOP does * if (error == 0) * zfs_log_*(...); // on success, make ZIL entry * dmu_tx_commit(tx); // commit DMU tx -- error or not * rw_exit(...); // drop locks * zfs_dirent_unlock(dl); // unlock directory entry * VN_RELE(...); // release held vnodes * zil_commit(zilog, foid); // synchronous when necessary * ZFS_EXIT(zfsvfs); // finished in zfs * return (error); // done, report error */ /* ARGSUSED */ static int zfs_open(vnode_t **vpp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(*vpp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((flag & FWRITE) && (zp->z_pflags & ZFS_APPENDONLY) && ((flag & FAPPEND) == 0)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (!zfs_has_ctldir(zp) && zp->z_zfsvfs->z_vscan && ZTOV(zp)->v_type == VREG && !(zp->z_pflags & ZFS_AV_QUARANTINED) && zp->z_size > 0) { if (fs_vscan(*vpp, cr, 0) != 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EACCES)); } } /* Keep a count of the synchronous opens in the znode */ if (flag & (FSYNC | FDSYNC)) atomic_inc_32(&zp->z_sync_cnt); ZFS_EXIT(zfsvfs); return (0); } /* ARGSUSED */ static int zfs_close(vnode_t *vp, int flag, int count, offset_t offset, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; /* * Clean up any locks held by this process on the vp. */ cleanlocks(vp, ddi_get_pid(), 0); cleanshares(vp, ddi_get_pid()); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* Decrement the synchronous opens in the znode */ if ((flag & (FSYNC | FDSYNC)) && (count == 1)) atomic_dec_32(&zp->z_sync_cnt); if (!zfs_has_ctldir(zp) && zp->z_zfsvfs->z_vscan && ZTOV(zp)->v_type == VREG && !(zp->z_pflags & ZFS_AV_QUARANTINED) && zp->z_size > 0) VERIFY(fs_vscan(vp, cr, 1) == 0); ZFS_EXIT(zfsvfs); return (0); } /* * Lseek support for finding holes (cmd == _FIO_SEEK_HOLE) and * data (cmd == _FIO_SEEK_DATA). "off" is an in/out parameter. */ static int zfs_holey(vnode_t *vp, u_long cmd, offset_t *off) { znode_t *zp = VTOZ(vp); uint64_t noff = (uint64_t)*off; /* new offset */ uint64_t file_sz; int error; boolean_t hole; file_sz = zp->z_size; if (noff >= file_sz) { return (SET_ERROR(ENXIO)); } if (cmd == _FIO_SEEK_HOLE) hole = B_TRUE; else hole = B_FALSE; error = dmu_offset_next(zp->z_zfsvfs->z_os, zp->z_id, hole, &noff); if (error == ESRCH) return (SET_ERROR(ENXIO)); /* * We could find a hole that begins after the logical end-of-file, * because dmu_offset_next() only works on whole blocks. If the * EOF falls mid-block, then indicate that the "virtual hole" * at the end of the file begins at the logical EOF, rather than * at the end of the last block. */ if (noff > file_sz) { ASSERT(hole); noff = file_sz; } if (noff < *off) return (error); *off = noff; return (error); } /* ARGSUSED */ static int zfs_ioctl(vnode_t *vp, u_long com, intptr_t data, int flag, cred_t *cred, int *rvalp, caller_context_t *ct) { offset_t off; offset_t ndata; dmu_object_info_t doi; int error; zfsvfs_t *zfsvfs; znode_t *zp; switch (com) { case _FIOFFS: { return (0); /* * The following two ioctls are used by bfu. Faking out, * necessary to avoid bfu errors. */ } case _FIOGDIO: case _FIOSDIO: { return (0); } case _FIO_SEEK_DATA: case _FIO_SEEK_HOLE: { #ifdef illumos if (ddi_copyin((void *)data, &off, sizeof (off), flag)) return (SET_ERROR(EFAULT)); #else off = *(offset_t *)data; #endif zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* offset parameter is in/out */ error = zfs_holey(vp, com, &off); ZFS_EXIT(zfsvfs); if (error) return (error); #ifdef illumos if (ddi_copyout(&off, (void *)data, sizeof (off), flag)) return (SET_ERROR(EFAULT)); #else *(offset_t *)data = off; #endif return (0); } #ifdef illumos case _FIO_COUNT_FILLED: { /* * _FIO_COUNT_FILLED adds a new ioctl command which * exposes the number of filled blocks in a * ZFS object. */ zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* * Wait for all dirty blocks for this object * to get synced out to disk, and the DMU info * updated. */ error = dmu_object_wait_synced(zfsvfs->z_os, zp->z_id); if (error) { ZFS_EXIT(zfsvfs); return (error); } /* * Retrieve fill count from DMU object. */ error = dmu_object_info(zfsvfs->z_os, zp->z_id, &doi); if (error) { ZFS_EXIT(zfsvfs); return (error); } ndata = doi.doi_fill_count; ZFS_EXIT(zfsvfs); if (ddi_copyout(&ndata, (void *)data, sizeof (ndata), flag)) return (SET_ERROR(EFAULT)); return (0); } #endif } return (SET_ERROR(ENOTTY)); } static vm_page_t page_busy(vnode_t *vp, int64_t start, int64_t off, int64_t nbytes) { vm_object_t obj; vm_page_t pp; int64_t end; /* * At present vm_page_clear_dirty extends the cleared range to DEV_BSIZE * aligned boundaries, if the range is not aligned. As a result a * DEV_BSIZE subrange with partially dirty data may get marked as clean. * It may happen that all DEV_BSIZE subranges are marked clean and thus * the whole page would be considred clean despite have some dirty data. * For this reason we should shrink the range to DEV_BSIZE aligned * boundaries before calling vm_page_clear_dirty. */ end = rounddown2(off + nbytes, DEV_BSIZE); off = roundup2(off, DEV_BSIZE); nbytes = end - off; obj = vp->v_object; zfs_vmobject_assert_wlocked(obj); vm_page_grab_valid(&pp, obj, OFF_TO_IDX(start), VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_NORMAL | VM_ALLOC_IGN_SBUSY); if (pp != NULL) { ASSERT3U(pp->valid, ==, VM_PAGE_BITS_ALL); vm_object_pip_add(obj, 1); pmap_remove_write(pp); if (nbytes != 0) vm_page_clear_dirty(pp, off, nbytes); } return (pp); } static void page_unbusy(vm_page_t pp) { vm_page_sunbusy(pp); vm_object_pip_wakeup(pp->object); } static vm_page_t page_wire(vnode_t *vp, int64_t start) { vm_object_t obj; vm_page_t m; obj = vp->v_object; zfs_vmobject_assert_wlocked(obj); vm_page_grab_valid(&m, obj, OFF_TO_IDX(start), VM_ALLOC_NOCREAT | VM_ALLOC_WIRED | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOBUSY); return (m); } static void page_unwire(vm_page_t pp) { vm_page_unwire(pp, PQ_ACTIVE); } /* * When a file is memory mapped, we must keep the IO data synchronized * between the DMU cache and the memory mapped pages. What this means: * * On Write: If we find a memory mapped page, we write to *both* * the page and the dmu buffer. */ static void update_pages(vnode_t *vp, int64_t start, int len, objset_t *os, uint64_t oid, int segflg, dmu_tx_t *tx) { vm_object_t obj; struct sf_buf *sf; caddr_t va; int off; ASSERT(segflg != UIO_NOCOPY); ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); off = start & PAGEOFFSET; zfs_vmobject_wlock(obj); vm_object_pip_add(obj, 1); for (start &= PAGEMASK; len > 0; start += PAGESIZE) { vm_page_t pp; int nbytes = imin(PAGESIZE - off, len); if ((pp = page_busy(vp, start, off, nbytes)) != NULL) { zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); (void) dmu_read(os, oid, start+off, nbytes, va+off, DMU_READ_PREFETCH);; zfs_unmap_page(sf); zfs_vmobject_wlock(obj); page_unbusy(pp); } len -= nbytes; off = 0; } vm_object_pip_wakeup(obj); zfs_vmobject_wunlock(obj); } /* * Read with UIO_NOCOPY flag means that sendfile(2) requests * ZFS to populate a range of page cache pages with data. * * NOTE: this function could be optimized to pre-allocate * all pages in advance, drain exclusive busy on all of them, * map them into contiguous KVA region and populate them * in one single dmu_read() call. */ static int mappedread_sf(vnode_t *vp, int nbytes, uio_t *uio) { znode_t *zp = VTOZ(vp); objset_t *os = zp->z_zfsvfs->z_os; struct sf_buf *sf; vm_object_t obj; vm_page_t pp; int64_t start; caddr_t va; int len = nbytes; int off; int error = 0; ASSERT(uio->uio_segflg == UIO_NOCOPY); ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); ASSERT((uio->uio_loffset & PAGEOFFSET) == 0); zfs_vmobject_wlock(obj); for (start = uio->uio_loffset; len > 0; start += PAGESIZE) { int bytes = MIN(PAGESIZE, len); pp = vm_page_grab(obj, OFF_TO_IDX(start), VM_ALLOC_SBUSY | VM_ALLOC_NORMAL | VM_ALLOC_IGN_SBUSY); if (vm_page_none_valid(pp)) { zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); error = dmu_read(os, zp->z_id, start, bytes, va, DMU_READ_PREFETCH); if (bytes != PAGESIZE && error == 0) bzero(va + bytes, PAGESIZE - bytes); zfs_unmap_page(sf); zfs_vmobject_wlock(obj); if (error == 0) { vm_page_valid(pp); - vm_page_lock(pp); vm_page_activate(pp); - vm_page_unlock(pp); } vm_page_sunbusy(pp); if (error != 0 && !vm_page_wired(pp) && pp->valid == 0 && vm_page_tryxbusy(pp)) vm_page_free(pp); } else { ASSERT3U(pp->valid, ==, VM_PAGE_BITS_ALL); vm_page_sunbusy(pp); } if (error) break; uio->uio_resid -= bytes; uio->uio_offset += bytes; len -= bytes; } zfs_vmobject_wunlock(obj); return (error); } /* * When a file is memory mapped, we must keep the IO data synchronized * between the DMU cache and the memory mapped pages. What this means: * * On Read: We "read" preferentially from memory mapped pages, * else we default from the dmu buffer. * * NOTE: We will always "break up" the IO into PAGESIZE uiomoves when * the file is memory mapped. */ static int mappedread(vnode_t *vp, int nbytes, uio_t *uio) { znode_t *zp = VTOZ(vp); vm_object_t obj; int64_t start; caddr_t va; int len = nbytes; int off; int error = 0; ASSERT(vp->v_mount != NULL); obj = vp->v_object; ASSERT(obj != NULL); start = uio->uio_loffset; off = start & PAGEOFFSET; zfs_vmobject_wlock(obj); for (start &= PAGEMASK; len > 0; start += PAGESIZE) { vm_page_t pp; uint64_t bytes = MIN(PAGESIZE - off, len); if (pp = page_wire(vp, start)) { struct sf_buf *sf; caddr_t va; zfs_vmobject_wunlock(obj); va = zfs_map_page(pp, &sf); #ifdef illumos error = uiomove(va + off, bytes, UIO_READ, uio); #else error = vn_io_fault_uiomove(va + off, bytes, uio); #endif zfs_unmap_page(sf); zfs_vmobject_wlock(obj); page_unwire(pp); } else { zfs_vmobject_wunlock(obj); error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, bytes); zfs_vmobject_wlock(obj); } len -= bytes; off = 0; if (error) break; } zfs_vmobject_wunlock(obj); return (error); } offset_t zfs_read_chunk_size = 1024 * 1024; /* Tunable */ /* * Read bytes from specified file into supplied buffer. * * IN: vp - vnode of file to be read from. * uio - structure supplying read location, range info, * and return buffer. * ioflag - SYNC flags; used to provide FRSYNC semantics. * cr - credentials of caller. * ct - caller context * * OUT: uio - updated offset and range, buffer filled. * * RETURN: 0 on success, error code on failure. * * Side Effects: * vp - atime updated if byte count > 0 */ /* ARGSUSED */ static int zfs_read(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ssize_t n, nbytes; int error = 0; xuio_t *xuio = NULL; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (zp->z_pflags & ZFS_AV_QUARANTINED) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EACCES)); } /* * Validate file offset */ if (uio->uio_loffset < (offset_t)0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Fasttrack empty reads */ if (uio->uio_resid == 0) { ZFS_EXIT(zfsvfs); return (0); } /* * Check for mandatory locks */ if (MANDMODE(zp->z_mode)) { if (error = chklock(vp, FREAD, uio->uio_loffset, uio->uio_resid, uio->uio_fmode, ct)) { ZFS_EXIT(zfsvfs); return (error); } } /* * If we're in FRSYNC mode, sync out this znode before reading it. */ if (zfsvfs->z_log && (ioflag & FRSYNC || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)) zil_commit(zfsvfs->z_log, zp->z_id); /* * Lock the range against changes. */ locked_range_t *lr = rangelock_enter(&zp->z_rangelock, uio->uio_loffset, uio->uio_resid, RL_READER); /* * If we are reading past end-of-file we can skip * to the end; but we might still need to set atime. */ if (uio->uio_loffset >= zp->z_size) { error = 0; goto out; } ASSERT(uio->uio_loffset < zp->z_size); n = MIN(uio->uio_resid, zp->z_size - uio->uio_loffset); #ifdef illumos if ((uio->uio_extflg == UIO_XUIO) && (((xuio_t *)uio)->xu_type == UIOTYPE_ZEROCOPY)) { int nblk; int blksz = zp->z_blksz; uint64_t offset = uio->uio_loffset; xuio = (xuio_t *)uio; if ((ISP2(blksz))) { nblk = (P2ROUNDUP(offset + n, blksz) - P2ALIGN(offset, blksz)) / blksz; } else { ASSERT(offset + n <= blksz); nblk = 1; } (void) dmu_xuio_init(xuio, nblk); if (vn_has_cached_data(vp)) { /* * For simplicity, we always allocate a full buffer * even if we only expect to read a portion of a block. */ while (--nblk >= 0) { (void) dmu_xuio_add(xuio, dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl), blksz), 0, blksz); } } } #endif /* illumos */ while (n > 0) { nbytes = MIN(n, zfs_read_chunk_size - P2PHASE(uio->uio_loffset, zfs_read_chunk_size)); #ifdef __FreeBSD__ if (uio->uio_segflg == UIO_NOCOPY) error = mappedread_sf(vp, nbytes, uio); else #endif /* __FreeBSD__ */ if (vn_has_cached_data(vp)) { error = mappedread(vp, nbytes, uio); } else { error = dmu_read_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, nbytes); } if (error) { /* convert checksum errors into IO errors */ if (error == ECKSUM) error = SET_ERROR(EIO); break; } n -= nbytes; } out: rangelock_exit(lr); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); return (error); } /* * Write the bytes to a file. * * IN: vp - vnode of file to be written to. * uio - structure supplying write location, range info, * and data buffer. * ioflag - FAPPEND, FSYNC, and/or FDSYNC. FAPPEND is * set if in append mode. * cr - credentials of caller. * ct - caller context (NFS/CIFS fem monitor only) * * OUT: uio - updated offset and range. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - ctime|mtime updated if byte count > 0 */ /* ARGSUSED */ static int zfs_write(vnode_t *vp, uio_t *uio, int ioflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); rlim64_t limit = MAXOFFSET_T; ssize_t start_resid = uio->uio_resid; ssize_t tx_bytes; uint64_t end_size; dmu_tx_t *tx; zfsvfs_t *zfsvfs = zp->z_zfsvfs; zilog_t *zilog; offset_t woff; ssize_t n, nbytes; int max_blksz = zfsvfs->z_max_blksz; int error = 0; arc_buf_t *abuf; iovec_t *aiov = NULL; xuio_t *xuio = NULL; int i_iov = 0; int iovcnt = uio->uio_iovcnt; iovec_t *iovp = uio->uio_iov; int write_eof; int count = 0; sa_bulk_attr_t bulk[4]; uint64_t mtime[2], ctime[2]; /* * Fasttrack empty write */ n = start_resid; if (n == 0) return (0); if (limit == RLIM64_INFINITY || limit > MAXOFFSET_T) limit = MAXOFFSET_T; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_SIZE(zfsvfs), NULL, &zp->z_size, 8); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, 8); /* * In a case vp->v_vfsp != zp->z_zfsvfs->z_vfs (e.g. snapshots) our * callers might not be able to detect properly that we are read-only, * so check it explicitly here. */ if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EROFS)); } /* * If immutable or not appending then return EPERM. * Intentionally allow ZFS_READONLY through here. * See zfs_zaccess_common() */ if ((zp->z_pflags & ZFS_IMMUTABLE) || ((zp->z_pflags & ZFS_APPENDONLY) && !(ioflag & FAPPEND) && (uio->uio_loffset < zp->z_size))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } zilog = zfsvfs->z_log; /* * Validate file offset */ woff = ioflag & FAPPEND ? zp->z_size : uio->uio_loffset; if (woff < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Check for mandatory locks before calling rangelock_enter() * in order to prevent a deadlock with locks set via fcntl(). */ if (MANDMODE((mode_t)zp->z_mode) && (error = chklock(vp, FWRITE, woff, n, uio->uio_fmode, ct)) != 0) { ZFS_EXIT(zfsvfs); return (error); } #ifdef illumos /* * Pre-fault the pages to ensure slow (eg NFS) pages * don't hold up txg. * Skip this if uio contains loaned arc_buf. */ if ((uio->uio_extflg == UIO_XUIO) && (((xuio_t *)uio)->xu_type == UIOTYPE_ZEROCOPY)) xuio = (xuio_t *)uio; else uio_prefaultpages(MIN(n, max_blksz), uio); #endif /* * If in append mode, set the io offset pointer to eof. */ locked_range_t *lr; if (ioflag & FAPPEND) { /* * Obtain an appending range lock to guarantee file append * semantics. We reset the write offset once we have the lock. */ lr = rangelock_enter(&zp->z_rangelock, 0, n, RL_APPEND); woff = lr->lr_offset; if (lr->lr_length == UINT64_MAX) { /* * We overlocked the file because this write will cause * the file block size to increase. * Note that zp_size cannot change with this lock held. */ woff = zp->z_size; } uio->uio_loffset = woff; } else { /* * Note that if the file block size will change as a result of * this write, then this range lock will lock the entire file * so that we can re-write the block safely. */ lr = rangelock_enter(&zp->z_rangelock, woff, n, RL_WRITER); } if (vn_rlimit_fsize(vp, uio, uio->uio_td)) { rangelock_exit(lr); ZFS_EXIT(zfsvfs); return (EFBIG); } if (woff >= limit) { rangelock_exit(lr); ZFS_EXIT(zfsvfs); return (SET_ERROR(EFBIG)); } if ((woff + n) > limit || woff > (limit - n)) n = limit - woff; /* Will this write extend the file length? */ write_eof = (woff + n > zp->z_size); end_size = MAX(zp->z_size, woff + n); /* * Write the file in reasonable size chunks. Each chunk is written * in a separate transaction; this keeps the intent log records small * and allows us to do more fine-grained space accounting. */ while (n > 0) { abuf = NULL; woff = uio->uio_loffset; if (zfs_owner_overquota(zfsvfs, zp, B_FALSE) || zfs_owner_overquota(zfsvfs, zp, B_TRUE)) { if (abuf != NULL) dmu_return_arcbuf(abuf); error = SET_ERROR(EDQUOT); break; } if (xuio && abuf == NULL) { ASSERT(i_iov < iovcnt); aiov = &iovp[i_iov]; abuf = dmu_xuio_arcbuf(xuio, i_iov); dmu_xuio_clear(xuio, i_iov); DTRACE_PROBE3(zfs_cp_write, int, i_iov, iovec_t *, aiov, arc_buf_t *, abuf); ASSERT((aiov->iov_base == abuf->b_data) || ((char *)aiov->iov_base - (char *)abuf->b_data + aiov->iov_len == arc_buf_size(abuf))); i_iov++; } else if (abuf == NULL && n >= max_blksz && woff >= zp->z_size && P2PHASE(woff, max_blksz) == 0 && zp->z_blksz == max_blksz) { /* * This write covers a full block. "Borrow" a buffer * from the dmu so that we can fill it before we enter * a transaction. This avoids the possibility of * holding up the transaction if the data copy hangs * up on a pagefault (e.g., from an NFS server mapping). */ size_t cbytes; abuf = dmu_request_arcbuf(sa_get_db(zp->z_sa_hdl), max_blksz); ASSERT(abuf != NULL); ASSERT(arc_buf_size(abuf) == max_blksz); if (error = uiocopy(abuf->b_data, max_blksz, UIO_WRITE, uio, &cbytes)) { dmu_return_arcbuf(abuf); break; } ASSERT(cbytes == max_blksz); } /* * Start a transaction. */ tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); dmu_tx_hold_write(tx, zp->z_id, woff, MIN(n, max_blksz)); zfs_sa_upgrade_txholds(tx, zp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); if (abuf != NULL) dmu_return_arcbuf(abuf); break; } /* * If rangelock_enter() over-locked we grow the blocksize * and then reduce the lock range. This will only happen * on the first iteration since rangelock_reduce() will * shrink down lr_length to the appropriate size. */ if (lr->lr_length == UINT64_MAX) { uint64_t new_blksz; if (zp->z_blksz > max_blksz) { /* * File's blocksize is already larger than the * "recordsize" property. Only let it grow to * the next power of 2. */ ASSERT(!ISP2(zp->z_blksz)); new_blksz = MIN(end_size, 1 << highbit64(zp->z_blksz)); } else { new_blksz = MIN(end_size, max_blksz); } zfs_grow_blocksize(zp, new_blksz, tx); rangelock_reduce(lr, woff, n); } /* * XXX - should we really limit each write to z_max_blksz? * Perhaps we should use SPA_MAXBLOCKSIZE chunks? */ nbytes = MIN(n, max_blksz - P2PHASE(woff, max_blksz)); if (woff + nbytes > zp->z_size) vnode_pager_setsize(vp, woff + nbytes); if (abuf == NULL) { tx_bytes = uio->uio_resid; error = dmu_write_uio_dbuf(sa_get_db(zp->z_sa_hdl), uio, nbytes, tx); tx_bytes -= uio->uio_resid; } else { tx_bytes = nbytes; ASSERT(xuio == NULL || tx_bytes == aiov->iov_len); /* * If this is not a full block write, but we are * extending the file past EOF and this data starts * block-aligned, use assign_arcbuf(). Otherwise, * write via dmu_write(). */ if (tx_bytes < max_blksz && (!write_eof || aiov->iov_base != abuf->b_data)) { ASSERT(xuio); dmu_write(zfsvfs->z_os, zp->z_id, woff, aiov->iov_len, aiov->iov_base, tx); dmu_return_arcbuf(abuf); xuio_stat_wbuf_copied(); } else { ASSERT(xuio || tx_bytes == max_blksz); dmu_assign_arcbuf(sa_get_db(zp->z_sa_hdl), woff, abuf, tx); } ASSERT(tx_bytes <= uio->uio_resid); uioskip(uio, tx_bytes); } if (tx_bytes && vn_has_cached_data(vp)) { update_pages(vp, woff, tx_bytes, zfsvfs->z_os, zp->z_id, uio->uio_segflg, tx); } /* * If we made no progress, we're done. If we made even * partial progress, update the znode and ZIL accordingly. */ if (tx_bytes == 0) { (void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs), (void *)&zp->z_size, sizeof (uint64_t), tx); dmu_tx_commit(tx); ASSERT(error != 0); break; } /* * Clear Set-UID/Set-GID bits on successful write if not * privileged and at least one of the excute bits is set. * * It would be nice to to this after all writes have * been done, but that would still expose the ISUID/ISGID * to another app after the partial write is committed. * * Note: we don't call zfs_fuid_map_id() here because * user 0 is not an ephemeral uid. */ mutex_enter(&zp->z_acl_lock); if ((zp->z_mode & (S_IXUSR | (S_IXUSR >> 3) | (S_IXUSR >> 6))) != 0 && (zp->z_mode & (S_ISUID | S_ISGID)) != 0 && secpolicy_vnode_setid_retain(vp, cr, (zp->z_mode & S_ISUID) != 0 && zp->z_uid == 0) != 0) { uint64_t newmode; zp->z_mode &= ~(S_ISUID | S_ISGID); newmode = zp->z_mode; (void) sa_update(zp->z_sa_hdl, SA_ZPL_MODE(zfsvfs), (void *)&newmode, sizeof (uint64_t), tx); } mutex_exit(&zp->z_acl_lock); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); /* * Update the file size (zp_size) if it has changed; * account for possible concurrent updates. */ while ((end_size = zp->z_size) < uio->uio_loffset) { (void) atomic_cas_64(&zp->z_size, end_size, uio->uio_loffset); #ifdef illumos ASSERT(error == 0); #else ASSERT(error == 0 || error == EFAULT); #endif } /* * If we are replaying and eof is non zero then force * the file size to the specified eof. Note, there's no * concurrency during replay. */ if (zfsvfs->z_replay && zfsvfs->z_replay_eof != 0) zp->z_size = zfsvfs->z_replay_eof; if (error == 0) error = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); else (void) sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); zfs_log_write(zilog, tx, TX_WRITE, zp, woff, tx_bytes, ioflag); dmu_tx_commit(tx); if (error != 0) break; ASSERT(tx_bytes == nbytes); n -= nbytes; #ifdef illumos if (!xuio && n > 0) uio_prefaultpages(MIN(n, max_blksz), uio); #endif } rangelock_exit(lr); /* * If we're in replay mode, or we made no progress, return error. * Otherwise, it's at least a partial write, so it's successful. */ if (zfsvfs->z_replay || uio->uio_resid == start_resid) { ZFS_EXIT(zfsvfs); return (error); } #ifdef __FreeBSD__ /* * EFAULT means that at least one page of the source buffer was not * available. VFS will re-try remaining I/O upon this error. */ if (error == EFAULT) { ZFS_EXIT(zfsvfs); return (error); } #endif if (ioflag & (FSYNC | FDSYNC) || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, zp->z_id); ZFS_EXIT(zfsvfs); return (0); } /* ARGSUSED */ void zfs_get_done(zgd_t *zgd, int error) { znode_t *zp = zgd->zgd_private; objset_t *os = zp->z_zfsvfs->z_os; if (zgd->zgd_db) dmu_buf_rele(zgd->zgd_db, zgd); rangelock_exit(zgd->zgd_lr); /* * Release the vnode asynchronously as we currently have the * txg stopped from syncing. */ VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(os))); kmem_free(zgd, sizeof (zgd_t)); } #ifdef DEBUG static int zil_fault_io = 0; #endif /* * Get data to generate a TX_WRITE intent log record. */ int zfs_get_data(void *arg, lr_write_t *lr, char *buf, struct lwb *lwb, zio_t *zio) { zfsvfs_t *zfsvfs = arg; objset_t *os = zfsvfs->z_os; znode_t *zp; uint64_t object = lr->lr_foid; uint64_t offset = lr->lr_offset; uint64_t size = lr->lr_length; dmu_buf_t *db; zgd_t *zgd; int error = 0; ASSERT3P(lwb, !=, NULL); ASSERT3P(zio, !=, NULL); ASSERT3U(size, !=, 0); /* * Nothing to do if the file has been removed */ if (zfs_zget(zfsvfs, object, &zp) != 0) return (SET_ERROR(ENOENT)); if (zp->z_unlinked) { /* * Release the vnode asynchronously as we currently have the * txg stopped from syncing. */ VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(os))); return (SET_ERROR(ENOENT)); } zgd = (zgd_t *)kmem_zalloc(sizeof (zgd_t), KM_SLEEP); zgd->zgd_lwb = lwb; zgd->zgd_private = zp; /* * Write records come in two flavors: immediate and indirect. * For small writes it's cheaper to store the data with the * log record (immediate); for large writes it's cheaper to * sync the data and get a pointer to it (indirect) so that * we don't have to write the data twice. */ if (buf != NULL) { /* immediate write */ zgd->zgd_lr = rangelock_enter(&zp->z_rangelock, offset, size, RL_READER); /* test for truncation needs to be done while range locked */ if (offset >= zp->z_size) { error = SET_ERROR(ENOENT); } else { error = dmu_read(os, object, offset, size, buf, DMU_READ_NO_PREFETCH); } ASSERT(error == 0 || error == ENOENT); } else { /* indirect write */ /* * Have to lock the whole block to ensure when it's * written out and its checksum is being calculated * that no one can change the data. We need to re-check * blocksize after we get the lock in case it's changed! */ for (;;) { uint64_t blkoff; size = zp->z_blksz; blkoff = ISP2(size) ? P2PHASE(offset, size) : offset; offset -= blkoff; zgd->zgd_lr = rangelock_enter(&zp->z_rangelock, offset, size, RL_READER); if (zp->z_blksz == size) break; offset += blkoff; rangelock_exit(zgd->zgd_lr); } /* test for truncation needs to be done while range locked */ if (lr->lr_offset >= zp->z_size) error = SET_ERROR(ENOENT); #ifdef DEBUG if (zil_fault_io) { error = SET_ERROR(EIO); zil_fault_io = 0; } #endif if (error == 0) error = dmu_buf_hold(os, object, offset, zgd, &db, DMU_READ_NO_PREFETCH); if (error == 0) { blkptr_t *bp = &lr->lr_blkptr; zgd->zgd_db = db; zgd->zgd_bp = bp; ASSERT(db->db_offset == offset); ASSERT(db->db_size == size); error = dmu_sync(zio, lr->lr_common.lrc_txg, zfs_get_done, zgd); ASSERT(error || lr->lr_length <= size); /* * On success, we need to wait for the write I/O * initiated by dmu_sync() to complete before we can * release this dbuf. We will finish everything up * in the zfs_get_done() callback. */ if (error == 0) return (0); if (error == EALREADY) { lr->lr_common.lrc_txtype = TX_WRITE2; /* * TX_WRITE2 relies on the data previously * written by the TX_WRITE that caused * EALREADY. We zero out the BP because * it is the old, currently-on-disk BP. */ zgd->zgd_bp = NULL; BP_ZERO(bp); error = 0; } } } zfs_get_done(zgd, error); return (error); } /*ARGSUSED*/ static int zfs_access(vnode_t *vp, int mode, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (flag & V_ACE_MASK) error = zfs_zaccess(zp, mode, flag, B_FALSE, cr); else error = zfs_zaccess_rwx(zp, mode, flag, cr); ZFS_EXIT(zfsvfs); return (error); } static int zfs_dd_callback(struct mount *mp, void *arg, int lkflags, struct vnode **vpp) { int error; *vpp = arg; error = vn_lock(*vpp, lkflags); if (error != 0) vrele(*vpp); return (error); } static int zfs_lookup_lock(vnode_t *dvp, vnode_t *vp, const char *name, int lkflags) { znode_t *zdp = VTOZ(dvp); zfsvfs_t *zfsvfs = zdp->z_zfsvfs; int error; int ltype; ASSERT_VOP_LOCKED(dvp, __func__); #ifdef DIAGNOSTIC if ((zdp->z_pflags & ZFS_XATTR) == 0) VERIFY(!RRM_LOCK_HELD(&zfsvfs->z_teardown_lock)); #endif if (name[0] == 0 || (name[0] == '.' && name[1] == 0)) { ASSERT3P(dvp, ==, vp); vref(dvp); ltype = lkflags & LK_TYPE_MASK; if (ltype != VOP_ISLOCKED(dvp)) { if (ltype == LK_EXCLUSIVE) vn_lock(dvp, LK_UPGRADE | LK_RETRY); else /* if (ltype == LK_SHARED) */ vn_lock(dvp, LK_DOWNGRADE | LK_RETRY); /* * Relock for the "." case could leave us with * reclaimed vnode. */ if (VN_IS_DOOMED(dvp)) { vrele(dvp); return (SET_ERROR(ENOENT)); } } return (0); } else if (name[0] == '.' && name[1] == '.' && name[2] == 0) { /* * Note that in this case, dvp is the child vnode, and we * are looking up the parent vnode - exactly reverse from * normal operation. Unlocking dvp requires some rather * tricky unlock/relock dance to prevent mp from being freed; * use vn_vget_ino_gen() which takes care of all that. * * XXX Note that there is a time window when both vnodes are * unlocked. It is possible, although highly unlikely, that * during that window the parent-child relationship between * the vnodes may change, for example, get reversed. * In that case we would have a wrong lock order for the vnodes. * All other filesystems seem to ignore this problem, so we * do the same here. * A potential solution could be implemented as follows: * - using LK_NOWAIT when locking the second vnode and retrying * if necessary * - checking that the parent-child relationship still holds * after locking both vnodes and retrying if it doesn't */ error = vn_vget_ino_gen(dvp, zfs_dd_callback, vp, lkflags, &vp); return (error); } else { error = vn_lock(vp, lkflags); if (error != 0) vrele(vp); return (error); } } /* * Lookup an entry in a directory, or an extended attribute directory. * If it exists, return a held vnode reference for it. * * IN: dvp - vnode of directory to search. * nm - name of entry to lookup. * pnp - full pathname to lookup [UNUSED]. * flags - LOOKUP_XATTR set if looking for an attribute. * rdir - root directory vnode [UNUSED]. * cr - credentials of caller. * ct - caller context * * OUT: vpp - vnode of located entry, NULL if not found. * * RETURN: 0 on success, error code on failure. * * Timestamps: * NA */ /* ARGSUSED */ static int zfs_lookup(vnode_t *dvp, char *nm, vnode_t **vpp, struct componentname *cnp, int nameiop, cred_t *cr, kthread_t *td, int flags) { znode_t *zdp = VTOZ(dvp); znode_t *zp; zfsvfs_t *zfsvfs = zdp->z_zfsvfs; int error = 0; /* * Fast path lookup, however we must skip DNLC lookup * for case folding or normalizing lookups because the * DNLC code only stores the passed in name. This means * creating 'a' and removing 'A' on a case insensitive * file system would work, but DNLC still thinks 'a' * exists and won't let you create it again on the next * pass through fast path. */ if (!(flags & LOOKUP_XATTR)) { if (dvp->v_type != VDIR) { return (SET_ERROR(ENOTDIR)); } else if (zdp->z_sa_hdl == NULL) { return (SET_ERROR(EIO)); } } DTRACE_PROBE2(zfs__fastpath__lookup__miss, vnode_t *, dvp, char *, nm); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zdp); *vpp = NULL; if (flags & LOOKUP_XATTR) { #ifdef TODO /* * If the xattr property is off, refuse the lookup request. */ if (!(zfsvfs->z_vfs->vfs_flag & VFS_XATTR)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } #endif /* * We don't allow recursive attributes.. * Maybe someday we will. */ if (zdp->z_pflags & ZFS_XATTR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (error = zfs_get_xattrdir(VTOZ(dvp), vpp, cr, flags)) { ZFS_EXIT(zfsvfs); return (error); } /* * Do we have permission to get into attribute directory? */ if (error = zfs_zaccess(VTOZ(*vpp), ACE_EXECUTE, 0, B_FALSE, cr)) { vrele(*vpp); *vpp = NULL; } ZFS_EXIT(zfsvfs); return (error); } /* * Check accessibility of directory. */ if (error = zfs_zaccess(zdp, ACE_EXECUTE, 0, B_FALSE, cr)) { ZFS_EXIT(zfsvfs); return (error); } if (zfsvfs->z_utf8 && u8_validate(nm, strlen(nm), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } /* * First handle the special cases. */ if ((cnp->cn_flags & ISDOTDOT) != 0) { /* * If we are a snapshot mounted under .zfs, return * the vp for the snapshot directory. */ if (zdp->z_id == zfsvfs->z_root && zfsvfs->z_parent != zfsvfs) { struct componentname cn; vnode_t *zfsctl_vp; int ltype; ZFS_EXIT(zfsvfs); ltype = VOP_ISLOCKED(dvp); VOP_UNLOCK(dvp, 0); error = zfsctl_root(zfsvfs->z_parent, LK_SHARED, &zfsctl_vp); if (error == 0) { cn.cn_nameptr = "snapshot"; cn.cn_namelen = strlen(cn.cn_nameptr); cn.cn_nameiop = cnp->cn_nameiop; cn.cn_flags = cnp->cn_flags & ~ISDOTDOT; cn.cn_lkflags = cnp->cn_lkflags; error = VOP_LOOKUP(zfsctl_vp, vpp, &cn); vput(zfsctl_vp); } vn_lock(dvp, ltype | LK_RETRY); return (error); } } if (zfs_has_ctldir(zdp) && strcmp(nm, ZFS_CTLDIR_NAME) == 0) { ZFS_EXIT(zfsvfs); if ((cnp->cn_flags & ISLASTCN) != 0 && nameiop != LOOKUP) return (SET_ERROR(ENOTSUP)); error = zfsctl_root(zfsvfs, cnp->cn_lkflags, vpp); return (error); } /* * The loop is retry the lookup if the parent-child relationship * changes during the dot-dot locking complexities. */ for (;;) { uint64_t parent; error = zfs_dirlook(zdp, nm, &zp); if (error == 0) *vpp = ZTOV(zp); ZFS_EXIT(zfsvfs); if (error != 0) break; error = zfs_lookup_lock(dvp, *vpp, nm, cnp->cn_lkflags); if (error != 0) { /* * If we've got a locking error, then the vnode * got reclaimed because of a force unmount. * We never enter doomed vnodes into the name cache. */ *vpp = NULL; return (error); } if ((cnp->cn_flags & ISDOTDOT) == 0) break; ZFS_ENTER(zfsvfs); if (zdp->z_sa_hdl == NULL) { error = SET_ERROR(EIO); } else { error = sa_lookup(zdp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent)); } if (error != 0) { ZFS_EXIT(zfsvfs); vput(ZTOV(zp)); break; } if (zp->z_id == parent) { ZFS_EXIT(zfsvfs); break; } vput(ZTOV(zp)); } out: if (error != 0) *vpp = NULL; /* Translate errors and add SAVENAME when needed. */ if (cnp->cn_flags & ISLASTCN) { switch (nameiop) { case CREATE: case RENAME: if (error == ENOENT) { error = EJUSTRETURN; cnp->cn_flags |= SAVENAME; break; } /* FALLTHROUGH */ case DELETE: if (error == 0) cnp->cn_flags |= SAVENAME; break; } } /* Insert name into cache (as non-existent) if appropriate. */ if (zfsvfs->z_use_namecache && error == ENOENT && (cnp->cn_flags & MAKEENTRY) != 0) cache_enter(dvp, NULL, cnp); /* Insert name into cache if appropriate. */ if (zfsvfs->z_use_namecache && error == 0 && (cnp->cn_flags & MAKEENTRY)) { if (!(cnp->cn_flags & ISLASTCN) || (nameiop != DELETE && nameiop != RENAME)) { cache_enter(dvp, *vpp, cnp); } } return (error); } /* * Attempt to create a new entry in a directory. If the entry * already exists, truncate the file if permissible, else return * an error. Return the vp of the created or trunc'd file. * * IN: dvp - vnode of directory to put new file entry in. * name - name of new file entry. * vap - attributes of new file. * excl - flag indicating exclusive or non-exclusive mode. * mode - mode to open file with. * cr - credentials of caller. * flag - large file flag [UNUSED]. * ct - caller context * vsecp - ACL to be set * * OUT: vpp - vnode of created or trunc'd entry. * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated if new entry created * vp - ctime|mtime always, atime if new */ /* ARGSUSED */ static int zfs_create(vnode_t *dvp, char *name, vattr_t *vap, int excl, int mode, vnode_t **vpp, cred_t *cr, kthread_t *td) { znode_t *zp, *dzp = VTOZ(dvp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; objset_t *os; dmu_tx_t *tx; int error; ksid_t *ksid; uid_t uid; gid_t gid = crgetgid(cr); zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; void *vsecp = NULL; int flag = 0; uint64_t txtype; /* * If we have an ephemeral id, ACL, or XVATTR then * make sure file system is at proper version */ ksid = crgetsid(cr, KSID_OWNER); if (ksid) uid = ksid_getid(ksid); else uid = crgetuid(cr); if (zfsvfs->z_use_fuids == B_FALSE && (vsecp || (vap->va_mask & AT_XVATTR) || IS_EPHEMERAL(uid) || IS_EPHEMERAL(gid))) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); os = zfsvfs->z_os; zilog = zfsvfs->z_log; if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (vap->va_mask & AT_XVATTR) { if ((error = secpolicy_xvattr(dvp, (xvattr_t *)vap, crgetuid(cr), cr, vap->va_type)) != 0) { ZFS_EXIT(zfsvfs); return (error); } } *vpp = NULL; if ((vap->va_mode & S_ISVTX) && secpolicy_vnode_stky_modify(cr)) vap->va_mode &= ~S_ISVTX; error = zfs_dirent_lookup(dzp, name, &zp, ZNEW); if (error) { ZFS_EXIT(zfsvfs); return (error); } ASSERT3P(zp, ==, NULL); /* * Create a new file object and update the directory * to reference it. */ if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { goto out; } /* * We only support the creation of regular files in * extended attribute directories. */ if ((dzp->z_pflags & ZFS_XATTR) && (vap->va_type != VREG)) { error = SET_ERROR(EINVAL); goto out; } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, vsecp, &acl_ids)) != 0) goto out; if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); error = SET_ERROR(EDQUOT); goto out; } getnewvnode_reserve(1); tx = dmu_tx_create(os); dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE); fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); dmu_tx_hold_sa(tx, dzp->z_sa_hdl, B_FALSE); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); (void) zfs_link_create(dzp, name, zp, tx, ZNEW); txtype = zfs_log_create_txtype(Z_FILE, vsecp, vap); zfs_log_create(zilog, tx, txtype, dzp, zp, name, vsecp, acl_ids.z_fuidp, vap); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); out: if (error == 0) { *vpp = ZTOV(zp); } if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Remove an entry from a directory. * * IN: dvp - vnode of directory to remove entry from. * name - name of entry to remove. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime * vp - ctime (if nlink > 0) */ /*ARGSUSED*/ static int zfs_remove(vnode_t *dvp, vnode_t *vp, char *name, cred_t *cr) { znode_t *dzp = VTOZ(dvp); znode_t *zp = VTOZ(vp); znode_t *xzp; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t acl_obj, xattr_obj; uint64_t obj = 0; dmu_tx_t *tx; boolean_t unlinked, toobig = FALSE; uint64_t txtype; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; zp = VTOZ(vp); xattr_obj = 0; xzp = NULL; if (error = zfs_zaccess_delete(dzp, zp, cr)) { goto out; } /* * Need to use rmdir for removing directories. */ if (vp->v_type == VDIR) { error = SET_ERROR(EPERM); goto out; } vnevent_remove(vp, dvp, name, ct); obj = zp->z_id; /* are there any extended attributes? */ error = sa_lookup(zp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &xattr_obj, sizeof (xattr_obj)); if (error == 0 && xattr_obj) { error = zfs_zget(zfsvfs, xattr_obj, &xzp); ASSERT0(error); } /* * We may delete the znode now, or we may put it in the unlinked set; * it depends on whether we're the last link, and on whether there are * other holds on the vnode. So we dmu_tx_hold() the right things to * allow for either case. */ tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, FALSE, name); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); zfs_sa_upgrade_txholds(tx, dzp); if (xzp) { dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); dmu_tx_hold_sa(tx, xzp->z_sa_hdl, B_FALSE); } /* charge as an update -- would be nice not to charge at all */ dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); /* * Mark this transaction as typically resulting in a net free of space */ dmu_tx_mark_netfree(tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } /* * Remove the directory entry. */ error = zfs_link_destroy(dzp, name, zp, tx, ZEXISTS, &unlinked); if (error) { dmu_tx_commit(tx); goto out; } if (unlinked) { zfs_unlinked_add(zp, tx); vp->v_vflag |= VV_NOSYNC; } txtype = TX_REMOVE; zfs_log_remove(zilog, tx, txtype, dzp, name, obj); dmu_tx_commit(tx); out: if (xzp) vrele(ZTOV(xzp)); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Create a new directory and insert it into dvp using the name * provided. Return a pointer to the inserted directory. * * IN: dvp - vnode of directory to add subdir to. * dirname - name of new directory. * vap - attributes of new directory. * cr - credentials of caller. * ct - caller context * flags - case flags * vsecp - ACL to be set * * OUT: vpp - vnode of created directory. * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated * vp - ctime|mtime|atime updated */ /*ARGSUSED*/ static int zfs_mkdir(vnode_t *dvp, char *dirname, vattr_t *vap, vnode_t **vpp, cred_t *cr) { znode_t *zp, *dzp = VTOZ(dvp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t txtype; dmu_tx_t *tx; int error; ksid_t *ksid; uid_t uid; gid_t gid = crgetgid(cr); zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; ASSERT(vap->va_type == VDIR); /* * If we have an ephemeral id, ACL, or XVATTR then * make sure file system is at proper version */ ksid = crgetsid(cr, KSID_OWNER); if (ksid) uid = ksid_getid(ksid); else uid = crgetuid(cr); if (zfsvfs->z_use_fuids == B_FALSE && ((vap->va_mask & AT_XVATTR) || IS_EPHEMERAL(uid) || IS_EPHEMERAL(gid))) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; if (dzp->z_pflags & ZFS_XATTR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (zfsvfs->z_utf8 && u8_validate(dirname, strlen(dirname), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (vap->va_mask & AT_XVATTR) { if ((error = secpolicy_xvattr(dvp, (xvattr_t *)vap, crgetuid(cr), cr, vap->va_type)) != 0) { ZFS_EXIT(zfsvfs); return (error); } } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, NULL, &acl_ids)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * First make sure the new directory doesn't exist. * * Existence is checked first to make sure we don't return * EACCES instead of EEXIST which can cause some applications * to fail. */ *vpp = NULL; if (error = zfs_dirent_lookup(dzp, dirname, &zp, ZNEW)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } ASSERT3P(zp, ==, NULL); if (error = zfs_zaccess(dzp, ACE_ADD_SUBDIRECTORY, 0, B_FALSE, cr)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (SET_ERROR(EDQUOT)); } /* * Add a new entry to the directory. */ getnewvnode_reserve(1); tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, dirname); dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, FALSE, NULL); fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } /* * Create new node. */ zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); /* * Now put new name in parent dir. */ (void) zfs_link_create(dzp, dirname, zp, tx, ZNEW); *vpp = ZTOV(zp); txtype = zfs_log_create_txtype(Z_DIR, NULL, vap); zfs_log_create(zilog, tx, txtype, dzp, zp, dirname, NULL, acl_ids.z_fuidp, vap); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (0); } /* * Remove a directory subdir entry. If the current working * directory is the same as the subdir to be removed, the * remove will fail. * * IN: dvp - vnode of directory to remove from. * name - name of directory to be removed. * cwd - vnode of current working directory. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_rmdir(vnode_t *dvp, vnode_t *vp, char *name, cred_t *cr) { znode_t *dzp = VTOZ(dvp); znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; if (error = zfs_zaccess_delete(dzp, zp, cr)) { goto out; } if (vp->v_type != VDIR) { error = SET_ERROR(ENOTDIR); goto out; } vnevent_rmdir(vp, dvp, name, ct); tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_zap(tx, dzp->z_id, FALSE, name); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); zfs_sa_upgrade_txholds(tx, zp); zfs_sa_upgrade_txholds(tx, dzp); dmu_tx_mark_netfree(tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } cache_purge(dvp); error = zfs_link_destroy(dzp, name, zp, tx, ZEXISTS, NULL); if (error == 0) { uint64_t txtype = TX_RMDIR; zfs_log_remove(zilog, tx, txtype, dzp, name, ZFS_NO_OBJECT); } dmu_tx_commit(tx); cache_purge(vp); out: if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Read as many directory entries as will fit into the provided * buffer from the given directory cursor position (specified in * the uio structure). * * IN: vp - vnode of directory to read. * uio - structure supplying read location, range info, * and return buffer. * cr - credentials of caller. * ct - caller context * flags - case flags * * OUT: uio - updated offset and range, buffer filled. * eofp - set to true if end-of-file detected. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - atime updated * * Note that the low 4 bits of the cookie returned by zap is always zero. * This allows us to use the low range for "special" directory entries: * We use 0 for '.', and 1 for '..'. If this is the root of the filesystem, * we use the offset 2 for the '.zfs' directory. */ /* ARGSUSED */ static int zfs_readdir(vnode_t *vp, uio_t *uio, cred_t *cr, int *eofp, int *ncookies, u_long **cookies) { znode_t *zp = VTOZ(vp); iovec_t *iovp; edirent_t *eodp; dirent64_t *odp; zfsvfs_t *zfsvfs = zp->z_zfsvfs; objset_t *os; caddr_t outbuf; size_t bufsize; zap_cursor_t zc; zap_attribute_t zap; uint_t bytes_wanted; uint64_t offset; /* must be unsigned; checks for < 1 */ uint64_t parent; int local_eof; int outcount; int error; uint8_t prefetch; boolean_t check_sysattrs; uint8_t type; int ncooks; u_long *cooks = NULL; int flags = 0; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent))) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * If we are not given an eof variable, * use a local one. */ if (eofp == NULL) eofp = &local_eof; /* * Check for valid iov_len. */ if (uio->uio_iov->iov_len <= 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * Quit if directory has been removed (posix) */ if ((*eofp = zp->z_unlinked) != 0) { ZFS_EXIT(zfsvfs); return (0); } error = 0; os = zfsvfs->z_os; offset = uio->uio_loffset; prefetch = zp->z_zn_prefetch; /* * Initialize the iterator cursor. */ if (offset <= 3) { /* * Start iteration from the beginning of the directory. */ zap_cursor_init(&zc, os, zp->z_id); } else { /* * The offset is a serialized cursor. */ zap_cursor_init_serialized(&zc, os, zp->z_id, offset); } /* * Get space to change directory entries into fs independent format. */ iovp = uio->uio_iov; bytes_wanted = iovp->iov_len; if (uio->uio_segflg != UIO_SYSSPACE || uio->uio_iovcnt != 1) { bufsize = bytes_wanted; outbuf = kmem_alloc(bufsize, KM_SLEEP); odp = (struct dirent64 *)outbuf; } else { bufsize = bytes_wanted; outbuf = NULL; odp = (struct dirent64 *)iovp->iov_base; } eodp = (struct edirent *)odp; if (ncookies != NULL) { /* * Minimum entry size is dirent size and 1 byte for a file name. */ ncooks = uio->uio_resid / (sizeof(struct dirent) - sizeof(((struct dirent *)NULL)->d_name) + 1); cooks = malloc(ncooks * sizeof(u_long), M_TEMP, M_WAITOK); *cookies = cooks; *ncookies = ncooks; } /* * If this VFS supports the system attribute view interface; and * we're looking at an extended attribute directory; and we care * about normalization conflicts on this vfs; then we must check * for normalization conflicts with the sysattr name space. */ #ifdef TODO check_sysattrs = vfs_has_feature(vp->v_vfsp, VFSFT_SYSATTR_VIEWS) && (vp->v_flag & V_XATTRDIR) && zfsvfs->z_norm && (flags & V_RDDIR_ENTFLAGS); #else check_sysattrs = 0; #endif /* * Transform to file-system independent format */ outcount = 0; while (outcount < bytes_wanted) { ino64_t objnum; ushort_t reclen; off64_t *next = NULL; /* * Special case `.', `..', and `.zfs'. */ if (offset == 0) { (void) strcpy(zap.za_name, "."); zap.za_normalization_conflict = 0; objnum = zp->z_id; type = DT_DIR; } else if (offset == 1) { (void) strcpy(zap.za_name, ".."); zap.za_normalization_conflict = 0; objnum = parent; type = DT_DIR; } else if (offset == 2 && zfs_show_ctldir(zp)) { (void) strcpy(zap.za_name, ZFS_CTLDIR_NAME); zap.za_normalization_conflict = 0; objnum = ZFSCTL_INO_ROOT; type = DT_DIR; } else { /* * Grab next entry. */ if (error = zap_cursor_retrieve(&zc, &zap)) { if ((*eofp = (error == ENOENT)) != 0) break; else goto update; } if (zap.za_integer_length != 8 || zap.za_num_integers != 1) { cmn_err(CE_WARN, "zap_readdir: bad directory " "entry, obj = %lld, offset = %lld\n", (u_longlong_t)zp->z_id, (u_longlong_t)offset); error = SET_ERROR(ENXIO); goto update; } objnum = ZFS_DIRENT_OBJ(zap.za_first_integer); /* * MacOS X can extract the object type here such as: * uint8_t type = ZFS_DIRENT_TYPE(zap.za_first_integer); */ type = ZFS_DIRENT_TYPE(zap.za_first_integer); if (check_sysattrs && !zap.za_normalization_conflict) { #ifdef TODO zap.za_normalization_conflict = xattr_sysattr_casechk(zap.za_name); #else panic("%s:%u: TODO", __func__, __LINE__); #endif } } if (flags & V_RDDIR_ACCFILTER) { /* * If we have no access at all, don't include * this entry in the returned information */ znode_t *ezp; if (zfs_zget(zp->z_zfsvfs, objnum, &ezp) != 0) goto skip_entry; if (!zfs_has_access(ezp, cr)) { vrele(ZTOV(ezp)); goto skip_entry; } vrele(ZTOV(ezp)); } if (flags & V_RDDIR_ENTFLAGS) reclen = EDIRENT_RECLEN(strlen(zap.za_name)); else reclen = DIRENT64_RECLEN(strlen(zap.za_name)); /* * Will this entry fit in the buffer? */ if (outcount + reclen > bufsize) { /* * Did we manage to fit anything in the buffer? */ if (!outcount) { error = SET_ERROR(EINVAL); goto update; } break; } if (flags & V_RDDIR_ENTFLAGS) { /* * Add extended flag entry: */ eodp->ed_ino = objnum; eodp->ed_reclen = reclen; /* NOTE: ed_off is the offset for the *next* entry. */ next = &eodp->ed_off; eodp->ed_eflags = zap.za_normalization_conflict ? ED_CASE_CONFLICT : 0; (void) strncpy(eodp->ed_name, zap.za_name, EDIRENT_NAMELEN(reclen)); eodp = (edirent_t *)((intptr_t)eodp + reclen); } else { /* * Add normal entry: */ odp->d_ino = objnum; odp->d_reclen = reclen; odp->d_namlen = strlen(zap.za_name); /* NOTE: d_off is the offset for the *next* entry. */ next = &odp->d_off; (void) strlcpy(odp->d_name, zap.za_name, odp->d_namlen + 1); odp->d_type = type; dirent_terminate(odp); odp = (dirent64_t *)((intptr_t)odp + reclen); } outcount += reclen; ASSERT(outcount <= bufsize); /* Prefetch znode */ if (prefetch) dmu_prefetch(os, objnum, 0, 0, 0, ZIO_PRIORITY_SYNC_READ); skip_entry: /* * Move to the next entry, fill in the previous offset. */ if (offset > 2 || (offset == 2 && !zfs_show_ctldir(zp))) { zap_cursor_advance(&zc); offset = zap_cursor_serialize(&zc); } else { offset += 1; } /* Fill the offset right after advancing the cursor. */ if (next != NULL) *next = offset; if (cooks != NULL) { *cooks++ = offset; ncooks--; KASSERT(ncooks >= 0, ("ncookies=%d", ncooks)); } } zp->z_zn_prefetch = B_FALSE; /* a lookup will re-enable pre-fetching */ /* Subtract unused cookies */ if (ncookies != NULL) *ncookies -= ncooks; if (uio->uio_segflg == UIO_SYSSPACE && uio->uio_iovcnt == 1) { iovp->iov_base += outcount; iovp->iov_len -= outcount; uio->uio_resid -= outcount; } else if (error = uiomove(outbuf, (long)outcount, UIO_READ, uio)) { /* * Reset the pointer. */ offset = uio->uio_loffset; } update: zap_cursor_fini(&zc); if (uio->uio_segflg != UIO_SYSSPACE || uio->uio_iovcnt != 1) kmem_free(outbuf, bufsize); if (error == ENOENT) error = 0; ZFS_ACCESSTIME_STAMP(zfsvfs, zp); uio->uio_loffset = offset; ZFS_EXIT(zfsvfs); if (error != 0 && cookies != NULL) { free(*cookies, M_TEMP); *cookies = NULL; *ncookies = 0; } return (error); } ulong_t zfs_fsync_sync_cnt = 4; static int zfs_fsync(vnode_t *vp, int syncflag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; (void) tsd_set(zfs_fsyncer_key, (void *)zfs_fsync_sync_cnt); if (zfsvfs->z_os->os_sync != ZFS_SYNC_DISABLED) { ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zil_commit(zfsvfs->z_log, zp->z_id); ZFS_EXIT(zfsvfs); } return (0); } /* * Get the requested file attributes and place them in the provided * vattr structure. * * IN: vp - vnode of file. * vap - va_mask identifies requested attributes. * If AT_XVATTR set, then optional attrs are requested * flags - ATTR_NOACLCHECK (CIFS server context) * cr - credentials of caller. * ct - caller context * * OUT: vap - attribute values. * * RETURN: 0 (always succeeds). */ /* ARGSUSED */ static int zfs_getattr(vnode_t *vp, vattr_t *vap, int flags, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error = 0; uint32_t blksize; u_longlong_t nblocks; uint64_t mtime[2], ctime[2], crtime[2], rdev; xvattr_t *xvap = (xvattr_t *)vap; /* vap may be an xvattr_t * */ xoptattr_t *xoap = NULL; boolean_t skipaclchk = (flags & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; sa_bulk_attr_t bulk[4]; int count = 0; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zfs_fuid_map_ids(zp, cr, &vap->va_uid, &vap->va_gid); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CRTIME(zfsvfs), NULL, &crtime, 16); if (vp->v_type == VBLK || vp->v_type == VCHR) SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_RDEV(zfsvfs), NULL, &rdev, 8); if ((error = sa_bulk_lookup(zp->z_sa_hdl, bulk, count)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * If ACL is trivial don't bother looking for ACE_READ_ATTRIBUTES. * Also, if we are the owner don't bother, since owner should * always be allowed to read basic attributes of file. */ if (!(zp->z_pflags & ZFS_ACL_TRIVIAL) && (vap->va_uid != crgetuid(cr))) { if (error = zfs_zaccess(zp, ACE_READ_ATTRIBUTES, 0, skipaclchk, cr)) { ZFS_EXIT(zfsvfs); return (error); } } /* * Return all attributes. It's cheaper to provide the answer * than to determine whether we were asked the question. */ vap->va_type = IFTOVT(zp->z_mode); vap->va_mode = zp->z_mode & ~S_IFMT; #ifdef illumos vap->va_fsid = zp->z_zfsvfs->z_vfs->vfs_dev; #else vn_fsid(vp, vap); #endif vap->va_nodeid = zp->z_id; vap->va_nlink = zp->z_links; if ((vp->v_flag & VROOT) && zfs_show_ctldir(zp) && zp->z_links < ZFS_LINK_MAX) vap->va_nlink++; vap->va_size = zp->z_size; #ifdef illumos vap->va_rdev = vp->v_rdev; #else if (vp->v_type == VBLK || vp->v_type == VCHR) vap->va_rdev = zfs_cmpldev(rdev); #endif vap->va_seq = zp->z_seq; vap->va_flags = 0; /* FreeBSD: Reset chflags(2) flags. */ vap->va_filerev = zp->z_seq; /* * Add in any requested optional attributes and the create time. * Also set the corresponding bits in the returned attribute bitmap. */ if ((xoap = xva_getxoptattr(xvap)) != NULL && zfsvfs->z_use_fuids) { if (XVA_ISSET_REQ(xvap, XAT_ARCHIVE)) { xoap->xoa_archive = ((zp->z_pflags & ZFS_ARCHIVE) != 0); XVA_SET_RTN(xvap, XAT_ARCHIVE); } if (XVA_ISSET_REQ(xvap, XAT_READONLY)) { xoap->xoa_readonly = ((zp->z_pflags & ZFS_READONLY) != 0); XVA_SET_RTN(xvap, XAT_READONLY); } if (XVA_ISSET_REQ(xvap, XAT_SYSTEM)) { xoap->xoa_system = ((zp->z_pflags & ZFS_SYSTEM) != 0); XVA_SET_RTN(xvap, XAT_SYSTEM); } if (XVA_ISSET_REQ(xvap, XAT_HIDDEN)) { xoap->xoa_hidden = ((zp->z_pflags & ZFS_HIDDEN) != 0); XVA_SET_RTN(xvap, XAT_HIDDEN); } if (XVA_ISSET_REQ(xvap, XAT_NOUNLINK)) { xoap->xoa_nounlink = ((zp->z_pflags & ZFS_NOUNLINK) != 0); XVA_SET_RTN(xvap, XAT_NOUNLINK); } if (XVA_ISSET_REQ(xvap, XAT_IMMUTABLE)) { xoap->xoa_immutable = ((zp->z_pflags & ZFS_IMMUTABLE) != 0); XVA_SET_RTN(xvap, XAT_IMMUTABLE); } if (XVA_ISSET_REQ(xvap, XAT_APPENDONLY)) { xoap->xoa_appendonly = ((zp->z_pflags & ZFS_APPENDONLY) != 0); XVA_SET_RTN(xvap, XAT_APPENDONLY); } if (XVA_ISSET_REQ(xvap, XAT_NODUMP)) { xoap->xoa_nodump = ((zp->z_pflags & ZFS_NODUMP) != 0); XVA_SET_RTN(xvap, XAT_NODUMP); } if (XVA_ISSET_REQ(xvap, XAT_OPAQUE)) { xoap->xoa_opaque = ((zp->z_pflags & ZFS_OPAQUE) != 0); XVA_SET_RTN(xvap, XAT_OPAQUE); } if (XVA_ISSET_REQ(xvap, XAT_AV_QUARANTINED)) { xoap->xoa_av_quarantined = ((zp->z_pflags & ZFS_AV_QUARANTINED) != 0); XVA_SET_RTN(xvap, XAT_AV_QUARANTINED); } if (XVA_ISSET_REQ(xvap, XAT_AV_MODIFIED)) { xoap->xoa_av_modified = ((zp->z_pflags & ZFS_AV_MODIFIED) != 0); XVA_SET_RTN(xvap, XAT_AV_MODIFIED); } if (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP) && vp->v_type == VREG) { zfs_sa_get_scanstamp(zp, xvap); } if (XVA_ISSET_REQ(xvap, XAT_REPARSE)) { xoap->xoa_reparse = ((zp->z_pflags & ZFS_REPARSE) != 0); XVA_SET_RTN(xvap, XAT_REPARSE); } if (XVA_ISSET_REQ(xvap, XAT_GEN)) { xoap->xoa_generation = zp->z_gen; XVA_SET_RTN(xvap, XAT_GEN); } if (XVA_ISSET_REQ(xvap, XAT_OFFLINE)) { xoap->xoa_offline = ((zp->z_pflags & ZFS_OFFLINE) != 0); XVA_SET_RTN(xvap, XAT_OFFLINE); } if (XVA_ISSET_REQ(xvap, XAT_SPARSE)) { xoap->xoa_sparse = ((zp->z_pflags & ZFS_SPARSE) != 0); XVA_SET_RTN(xvap, XAT_SPARSE); } } ZFS_TIME_DECODE(&vap->va_atime, zp->z_atime); ZFS_TIME_DECODE(&vap->va_mtime, mtime); ZFS_TIME_DECODE(&vap->va_ctime, ctime); ZFS_TIME_DECODE(&vap->va_birthtime, crtime); sa_object_size(zp->z_sa_hdl, &blksize, &nblocks); vap->va_blksize = blksize; vap->va_bytes = nblocks << 9; /* nblocks * 512 */ if (zp->z_blksz == 0) { /* * Block size hasn't been set; suggest maximal I/O transfers. */ vap->va_blksize = zfsvfs->z_max_blksz; } ZFS_EXIT(zfsvfs); return (0); } /* * Set the file attributes to the values contained in the * vattr structure. * * IN: vp - vnode of file to be modified. * vap - new attribute values. * If AT_XVATTR set, then optional attrs are being set * flags - ATTR_UTIME set if non-default time values provided. * - ATTR_NOACLCHECK (CIFS context only). * cr - credentials of caller. * ct - caller context * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - ctime updated, mtime updated if size changed. */ /* ARGSUSED */ static int zfs_setattr(vnode_t *vp, vattr_t *vap, int flags, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; vattr_t oldva; xvattr_t tmpxvattr; uint_t mask = vap->va_mask; uint_t saved_mask = 0; uint64_t saved_mode; int trim_mask = 0; uint64_t new_mode; uint64_t new_uid, new_gid; uint64_t xattr_obj; uint64_t mtime[2], ctime[2]; znode_t *attrzp; int need_policy = FALSE; int err, err2; zfs_fuid_info_t *fuidp = NULL; xvattr_t *xvap = (xvattr_t *)vap; /* vap may be an xvattr_t * */ xoptattr_t *xoap; zfs_acl_t *aclp; boolean_t skipaclchk = (flags & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; boolean_t fuid_dirtied = B_FALSE; sa_bulk_attr_t bulk[7], xattr_bulk[7]; int count = 0, xattr_count = 0; if (mask == 0) return (0); if (mask & AT_NOSET) return (SET_ERROR(EINVAL)); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); zilog = zfsvfs->z_log; /* * Make sure that if we have ephemeral uid/gid or xvattr specified * that file system is at proper version level */ if (zfsvfs->z_use_fuids == B_FALSE && (((mask & AT_UID) && IS_EPHEMERAL(vap->va_uid)) || ((mask & AT_GID) && IS_EPHEMERAL(vap->va_gid)) || (mask & AT_XVATTR))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } if (mask & AT_SIZE && vp->v_type == VDIR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EISDIR)); } if (mask & AT_SIZE && vp->v_type != VREG && vp->v_type != VFIFO) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } /* * If this is an xvattr_t, then get a pointer to the structure of * optional attributes. If this is NULL, then we have a vattr_t. */ xoap = xva_getxoptattr(xvap); xva_init(&tmpxvattr); /* * Immutable files can only alter immutable bit and atime */ if ((zp->z_pflags & ZFS_IMMUTABLE) && ((mask & (AT_SIZE|AT_UID|AT_GID|AT_MTIME|AT_MODE)) || ((mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_CREATETIME)))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } /* * Note: ZFS_READONLY is handled in zfs_zaccess_common. */ /* * Verify timestamps doesn't overflow 32 bits. * ZFS can handle large timestamps, but 32bit syscalls can't * handle times greater than 2039. This check should be removed * once large timestamps are fully supported. */ if (mask & (AT_ATIME | AT_MTIME)) { if (((mask & AT_ATIME) && TIMESPEC_OVERFLOW(&vap->va_atime)) || ((mask & AT_MTIME) && TIMESPEC_OVERFLOW(&vap->va_mtime))) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EOVERFLOW)); } } if (xoap && (mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_CREATETIME) && TIMESPEC_OVERFLOW(&vap->va_birthtime)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EOVERFLOW)); } attrzp = NULL; aclp = NULL; /* Can this be moved to before the top label? */ if (zfsvfs->z_vfs->vfs_flag & VFS_RDONLY) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EROFS)); } /* * First validate permissions */ if (mask & AT_SIZE) { /* * XXX - Note, we are not providing any open * mode flags here (like FNDELAY), so we may * block if there are locks present... this * should be addressed in openat(). */ /* XXX - would it be OK to generate a log record here? */ err = zfs_freesp(zp, vap->va_size, 0, 0, FALSE); if (err) { ZFS_EXIT(zfsvfs); return (err); } } if (mask & (AT_ATIME|AT_MTIME) || ((mask & AT_XVATTR) && (XVA_ISSET_REQ(xvap, XAT_HIDDEN) || XVA_ISSET_REQ(xvap, XAT_READONLY) || XVA_ISSET_REQ(xvap, XAT_ARCHIVE) || XVA_ISSET_REQ(xvap, XAT_OFFLINE) || XVA_ISSET_REQ(xvap, XAT_SPARSE) || XVA_ISSET_REQ(xvap, XAT_CREATETIME) || XVA_ISSET_REQ(xvap, XAT_SYSTEM)))) { need_policy = zfs_zaccess(zp, ACE_WRITE_ATTRIBUTES, 0, skipaclchk, cr); } if (mask & (AT_UID|AT_GID)) { int idmask = (mask & (AT_UID|AT_GID)); int take_owner; int take_group; /* * NOTE: even if a new mode is being set, * we may clear S_ISUID/S_ISGID bits. */ if (!(mask & AT_MODE)) vap->va_mode = zp->z_mode; /* * Take ownership or chgrp to group we are a member of */ take_owner = (mask & AT_UID) && (vap->va_uid == crgetuid(cr)); take_group = (mask & AT_GID) && zfs_groupmember(zfsvfs, vap->va_gid, cr); /* * If both AT_UID and AT_GID are set then take_owner and * take_group must both be set in order to allow taking * ownership. * * Otherwise, send the check through secpolicy_vnode_setattr() * */ if (((idmask == (AT_UID|AT_GID)) && take_owner && take_group) || ((idmask == AT_UID) && take_owner) || ((idmask == AT_GID) && take_group)) { if (zfs_zaccess(zp, ACE_WRITE_OWNER, 0, skipaclchk, cr) == 0) { /* * Remove setuid/setgid for non-privileged users */ secpolicy_setid_clear(vap, vp, cr); trim_mask = (mask & (AT_UID|AT_GID)); } else { need_policy = TRUE; } } else { need_policy = TRUE; } } oldva.va_mode = zp->z_mode; zfs_fuid_map_ids(zp, cr, &oldva.va_uid, &oldva.va_gid); if (mask & AT_XVATTR) { /* * Update xvattr mask to include only those attributes * that are actually changing. * * the bits will be restored prior to actually setting * the attributes so the caller thinks they were set. */ if (XVA_ISSET_REQ(xvap, XAT_APPENDONLY)) { if (xoap->xoa_appendonly != ((zp->z_pflags & ZFS_APPENDONLY) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_APPENDONLY); XVA_SET_REQ(&tmpxvattr, XAT_APPENDONLY); } } if (XVA_ISSET_REQ(xvap, XAT_NOUNLINK)) { if (xoap->xoa_nounlink != ((zp->z_pflags & ZFS_NOUNLINK) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_NOUNLINK); XVA_SET_REQ(&tmpxvattr, XAT_NOUNLINK); } } if (XVA_ISSET_REQ(xvap, XAT_IMMUTABLE)) { if (xoap->xoa_immutable != ((zp->z_pflags & ZFS_IMMUTABLE) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_IMMUTABLE); XVA_SET_REQ(&tmpxvattr, XAT_IMMUTABLE); } } if (XVA_ISSET_REQ(xvap, XAT_NODUMP)) { if (xoap->xoa_nodump != ((zp->z_pflags & ZFS_NODUMP) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_NODUMP); XVA_SET_REQ(&tmpxvattr, XAT_NODUMP); } } if (XVA_ISSET_REQ(xvap, XAT_AV_MODIFIED)) { if (xoap->xoa_av_modified != ((zp->z_pflags & ZFS_AV_MODIFIED) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_AV_MODIFIED); XVA_SET_REQ(&tmpxvattr, XAT_AV_MODIFIED); } } if (XVA_ISSET_REQ(xvap, XAT_AV_QUARANTINED)) { if ((vp->v_type != VREG && xoap->xoa_av_quarantined) || xoap->xoa_av_quarantined != ((zp->z_pflags & ZFS_AV_QUARANTINED) != 0)) { need_policy = TRUE; } else { XVA_CLR_REQ(xvap, XAT_AV_QUARANTINED); XVA_SET_REQ(&tmpxvattr, XAT_AV_QUARANTINED); } } if (XVA_ISSET_REQ(xvap, XAT_REPARSE)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (need_policy == FALSE && (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP) || XVA_ISSET_REQ(xvap, XAT_OPAQUE))) { need_policy = TRUE; } } if (mask & AT_MODE) { if (zfs_zaccess(zp, ACE_WRITE_ACL, 0, skipaclchk, cr) == 0) { err = secpolicy_setid_setsticky_clear(vp, vap, &oldva, cr); if (err) { ZFS_EXIT(zfsvfs); return (err); } trim_mask |= AT_MODE; } else { need_policy = TRUE; } } if (need_policy) { /* * If trim_mask is set then take ownership * has been granted or write_acl is present and user * has the ability to modify mode. In that case remove * UID|GID and or MODE from mask so that * secpolicy_vnode_setattr() doesn't revoke it. */ if (trim_mask) { saved_mask = vap->va_mask; vap->va_mask &= ~trim_mask; if (trim_mask & AT_MODE) { /* * Save the mode, as secpolicy_vnode_setattr() * will overwrite it with ova.va_mode. */ saved_mode = vap->va_mode; } } err = secpolicy_vnode_setattr(cr, vp, vap, &oldva, flags, (int (*)(void *, int, cred_t *))zfs_zaccess_unix, zp); if (err) { ZFS_EXIT(zfsvfs); return (err); } if (trim_mask) { vap->va_mask |= saved_mask; if (trim_mask & AT_MODE) { /* * Recover the mode after * secpolicy_vnode_setattr(). */ vap->va_mode = saved_mode; } } } /* * secpolicy_vnode_setattr, or take ownership may have * changed va_mask */ mask = vap->va_mask; if ((mask & (AT_UID | AT_GID))) { err = sa_lookup(zp->z_sa_hdl, SA_ZPL_XATTR(zfsvfs), &xattr_obj, sizeof (xattr_obj)); if (err == 0 && xattr_obj) { err = zfs_zget(zp->z_zfsvfs, xattr_obj, &attrzp); if (err == 0) { err = vn_lock(ZTOV(attrzp), LK_EXCLUSIVE); if (err != 0) vrele(ZTOV(attrzp)); } if (err) goto out2; } if (mask & AT_UID) { new_uid = zfs_fuid_create(zfsvfs, (uint64_t)vap->va_uid, cr, ZFS_OWNER, &fuidp); if (new_uid != zp->z_uid && zfs_fuid_overquota(zfsvfs, B_FALSE, new_uid)) { if (attrzp) vput(ZTOV(attrzp)); err = SET_ERROR(EDQUOT); goto out2; } } if (mask & AT_GID) { new_gid = zfs_fuid_create(zfsvfs, (uint64_t)vap->va_gid, cr, ZFS_GROUP, &fuidp); if (new_gid != zp->z_gid && zfs_fuid_overquota(zfsvfs, B_TRUE, new_gid)) { if (attrzp) vput(ZTOV(attrzp)); err = SET_ERROR(EDQUOT); goto out2; } } } tx = dmu_tx_create(zfsvfs->z_os); if (mask & AT_MODE) { uint64_t pmode = zp->z_mode; uint64_t acl_obj; new_mode = (pmode & S_IFMT) | (vap->va_mode & ~S_IFMT); if (zp->z_zfsvfs->z_acl_mode == ZFS_ACL_RESTRICTED && !(zp->z_pflags & ZFS_ACL_TRIVIAL)) { err = SET_ERROR(EPERM); goto out; } if (err = zfs_acl_chmod_setattr(zp, &aclp, new_mode)) goto out; if (!zp->z_is_sa && ((acl_obj = zfs_external_acl(zp)) != 0)) { /* * Are we upgrading ACL from old V0 format * to V1 format? */ if (zfsvfs->z_version >= ZPL_VERSION_FUID && zfs_znode_acl_version(zp) == ZFS_ACL_VERSION_INITIAL) { dmu_tx_hold_free(tx, acl_obj, 0, DMU_OBJECT_END); dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, aclp->z_acl_bytes); } else { dmu_tx_hold_write(tx, acl_obj, 0, aclp->z_acl_bytes); } } else if (!zp->z_is_sa && aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, aclp->z_acl_bytes); } dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); } else { if ((mask & AT_XVATTR) && XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP)) dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_TRUE); else dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); } if (attrzp) { dmu_tx_hold_sa(tx, attrzp->z_sa_hdl, B_FALSE); } fuid_dirtied = zfsvfs->z_fuid_dirty; if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); zfs_sa_upgrade_txholds(tx, zp); err = dmu_tx_assign(tx, TXG_WAIT); if (err) goto out; count = 0; /* * Set each attribute requested. * We group settings according to the locks they need to acquire. * * Note: you cannot set ctime directly, although it will be * updated as a side-effect of calling this function. */ if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_enter(&zp->z_acl_lock); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, sizeof (zp->z_pflags)); if (attrzp) { if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_enter(&attrzp->z_acl_lock); SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_FLAGS(zfsvfs), NULL, &attrzp->z_pflags, sizeof (attrzp->z_pflags)); } if (mask & (AT_UID|AT_GID)) { if (mask & AT_UID) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_UID(zfsvfs), NULL, &new_uid, sizeof (new_uid)); zp->z_uid = new_uid; if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_UID(zfsvfs), NULL, &new_uid, sizeof (new_uid)); attrzp->z_uid = new_uid; } } if (mask & AT_GID) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_GID(zfsvfs), NULL, &new_gid, sizeof (new_gid)); zp->z_gid = new_gid; if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_GID(zfsvfs), NULL, &new_gid, sizeof (new_gid)); attrzp->z_gid = new_gid; } } if (!(mask & AT_MODE)) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MODE(zfsvfs), NULL, &new_mode, sizeof (new_mode)); new_mode = zp->z_mode; } err = zfs_acl_chown_setattr(zp); ASSERT(err == 0); if (attrzp) { err = zfs_acl_chown_setattr(attrzp); ASSERT(err == 0); } } if (mask & AT_MODE) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MODE(zfsvfs), NULL, &new_mode, sizeof (new_mode)); zp->z_mode = new_mode; ASSERT3U((uintptr_t)aclp, !=, 0); err = zfs_aclset_common(zp, aclp, cr, tx); ASSERT0(err); if (zp->z_acl_cached) zfs_acl_free(zp->z_acl_cached); zp->z_acl_cached = aclp; aclp = NULL; } if (mask & AT_ATIME) { ZFS_TIME_ENCODE(&vap->va_atime, zp->z_atime); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_ATIME(zfsvfs), NULL, &zp->z_atime, sizeof (zp->z_atime)); } if (mask & AT_MTIME) { ZFS_TIME_ENCODE(&vap->va_mtime, mtime); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, mtime, sizeof (mtime)); } /* XXX - shouldn't this be done *before* the ATIME/MTIME checks? */ if (mask & AT_SIZE && !(mask & AT_MTIME)) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, mtime, sizeof (mtime)); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); } else if (mask != 0) { SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(zp, STATE_CHANGED, mtime, ctime, B_TRUE); if (attrzp) { SA_ADD_BULK_ATTR(xattr_bulk, xattr_count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, sizeof (ctime)); zfs_tstamp_update_setup(attrzp, STATE_CHANGED, mtime, ctime, B_TRUE); } } /* * Do this after setting timestamps to prevent timestamp * update from toggling bit */ if (xoap && (mask & AT_XVATTR)) { if (XVA_ISSET_REQ(xvap, XAT_CREATETIME)) xoap->xoa_createtime = vap->va_birthtime; /* * restore trimmed off masks * so that return masks can be set for caller. */ if (XVA_ISSET_REQ(&tmpxvattr, XAT_APPENDONLY)) { XVA_SET_REQ(xvap, XAT_APPENDONLY); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_NOUNLINK)) { XVA_SET_REQ(xvap, XAT_NOUNLINK); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_IMMUTABLE)) { XVA_SET_REQ(xvap, XAT_IMMUTABLE); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_NODUMP)) { XVA_SET_REQ(xvap, XAT_NODUMP); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_AV_MODIFIED)) { XVA_SET_REQ(xvap, XAT_AV_MODIFIED); } if (XVA_ISSET_REQ(&tmpxvattr, XAT_AV_QUARANTINED)) { XVA_SET_REQ(xvap, XAT_AV_QUARANTINED); } if (XVA_ISSET_REQ(xvap, XAT_AV_SCANSTAMP)) ASSERT(vp->v_type == VREG); zfs_xvattr_set(zp, xvap, tx); } if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); if (mask != 0) zfs_log_setattr(zilog, tx, TX_SETATTR, zp, vap, mask, fuidp); if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_exit(&zp->z_acl_lock); if (attrzp) { if (mask & (AT_UID|AT_GID|AT_MODE)) mutex_exit(&attrzp->z_acl_lock); } out: if (err == 0 && attrzp) { err2 = sa_bulk_update(attrzp->z_sa_hdl, xattr_bulk, xattr_count, tx); ASSERT(err2 == 0); } if (attrzp) vput(ZTOV(attrzp)); if (aclp) zfs_acl_free(aclp); if (fuidp) { zfs_fuid_info_free(fuidp); fuidp = NULL; } if (err) { dmu_tx_abort(tx); } else { err2 = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); dmu_tx_commit(tx); } out2: if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (err); } /* * We acquire all but fdvp locks using non-blocking acquisitions. If we * fail to acquire any lock in the path we will drop all held locks, * acquire the new lock in a blocking fashion, and then release it and * restart the rename. This acquire/release step ensures that we do not * spin on a lock waiting for release. On error release all vnode locks * and decrement references the way tmpfs_rename() would do. */ static int zfs_rename_relock(struct vnode *sdvp, struct vnode **svpp, struct vnode *tdvp, struct vnode **tvpp, const struct componentname *scnp, const struct componentname *tcnp) { zfsvfs_t *zfsvfs; struct vnode *nvp, *svp, *tvp; znode_t *sdzp, *tdzp, *szp, *tzp; const char *snm = scnp->cn_nameptr; const char *tnm = tcnp->cn_nameptr; int error; VOP_UNLOCK(tdvp, 0); if (*tvpp != NULL && *tvpp != tdvp) VOP_UNLOCK(*tvpp, 0); relock: error = vn_lock(sdvp, LK_EXCLUSIVE); if (error) goto out; sdzp = VTOZ(sdvp); error = vn_lock(tdvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); if (error != EBUSY) goto out; error = vn_lock(tdvp, LK_EXCLUSIVE); if (error) goto out; VOP_UNLOCK(tdvp, 0); goto relock; } tdzp = VTOZ(tdvp); /* * Before using sdzp and tdzp we must ensure that they are live. * As a porting legacy from illumos we have two things to worry * about. One is typical for FreeBSD and it is that the vnode is * not reclaimed (doomed). The other is that the znode is live. * The current code can invalidate the znode without acquiring the * corresponding vnode lock if the object represented by the znode * and vnode is no longer valid after a rollback or receive operation. * z_teardown_lock hidden behind ZFS_ENTER and ZFS_EXIT is the lock * that protects the znodes from the invalidation. */ zfsvfs = sdzp->z_zfsvfs; ASSERT3P(zfsvfs, ==, tdzp->z_zfsvfs); ZFS_ENTER(zfsvfs); /* * We can not use ZFS_VERIFY_ZP() here because it could directly return * bypassing the cleanup code in the case of an error. */ if (tdzp->z_sa_hdl == NULL || sdzp->z_sa_hdl == NULL) { ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); error = SET_ERROR(EIO); goto out; } /* * Re-resolve svp to be certain it still exists and fetch the * correct vnode. */ error = zfs_dirent_lookup(sdzp, snm, &szp, ZEXISTS); if (error != 0) { /* Source entry invalid or not there. */ ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); if ((scnp->cn_flags & ISDOTDOT) != 0 || (scnp->cn_namelen == 1 && scnp->cn_nameptr[0] == '.')) error = SET_ERROR(EINVAL); goto out; } svp = ZTOV(szp); /* * Re-resolve tvp, if it disappeared we just carry on. */ error = zfs_dirent_lookup(tdzp, tnm, &tzp, 0); if (error != 0) { ZFS_EXIT(zfsvfs); VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); vrele(svp); if ((tcnp->cn_flags & ISDOTDOT) != 0) error = SET_ERROR(EINVAL); goto out; } if (tzp != NULL) tvp = ZTOV(tzp); else tvp = NULL; /* * At present the vnode locks must be acquired before z_teardown_lock, * although it would be more logical to use the opposite order. */ ZFS_EXIT(zfsvfs); /* * Now try acquire locks on svp and tvp. */ nvp = svp; error = vn_lock(nvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); if (tvp != NULL) vrele(tvp); if (error != EBUSY) { vrele(nvp); goto out; } error = vn_lock(nvp, LK_EXCLUSIVE); if (error != 0) { vrele(nvp); goto out; } VOP_UNLOCK(nvp, 0); /* * Concurrent rename race. * XXX ? */ if (nvp == tdvp) { vrele(nvp); error = SET_ERROR(EINVAL); goto out; } vrele(*svpp); *svpp = nvp; goto relock; } vrele(*svpp); *svpp = nvp; if (*tvpp != NULL) vrele(*tvpp); *tvpp = NULL; if (tvp != NULL) { nvp = tvp; error = vn_lock(nvp, LK_EXCLUSIVE | LK_NOWAIT); if (error != 0) { VOP_UNLOCK(sdvp, 0); VOP_UNLOCK(tdvp, 0); VOP_UNLOCK(*svpp, 0); if (error != EBUSY) { vrele(nvp); goto out; } error = vn_lock(nvp, LK_EXCLUSIVE); if (error != 0) { vrele(nvp); goto out; } vput(nvp); goto relock; } *tvpp = nvp; } return (0); out: return (error); } /* * Note that we must use VRELE_ASYNC in this function as it walks * up the directory tree and vrele may need to acquire an exclusive * lock if a last reference to a vnode is dropped. */ static int zfs_rename_check(znode_t *szp, znode_t *sdzp, znode_t *tdzp) { zfsvfs_t *zfsvfs; znode_t *zp, *zp1; uint64_t parent; int error; zfsvfs = tdzp->z_zfsvfs; if (tdzp == szp) return (SET_ERROR(EINVAL)); if (tdzp == sdzp) return (0); if (tdzp->z_id == zfsvfs->z_root) return (0); zp = tdzp; for (;;) { ASSERT(!zp->z_unlinked); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (parent))) != 0) break; if (parent == szp->z_id) { error = SET_ERROR(EINVAL); break; } if (parent == zfsvfs->z_root) break; if (parent == sdzp->z_id) break; error = zfs_zget(zfsvfs, parent, &zp1); if (error != 0) break; if (zp != tdzp) VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(zfsvfs->z_os))); zp = zp1; } if (error == ENOTDIR) panic("checkpath: .. not a directory\n"); if (zp != tdzp) VN_RELE_ASYNC(ZTOV(zp), dsl_pool_vnrele_taskq(dmu_objset_pool(zfsvfs->z_os))); return (error); } /* * Move an entry from the provided source directory to the target * directory. Change the entry name as indicated. * * IN: sdvp - Source directory containing the "old entry". * snm - Old entry name. * tdvp - Target directory to contain the "new entry". * tnm - New entry name. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * sdvp,tdvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_rename(vnode_t *sdvp, vnode_t **svpp, struct componentname *scnp, vnode_t *tdvp, vnode_t **tvpp, struct componentname *tcnp, cred_t *cr) { zfsvfs_t *zfsvfs; znode_t *sdzp, *tdzp, *szp, *tzp; zilog_t *zilog = NULL; dmu_tx_t *tx; char *snm = scnp->cn_nameptr; char *tnm = tcnp->cn_nameptr; int error = 0; /* Reject renames across filesystems. */ if ((*svpp)->v_mount != tdvp->v_mount || ((*tvpp) != NULL && (*svpp)->v_mount != (*tvpp)->v_mount)) { error = SET_ERROR(EXDEV); goto out; } if (zfsctl_is_node(tdvp)) { error = SET_ERROR(EXDEV); goto out; } /* * Lock all four vnodes to ensure safety and semantics of renaming. */ error = zfs_rename_relock(sdvp, svpp, tdvp, tvpp, scnp, tcnp); if (error != 0) { /* no vnodes are locked in the case of error here */ return (error); } tdzp = VTOZ(tdvp); sdzp = VTOZ(sdvp); zfsvfs = tdzp->z_zfsvfs; zilog = zfsvfs->z_log; /* * After we re-enter ZFS_ENTER() we will have to revalidate all * znodes involved. */ ZFS_ENTER(zfsvfs); if (zfsvfs->z_utf8 && u8_validate(tnm, strlen(tnm), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { error = SET_ERROR(EILSEQ); goto unlockout; } /* If source and target are the same file, there is nothing to do. */ if ((*svpp) == (*tvpp)) { error = 0; goto unlockout; } if (((*svpp)->v_type == VDIR && (*svpp)->v_mountedhere != NULL) || ((*tvpp) != NULL && (*tvpp)->v_type == VDIR && (*tvpp)->v_mountedhere != NULL)) { error = SET_ERROR(EXDEV); goto unlockout; } /* * We can not use ZFS_VERIFY_ZP() here because it could directly return * bypassing the cleanup code in the case of an error. */ if (tdzp->z_sa_hdl == NULL || sdzp->z_sa_hdl == NULL) { error = SET_ERROR(EIO); goto unlockout; } szp = VTOZ(*svpp); tzp = *tvpp == NULL ? NULL : VTOZ(*tvpp); if (szp->z_sa_hdl == NULL || (tzp != NULL && tzp->z_sa_hdl == NULL)) { error = SET_ERROR(EIO); goto unlockout; } /* * This is to prevent the creation of links into attribute space * by renaming a linked file into/outof an attribute directory. * See the comment in zfs_link() for why this is considered bad. */ if ((tdzp->z_pflags & ZFS_XATTR) != (sdzp->z_pflags & ZFS_XATTR)) { error = SET_ERROR(EINVAL); goto unlockout; } /* * Must have write access at the source to remove the old entry * and write access at the target to create the new entry. * Note that if target and source are the same, this can be * done in a single check. */ if (error = zfs_zaccess_rename(sdzp, szp, tdzp, tzp, cr)) goto unlockout; if ((*svpp)->v_type == VDIR) { /* * Avoid ".", "..", and aliases of "." for obvious reasons. */ if ((scnp->cn_namelen == 1 && scnp->cn_nameptr[0] == '.') || sdzp == szp || (scnp->cn_flags | tcnp->cn_flags) & ISDOTDOT) { error = EINVAL; goto unlockout; } /* * Check to make sure rename is valid. * Can't do a move like this: /usr/a/b to /usr/a/b/c/d */ if (error = zfs_rename_check(szp, sdzp, tdzp)) goto unlockout; } /* * Does target exist? */ if (tzp) { /* * Source and target must be the same type. */ if ((*svpp)->v_type == VDIR) { if ((*tvpp)->v_type != VDIR) { error = SET_ERROR(ENOTDIR); goto unlockout; } else { cache_purge(tdvp); if (sdvp != tdvp) cache_purge(sdvp); } } else { if ((*tvpp)->v_type == VDIR) { error = SET_ERROR(EISDIR); goto unlockout; } } } vnevent_rename_src(*svpp, sdvp, scnp->cn_nameptr, ct); if (tzp) vnevent_rename_dest(*tvpp, tdvp, tnm, ct); /* * notify the target directory if it is not the same * as source directory. */ if (tdvp != sdvp) { vnevent_rename_dest_dir(tdvp, ct); } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, szp->z_sa_hdl, B_FALSE); dmu_tx_hold_sa(tx, sdzp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, sdzp->z_id, FALSE, snm); dmu_tx_hold_zap(tx, tdzp->z_id, TRUE, tnm); if (sdzp != tdzp) { dmu_tx_hold_sa(tx, tdzp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, tdzp); } if (tzp) { dmu_tx_hold_sa(tx, tzp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, tzp); } zfs_sa_upgrade_txholds(tx, szp); dmu_tx_hold_zap(tx, zfsvfs->z_unlinkedobj, FALSE, NULL); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); goto unlockout; } if (tzp) /* Attempt to remove the existing target */ error = zfs_link_destroy(tdzp, tnm, tzp, tx, 0, NULL); if (error == 0) { error = zfs_link_create(tdzp, tnm, szp, tx, ZRENAMING); if (error == 0) { szp->z_pflags |= ZFS_AV_MODIFIED; error = sa_update(szp->z_sa_hdl, SA_ZPL_FLAGS(zfsvfs), (void *)&szp->z_pflags, sizeof (uint64_t), tx); ASSERT0(error); error = zfs_link_destroy(sdzp, snm, szp, tx, ZRENAMING, NULL); if (error == 0) { zfs_log_rename(zilog, tx, TX_RENAME, sdzp, snm, tdzp, tnm, szp); /* * Update path information for the target vnode */ vn_renamepath(tdvp, *svpp, tnm, strlen(tnm)); } else { /* * At this point, we have successfully created * the target name, but have failed to remove * the source name. Since the create was done * with the ZRENAMING flag, there are * complications; for one, the link count is * wrong. The easiest way to deal with this * is to remove the newly created target, and * return the original error. This must * succeed; fortunately, it is very unlikely to * fail, since we just created it. */ VERIFY3U(zfs_link_destroy(tdzp, tnm, szp, tx, ZRENAMING, NULL), ==, 0); } } if (error == 0) { cache_purge(*svpp); if (*tvpp != NULL) cache_purge(*tvpp); cache_purge_negative(tdvp); } } dmu_tx_commit(tx); unlockout: /* all 4 vnodes are locked, ZFS_ENTER called */ ZFS_EXIT(zfsvfs); VOP_UNLOCK(*svpp, 0); VOP_UNLOCK(sdvp, 0); out: /* original two vnodes are locked */ if (error == 0 && zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); if (*tvpp != NULL) VOP_UNLOCK(*tvpp, 0); if (tdvp != *tvpp) VOP_UNLOCK(tdvp, 0); return (error); } /* * Insert the indicated symbolic reference entry into the directory. * * IN: dvp - Directory to contain new symbolic link. * link - Name for new symlink entry. * vap - Attributes of new entry. * cr - credentials of caller. * ct - caller context * flags - case flags * * RETURN: 0 on success, error code on failure. * * Timestamps: * dvp - ctime|mtime updated */ /*ARGSUSED*/ static int zfs_symlink(vnode_t *dvp, vnode_t **vpp, char *name, vattr_t *vap, char *link, cred_t *cr, kthread_t *td) { znode_t *zp, *dzp = VTOZ(dvp); dmu_tx_t *tx; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; uint64_t len = strlen(link); int error; zfs_acl_ids_t acl_ids; boolean_t fuid_dirtied; uint64_t txtype = TX_SYMLINK; int flags = 0; ASSERT(vap->va_type == VLNK); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } if (len > MAXPATHLEN) { ZFS_EXIT(zfsvfs); return (SET_ERROR(ENAMETOOLONG)); } if ((error = zfs_acl_ids_create(dzp, 0, vap, cr, NULL, &acl_ids)) != 0) { ZFS_EXIT(zfsvfs); return (error); } /* * Attempt to lock directory; fail if entry already exists. */ error = zfs_dirent_lookup(dzp, name, &zp, ZNEW); if (error) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (error); } if (zfs_acl_ids_overquota(zfsvfs, &acl_ids)) { zfs_acl_ids_free(&acl_ids); ZFS_EXIT(zfsvfs); return (SET_ERROR(EDQUOT)); } getnewvnode_reserve(1); tx = dmu_tx_create(zfsvfs->z_os); fuid_dirtied = zfsvfs->z_fuid_dirty; dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, MAX(1, len)); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); dmu_tx_hold_sa_create(tx, acl_ids.z_aclp->z_acl_bytes + ZFS_SA_BASE_ATTR_SIZE + len); dmu_tx_hold_sa(tx, dzp->z_sa_hdl, B_FALSE); if (!zfsvfs->z_use_sa && acl_ids.z_aclp->z_acl_bytes > ZFS_ACE_SPACE) { dmu_tx_hold_write(tx, DMU_NEW_OBJECT, 0, acl_ids.z_aclp->z_acl_bytes); } if (fuid_dirtied) zfs_fuid_txhold(zfsvfs, tx); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { zfs_acl_ids_free(&acl_ids); dmu_tx_abort(tx); getnewvnode_drop_reserve(); ZFS_EXIT(zfsvfs); return (error); } /* * Create a new object for the symlink. * for version 4 ZPL datsets the symlink will be an SA attribute */ zfs_mknode(dzp, vap, tx, cr, 0, &zp, &acl_ids); if (fuid_dirtied) zfs_fuid_sync(zfsvfs, tx); if (zp->z_is_sa) error = sa_update(zp->z_sa_hdl, SA_ZPL_SYMLINK(zfsvfs), link, len, tx); else zfs_sa_symlink(zp, link, len, tx); zp->z_size = len; (void) sa_update(zp->z_sa_hdl, SA_ZPL_SIZE(zfsvfs), &zp->z_size, sizeof (zp->z_size), tx); /* * Insert the new object into the directory. */ (void) zfs_link_create(dzp, name, zp, tx, ZNEW); zfs_log_symlink(zilog, tx, txtype, dzp, zp, name, link); *vpp = ZTOV(zp); zfs_acl_ids_free(&acl_ids); dmu_tx_commit(tx); getnewvnode_drop_reserve(); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /* * Return, in the buffer contained in the provided uio structure, * the symbolic path referred to by vp. * * IN: vp - vnode of symbolic link. * uio - structure to contain the link path. * cr - credentials of caller. * ct - caller context * * OUT: uio - structure containing the link path. * * RETURN: 0 on success, error code on failure. * * Timestamps: * vp - atime updated */ /* ARGSUSED */ static int zfs_readlink(vnode_t *vp, uio_t *uio, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if (zp->z_is_sa) error = sa_lookup_uio(zp->z_sa_hdl, SA_ZPL_SYMLINK(zfsvfs), uio); else error = zfs_sa_readlink(zp, uio); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); return (error); } /* * Insert a new entry into directory tdvp referencing svp. * * IN: tdvp - Directory to contain new entry. * svp - vnode of new entry. * name - name of new entry. * cr - credentials of caller. * ct - caller context * * RETURN: 0 on success, error code on failure. * * Timestamps: * tdvp - ctime|mtime updated * svp - ctime updated */ /* ARGSUSED */ static int zfs_link(vnode_t *tdvp, vnode_t *svp, char *name, cred_t *cr, caller_context_t *ct, int flags) { znode_t *dzp = VTOZ(tdvp); znode_t *tzp, *szp; zfsvfs_t *zfsvfs = dzp->z_zfsvfs; zilog_t *zilog; dmu_tx_t *tx; int error; uint64_t parent; uid_t owner; ASSERT(tdvp->v_type == VDIR); ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(dzp); zilog = zfsvfs->z_log; /* * POSIX dictates that we return EPERM here. * Better choices include ENOTSUP or EISDIR. */ if (svp->v_type == VDIR) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } szp = VTOZ(svp); ZFS_VERIFY_ZP(szp); if (szp->z_pflags & (ZFS_APPENDONLY | ZFS_IMMUTABLE | ZFS_READONLY)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } /* Prevent links to .zfs/shares files */ if ((error = sa_lookup(szp->z_sa_hdl, SA_ZPL_PARENT(zfsvfs), &parent, sizeof (uint64_t))) != 0) { ZFS_EXIT(zfsvfs); return (error); } if (parent == zfsvfs->z_shares_dir) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (zfsvfs->z_utf8 && u8_validate(name, strlen(name), NULL, U8_VALIDATE_ENTIRE, &error) < 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EILSEQ)); } /* * We do not support links between attributes and non-attributes * because of the potential security risk of creating links * into "normal" file space in order to circumvent restrictions * imposed in attribute space. */ if ((szp->z_pflags & ZFS_XATTR) != (dzp->z_pflags & ZFS_XATTR)) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EINVAL)); } owner = zfs_fuid_map_id(zfsvfs, szp->z_uid, cr, ZFS_OWNER); if (owner != crgetuid(cr) && secpolicy_basic_link(svp, cr) != 0) { ZFS_EXIT(zfsvfs); return (SET_ERROR(EPERM)); } if (error = zfs_zaccess(dzp, ACE_ADD_FILE, 0, B_FALSE, cr)) { ZFS_EXIT(zfsvfs); return (error); } /* * Attempt to lock directory; fail if entry already exists. */ error = zfs_dirent_lookup(dzp, name, &tzp, ZNEW); if (error) { ZFS_EXIT(zfsvfs); return (error); } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, szp->z_sa_hdl, B_FALSE); dmu_tx_hold_zap(tx, dzp->z_id, TRUE, name); zfs_sa_upgrade_txholds(tx, szp); zfs_sa_upgrade_txholds(tx, dzp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); ZFS_EXIT(zfsvfs); return (error); } error = zfs_link_create(dzp, name, szp, tx, 0); if (error == 0) { uint64_t txtype = TX_LINK; zfs_log_link(zilog, tx, txtype, dzp, szp, name); } dmu_tx_commit(tx); if (error == 0) { vnevent_link(svp, ct); } if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } /*ARGSUSED*/ void zfs_inactive(vnode_t *vp, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; rw_enter(&zfsvfs->z_teardown_inactive_lock, RW_READER); if (zp->z_sa_hdl == NULL) { /* * The fs has been unmounted, or we did a * suspend/resume and this file no longer exists. */ rw_exit(&zfsvfs->z_teardown_inactive_lock); vrecycle(vp); return; } if (zp->z_unlinked) { /* * Fast path to recycle a vnode of a removed file. */ rw_exit(&zfsvfs->z_teardown_inactive_lock); vrecycle(vp); return; } if (zp->z_atime_dirty && zp->z_unlinked == 0) { dmu_tx_t *tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); error = dmu_tx_assign(tx, TXG_WAIT); if (error) { dmu_tx_abort(tx); } else { (void) sa_update(zp->z_sa_hdl, SA_ZPL_ATIME(zfsvfs), (void *)&zp->z_atime, sizeof (zp->z_atime), tx); zp->z_atime_dirty = 0; dmu_tx_commit(tx); } } rw_exit(&zfsvfs->z_teardown_inactive_lock); } CTASSERT(sizeof(struct zfid_short) <= sizeof(struct fid)); CTASSERT(sizeof(struct zfid_long) <= sizeof(struct fid)); /*ARGSUSED*/ static int zfs_fid(vnode_t *vp, fid_t *fidp, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; uint32_t gen; uint64_t gen64; uint64_t object = zp->z_id; zfid_short_t *zfid; int size, i, error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); if ((error = sa_lookup(zp->z_sa_hdl, SA_ZPL_GEN(zfsvfs), &gen64, sizeof (uint64_t))) != 0) { ZFS_EXIT(zfsvfs); return (error); } gen = (uint32_t)gen64; size = (zfsvfs->z_parent != zfsvfs) ? LONG_FID_LEN : SHORT_FID_LEN; #ifdef illumos if (fidp->fid_len < size) { fidp->fid_len = size; ZFS_EXIT(zfsvfs); return (SET_ERROR(ENOSPC)); } #else fidp->fid_len = size; #endif zfid = (zfid_short_t *)fidp; zfid->zf_len = size; for (i = 0; i < sizeof (zfid->zf_object); i++) zfid->zf_object[i] = (uint8_t)(object >> (8 * i)); /* Must have a non-zero generation number to distinguish from .zfs */ if (gen == 0) gen = 1; for (i = 0; i < sizeof (zfid->zf_gen); i++) zfid->zf_gen[i] = (uint8_t)(gen >> (8 * i)); if (size == LONG_FID_LEN) { uint64_t objsetid = dmu_objset_id(zfsvfs->z_os); zfid_long_t *zlfid; zlfid = (zfid_long_t *)fidp; for (i = 0; i < sizeof (zlfid->zf_setid); i++) zlfid->zf_setid[i] = (uint8_t)(objsetid >> (8 * i)); /* XXX - this should be the generation number for the objset */ for (i = 0; i < sizeof (zlfid->zf_setgen); i++) zlfid->zf_setgen[i] = 0; } ZFS_EXIT(zfsvfs); return (0); } static int zfs_pathconf(vnode_t *vp, int cmd, ulong_t *valp, cred_t *cr, caller_context_t *ct) { znode_t *zp, *xzp; zfsvfs_t *zfsvfs; int error; switch (cmd) { case _PC_LINK_MAX: *valp = MIN(LONG_MAX, ZFS_LINK_MAX); return (0); case _PC_FILESIZEBITS: *valp = 64; return (0); #ifdef illumos case _PC_XATTR_EXISTS: zp = VTOZ(vp); zfsvfs = zp->z_zfsvfs; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); *valp = 0; error = zfs_dirent_lookup(zp, "", &xzp, ZXATTR | ZEXISTS | ZSHARED); if (error == 0) { if (!zfs_dirempty(xzp)) *valp = 1; vrele(ZTOV(xzp)); } else if (error == ENOENT) { /* * If there aren't extended attributes, it's the * same as having zero of them. */ error = 0; } ZFS_EXIT(zfsvfs); return (error); case _PC_SATTR_ENABLED: case _PC_SATTR_EXISTS: *valp = vfs_has_feature(vp->v_vfsp, VFSFT_SYSATTR_VIEWS) && (vp->v_type == VREG || vp->v_type == VDIR); return (0); case _PC_ACCESS_FILTERING: *valp = vfs_has_feature(vp->v_vfsp, VFSFT_ACCESS_FILTER) && vp->v_type == VDIR; return (0); case _PC_ACL_ENABLED: *valp = _ACL_ACE_ENABLED; return (0); #endif /* illumos */ case _PC_MIN_HOLE_SIZE: *valp = (int)SPA_MINBLOCKSIZE; return (0); #ifdef illumos case _PC_TIMESTAMP_RESOLUTION: /* nanosecond timestamp resolution */ *valp = 1L; return (0); #endif case _PC_ACL_EXTENDED: *valp = 0; return (0); case _PC_ACL_NFS4: *valp = 1; return (0); case _PC_ACL_PATH_MAX: *valp = ACL_MAX_ENTRIES; return (0); default: return (EOPNOTSUPP); } } /*ARGSUSED*/ static int zfs_getsecattr(vnode_t *vp, vsecattr_t *vsecp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); error = zfs_getacl(zp, vsecp, skipaclchk, cr); ZFS_EXIT(zfsvfs); return (error); } /*ARGSUSED*/ int zfs_setsecattr(vnode_t *vp, vsecattr_t *vsecp, int flag, cred_t *cr, caller_context_t *ct) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; int error; boolean_t skipaclchk = (flag & ATTR_NOACLCHECK) ? B_TRUE : B_FALSE; zilog_t *zilog = zfsvfs->z_log; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); error = zfs_setacl(zp, vsecp, skipaclchk, cr); if (zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zilog, 0); ZFS_EXIT(zfsvfs); return (error); } static int zfs_getpages(struct vnode *vp, vm_page_t *ma, int count, int *rbehind, int *rahead) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; objset_t *os = zp->z_zfsvfs->z_os; locked_range_t *lr; vm_object_t object; off_t start, end, obj_size; uint_t blksz; int pgsin_b, pgsin_a; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); start = IDX_TO_OFF(ma[0]->pindex); end = IDX_TO_OFF(ma[count - 1]->pindex + 1); /* * Lock a range covering all required and optional pages. * Note that we need to handle the case of the block size growing. */ for (;;) { blksz = zp->z_blksz; lr = rangelock_enter(&zp->z_rangelock, rounddown(start, blksz), roundup(end, blksz) - rounddown(start, blksz), RL_READER); if (blksz == zp->z_blksz) break; rangelock_exit(lr); } object = ma[0]->object; zfs_vmobject_wlock(object); obj_size = object->un_pager.vnp.vnp_size; zfs_vmobject_wunlock(object); if (IDX_TO_OFF(ma[count - 1]->pindex) >= obj_size) { rangelock_exit(lr); ZFS_EXIT(zfsvfs); return (zfs_vm_pagerret_bad); } pgsin_b = 0; if (rbehind != NULL) { pgsin_b = OFF_TO_IDX(start - rounddown(start, blksz)); pgsin_b = MIN(*rbehind, pgsin_b); } pgsin_a = 0; if (rahead != NULL) { pgsin_a = OFF_TO_IDX(roundup(end, blksz) - end); if (end + IDX_TO_OFF(pgsin_a) >= obj_size) pgsin_a = OFF_TO_IDX(round_page(obj_size) - end); pgsin_a = MIN(*rahead, pgsin_a); } /* * NB: we need to pass the exact byte size of the data that we expect * to read after accounting for the file size. This is required because * ZFS will panic if we request DMU to read beyond the end of the last * allocated block. */ error = dmu_read_pages(os, zp->z_id, ma, count, &pgsin_b, &pgsin_a, MIN(end, obj_size) - (end - PAGE_SIZE)); rangelock_exit(lr); ZFS_ACCESSTIME_STAMP(zfsvfs, zp); ZFS_EXIT(zfsvfs); if (error != 0) return (zfs_vm_pagerret_error); VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, count + pgsin_b + pgsin_a); if (rbehind != NULL) *rbehind = pgsin_b; if (rahead != NULL) *rahead = pgsin_a; return (zfs_vm_pagerret_ok); } static int zfs_freebsd_getpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int *a_rbehind; int *a_rahead; } */ *ap; { return (zfs_getpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead)); } static int zfs_putpages(struct vnode *vp, vm_page_t *ma, size_t len, int flags, int *rtvals) { znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; locked_range_t *lr; dmu_tx_t *tx; struct sf_buf *sf; vm_object_t object; vm_page_t m; caddr_t va; size_t tocopy; size_t lo_len; vm_ooffset_t lo_off; vm_ooffset_t off; uint_t blksz; int ncount; int pcount; int err; int i; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); object = vp->v_object; pcount = btoc(len); ncount = pcount; KASSERT(ma[0]->object == object, ("mismatching object")); KASSERT(len > 0 && (len & PAGE_MASK) == 0, ("unexpected length")); for (i = 0; i < pcount; i++) rtvals[i] = zfs_vm_pagerret_error; off = IDX_TO_OFF(ma[0]->pindex); blksz = zp->z_blksz; lo_off = rounddown(off, blksz); lo_len = roundup(len + (off - lo_off), blksz); lr = rangelock_enter(&zp->z_rangelock, lo_off, lo_len, RL_WRITER); zfs_vmobject_wlock(object); if (len + off > object->un_pager.vnp.vnp_size) { if (object->un_pager.vnp.vnp_size > off) { int pgoff; len = object->un_pager.vnp.vnp_size - off; ncount = btoc(len); if ((pgoff = (int)len & PAGE_MASK) != 0) { /* * If the object is locked and the following * conditions hold, then the page's dirty * field cannot be concurrently changed by a * pmap operation. */ m = ma[ncount - 1]; vm_page_assert_sbusied(m); KASSERT(!pmap_page_is_write_mapped(m), ("zfs_putpages: page %p is not read-only", m)); vm_page_clear_dirty(m, pgoff, PAGE_SIZE - pgoff); } } else { len = 0; ncount = 0; } if (ncount < pcount) { for (i = ncount; i < pcount; i++) { rtvals[i] = zfs_vm_pagerret_bad; } } } zfs_vmobject_wunlock(object); if (ncount == 0) goto out; if (zfs_owner_overquota(zfsvfs, zp, B_FALSE) || zfs_owner_overquota(zfsvfs, zp, B_TRUE)) { goto out; } tx = dmu_tx_create(zfsvfs->z_os); dmu_tx_hold_write(tx, zp->z_id, off, len); dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE); zfs_sa_upgrade_txholds(tx, zp); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); goto out; } if (zp->z_blksz < PAGE_SIZE) { for (i = 0; len > 0; off += tocopy, len -= tocopy, i++) { tocopy = len > PAGE_SIZE ? PAGE_SIZE : len; va = zfs_map_page(ma[i], &sf); dmu_write(zfsvfs->z_os, zp->z_id, off, tocopy, va, tx); zfs_unmap_page(sf); } } else { err = dmu_write_pages(zfsvfs->z_os, zp->z_id, off, len, ma, tx); } if (err == 0) { uint64_t mtime[2], ctime[2]; sa_bulk_attr_t bulk[3]; int count = 0; SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_MTIME(zfsvfs), NULL, &mtime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_CTIME(zfsvfs), NULL, &ctime, 16); SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_FLAGS(zfsvfs), NULL, &zp->z_pflags, 8); zfs_tstamp_update_setup(zp, CONTENT_MODIFIED, mtime, ctime, B_TRUE); err = sa_bulk_update(zp->z_sa_hdl, bulk, count, tx); ASSERT0(err); zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, off, len, 0); zfs_vmobject_wlock(object); for (i = 0; i < ncount; i++) { rtvals[i] = zfs_vm_pagerret_ok; vm_page_undirty(ma[i]); } zfs_vmobject_wunlock(object); VM_CNT_INC(v_vnodeout); VM_CNT_ADD(v_vnodepgsout, ncount); } dmu_tx_commit(tx); out: rangelock_exit(lr); if ((flags & (zfs_vm_pagerput_sync | zfs_vm_pagerput_inval)) != 0 || zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS) zil_commit(zfsvfs->z_log, zp->z_id); ZFS_EXIT(zfsvfs); return (rtvals[0]); } int zfs_freebsd_putpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; } */ *ap; { return (zfs_putpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_sync, ap->a_rtvals)); } static int zfs_freebsd_bmap(ap) struct vop_bmap_args /* { struct vnode *a_vp; daddr_t a_bn; struct bufobj **a_bop; daddr_t *a_bnp; int *a_runp; int *a_runb; } */ *ap; { if (ap->a_bop != NULL) *ap->a_bop = &ap->a_vp->v_bufobj; if (ap->a_bnp != NULL) *ap->a_bnp = ap->a_bn; if (ap->a_runp != NULL) *ap->a_runp = 0; if (ap->a_runb != NULL) *ap->a_runb = 0; return (0); } static int zfs_freebsd_open(ap) struct vop_open_args /* { struct vnode *a_vp; int a_mode; struct ucred *a_cred; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); int error; error = zfs_open(&vp, ap->a_mode, ap->a_cred, NULL); if (error == 0) vnode_create_vobject(vp, zp->z_size, ap->a_td); return (error); } static int zfs_freebsd_close(ap) struct vop_close_args /* { struct vnode *a_vp; int a_fflag; struct ucred *a_cred; struct thread *a_td; } */ *ap; { return (zfs_close(ap->a_vp, ap->a_fflag, 1, 0, ap->a_cred, NULL)); } static int zfs_freebsd_ioctl(ap) struct vop_ioctl_args /* { struct vnode *a_vp; u_long a_command; caddr_t a_data; int a_fflag; struct ucred *cred; struct thread *td; } */ *ap; { return (zfs_ioctl(ap->a_vp, ap->a_command, (intptr_t)ap->a_data, ap->a_fflag, ap->a_cred, NULL, NULL)); } static int ioflags(int ioflags) { int flags = 0; if (ioflags & IO_APPEND) flags |= FAPPEND; if (ioflags & IO_NDELAY) flags |= FNONBLOCK; if (ioflags & IO_SYNC) flags |= (FSYNC | FDSYNC | FRSYNC); return (flags); } static int zfs_freebsd_read(ap) struct vop_read_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { return (zfs_read(ap->a_vp, ap->a_uio, ioflags(ap->a_ioflag), ap->a_cred, NULL)); } static int zfs_freebsd_write(ap) struct vop_write_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { return (zfs_write(ap->a_vp, ap->a_uio, ioflags(ap->a_ioflag), ap->a_cred, NULL)); } static int zfs_freebsd_access(ap) struct vop_access_args /* { struct vnode *a_vp; accmode_t a_accmode; struct ucred *a_cred; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); accmode_t accmode; int error = 0; /* * ZFS itself only knowns about VREAD, VWRITE, VEXEC and VAPPEND, */ accmode = ap->a_accmode & (VREAD|VWRITE|VEXEC|VAPPEND); if (accmode != 0) error = zfs_access(ap->a_vp, accmode, 0, ap->a_cred, NULL); /* * VADMIN has to be handled by vaccess(). */ if (error == 0) { accmode = ap->a_accmode & ~(VREAD|VWRITE|VEXEC|VAPPEND); if (accmode != 0) { error = vaccess(vp->v_type, zp->z_mode, zp->z_uid, zp->z_gid, accmode, ap->a_cred, NULL); } } /* * For VEXEC, ensure that at least one execute bit is set for * non-directories. */ if (error == 0 && (ap->a_accmode & VEXEC) != 0 && vp->v_type != VDIR && (zp->z_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) == 0) { error = EACCES; } return (error); } static int zfs_freebsd_lookup(ap) struct vop_lookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; char nm[NAME_MAX + 1]; ASSERT(cnp->cn_namelen < sizeof(nm)); strlcpy(nm, cnp->cn_nameptr, MIN(cnp->cn_namelen + 1, sizeof(nm))); return (zfs_lookup(ap->a_dvp, nm, ap->a_vpp, cnp, cnp->cn_nameiop, cnp->cn_cred, cnp->cn_thread, 0)); } static int zfs_cache_lookup(ap) struct vop_lookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { zfsvfs_t *zfsvfs; zfsvfs = ap->a_dvp->v_mount->mnt_data; if (zfsvfs->z_use_namecache) return (vfs_cache_lookup(ap)); else return (zfs_freebsd_lookup(ap)); } static int zfs_freebsd_create(ap) struct vop_create_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; } */ *ap; { zfsvfs_t *zfsvfs; struct componentname *cnp = ap->a_cnp; vattr_t *vap = ap->a_vap; int error, mode; ASSERT(cnp->cn_flags & SAVENAME); vattr_init_mask(vap); mode = vap->va_mode & ALLPERMS; zfsvfs = ap->a_dvp->v_mount->mnt_data; error = zfs_create(ap->a_dvp, cnp->cn_nameptr, vap, !EXCL, mode, ap->a_vpp, cnp->cn_cred, cnp->cn_thread); if (zfsvfs->z_use_namecache && error == 0 && (cnp->cn_flags & MAKEENTRY) != 0) cache_enter(ap->a_dvp, *ap->a_vpp, cnp); return (error); } static int zfs_freebsd_remove(ap) struct vop_remove_args /* { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { ASSERT(ap->a_cnp->cn_flags & SAVENAME); return (zfs_remove(ap->a_dvp, ap->a_vp, ap->a_cnp->cn_nameptr, ap->a_cnp->cn_cred)); } static int zfs_freebsd_mkdir(ap) struct vop_mkdir_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; } */ *ap; { vattr_t *vap = ap->a_vap; ASSERT(ap->a_cnp->cn_flags & SAVENAME); vattr_init_mask(vap); return (zfs_mkdir(ap->a_dvp, ap->a_cnp->cn_nameptr, vap, ap->a_vpp, ap->a_cnp->cn_cred)); } static int zfs_freebsd_rmdir(ap) struct vop_rmdir_args /* { struct vnode *a_dvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; ASSERT(cnp->cn_flags & SAVENAME); return (zfs_rmdir(ap->a_dvp, ap->a_vp, cnp->cn_nameptr, cnp->cn_cred)); } static int zfs_freebsd_readdir(ap) struct vop_readdir_args /* { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; int *a_eofflag; int *a_ncookies; u_long **a_cookies; } */ *ap; { return (zfs_readdir(ap->a_vp, ap->a_uio, ap->a_cred, ap->a_eofflag, ap->a_ncookies, ap->a_cookies)); } static int zfs_freebsd_fsync(ap) struct vop_fsync_args /* { struct vnode *a_vp; int a_waitfor; struct thread *a_td; } */ *ap; { vop_stdfsync(ap); return (zfs_fsync(ap->a_vp, 0, ap->a_td->td_ucred, NULL)); } static int zfs_freebsd_getattr(ap) struct vop_getattr_args /* { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; } */ *ap; { vattr_t *vap = ap->a_vap; xvattr_t xvap; u_long fflags = 0; int error; xva_init(&xvap); xvap.xva_vattr = *vap; xvap.xva_vattr.va_mask |= AT_XVATTR; /* Convert chflags into ZFS-type flags. */ /* XXX: what about SF_SETTABLE?. */ XVA_SET_REQ(&xvap, XAT_IMMUTABLE); XVA_SET_REQ(&xvap, XAT_APPENDONLY); XVA_SET_REQ(&xvap, XAT_NOUNLINK); XVA_SET_REQ(&xvap, XAT_NODUMP); XVA_SET_REQ(&xvap, XAT_READONLY); XVA_SET_REQ(&xvap, XAT_ARCHIVE); XVA_SET_REQ(&xvap, XAT_SYSTEM); XVA_SET_REQ(&xvap, XAT_HIDDEN); XVA_SET_REQ(&xvap, XAT_REPARSE); XVA_SET_REQ(&xvap, XAT_OFFLINE); XVA_SET_REQ(&xvap, XAT_SPARSE); error = zfs_getattr(ap->a_vp, (vattr_t *)&xvap, 0, ap->a_cred, NULL); if (error != 0) return (error); /* Convert ZFS xattr into chflags. */ #define FLAG_CHECK(fflag, xflag, xfield) do { \ if (XVA_ISSET_RTN(&xvap, (xflag)) && (xfield) != 0) \ fflags |= (fflag); \ } while (0) FLAG_CHECK(SF_IMMUTABLE, XAT_IMMUTABLE, xvap.xva_xoptattrs.xoa_immutable); FLAG_CHECK(SF_APPEND, XAT_APPENDONLY, xvap.xva_xoptattrs.xoa_appendonly); FLAG_CHECK(SF_NOUNLINK, XAT_NOUNLINK, xvap.xva_xoptattrs.xoa_nounlink); FLAG_CHECK(UF_ARCHIVE, XAT_ARCHIVE, xvap.xva_xoptattrs.xoa_archive); FLAG_CHECK(UF_NODUMP, XAT_NODUMP, xvap.xva_xoptattrs.xoa_nodump); FLAG_CHECK(UF_READONLY, XAT_READONLY, xvap.xva_xoptattrs.xoa_readonly); FLAG_CHECK(UF_SYSTEM, XAT_SYSTEM, xvap.xva_xoptattrs.xoa_system); FLAG_CHECK(UF_HIDDEN, XAT_HIDDEN, xvap.xva_xoptattrs.xoa_hidden); FLAG_CHECK(UF_REPARSE, XAT_REPARSE, xvap.xva_xoptattrs.xoa_reparse); FLAG_CHECK(UF_OFFLINE, XAT_OFFLINE, xvap.xva_xoptattrs.xoa_offline); FLAG_CHECK(UF_SPARSE, XAT_SPARSE, xvap.xva_xoptattrs.xoa_sparse); #undef FLAG_CHECK *vap = xvap.xva_vattr; vap->va_flags = fflags; return (0); } static int zfs_freebsd_setattr(ap) struct vop_setattr_args /* { struct vnode *a_vp; struct vattr *a_vap; struct ucred *a_cred; } */ *ap; { vnode_t *vp = ap->a_vp; vattr_t *vap = ap->a_vap; cred_t *cred = ap->a_cred; xvattr_t xvap; u_long fflags; uint64_t zflags; vattr_init_mask(vap); vap->va_mask &= ~AT_NOSET; xva_init(&xvap); xvap.xva_vattr = *vap; zflags = VTOZ(vp)->z_pflags; if (vap->va_flags != VNOVAL) { zfsvfs_t *zfsvfs = VTOZ(vp)->z_zfsvfs; int error; if (zfsvfs->z_use_fuids == B_FALSE) return (EOPNOTSUPP); fflags = vap->va_flags; /* * XXX KDM * We need to figure out whether it makes sense to allow * UF_REPARSE through, since we don't really have other * facilities to handle reparse points and zfs_setattr() * doesn't currently allow setting that attribute anyway. */ if ((fflags & ~(SF_IMMUTABLE|SF_APPEND|SF_NOUNLINK|UF_ARCHIVE| UF_NODUMP|UF_SYSTEM|UF_HIDDEN|UF_READONLY|UF_REPARSE| UF_OFFLINE|UF_SPARSE)) != 0) return (EOPNOTSUPP); /* * Unprivileged processes are not permitted to unset system * flags, or modify flags if any system flags are set. * Privileged non-jail processes may not modify system flags * if securelevel > 0 and any existing system flags are set. * Privileged jail processes behave like privileged non-jail * processes if the PR_ALLOW_CHFLAGS permission bit is set; * otherwise, they behave like unprivileged processes. */ if (secpolicy_fs_owner(vp->v_mount, cred) == 0 || priv_check_cred(cred, PRIV_VFS_SYSFLAGS) == 0) { if (zflags & (ZFS_IMMUTABLE | ZFS_APPENDONLY | ZFS_NOUNLINK)) { error = securelevel_gt(cred, 0); if (error != 0) return (error); } } else { /* * Callers may only modify the file flags on objects they * have VADMIN rights for. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, curthread)) != 0) return (error); if (zflags & (ZFS_IMMUTABLE | ZFS_APPENDONLY | ZFS_NOUNLINK)) { return (EPERM); } if (fflags & (SF_IMMUTABLE | SF_APPEND | SF_NOUNLINK)) { return (EPERM); } } #define FLAG_CHANGE(fflag, zflag, xflag, xfield) do { \ if (((fflags & (fflag)) && !(zflags & (zflag))) || \ ((zflags & (zflag)) && !(fflags & (fflag)))) { \ XVA_SET_REQ(&xvap, (xflag)); \ (xfield) = ((fflags & (fflag)) != 0); \ } \ } while (0) /* Convert chflags into ZFS-type flags. */ /* XXX: what about SF_SETTABLE?. */ FLAG_CHANGE(SF_IMMUTABLE, ZFS_IMMUTABLE, XAT_IMMUTABLE, xvap.xva_xoptattrs.xoa_immutable); FLAG_CHANGE(SF_APPEND, ZFS_APPENDONLY, XAT_APPENDONLY, xvap.xva_xoptattrs.xoa_appendonly); FLAG_CHANGE(SF_NOUNLINK, ZFS_NOUNLINK, XAT_NOUNLINK, xvap.xva_xoptattrs.xoa_nounlink); FLAG_CHANGE(UF_ARCHIVE, ZFS_ARCHIVE, XAT_ARCHIVE, xvap.xva_xoptattrs.xoa_archive); FLAG_CHANGE(UF_NODUMP, ZFS_NODUMP, XAT_NODUMP, xvap.xva_xoptattrs.xoa_nodump); FLAG_CHANGE(UF_READONLY, ZFS_READONLY, XAT_READONLY, xvap.xva_xoptattrs.xoa_readonly); FLAG_CHANGE(UF_SYSTEM, ZFS_SYSTEM, XAT_SYSTEM, xvap.xva_xoptattrs.xoa_system); FLAG_CHANGE(UF_HIDDEN, ZFS_HIDDEN, XAT_HIDDEN, xvap.xva_xoptattrs.xoa_hidden); FLAG_CHANGE(UF_REPARSE, ZFS_REPARSE, XAT_REPARSE, xvap.xva_xoptattrs.xoa_reparse); FLAG_CHANGE(UF_OFFLINE, ZFS_OFFLINE, XAT_OFFLINE, xvap.xva_xoptattrs.xoa_offline); FLAG_CHANGE(UF_SPARSE, ZFS_SPARSE, XAT_SPARSE, xvap.xva_xoptattrs.xoa_sparse); #undef FLAG_CHANGE } if (vap->va_birthtime.tv_sec != VNOVAL) { xvap.xva_vattr.va_mask |= AT_XVATTR; XVA_SET_REQ(&xvap, XAT_CREATETIME); } return (zfs_setattr(vp, (vattr_t *)&xvap, 0, cred, NULL)); } static int zfs_freebsd_rename(ap) struct vop_rename_args /* { struct vnode *a_fdvp; struct vnode *a_fvp; struct componentname *a_fcnp; struct vnode *a_tdvp; struct vnode *a_tvp; struct componentname *a_tcnp; } */ *ap; { vnode_t *fdvp = ap->a_fdvp; vnode_t *fvp = ap->a_fvp; vnode_t *tdvp = ap->a_tdvp; vnode_t *tvp = ap->a_tvp; int error; ASSERT(ap->a_fcnp->cn_flags & (SAVENAME|SAVESTART)); ASSERT(ap->a_tcnp->cn_flags & (SAVENAME|SAVESTART)); error = zfs_rename(fdvp, &fvp, ap->a_fcnp, tdvp, &tvp, ap->a_tcnp, ap->a_fcnp->cn_cred); vrele(fdvp); vrele(fvp); vrele(tdvp); if (tvp != NULL) vrele(tvp); return (error); } static int zfs_freebsd_symlink(ap) struct vop_symlink_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; struct vattr *a_vap; char *a_target; } */ *ap; { struct componentname *cnp = ap->a_cnp; vattr_t *vap = ap->a_vap; ASSERT(cnp->cn_flags & SAVENAME); vap->va_type = VLNK; /* FreeBSD: Syscall only sets va_mode. */ vattr_init_mask(vap); return (zfs_symlink(ap->a_dvp, ap->a_vpp, cnp->cn_nameptr, vap, __DECONST(char *, ap->a_target), cnp->cn_cred, cnp->cn_thread)); } static int zfs_freebsd_readlink(ap) struct vop_readlink_args /* { struct vnode *a_vp; struct uio *a_uio; struct ucred *a_cred; } */ *ap; { return (zfs_readlink(ap->a_vp, ap->a_uio, ap->a_cred, NULL)); } static int zfs_freebsd_link(ap) struct vop_link_args /* { struct vnode *a_tdvp; struct vnode *a_vp; struct componentname *a_cnp; } */ *ap; { struct componentname *cnp = ap->a_cnp; vnode_t *vp = ap->a_vp; vnode_t *tdvp = ap->a_tdvp; if (tdvp->v_mount != vp->v_mount) return (EXDEV); ASSERT(cnp->cn_flags & SAVENAME); return (zfs_link(tdvp, vp, cnp->cn_nameptr, cnp->cn_cred, NULL, 0)); } static int zfs_freebsd_inactive(ap) struct vop_inactive_args /* { struct vnode *a_vp; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; zfs_inactive(vp, ap->a_td->td_ucred, NULL); return (0); } static int zfs_freebsd_reclaim(ap) struct vop_reclaim_args /* { struct vnode *a_vp; struct thread *a_td; } */ *ap; { vnode_t *vp = ap->a_vp; znode_t *zp = VTOZ(vp); zfsvfs_t *zfsvfs = zp->z_zfsvfs; ASSERT(zp != NULL); /* * z_teardown_inactive_lock protects from a race with * zfs_znode_dmu_fini in zfsvfs_teardown during * force unmount. */ rw_enter(&zfsvfs->z_teardown_inactive_lock, RW_READER); if (zp->z_sa_hdl == NULL) zfs_znode_free(zp); else zfs_zinactive(zp); rw_exit(&zfsvfs->z_teardown_inactive_lock); vp->v_data = NULL; return (0); } static int zfs_freebsd_fid(ap) struct vop_fid_args /* { struct vnode *a_vp; struct fid *a_fid; } */ *ap; { return (zfs_fid(ap->a_vp, (void *)ap->a_fid, NULL)); } static int zfs_freebsd_pathconf(ap) struct vop_pathconf_args /* { struct vnode *a_vp; int a_name; register_t *a_retval; } */ *ap; { ulong_t val; int error; error = zfs_pathconf(ap->a_vp, ap->a_name, &val, curthread->td_ucred, NULL); if (error == 0) { *ap->a_retval = val; return (error); } if (error != EOPNOTSUPP) return (error); switch (ap->a_name) { case _PC_NAME_MAX: *ap->a_retval = NAME_MAX; return (0); case _PC_PIPE_BUF: if (ap->a_vp->v_type == VDIR || ap->a_vp->v_type == VFIFO) { *ap->a_retval = PIPE_BUF; return (0); } return (EINVAL); default: return (vop_stdpathconf(ap)); } } /* * FreeBSD's extended attributes namespace defines file name prefix for ZFS' * extended attribute name: * * NAMESPACE PREFIX * system freebsd:system: * user (none, can be used to access ZFS fsattr(5) attributes * created on Solaris) */ static int zfs_create_attrname(int attrnamespace, const char *name, char *attrname, size_t size) { const char *namespace, *prefix, *suffix; /* We don't allow '/' character in attribute name. */ if (strchr(name, '/') != NULL) return (EINVAL); /* We don't allow attribute names that start with "freebsd:" string. */ if (strncmp(name, "freebsd:", 8) == 0) return (EINVAL); bzero(attrname, size); switch (attrnamespace) { case EXTATTR_NAMESPACE_USER: #if 0 prefix = "freebsd:"; namespace = EXTATTR_NAMESPACE_USER_STRING; suffix = ":"; #else /* * This is the default namespace by which we can access all * attributes created on Solaris. */ prefix = namespace = suffix = ""; #endif break; case EXTATTR_NAMESPACE_SYSTEM: prefix = "freebsd:"; namespace = EXTATTR_NAMESPACE_SYSTEM_STRING; suffix = ":"; break; case EXTATTR_NAMESPACE_EMPTY: default: return (EINVAL); } if (snprintf(attrname, size, "%s%s%s%s", prefix, namespace, suffix, name) >= size) { return (ENAMETOOLONG); } return (0); } /* * Vnode operating to retrieve a named extended attribute. */ static int zfs_getextattr(struct vop_getextattr_args *ap) /* vop_getextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } flags = FREAD; NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, attrname, xvp, td); error = vn_open_cred(&nd, &flags, VN_OPEN_INVFS, 0, ap->a_cred, NULL); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); if (error == ENOENT) error = ENOATTR; return (error); } if (ap->a_size != NULL) { error = VOP_GETATTR(vp, &va, ap->a_cred); if (error == 0) *ap->a_size = (size_t)va.va_size; } else if (ap->a_uio != NULL) error = VOP_READ(vp, ap->a_uio, IO_UNIT, ap->a_cred); VOP_UNLOCK(vp, 0); vn_close(vp, flags, ap->a_cred, td); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to remove a named attribute. */ int zfs_deleteextattr(struct vop_deleteextattr_args *ap) /* vop_deleteextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } NDINIT_ATVP(&nd, DELETE, NOFOLLOW | LOCKPARENT | LOCKLEAF, UIO_SYSSPACE, attrname, xvp, td); error = namei(&nd); vp = nd.ni_vp; if (error != 0) { ZFS_EXIT(zfsvfs); NDFREE(&nd, NDF_ONLY_PNBUF); if (error == ENOENT) error = ENOATTR; return (error); } error = VOP_REMOVE(nd.ni_dvp, vp, &nd.ni_cnd); NDFREE(&nd, NDF_ONLY_PNBUF); vput(nd.ni_dvp); if (vp == nd.ni_dvp) vrele(vp); else vput(vp); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to set a named attribute. */ static int zfs_setextattr(struct vop_setextattr_args *ap) /* vop_setextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrname[255]; struct vattr va; vnode_t *xvp = NULL, *vp; int error, flags; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, ap->a_name, attrname, sizeof(attrname)); if (error != 0) return (error); ZFS_ENTER(zfsvfs); error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR | CREATE_XATTR_DIR); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } flags = FFLAGS(O_WRONLY | O_CREAT); NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, attrname, xvp, td); error = vn_open_cred(&nd, &flags, 0600, VN_OPEN_INVFS, ap->a_cred, NULL); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } VATTR_NULL(&va); va.va_size = 0; error = VOP_SETATTR(vp, &va, ap->a_cred); if (error == 0) VOP_WRITE(vp, ap->a_uio, IO_UNIT, ap->a_cred); VOP_UNLOCK(vp, 0); vn_close(vp, flags, ap->a_cred, td); ZFS_EXIT(zfsvfs); return (error); } /* * Vnode operation to retrieve extended attributes on a vnode. */ static int zfs_listextattr(struct vop_listextattr_args *ap) /* vop_listextattr { IN struct vnode *a_vp; IN int a_attrnamespace; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { zfsvfs_t *zfsvfs = VTOZ(ap->a_vp)->z_zfsvfs; struct thread *td = ap->a_td; struct nameidata nd; char attrprefix[16]; u_char dirbuf[sizeof(struct dirent)]; struct dirent *dp; struct iovec aiov; struct uio auio, *uio = ap->a_uio; size_t *sizep = ap->a_size; size_t plen; vnode_t *xvp = NULL, *vp; int done, error, eof, pos; error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error != 0) return (error); error = zfs_create_attrname(ap->a_attrnamespace, "", attrprefix, sizeof(attrprefix)); if (error != 0) return (error); plen = strlen(attrprefix); ZFS_ENTER(zfsvfs); if (sizep != NULL) *sizep = 0; error = zfs_lookup(ap->a_vp, NULL, &xvp, NULL, 0, ap->a_cred, td, LOOKUP_XATTR); if (error != 0) { ZFS_EXIT(zfsvfs); /* * ENOATTR means that the EA directory does not yet exist, * i.e. there are no extended attributes there. */ if (error == ENOATTR) error = 0; return (error); } NDINIT_ATVP(&nd, LOOKUP, NOFOLLOW | LOCKLEAF | LOCKSHARED, UIO_SYSSPACE, ".", xvp, td); error = namei(&nd); vp = nd.ni_vp; NDFREE(&nd, NDF_ONLY_PNBUF); if (error != 0) { ZFS_EXIT(zfsvfs); return (error); } auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_rw = UIO_READ; auio.uio_offset = 0; do { u_char nlen; aiov.iov_base = (void *)dirbuf; aiov.iov_len = sizeof(dirbuf); auio.uio_resid = sizeof(dirbuf); error = VOP_READDIR(vp, &auio, ap->a_cred, &eof, NULL, NULL); done = sizeof(dirbuf) - auio.uio_resid; if (error != 0) break; for (pos = 0; pos < done;) { dp = (struct dirent *)(dirbuf + pos); pos += dp->d_reclen; /* * XXX: Temporarily we also accept DT_UNKNOWN, as this * is what we get when attribute was created on Solaris. */ if (dp->d_type != DT_REG && dp->d_type != DT_UNKNOWN) continue; if (plen == 0 && strncmp(dp->d_name, "freebsd:", 8) == 0) continue; else if (strncmp(dp->d_name, attrprefix, plen) != 0) continue; nlen = dp->d_namlen - plen; if (sizep != NULL) *sizep += 1 + nlen; else if (uio != NULL) { /* * Format of extattr name entry is one byte for * length and the rest for name. */ error = uiomove(&nlen, 1, uio->uio_rw, uio); if (error == 0) { error = uiomove(dp->d_name + plen, nlen, uio->uio_rw, uio); } if (error != 0) break; } } } while (!eof && error == 0); vput(vp); ZFS_EXIT(zfsvfs); return (error); } int zfs_freebsd_getacl(ap) struct vop_getacl_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { int error; vsecattr_t vsecattr; if (ap->a_type != ACL_TYPE_NFS4) return (EINVAL); vsecattr.vsa_mask = VSA_ACE | VSA_ACECNT; if (error = zfs_getsecattr(ap->a_vp, &vsecattr, 0, ap->a_cred, NULL)) return (error); error = acl_from_aces(ap->a_aclp, vsecattr.vsa_aclentp, vsecattr.vsa_aclcnt); if (vsecattr.vsa_aclentp != NULL) kmem_free(vsecattr.vsa_aclentp, vsecattr.vsa_aclentsz); return (error); } int zfs_freebsd_setacl(ap) struct vop_setacl_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { int error; vsecattr_t vsecattr; int aclbsize; /* size of acl list in bytes */ aclent_t *aaclp; if (ap->a_type != ACL_TYPE_NFS4) return (EINVAL); if (ap->a_aclp == NULL) return (EINVAL); if (ap->a_aclp->acl_cnt < 1 || ap->a_aclp->acl_cnt > MAX_ACL_ENTRIES) return (EINVAL); /* * With NFSv4 ACLs, chmod(2) may need to add additional entries, * splitting every entry into two and appending "canonical six" * entries at the end. Don't allow for setting an ACL that would * cause chmod(2) to run out of ACL entries. */ if (ap->a_aclp->acl_cnt * 2 + 6 > ACL_MAX_ENTRIES) return (ENOSPC); error = acl_nfs4_check(ap->a_aclp, ap->a_vp->v_type == VDIR); if (error != 0) return (error); vsecattr.vsa_mask = VSA_ACE; aclbsize = ap->a_aclp->acl_cnt * sizeof(ace_t); vsecattr.vsa_aclentp = kmem_alloc(aclbsize, KM_SLEEP); aaclp = vsecattr.vsa_aclentp; vsecattr.vsa_aclentsz = aclbsize; aces_from_acl(vsecattr.vsa_aclentp, &vsecattr.vsa_aclcnt, ap->a_aclp); error = zfs_setsecattr(ap->a_vp, &vsecattr, 0, ap->a_cred, NULL); kmem_free(aaclp, aclbsize); return (error); } int zfs_freebsd_aclcheck(ap) struct vop_aclcheck_args /* { struct vnode *vp; acl_type_t type; struct acl *aclp; struct ucred *cred; struct thread *td; } */ *ap; { return (EOPNOTSUPP); } static int zfs_vptocnp(struct vop_vptocnp_args *ap) { vnode_t *covered_vp; vnode_t *vp = ap->a_vp;; zfsvfs_t *zfsvfs = vp->v_vfsp->vfs_data; znode_t *zp = VTOZ(vp); enum vgetstate vs; int ltype; int error; ZFS_ENTER(zfsvfs); ZFS_VERIFY_ZP(zp); /* * If we are a snapshot mounted under .zfs, run the operation * on the covered vnode. */ if (zp->z_id != zfsvfs->z_root || zfsvfs->z_parent == zfsvfs) { char name[MAXNAMLEN + 1]; znode_t *dzp; size_t len; error = zfs_znode_parent_and_name(zp, &dzp, name); if (error == 0) { len = strlen(name); if (*ap->a_buflen < len) error = SET_ERROR(ENOMEM); } if (error == 0) { *ap->a_buflen -= len; bcopy(name, ap->a_buf + *ap->a_buflen, len); *ap->a_vpp = ZTOV(dzp); } ZFS_EXIT(zfsvfs); return (error); } ZFS_EXIT(zfsvfs); covered_vp = vp->v_mount->mnt_vnodecovered; vs = vget_prep(covered_vp); ltype = VOP_ISLOCKED(vp); VOP_UNLOCK(vp, 0); error = vget_finish(covered_vp, LK_SHARED, vs); if (error == 0) { error = VOP_VPTOCNP(covered_vp, ap->a_vpp, ap->a_cred, ap->a_buf, ap->a_buflen); vput(covered_vp); } vn_lock(vp, ltype | LK_RETRY); if (VN_IS_DOOMED(vp)) error = SET_ERROR(ENOENT); return (error); } #ifdef DIAGNOSTIC static int zfs_lock(ap) struct vop_lock1_args /* { struct vnode *a_vp; int a_flags; char *file; int line; } */ *ap; { vnode_t *vp; znode_t *zp; int err; err = vop_lock(ap); if (err == 0 && (ap->a_flags & LK_NOWAIT) == 0) { vp = ap->a_vp; zp = vp->v_data; if (vp->v_mount != NULL && !VN_IS_DOOMED(vp) && zp != NULL && (zp->z_pflags & ZFS_XATTR) == 0) VERIFY(!RRM_LOCK_HELD(&zp->z_zfsvfs->z_teardown_lock)); } return (err); } #endif struct vop_vector zfs_vnodeops; struct vop_vector zfs_fifoops; struct vop_vector zfs_shareops; struct vop_vector zfs_vnodeops = { .vop_default = &default_vnodeops, .vop_inactive = zfs_freebsd_inactive, .vop_reclaim = zfs_freebsd_reclaim, .vop_access = zfs_freebsd_access, .vop_allocate = VOP_EINVAL, .vop_lookup = zfs_cache_lookup, .vop_cachedlookup = zfs_freebsd_lookup, .vop_getattr = zfs_freebsd_getattr, .vop_setattr = zfs_freebsd_setattr, .vop_create = zfs_freebsd_create, .vop_mknod = zfs_freebsd_create, .vop_mkdir = zfs_freebsd_mkdir, .vop_readdir = zfs_freebsd_readdir, .vop_fsync = zfs_freebsd_fsync, .vop_open = zfs_freebsd_open, .vop_close = zfs_freebsd_close, .vop_rmdir = zfs_freebsd_rmdir, .vop_ioctl = zfs_freebsd_ioctl, .vop_link = zfs_freebsd_link, .vop_symlink = zfs_freebsd_symlink, .vop_readlink = zfs_freebsd_readlink, .vop_read = zfs_freebsd_read, .vop_write = zfs_freebsd_write, .vop_remove = zfs_freebsd_remove, .vop_rename = zfs_freebsd_rename, .vop_pathconf = zfs_freebsd_pathconf, .vop_bmap = zfs_freebsd_bmap, .vop_fid = zfs_freebsd_fid, .vop_getextattr = zfs_getextattr, .vop_deleteextattr = zfs_deleteextattr, .vop_setextattr = zfs_setextattr, .vop_listextattr = zfs_listextattr, .vop_getacl = zfs_freebsd_getacl, .vop_setacl = zfs_freebsd_setacl, .vop_aclcheck = zfs_freebsd_aclcheck, .vop_getpages = zfs_freebsd_getpages, .vop_putpages = zfs_freebsd_putpages, .vop_vptocnp = zfs_vptocnp, #ifdef DIAGNOSTIC .vop_lock1 = zfs_lock, #else .vop_lock1 = vop_lock, #endif .vop_unlock = vop_unlock, .vop_islocked = vop_islocked, }; VFS_VOP_VECTOR_REGISTER(zfs_vnodeops); struct vop_vector zfs_fifoops = { .vop_default = &fifo_specops, .vop_fsync = zfs_freebsd_fsync, .vop_access = zfs_freebsd_access, .vop_getattr = zfs_freebsd_getattr, .vop_inactive = zfs_freebsd_inactive, .vop_read = VOP_PANIC, .vop_reclaim = zfs_freebsd_reclaim, .vop_setattr = zfs_freebsd_setattr, .vop_write = VOP_PANIC, .vop_pathconf = zfs_freebsd_pathconf, .vop_fid = zfs_freebsd_fid, .vop_getacl = zfs_freebsd_getacl, .vop_setacl = zfs_freebsd_setacl, .vop_aclcheck = zfs_freebsd_aclcheck, }; VFS_VOP_VECTOR_REGISTER(zfs_fifoops); /* * special share hidden files vnode operations template */ struct vop_vector zfs_shareops = { .vop_default = &default_vnodeops, .vop_access = zfs_freebsd_access, .vop_inactive = zfs_freebsd_inactive, .vop_reclaim = zfs_freebsd_reclaim, .vop_fid = zfs_freebsd_fid, .vop_pathconf = zfs_freebsd_pathconf, }; VFS_VOP_VECTOR_REGISTER(zfs_shareops); Index: head/sys/dev/md/md.c =================================================================== --- head/sys/dev/md/md.c (revision 356156) +++ head/sys/dev/md/md.c (revision 356157) @@ -1,2168 +1,2163 @@ /*- * SPDX-License-Identifier: (Beerware AND BSD-3-Clause) * * ---------------------------------------------------------------------------- * "THE BEER-WARE LICENSE" (Revision 42): * wrote this file. As long as you retain this notice you * can do whatever you want with this stuff. If we meet some day, and you think * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp * ---------------------------------------------------------------------------- * * $FreeBSD$ * */ /*- * The following functions are based in the vn(4) driver: mdstart_swap(), * mdstart_vnode(), mdcreate_swap(), mdcreate_vnode() and mddestroy(), * and as such under the following copyright: * * Copyright (c) 1988 University of Utah. * Copyright (c) 1990, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: Utah Hdr: vn.c 1.13 94/04/02 * * from: @(#)vn.c 8.6 (Berkeley) 4/1/94 * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03 */ #include "opt_rootdevname.h" #include "opt_geom.h" #include "opt_md.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define MD_MODVER 1 #define MD_SHUTDOWN 0x10000 /* Tell worker thread to terminate. */ #define MD_EXITING 0x20000 /* Worker thread is exiting. */ #define MD_PROVIDERGONE 0x40000 /* Safe to free the softc */ #ifndef MD_NSECT #define MD_NSECT (10000 * 2) #endif struct md_req { unsigned md_unit; /* unit number */ enum md_types md_type; /* type of disk */ off_t md_mediasize; /* size of disk in bytes */ unsigned md_sectorsize; /* sectorsize */ unsigned md_options; /* options */ int md_fwheads; /* firmware heads */ int md_fwsectors; /* firmware sectors */ char *md_file; /* pathname of file to mount */ enum uio_seg md_file_seg; /* location of md_file */ char *md_label; /* label of the device (userspace) */ int *md_units; /* pointer to units array (kernel) */ size_t md_units_nitems; /* items in md_units array */ }; #ifdef COMPAT_FREEBSD32 struct md_ioctl32 { unsigned md_version; unsigned md_unit; enum md_types md_type; uint32_t md_file; off_t md_mediasize; unsigned md_sectorsize; unsigned md_options; uint64_t md_base; int md_fwheads; int md_fwsectors; uint32_t md_label; int md_pad[MDNPAD]; } __attribute__((__packed__)); CTASSERT((sizeof(struct md_ioctl32)) == 436); #define MDIOCATTACH_32 _IOC_NEWTYPE(MDIOCATTACH, struct md_ioctl32) #define MDIOCDETACH_32 _IOC_NEWTYPE(MDIOCDETACH, struct md_ioctl32) #define MDIOCQUERY_32 _IOC_NEWTYPE(MDIOCQUERY, struct md_ioctl32) #define MDIOCRESIZE_32 _IOC_NEWTYPE(MDIOCRESIZE, struct md_ioctl32) #endif /* COMPAT_FREEBSD32 */ static MALLOC_DEFINE(M_MD, "md_disk", "Memory Disk"); static MALLOC_DEFINE(M_MDSECT, "md_sectors", "Memory Disk Sectors"); static int md_debug; SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0, "Enable md(4) debug messages"); static int md_malloc_wait; SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0, "Allow malloc to wait for memory allocations"); #if defined(MD_ROOT) && !defined(MD_ROOT_FSTYPE) #define MD_ROOT_FSTYPE "ufs" #endif #if defined(MD_ROOT) /* * Preloaded image gets put here. */ #if defined(MD_ROOT_SIZE) /* * We put the mfs_root symbol into the oldmfs section of the kernel object file. * Applications that patch the object with the image can determine * the size looking at the oldmfs section size within the kernel. */ u_char mfs_root[MD_ROOT_SIZE*1024] __attribute__ ((section ("oldmfs"))); const int mfs_root_size = sizeof(mfs_root); #elif defined(MD_ROOT_MEM) /* MD region already mapped in the memory */ u_char *mfs_root; int mfs_root_size; #else extern volatile u_char __weak_symbol mfs_root; extern volatile u_char __weak_symbol mfs_root_end; __GLOBL(mfs_root); __GLOBL(mfs_root_end); #define mfs_root_size ((uintptr_t)(&mfs_root_end - &mfs_root)) #endif #endif static g_init_t g_md_init; static g_fini_t g_md_fini; static g_start_t g_md_start; static g_access_t g_md_access; static void g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp); static g_provgone_t g_md_providergone; static struct cdev *status_dev = NULL; static struct sx md_sx; static struct unrhdr *md_uh; static d_ioctl_t mdctlioctl; static struct cdevsw mdctl_cdevsw = { .d_version = D_VERSION, .d_ioctl = mdctlioctl, .d_name = MD_NAME, }; struct g_class g_md_class = { .name = "MD", .version = G_VERSION, .init = g_md_init, .fini = g_md_fini, .start = g_md_start, .access = g_md_access, .dumpconf = g_md_dumpconf, .providergone = g_md_providergone, }; DECLARE_GEOM_CLASS(g_md_class, g_md); static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(md_softc_list); #define NINDIR (PAGE_SIZE / sizeof(uintptr_t)) #define NMASK (NINDIR-1) static int nshift; static uma_zone_t md_pbuf_zone; struct indir { uintptr_t *array; u_int total; u_int used; u_int shift; }; struct md_s { int unit; LIST_ENTRY(md_s) list; struct bio_queue_head bio_queue; struct mtx queue_mtx; struct mtx stat_mtx; struct cdev *dev; enum md_types type; off_t mediasize; unsigned sectorsize; unsigned opencount; unsigned fwheads; unsigned fwsectors; char ident[32]; unsigned flags; char name[20]; struct proc *procp; struct g_geom *gp; struct g_provider *pp; int (*start)(struct md_s *sc, struct bio *bp); struct devstat *devstat; /* MD_MALLOC related fields */ struct indir *indir; uma_zone_t uma; /* MD_PRELOAD related fields */ u_char *pl_ptr; size_t pl_len; /* MD_VNODE related fields */ struct vnode *vnode; char file[PATH_MAX]; char label[PATH_MAX]; struct ucred *cred; /* MD_SWAP related fields */ vm_object_t object; }; static struct indir * new_indir(u_int shift) { struct indir *ip; ip = malloc(sizeof *ip, M_MD, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (ip == NULL) return (NULL); ip->array = malloc(sizeof(uintptr_t) * NINDIR, M_MDSECT, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (ip->array == NULL) { free(ip, M_MD); return (NULL); } ip->total = NINDIR; ip->shift = shift; return (ip); } static void del_indir(struct indir *ip) { free(ip->array, M_MDSECT); free(ip, M_MD); } static void destroy_indir(struct md_s *sc, struct indir *ip) { int i; for (i = 0; i < NINDIR; i++) { if (!ip->array[i]) continue; if (ip->shift) destroy_indir(sc, (struct indir*)(ip->array[i])); else if (ip->array[i] > 255) uma_zfree(sc->uma, (void *)(ip->array[i])); } del_indir(ip); } /* * This function does the math and allocates the top level "indir" structure * for a device of "size" sectors. */ static struct indir * dimension(off_t size) { off_t rcnt; struct indir *ip; int layer; rcnt = size; layer = 0; while (rcnt > NINDIR) { rcnt /= NINDIR; layer++; } /* * XXX: the top layer is probably not fully populated, so we allocate * too much space for ip->array in here. */ ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO); ip->array = malloc(sizeof(uintptr_t) * NINDIR, M_MDSECT, M_WAITOK | M_ZERO); ip->total = NINDIR; ip->shift = layer * nshift; return (ip); } /* * Read a given sector */ static uintptr_t s_read(struct indir *ip, off_t offset) { struct indir *cip; int idx; uintptr_t up; if (md_debug > 1) printf("s_read(%jd)\n", (intmax_t)offset); up = 0; for (cip = ip; cip != NULL;) { if (cip->shift) { idx = (offset >> cip->shift) & NMASK; up = cip->array[idx]; cip = (struct indir *)up; continue; } idx = offset & NMASK; return (cip->array[idx]); } return (0); } /* * Write a given sector, prune the tree if the value is 0 */ static int s_write(struct indir *ip, off_t offset, uintptr_t ptr) { struct indir *cip, *lip[10]; int idx, li; uintptr_t up; if (md_debug > 1) printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr); up = 0; li = 0; cip = ip; for (;;) { lip[li++] = cip; if (cip->shift) { idx = (offset >> cip->shift) & NMASK; up = cip->array[idx]; if (up != 0) { cip = (struct indir *)up; continue; } /* Allocate branch */ cip->array[idx] = (uintptr_t)new_indir(cip->shift - nshift); if (cip->array[idx] == 0) return (ENOSPC); cip->used++; up = cip->array[idx]; cip = (struct indir *)up; continue; } /* leafnode */ idx = offset & NMASK; up = cip->array[idx]; if (up != 0) cip->used--; cip->array[idx] = ptr; if (ptr != 0) cip->used++; break; } if (cip->used != 0 || li == 1) return (0); li--; while (cip->used == 0 && cip != ip) { li--; idx = (offset >> lip[li]->shift) & NMASK; up = lip[li]->array[idx]; KASSERT(up == (uintptr_t)cip, ("md screwed up")); del_indir(cip); lip[li]->array[idx] = 0; lip[li]->used--; cip = lip[li]; } return (0); } static int g_md_access(struct g_provider *pp, int r, int w, int e) { struct md_s *sc; sc = pp->geom->softc; if (sc == NULL) { if (r <= 0 && w <= 0 && e <= 0) return (0); return (ENXIO); } r += pp->acr; w += pp->acw; e += pp->ace; if ((sc->flags & MD_READONLY) != 0 && w > 0) return (EROFS); if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) { sc->opencount = 1; } else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) { sc->opencount = 0; } return (0); } static void g_md_start(struct bio *bp) { struct md_s *sc; sc = bp->bio_to->geom->softc; if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE)) { mtx_lock(&sc->stat_mtx); devstat_start_transaction_bio(sc->devstat, bp); mtx_unlock(&sc->stat_mtx); } mtx_lock(&sc->queue_mtx); bioq_disksort(&sc->bio_queue, bp); wakeup(sc); mtx_unlock(&sc->queue_mtx); } #define MD_MALLOC_MOVE_ZERO 1 #define MD_MALLOC_MOVE_FILL 2 #define MD_MALLOC_MOVE_READ 3 #define MD_MALLOC_MOVE_WRITE 4 #define MD_MALLOC_MOVE_CMP 5 static int md_malloc_move_ma(vm_page_t **mp, int *ma_offs, unsigned sectorsize, void *ptr, u_char fill, int op) { struct sf_buf *sf; vm_page_t m, *mp1; char *p, first; off_t *uc; unsigned n; int error, i, ma_offs1, sz, first_read; m = NULL; error = 0; sf = NULL; /* if (op == MD_MALLOC_MOVE_CMP) { gcc */ first = 0; first_read = 0; uc = ptr; mp1 = *mp; ma_offs1 = *ma_offs; /* } */ sched_pin(); for (n = sectorsize; n != 0; n -= sz) { sz = imin(PAGE_SIZE - *ma_offs, n); if (m != **mp) { if (sf != NULL) sf_buf_free(sf); m = **mp; sf = sf_buf_alloc(m, SFB_CPUPRIVATE | (md_malloc_wait ? 0 : SFB_NOWAIT)); if (sf == NULL) { error = ENOMEM; break; } } p = (char *)sf_buf_kva(sf) + *ma_offs; switch (op) { case MD_MALLOC_MOVE_ZERO: bzero(p, sz); break; case MD_MALLOC_MOVE_FILL: memset(p, fill, sz); break; case MD_MALLOC_MOVE_READ: bcopy(ptr, p, sz); cpu_flush_dcache(p, sz); break; case MD_MALLOC_MOVE_WRITE: bcopy(p, ptr, sz); break; case MD_MALLOC_MOVE_CMP: for (i = 0; i < sz; i++, p++) { if (!first_read) { *uc = (u_char)*p; first = *p; first_read = 1; } else if (*p != first) { error = EDOOFUS; break; } } break; default: KASSERT(0, ("md_malloc_move_ma unknown op %d\n", op)); break; } if (error != 0) break; *ma_offs += sz; *ma_offs %= PAGE_SIZE; if (*ma_offs == 0) (*mp)++; ptr = (char *)ptr + sz; } if (sf != NULL) sf_buf_free(sf); sched_unpin(); if (op == MD_MALLOC_MOVE_CMP && error != 0) { *mp = mp1; *ma_offs = ma_offs1; } return (error); } static int md_malloc_move_vlist(bus_dma_segment_t **pvlist, int *pma_offs, unsigned len, void *ptr, u_char fill, int op) { bus_dma_segment_t *vlist; uint8_t *p, *end, first; off_t *uc; int ma_offs, seg_len; vlist = *pvlist; ma_offs = *pma_offs; uc = ptr; for (; len != 0; len -= seg_len) { seg_len = imin(vlist->ds_len - ma_offs, len); p = (uint8_t *)(uintptr_t)vlist->ds_addr + ma_offs; switch (op) { case MD_MALLOC_MOVE_ZERO: bzero(p, seg_len); break; case MD_MALLOC_MOVE_FILL: memset(p, fill, seg_len); break; case MD_MALLOC_MOVE_READ: bcopy(ptr, p, seg_len); cpu_flush_dcache(p, seg_len); break; case MD_MALLOC_MOVE_WRITE: bcopy(p, ptr, seg_len); break; case MD_MALLOC_MOVE_CMP: end = p + seg_len; first = *uc = *p; /* Confirm all following bytes match the first */ while (++p < end) { if (*p != first) return (EDOOFUS); } break; default: KASSERT(0, ("md_malloc_move_vlist unknown op %d\n", op)); break; } ma_offs += seg_len; if (ma_offs == vlist->ds_len) { ma_offs = 0; vlist++; } ptr = (uint8_t *)ptr + seg_len; } *pvlist = vlist; *pma_offs = ma_offs; return (0); } static int mdstart_malloc(struct md_s *sc, struct bio *bp) { u_char *dst; vm_page_t *m; bus_dma_segment_t *vlist; int i, error, error1, ma_offs, notmapped; off_t secno, nsec, uc; uintptr_t sp, osp; switch (bp->bio_cmd) { case BIO_READ: case BIO_WRITE: case BIO_DELETE: break; default: return (EOPNOTSUPP); } notmapped = (bp->bio_flags & BIO_UNMAPPED) != 0; vlist = (bp->bio_flags & BIO_VLIST) != 0 ? (bus_dma_segment_t *)bp->bio_data : NULL; if (notmapped) { m = bp->bio_ma; ma_offs = bp->bio_ma_offset; dst = NULL; KASSERT(vlist == NULL, ("vlists cannot be unmapped")); } else if (vlist != NULL) { ma_offs = bp->bio_ma_offset; dst = NULL; } else { dst = bp->bio_data; } nsec = bp->bio_length / sc->sectorsize; secno = bp->bio_offset / sc->sectorsize; error = 0; while (nsec--) { osp = s_read(sc->indir, secno); if (bp->bio_cmd == BIO_DELETE) { if (osp != 0) error = s_write(sc->indir, secno, 0); } else if (bp->bio_cmd == BIO_READ) { if (osp == 0) { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, NULL, 0, MD_MALLOC_MOVE_ZERO); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, NULL, 0, MD_MALLOC_MOVE_ZERO); } else bzero(dst, sc->sectorsize); } else if (osp <= 255) { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, NULL, osp, MD_MALLOC_MOVE_FILL); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, NULL, osp, MD_MALLOC_MOVE_FILL); } else memset(dst, osp, sc->sectorsize); } else { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_READ); } else if (vlist != NULL) { error = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_READ); } else { bcopy((void *)osp, dst, sc->sectorsize); cpu_flush_dcache(dst, sc->sectorsize); } } osp = 0; } else if (bp->bio_cmd == BIO_WRITE) { if (sc->flags & MD_COMPRESS) { if (notmapped) { error1 = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, &uc, 0, MD_MALLOC_MOVE_CMP); i = error1 == 0 ? sc->sectorsize : 0; } else if (vlist != NULL) { error1 = md_malloc_move_vlist(&vlist, &ma_offs, sc->sectorsize, &uc, 0, MD_MALLOC_MOVE_CMP); i = error1 == 0 ? sc->sectorsize : 0; } else { uc = dst[0]; for (i = 1; i < sc->sectorsize; i++) { if (dst[i] != uc) break; } } } else { i = 0; uc = 0; } if (i == sc->sectorsize) { if (osp != uc) error = s_write(sc->indir, secno, uc); } else { if (osp <= 255) { sp = (uintptr_t)uma_zalloc(sc->uma, md_malloc_wait ? M_WAITOK : M_NOWAIT); if (sp == 0) { error = ENOSPC; break; } if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)sp, 0, MD_MALLOC_MOVE_WRITE); } else if (vlist != NULL) { error = md_malloc_move_vlist( &vlist, &ma_offs, sc->sectorsize, (void *)sp, 0, MD_MALLOC_MOVE_WRITE); } else { bcopy(dst, (void *)sp, sc->sectorsize); } error = s_write(sc->indir, secno, sp); } else { if (notmapped) { error = md_malloc_move_ma(&m, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_WRITE); } else if (vlist != NULL) { error = md_malloc_move_vlist( &vlist, &ma_offs, sc->sectorsize, (void *)osp, 0, MD_MALLOC_MOVE_WRITE); } else { bcopy(dst, (void *)osp, sc->sectorsize); } osp = 0; } } } else { error = EOPNOTSUPP; } if (osp > 255) uma_zfree(sc->uma, (void*)osp); if (error != 0) break; secno++; if (!notmapped && vlist == NULL) dst += sc->sectorsize; } bp->bio_resid = 0; return (error); } static void mdcopyto_vlist(void *src, bus_dma_segment_t *vlist, off_t offset, off_t len) { off_t seg_len; while (offset >= vlist->ds_len) { offset -= vlist->ds_len; vlist++; } while (len != 0) { seg_len = omin(len, vlist->ds_len - offset); bcopy(src, (void *)(uintptr_t)(vlist->ds_addr + offset), seg_len); offset = 0; src = (uint8_t *)src + seg_len; len -= seg_len; vlist++; } } static void mdcopyfrom_vlist(bus_dma_segment_t *vlist, off_t offset, void *dst, off_t len) { off_t seg_len; while (offset >= vlist->ds_len) { offset -= vlist->ds_len; vlist++; } while (len != 0) { seg_len = omin(len, vlist->ds_len - offset); bcopy((void *)(uintptr_t)(vlist->ds_addr + offset), dst, seg_len); offset = 0; dst = (uint8_t *)dst + seg_len; len -= seg_len; vlist++; } } static int mdstart_preload(struct md_s *sc, struct bio *bp) { uint8_t *p; p = sc->pl_ptr + bp->bio_offset; switch (bp->bio_cmd) { case BIO_READ: if ((bp->bio_flags & BIO_VLIST) != 0) { mdcopyto_vlist(p, (bus_dma_segment_t *)bp->bio_data, bp->bio_ma_offset, bp->bio_length); } else { bcopy(p, bp->bio_data, bp->bio_length); } cpu_flush_dcache(bp->bio_data, bp->bio_length); break; case BIO_WRITE: if ((bp->bio_flags & BIO_VLIST) != 0) { mdcopyfrom_vlist((bus_dma_segment_t *)bp->bio_data, bp->bio_ma_offset, p, bp->bio_length); } else { bcopy(bp->bio_data, p, bp->bio_length); } break; } bp->bio_resid = 0; return (0); } static int mdstart_vnode(struct md_s *sc, struct bio *bp) { int error; struct uio auio; struct iovec aiov; struct iovec *piov; struct mount *mp; struct vnode *vp; struct buf *pb; bus_dma_segment_t *vlist; struct thread *td; off_t iolen, iostart, len, zerosize; int ma_offs, npages; switch (bp->bio_cmd) { case BIO_READ: auio.uio_rw = UIO_READ; break; case BIO_WRITE: case BIO_DELETE: auio.uio_rw = UIO_WRITE; break; case BIO_FLUSH: break; default: return (EOPNOTSUPP); } td = curthread; vp = sc->vnode; pb = NULL; piov = NULL; ma_offs = bp->bio_ma_offset; len = bp->bio_length; /* * VNODE I/O * * If an error occurs, we set BIO_ERROR but we do not set * B_INVAL because (for a write anyway), the buffer is * still valid. */ if (bp->bio_cmd == BIO_FLUSH) { (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_FSYNC(vp, MNT_WAIT, td); VOP_UNLOCK(vp, 0); vn_finished_write(mp); return (error); } auio.uio_offset = (vm_ooffset_t)bp->bio_offset; auio.uio_resid = bp->bio_length; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; if (bp->bio_cmd == BIO_DELETE) { /* * Emulate BIO_DELETE by writing zeros. */ zerosize = ZERO_REGION_SIZE - (ZERO_REGION_SIZE % sc->sectorsize); auio.uio_iovcnt = howmany(bp->bio_length, zerosize); piov = malloc(sizeof(*piov) * auio.uio_iovcnt, M_MD, M_WAITOK); auio.uio_iov = piov; while (len > 0) { piov->iov_base = __DECONST(void *, zero_region); piov->iov_len = len; if (len > zerosize) piov->iov_len = zerosize; len -= piov->iov_len; piov++; } piov = auio.uio_iov; } else if ((bp->bio_flags & BIO_VLIST) != 0) { piov = malloc(sizeof(*piov) * bp->bio_ma_n, M_MD, M_WAITOK); auio.uio_iov = piov; vlist = (bus_dma_segment_t *)bp->bio_data; while (len > 0) { piov->iov_base = (void *)(uintptr_t)(vlist->ds_addr + ma_offs); piov->iov_len = vlist->ds_len - ma_offs; if (piov->iov_len > len) piov->iov_len = len; len -= piov->iov_len; ma_offs = 0; vlist++; piov++; } auio.uio_iovcnt = piov - auio.uio_iov; piov = auio.uio_iov; } else if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pb = uma_zalloc(md_pbuf_zone, M_WAITOK); bp->bio_resid = len; unmapped_step: npages = atop(min(MAXPHYS, round_page(len + (ma_offs & PAGE_MASK)))); iolen = min(ptoa(npages) - (ma_offs & PAGE_MASK), len); KASSERT(iolen > 0, ("zero iolen")); pmap_qenter((vm_offset_t)pb->b_data, &bp->bio_ma[atop(ma_offs)], npages); aiov.iov_base = (void *)((vm_offset_t)pb->b_data + (ma_offs & PAGE_MASK)); aiov.iov_len = iolen; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_resid = iolen; } else { aiov.iov_base = bp->bio_data; aiov.iov_len = bp->bio_length; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; } iostart = auio.uio_offset; if (auio.uio_rw == UIO_READ) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_READ(vp, &auio, 0, sc->cred); VOP_UNLOCK(vp, 0); } else { (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); error = VOP_WRITE(vp, &auio, sc->flags & MD_ASYNC ? 0 : IO_SYNC, sc->cred); VOP_UNLOCK(vp, 0); vn_finished_write(mp); if (error == 0) sc->flags &= ~MD_VERIFY; } /* When MD_CACHE is set, try to avoid double-caching the data. */ if (error == 0 && (sc->flags & MD_CACHE) == 0) VOP_ADVISE(vp, iostart, auio.uio_offset - 1, POSIX_FADV_DONTNEED); if (pb != NULL) { pmap_qremove((vm_offset_t)pb->b_data, npages); if (error == 0) { len -= iolen; bp->bio_resid -= iolen; ma_offs += iolen; if (len > 0) goto unmapped_step; } uma_zfree(md_pbuf_zone, pb); } free(piov, M_MD); if (pb == NULL) bp->bio_resid = auio.uio_resid; return (error); } static int mdstart_swap(struct md_s *sc, struct bio *bp) { vm_page_t m; u_char *p; vm_pindex_t i, lastp; bus_dma_segment_t *vlist; int rv, ma_offs, offs, len, lastend; switch (bp->bio_cmd) { case BIO_READ: case BIO_WRITE: case BIO_DELETE: break; default: return (EOPNOTSUPP); } p = bp->bio_data; ma_offs = (bp->bio_flags & (BIO_UNMAPPED|BIO_VLIST)) != 0 ? bp->bio_ma_offset : 0; vlist = (bp->bio_flags & BIO_VLIST) != 0 ? (bus_dma_segment_t *)bp->bio_data : NULL; /* * offs is the offset at which to start operating on the * next (ie, first) page. lastp is the last page on * which we're going to operate. lastend is the ending * position within that last page (ie, PAGE_SIZE if * we're operating on complete aligned pages). */ offs = bp->bio_offset % PAGE_SIZE; lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE; lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1; rv = VM_PAGER_OK; VM_OBJECT_WLOCK(sc->object); vm_object_pip_add(sc->object, 1); for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) { len = ((i == lastp) ? lastend : PAGE_SIZE) - offs; m = vm_page_grab(sc->object, i, VM_ALLOC_SYSTEM); if (bp->bio_cmd == BIO_READ) { if (vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { vm_page_free(m); break; } else if (rv == VM_PAGER_FAIL) { /* * Pager does not have the page. Zero * the allocated page, and mark it as * valid. Do not set dirty, the page * can be recreated if thrown out. */ pmap_zero_page(m); vm_page_valid(m); } if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pmap_copy_pages(&m, offs, bp->bio_ma, ma_offs, len); } else if ((bp->bio_flags & BIO_VLIST) != 0) { physcopyout_vlist(VM_PAGE_TO_PHYS(m) + offs, vlist, ma_offs, len); cpu_flush_dcache(p, len); } else { physcopyout(VM_PAGE_TO_PHYS(m) + offs, p, len); cpu_flush_dcache(p, len); } } else if (bp->bio_cmd == BIO_WRITE) { if (len == PAGE_SIZE || vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { vm_page_free(m); break; } else if (rv == VM_PAGER_FAIL) pmap_zero_page(m); if ((bp->bio_flags & BIO_UNMAPPED) != 0) { pmap_copy_pages(bp->bio_ma, ma_offs, &m, offs, len); } else if ((bp->bio_flags & BIO_VLIST) != 0) { physcopyin_vlist(vlist, ma_offs, VM_PAGE_TO_PHYS(m) + offs, len); } else { physcopyin(p, VM_PAGE_TO_PHYS(m) + offs, len); } vm_page_valid(m); vm_page_set_dirty(m); } else if (bp->bio_cmd == BIO_DELETE) { if (len == PAGE_SIZE || vm_page_all_valid(m)) rv = VM_PAGER_OK; else rv = vm_pager_get_pages(sc->object, &m, 1, NULL, NULL); if (rv == VM_PAGER_ERROR) { vm_page_free(m); break; } else if (rv == VM_PAGER_FAIL) { vm_page_free(m); m = NULL; } else { /* Page is valid. */ if (len != PAGE_SIZE) { pmap_zero_page_area(m, offs, len); vm_page_set_dirty(m); } else { vm_pager_page_unswapped(m); vm_page_free(m); m = NULL; } } } if (m != NULL) { vm_page_xunbusy(m); - vm_page_lock(m); - if (vm_page_active(m)) - vm_page_reference(m); - else - vm_page_activate(m); - vm_page_unlock(m); + vm_page_reference(m); } /* Actions on further pages start at offset 0 */ p += PAGE_SIZE - offs; offs = 0; ma_offs += len; } vm_object_pip_wakeup(sc->object); VM_OBJECT_WUNLOCK(sc->object); return (rv != VM_PAGER_ERROR ? 0 : ENOSPC); } static int mdstart_null(struct md_s *sc, struct bio *bp) { switch (bp->bio_cmd) { case BIO_READ: bzero(bp->bio_data, bp->bio_length); cpu_flush_dcache(bp->bio_data, bp->bio_length); break; case BIO_WRITE: break; } bp->bio_resid = 0; return (0); } static void md_kthread(void *arg) { struct md_s *sc; struct bio *bp; int error; sc = arg; thread_lock(curthread); sched_prio(curthread, PRIBIO); thread_unlock(curthread); if (sc->type == MD_VNODE) curthread->td_pflags |= TDP_NORUNNINGBUF; for (;;) { mtx_lock(&sc->queue_mtx); if (sc->flags & MD_SHUTDOWN) { sc->flags |= MD_EXITING; mtx_unlock(&sc->queue_mtx); kproc_exit(0); } bp = bioq_takefirst(&sc->bio_queue); if (!bp) { msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0); continue; } mtx_unlock(&sc->queue_mtx); if (bp->bio_cmd == BIO_GETATTR) { int isv = ((sc->flags & MD_VERIFY) != 0); if ((sc->fwsectors && sc->fwheads && (g_handleattr_int(bp, "GEOM::fwsectors", sc->fwsectors) || g_handleattr_int(bp, "GEOM::fwheads", sc->fwheads))) || g_handleattr_int(bp, "GEOM::candelete", 1)) error = -1; else if (sc->ident[0] != '\0' && g_handleattr_str(bp, "GEOM::ident", sc->ident)) error = -1; else if (g_handleattr_int(bp, "MNT::verified", isv)) error = -1; else error = EOPNOTSUPP; } else { error = sc->start(sc, bp); } if (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE) { /* * Devstat uses (bio_bcount, bio_resid) for * determining the length of the completed part of * the i/o. g_io_deliver() will translate from * bio_completed to that, but it also destroys the * bio so we must do our own translation. */ bp->bio_bcount = bp->bio_length; bp->bio_resid = (error == -1 ? bp->bio_bcount : 0); devstat_end_transaction_bio(sc->devstat, bp); } if (error != -1) { bp->bio_completed = bp->bio_length; g_io_deliver(bp, error); } } } static struct md_s * mdfind(int unit) { struct md_s *sc; LIST_FOREACH(sc, &md_softc_list, list) { if (sc->unit == unit) break; } return (sc); } static struct md_s * mdnew(int unit, int *errp, enum md_types type) { struct md_s *sc; int error; *errp = 0; if (unit == -1) unit = alloc_unr(md_uh); else unit = alloc_unr_specific(md_uh, unit); if (unit == -1) { *errp = EBUSY; return (NULL); } sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO); sc->type = type; bioq_init(&sc->bio_queue); mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF); mtx_init(&sc->stat_mtx, "md stat", NULL, MTX_DEF); sc->unit = unit; sprintf(sc->name, "md%d", unit); LIST_INSERT_HEAD(&md_softc_list, sc, list); error = kproc_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name); if (error == 0) return (sc); LIST_REMOVE(sc, list); mtx_destroy(&sc->stat_mtx); mtx_destroy(&sc->queue_mtx); free_unr(md_uh, sc->unit); free(sc, M_MD); *errp = error; return (NULL); } static void mdinit(struct md_s *sc) { struct g_geom *gp; struct g_provider *pp; g_topology_lock(); gp = g_new_geomf(&g_md_class, "md%d", sc->unit); gp->softc = sc; pp = g_new_providerf(gp, "md%d", sc->unit); pp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE; pp->mediasize = sc->mediasize; pp->sectorsize = sc->sectorsize; switch (sc->type) { case MD_MALLOC: case MD_VNODE: case MD_SWAP: pp->flags |= G_PF_ACCEPT_UNMAPPED; break; case MD_PRELOAD: case MD_NULL: break; } sc->gp = gp; sc->pp = pp; g_error_provider(pp, 0); g_topology_unlock(); sc->devstat = devstat_new_entry("md", sc->unit, sc->sectorsize, DEVSTAT_ALL_SUPPORTED, DEVSTAT_TYPE_DIRECT, DEVSTAT_PRIORITY_MAX); } static int mdcreate_malloc(struct md_s *sc, struct md_req *mdr) { uintptr_t sp; int error; off_t u; error = 0; if (mdr->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE)) return (EINVAL); if (mdr->md_sectorsize != 0 && !powerof2(mdr->md_sectorsize)) return (EINVAL); /* Compression doesn't make sense if we have reserved space */ if (mdr->md_options & MD_RESERVE) mdr->md_options &= ~MD_COMPRESS; if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; sc->flags = mdr->md_options & (MD_COMPRESS | MD_FORCE); sc->indir = dimension(sc->mediasize / sc->sectorsize); sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL, 0x1ff, 0); if (mdr->md_options & MD_RESERVE) { off_t nsectors; nsectors = sc->mediasize / sc->sectorsize; for (u = 0; u < nsectors; u++) { sp = (uintptr_t)uma_zalloc(sc->uma, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO); if (sp != 0) error = s_write(sc->indir, u, sp); else error = ENOMEM; if (error != 0) break; } } return (error); } static int mdsetcred(struct md_s *sc, struct ucred *cred) { char *tmpbuf; int error = 0; /* * Set credits in our softc */ if (sc->cred) crfree(sc->cred); sc->cred = crhold(cred); /* * Horrible kludge to establish credentials for NFS XXX. */ if (sc->vnode) { struct uio auio; struct iovec aiov; tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK); bzero(&auio, sizeof(auio)); aiov.iov_base = tmpbuf; aiov.iov_len = sc->sectorsize; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = 0; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_resid = aiov.iov_len; vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY); error = VOP_READ(sc->vnode, &auio, 0, sc->cred); VOP_UNLOCK(sc->vnode, 0); free(tmpbuf, M_TEMP); } return (error); } static int mdcreate_vnode(struct md_s *sc, struct md_req *mdr, struct thread *td) { struct vattr vattr; struct nameidata nd; char *fname; int error, flags; fname = mdr->md_file; if (mdr->md_file_seg == UIO_USERSPACE) { error = copyinstr(fname, sc->file, sizeof(sc->file), NULL); if (error != 0) return (error); } else if (mdr->md_file_seg == UIO_SYSSPACE) strlcpy(sc->file, fname, sizeof(sc->file)); else return (EDOOFUS); /* * If the user specified that this is a read only device, don't * set the FWRITE mask before trying to open the backing store. */ flags = FREAD | ((mdr->md_options & MD_READONLY) ? 0 : FWRITE) \ | ((mdr->md_options & MD_VERIFY) ? O_VERIFY : 0); NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td); error = vn_open(&nd, &flags, 0, NULL); if (error != 0) return (error); NDFREE(&nd, NDF_ONLY_PNBUF); if (nd.ni_vp->v_type != VREG) { error = EINVAL; goto bad; } error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred); if (error != 0) goto bad; if (VOP_ISLOCKED(nd.ni_vp) != LK_EXCLUSIVE) { vn_lock(nd.ni_vp, LK_UPGRADE | LK_RETRY); if (VN_IS_DOOMED(nd.ni_vp)) { /* Forced unmount. */ error = EBADF; goto bad; } } nd.ni_vp->v_vflag |= VV_MD; VOP_UNLOCK(nd.ni_vp, 0); if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; snprintf(sc->ident, sizeof(sc->ident), "MD-DEV%ju-INO%ju", (uintmax_t)vattr.va_fsid, (uintmax_t)vattr.va_fileid); sc->flags = mdr->md_options & (MD_ASYNC | MD_CACHE | MD_FORCE | MD_VERIFY); if (!(flags & FWRITE)) sc->flags |= MD_READONLY; sc->vnode = nd.ni_vp; error = mdsetcred(sc, td->td_ucred); if (error != 0) { sc->vnode = NULL; vn_lock(nd.ni_vp, LK_EXCLUSIVE | LK_RETRY); nd.ni_vp->v_vflag &= ~VV_MD; goto bad; } return (0); bad: VOP_UNLOCK(nd.ni_vp, 0); (void)vn_close(nd.ni_vp, flags, td->td_ucred, td); return (error); } static void g_md_providergone(struct g_provider *pp) { struct md_s *sc = pp->geom->softc; mtx_lock(&sc->queue_mtx); sc->flags |= MD_PROVIDERGONE; wakeup(&sc->flags); mtx_unlock(&sc->queue_mtx); } static int mddestroy(struct md_s *sc, struct thread *td) { if (sc->gp) { g_topology_lock(); g_wither_geom(sc->gp, ENXIO); g_topology_unlock(); mtx_lock(&sc->queue_mtx); while (!(sc->flags & MD_PROVIDERGONE)) msleep(&sc->flags, &sc->queue_mtx, PRIBIO, "mddestroy", 0); mtx_unlock(&sc->queue_mtx); } if (sc->devstat) { devstat_remove_entry(sc->devstat); sc->devstat = NULL; } mtx_lock(&sc->queue_mtx); sc->flags |= MD_SHUTDOWN; wakeup(sc); while (!(sc->flags & MD_EXITING)) msleep(sc->procp, &sc->queue_mtx, PRIBIO, "mddestroy", hz / 10); mtx_unlock(&sc->queue_mtx); mtx_destroy(&sc->stat_mtx); mtx_destroy(&sc->queue_mtx); if (sc->vnode != NULL) { vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY); sc->vnode->v_vflag &= ~VV_MD; VOP_UNLOCK(sc->vnode, 0); (void)vn_close(sc->vnode, sc->flags & MD_READONLY ? FREAD : (FREAD|FWRITE), sc->cred, td); } if (sc->cred != NULL) crfree(sc->cred); if (sc->object != NULL) vm_object_deallocate(sc->object); if (sc->indir) destroy_indir(sc, sc->indir); if (sc->uma) uma_zdestroy(sc->uma); LIST_REMOVE(sc, list); free_unr(md_uh, sc->unit); free(sc, M_MD); return (0); } static int mdresize(struct md_s *sc, struct md_req *mdr) { int error, res; vm_pindex_t oldpages, newpages; switch (sc->type) { case MD_VNODE: case MD_NULL: break; case MD_SWAP: if (mdr->md_mediasize <= 0 || (mdr->md_mediasize % PAGE_SIZE) != 0) return (EDOM); oldpages = OFF_TO_IDX(round_page(sc->mediasize)); newpages = OFF_TO_IDX(round_page(mdr->md_mediasize)); if (newpages < oldpages) { VM_OBJECT_WLOCK(sc->object); vm_object_page_remove(sc->object, newpages, 0, 0); swap_pager_freespace(sc->object, newpages, oldpages - newpages); swap_release_by_cred(IDX_TO_OFF(oldpages - newpages), sc->cred); sc->object->charge = IDX_TO_OFF(newpages); sc->object->size = newpages; VM_OBJECT_WUNLOCK(sc->object); } else if (newpages > oldpages) { res = swap_reserve_by_cred(IDX_TO_OFF(newpages - oldpages), sc->cred); if (!res) return (ENOMEM); if ((mdr->md_options & MD_RESERVE) || (sc->flags & MD_RESERVE)) { error = swap_pager_reserve(sc->object, oldpages, newpages - oldpages); if (error < 0) { swap_release_by_cred( IDX_TO_OFF(newpages - oldpages), sc->cred); return (EDOM); } } VM_OBJECT_WLOCK(sc->object); sc->object->charge = IDX_TO_OFF(newpages); sc->object->size = newpages; VM_OBJECT_WUNLOCK(sc->object); } break; default: return (EOPNOTSUPP); } sc->mediasize = mdr->md_mediasize; g_topology_lock(); g_resize_provider(sc->pp, sc->mediasize); g_topology_unlock(); return (0); } static int mdcreate_swap(struct md_s *sc, struct md_req *mdr, struct thread *td) { vm_ooffset_t npage; int error; /* * Range check. Disallow negative sizes and sizes not being * multiple of page size. */ if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0) return (EDOM); /* * Allocate an OBJT_SWAP object. * * Note the truncation. */ if ((mdr->md_options & MD_VERIFY) != 0) return (EINVAL); npage = mdr->md_mediasize / PAGE_SIZE; if (mdr->md_fwsectors != 0) sc->fwsectors = mdr->md_fwsectors; if (mdr->md_fwheads != 0) sc->fwheads = mdr->md_fwheads; sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage, VM_PROT_DEFAULT, 0, td->td_ucred); if (sc->object == NULL) return (ENOMEM); sc->flags = mdr->md_options & (MD_FORCE | MD_RESERVE); if (mdr->md_options & MD_RESERVE) { if (swap_pager_reserve(sc->object, 0, npage) < 0) { error = EDOM; goto finish; } } error = mdsetcred(sc, td->td_ucred); finish: if (error != 0) { vm_object_deallocate(sc->object); sc->object = NULL; } return (error); } static int mdcreate_null(struct md_s *sc, struct md_req *mdr, struct thread *td) { /* * Range check. Disallow negative sizes and sizes not being * multiple of page size. */ if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0) return (EDOM); return (0); } static int kern_mdattach_locked(struct thread *td, struct md_req *mdr) { struct md_s *sc; unsigned sectsize; int error, i; sx_assert(&md_sx, SA_XLOCKED); switch (mdr->md_type) { case MD_MALLOC: case MD_PRELOAD: case MD_VNODE: case MD_SWAP: case MD_NULL: break; default: return (EINVAL); } if (mdr->md_sectorsize == 0) sectsize = DEV_BSIZE; else sectsize = mdr->md_sectorsize; if (sectsize > MAXPHYS || mdr->md_mediasize < sectsize) return (EINVAL); if (mdr->md_options & MD_AUTOUNIT) sc = mdnew(-1, &error, mdr->md_type); else { if (mdr->md_unit > INT_MAX) return (EINVAL); sc = mdnew(mdr->md_unit, &error, mdr->md_type); } if (sc == NULL) return (error); if (mdr->md_label != NULL) error = copyinstr(mdr->md_label, sc->label, sizeof(sc->label), NULL); if (error != 0) goto err_after_new; if (mdr->md_options & MD_AUTOUNIT) mdr->md_unit = sc->unit; sc->mediasize = mdr->md_mediasize; sc->sectorsize = sectsize; error = EDOOFUS; switch (sc->type) { case MD_MALLOC: sc->start = mdstart_malloc; error = mdcreate_malloc(sc, mdr); break; case MD_PRELOAD: /* * We disallow attaching preloaded memory disks via * ioctl. Preloaded memory disks are automatically * attached in g_md_init(). */ error = EOPNOTSUPP; break; case MD_VNODE: sc->start = mdstart_vnode; error = mdcreate_vnode(sc, mdr, td); break; case MD_SWAP: sc->start = mdstart_swap; error = mdcreate_swap(sc, mdr, td); break; case MD_NULL: sc->start = mdstart_null; error = mdcreate_null(sc, mdr, td); break; } err_after_new: if (error != 0) { mddestroy(sc, td); return (error); } /* Prune off any residual fractional sector */ i = sc->mediasize % sc->sectorsize; sc->mediasize -= i; mdinit(sc); return (0); } static int kern_mdattach(struct thread *td, struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdattach_locked(td, mdr); sx_xunlock(&md_sx); return (error); } static int kern_mddetach_locked(struct thread *td, struct md_req *mdr) { struct md_s *sc; sx_assert(&md_sx, SA_XLOCKED); if (mdr->md_mediasize != 0 || (mdr->md_options & ~MD_FORCE) != 0) return (EINVAL); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); if (sc->opencount != 0 && !(sc->flags & MD_FORCE) && !(mdr->md_options & MD_FORCE)) return (EBUSY); return (mddestroy(sc, td)); } static int kern_mddetach(struct thread *td, struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mddetach_locked(td, mdr); sx_xunlock(&md_sx); return (error); } static int kern_mdresize_locked(struct md_req *mdr) { struct md_s *sc; sx_assert(&md_sx, SA_XLOCKED); if ((mdr->md_options & ~(MD_FORCE | MD_RESERVE)) != 0) return (EINVAL); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); if (mdr->md_mediasize < sc->sectorsize) return (EINVAL); if (mdr->md_mediasize < sc->mediasize && !(sc->flags & MD_FORCE) && !(mdr->md_options & MD_FORCE)) return (EBUSY); return (mdresize(sc, mdr)); } static int kern_mdresize(struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdresize_locked(mdr); sx_xunlock(&md_sx); return (error); } static int kern_mdquery_locked(struct md_req *mdr) { struct md_s *sc; int error; sx_assert(&md_sx, SA_XLOCKED); sc = mdfind(mdr->md_unit); if (sc == NULL) return (ENOENT); mdr->md_type = sc->type; mdr->md_options = sc->flags; mdr->md_mediasize = sc->mediasize; mdr->md_sectorsize = sc->sectorsize; error = 0; if (mdr->md_label != NULL) { error = copyout(sc->label, mdr->md_label, strlen(sc->label) + 1); if (error != 0) return (error); } if (sc->type == MD_VNODE || (sc->type == MD_PRELOAD && mdr->md_file != NULL)) error = copyout(sc->file, mdr->md_file, strlen(sc->file) + 1); return (error); } static int kern_mdquery(struct md_req *mdr) { int error; sx_xlock(&md_sx); error = kern_mdquery_locked(mdr); sx_xunlock(&md_sx); return (error); } /* Copy members that are not userspace pointers. */ #define MD_IOCTL2REQ(mdio, mdr) do { \ (mdr)->md_unit = (mdio)->md_unit; \ (mdr)->md_type = (mdio)->md_type; \ (mdr)->md_mediasize = (mdio)->md_mediasize; \ (mdr)->md_sectorsize = (mdio)->md_sectorsize; \ (mdr)->md_options = (mdio)->md_options; \ (mdr)->md_fwheads = (mdio)->md_fwheads; \ (mdr)->md_fwsectors = (mdio)->md_fwsectors; \ (mdr)->md_units = &(mdio)->md_pad[0]; \ (mdr)->md_units_nitems = nitems((mdio)->md_pad); \ } while(0) /* Copy members that might have been updated */ #define MD_REQ2IOCTL(mdr, mdio) do { \ (mdio)->md_unit = (mdr)->md_unit; \ (mdio)->md_type = (mdr)->md_type; \ (mdio)->md_mediasize = (mdr)->md_mediasize; \ (mdio)->md_sectorsize = (mdr)->md_sectorsize; \ (mdio)->md_options = (mdr)->md_options; \ (mdio)->md_fwheads = (mdr)->md_fwheads; \ (mdio)->md_fwsectors = (mdr)->md_fwsectors; \ } while(0) static int mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td) { struct md_req mdr; int error; if (md_debug) printf("mdctlioctl(%s %lx %p %x %p)\n", devtoname(dev), cmd, addr, flags, td); bzero(&mdr, sizeof(mdr)); switch (cmd) { case MDIOCATTACH: case MDIOCDETACH: case MDIOCRESIZE: case MDIOCQUERY: { struct md_ioctl *mdio = (struct md_ioctl *)addr; if (mdio->md_version != MDIOVERSION) return (EINVAL); MD_IOCTL2REQ(mdio, &mdr); mdr.md_file = mdio->md_file; mdr.md_file_seg = UIO_USERSPACE; /* If the file is adjacent to the md_ioctl it's in kernel. */ if ((void *)mdio->md_file == (void *)(mdio + 1)) mdr.md_file_seg = UIO_SYSSPACE; mdr.md_label = mdio->md_label; break; } #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: case MDIOCDETACH_32: case MDIOCRESIZE_32: case MDIOCQUERY_32: { struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr; if (mdio->md_version != MDIOVERSION) return (EINVAL); MD_IOCTL2REQ(mdio, &mdr); mdr.md_file = (void *)(uintptr_t)mdio->md_file; mdr.md_file_seg = UIO_USERSPACE; mdr.md_label = (void *)(uintptr_t)mdio->md_label; break; } #endif default: /* Fall through to handler switch. */ break; } error = 0; switch (cmd) { case MDIOCATTACH: #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: #endif error = kern_mdattach(td, &mdr); break; case MDIOCDETACH: #ifdef COMPAT_FREEBSD32 case MDIOCDETACH_32: #endif error = kern_mddetach(td, &mdr); break; case MDIOCRESIZE: #ifdef COMPAT_FREEBSD32 case MDIOCRESIZE_32: #endif error = kern_mdresize(&mdr); break; case MDIOCQUERY: #ifdef COMPAT_FREEBSD32 case MDIOCQUERY_32: #endif error = kern_mdquery(&mdr); break; default: error = ENOIOCTL; } switch (cmd) { case MDIOCATTACH: case MDIOCQUERY: { struct md_ioctl *mdio = (struct md_ioctl *)addr; MD_REQ2IOCTL(&mdr, mdio); break; } #ifdef COMPAT_FREEBSD32 case MDIOCATTACH_32: case MDIOCQUERY_32: { struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr; MD_REQ2IOCTL(&mdr, mdio); break; } #endif default: /* Other commands to not alter mdr. */ break; } return (error); } static void md_preloaded(u_char *image, size_t length, const char *name) { struct md_s *sc; int error; sc = mdnew(-1, &error, MD_PRELOAD); if (sc == NULL) return; sc->mediasize = length; sc->sectorsize = DEV_BSIZE; sc->pl_ptr = image; sc->pl_len = length; sc->start = mdstart_preload; if (name != NULL) strlcpy(sc->file, name, sizeof(sc->file)); #ifdef MD_ROOT if (sc->unit == 0) { #ifndef ROOTDEVNAME rootdevnames[0] = MD_ROOT_FSTYPE ":/dev/md0"; #endif #ifdef MD_ROOT_READONLY sc->flags |= MD_READONLY; #endif } #endif mdinit(sc); if (name != NULL) { printf("%s%d: Preloaded image <%s> %zd bytes at %p\n", MD_NAME, sc->unit, name, length, image); } else { printf("%s%d: Embedded image %zd bytes at %p\n", MD_NAME, sc->unit, length, image); } } static void g_md_init(struct g_class *mp __unused) { caddr_t mod; u_char *ptr, *name, *type; unsigned len; int i; /* figure out log2(NINDIR) */ for (i = NINDIR, nshift = -1; i; nshift++) i >>= 1; mod = NULL; sx_init(&md_sx, "MD config lock"); g_topology_unlock(); md_uh = new_unrhdr(0, INT_MAX, NULL); #ifdef MD_ROOT if (mfs_root_size != 0) { sx_xlock(&md_sx); #ifdef MD_ROOT_MEM md_preloaded(mfs_root, mfs_root_size, NULL); #else md_preloaded(__DEVOLATILE(u_char *, &mfs_root), mfs_root_size, NULL); #endif sx_xunlock(&md_sx); } #endif /* XXX: are preload_* static or do they need Giant ? */ while ((mod = preload_search_next_name(mod)) != NULL) { name = (char *)preload_search_info(mod, MODINFO_NAME); if (name == NULL) continue; type = (char *)preload_search_info(mod, MODINFO_TYPE); if (type == NULL) continue; if (strcmp(type, "md_image") && strcmp(type, "mfs_root")) continue; ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); if (ptr != NULL && len != 0) { sx_xlock(&md_sx); md_preloaded(ptr, len, name); sx_xunlock(&md_sx); } } md_pbuf_zone = pbuf_zsecond_create("mdpbuf", nswbuf / 10); status_dev = make_dev(&mdctl_cdevsw, INT_MAX, UID_ROOT, GID_WHEEL, 0600, MDCTL_NAME); g_topology_lock(); } static void g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp) { struct md_s *mp; char *type; mp = gp->softc; if (mp == NULL) return; switch (mp->type) { case MD_MALLOC: type = "malloc"; break; case MD_PRELOAD: type = "preload"; break; case MD_VNODE: type = "vnode"; break; case MD_SWAP: type = "swap"; break; case MD_NULL: type = "null"; break; default: type = "unknown"; break; } if (pp != NULL) { if (indent == NULL) { sbuf_printf(sb, " u %d", mp->unit); sbuf_printf(sb, " s %ju", (uintmax_t) mp->sectorsize); sbuf_printf(sb, " f %ju", (uintmax_t) mp->fwheads); sbuf_printf(sb, " fs %ju", (uintmax_t) mp->fwsectors); sbuf_printf(sb, " l %ju", (uintmax_t) mp->mediasize); sbuf_printf(sb, " t %s", type); if ((mp->type == MD_VNODE && mp->vnode != NULL) || (mp->type == MD_PRELOAD && mp->file[0] != '\0')) sbuf_printf(sb, " file %s", mp->file); sbuf_printf(sb, " label %s", mp->label); } else { sbuf_printf(sb, "%s%d\n", indent, mp->unit); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->sectorsize); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->fwheads); sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->fwsectors); if (mp->ident[0] != '\0') { sbuf_printf(sb, "%s", indent); g_conf_printf_escaped(sb, "%s", mp->ident); sbuf_printf(sb, "\n"); } sbuf_printf(sb, "%s%ju\n", indent, (uintmax_t) mp->mediasize); sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_COMPRESS) == 0 ? "off": "on"); sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_READONLY) == 0 ? "read-write": "read-only"); sbuf_printf(sb, "%s%s\n", indent, type); if ((mp->type == MD_VNODE && mp->vnode != NULL) || (mp->type == MD_PRELOAD && mp->file[0] != '\0')) { sbuf_printf(sb, "%s", indent); g_conf_printf_escaped(sb, "%s", mp->file); sbuf_printf(sb, "\n"); } if (mp->type == MD_VNODE) sbuf_printf(sb, "%s%s\n", indent, (mp->flags & MD_CACHE) == 0 ? "off": "on"); sbuf_printf(sb, "%s\n"); } } } static void g_md_fini(struct g_class *mp __unused) { sx_destroy(&md_sx); if (status_dev != NULL) destroy_dev(status_dev); uma_zdestroy(md_pbuf_zone); delete_unrhdr(md_uh); } Index: head/sys/fs/tmpfs/tmpfs_subr.c =================================================================== --- head/sys/fs/tmpfs/tmpfs_subr.c (revision 356156) +++ head/sys/fs/tmpfs/tmpfs_subr.c (revision 356157) @@ -1,1937 +1,1935 @@ /* $NetBSD: tmpfs_subr.c,v 1.35 2007/07/09 21:10:50 ad Exp $ */ /*- * SPDX-License-Identifier: BSD-2-Clause-NetBSD * * Copyright (c) 2005 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation * by Julio M. Merino Vidal, developed as part of Google's Summer of Code * 2005 program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ /* * Efficient memory file system supporting functions. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include SYSCTL_NODE(_vfs, OID_AUTO, tmpfs, CTLFLAG_RW, 0, "tmpfs file system"); static long tmpfs_pages_reserved = TMPFS_PAGES_MINRESERVED; static uma_zone_t tmpfs_dirent_pool; static uma_zone_t tmpfs_node_pool; static int tmpfs_node_ctor(void *mem, int size, void *arg, int flags) { struct tmpfs_node *node; node = mem; node->tn_gen++; node->tn_size = 0; node->tn_status = 0; node->tn_flags = 0; node->tn_links = 0; node->tn_vnode = NULL; node->tn_vpstate = 0; return (0); } static void tmpfs_node_dtor(void *mem, int size, void *arg) { struct tmpfs_node *node; node = mem; node->tn_type = VNON; } static int tmpfs_node_init(void *mem, int size, int flags) { struct tmpfs_node *node; node = mem; node->tn_id = 0; mtx_init(&node->tn_interlock, "tmpfsni", NULL, MTX_DEF); node->tn_gen = arc4random(); return (0); } static void tmpfs_node_fini(void *mem, int size) { struct tmpfs_node *node; node = mem; mtx_destroy(&node->tn_interlock); } void tmpfs_subr_init(void) { tmpfs_dirent_pool = uma_zcreate("TMPFS dirent", sizeof(struct tmpfs_dirent), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); tmpfs_node_pool = uma_zcreate("TMPFS node", sizeof(struct tmpfs_node), tmpfs_node_ctor, tmpfs_node_dtor, tmpfs_node_init, tmpfs_node_fini, UMA_ALIGN_PTR, 0); } void tmpfs_subr_uninit(void) { uma_zdestroy(tmpfs_node_pool); uma_zdestroy(tmpfs_dirent_pool); } static int sysctl_mem_reserved(SYSCTL_HANDLER_ARGS) { int error; long pages, bytes; pages = *(long *)arg1; bytes = pages * PAGE_SIZE; error = sysctl_handle_long(oidp, &bytes, 0, req); if (error || !req->newptr) return (error); pages = bytes / PAGE_SIZE; if (pages < TMPFS_PAGES_MINRESERVED) return (EINVAL); *(long *)arg1 = pages; return (0); } SYSCTL_PROC(_vfs_tmpfs, OID_AUTO, memory_reserved, CTLTYPE_LONG|CTLFLAG_RW, &tmpfs_pages_reserved, 0, sysctl_mem_reserved, "L", "Amount of available memory and swap below which tmpfs growth stops"); static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b); RB_PROTOTYPE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); size_t tmpfs_mem_avail(void) { vm_ooffset_t avail; avail = swap_pager_avail + vm_free_count() - tmpfs_pages_reserved; if (__predict_false(avail < 0)) avail = 0; return (avail); } size_t tmpfs_pages_used(struct tmpfs_mount *tmp) { const size_t node_size = sizeof(struct tmpfs_node) + sizeof(struct tmpfs_dirent); size_t meta_pages; meta_pages = howmany((uintmax_t)tmp->tm_nodes_inuse * node_size, PAGE_SIZE); return (meta_pages + tmp->tm_pages_used); } static size_t tmpfs_pages_check_avail(struct tmpfs_mount *tmp, size_t req_pages) { if (tmpfs_mem_avail() < req_pages) return (0); if (tmp->tm_pages_max != ULONG_MAX && tmp->tm_pages_max < req_pages + tmpfs_pages_used(tmp)) return (0); return (1); } void tmpfs_ref_node(struct tmpfs_node *node) { TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); TMPFS_NODE_UNLOCK(node); } void tmpfs_ref_node_locked(struct tmpfs_node *node) { TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p zero refcount", node)); KASSERT(node->tn_refcount < UINT_MAX, ("node %p refcount %u", node, node->tn_refcount)); node->tn_refcount++; } /* * Allocates a new node of type 'type' inside the 'tmp' mount point, with * its owner set to 'uid', its group to 'gid' and its mode set to 'mode', * using the credentials of the process 'p'. * * If the node type is set to 'VDIR', then the parent parameter must point * to the parent directory of the node being created. It may only be NULL * while allocating the root node. * * If the node type is set to 'VBLK' or 'VCHR', then the rdev parameter * specifies the device the node represents. * * If the node type is set to 'VLNK', then the parameter target specifies * the file name of the target file for the symbolic link that is being * created. * * Note that new nodes are retrieved from the available list if it has * items or, if it is empty, from the node pool as long as there is enough * space to create them. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_node(struct mount *mp, struct tmpfs_mount *tmp, enum vtype type, uid_t uid, gid_t gid, mode_t mode, struct tmpfs_node *parent, const char *target, dev_t rdev, struct tmpfs_node **node) { struct tmpfs_node *nnode; vm_object_t obj; /* If the root directory of the 'tmp' file system is not yet * allocated, this must be the request to do it. */ MPASS(IMPLIES(tmp->tm_root == NULL, parent == NULL && type == VDIR)); MPASS(IFF(type == VLNK, target != NULL)); MPASS(IFF(type == VBLK || type == VCHR, rdev != VNOVAL)); if (tmp->tm_nodes_inuse >= tmp->tm_nodes_max) return (ENOSPC); if (tmpfs_pages_check_avail(tmp, 1) == 0) return (ENOSPC); if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { /* * When a new tmpfs node is created for fully * constructed mount point, there must be a parent * node, which vnode is locked exclusively. As * consequence, if the unmount is executing in * parallel, vflush() cannot reclaim the parent vnode. * Due to this, the check for MNTK_UNMOUNT flag is not * racy: if we did not see MNTK_UNMOUNT flag, then tmp * cannot be destroyed until node construction is * finished and the parent vnode unlocked. * * Tmpfs does not need to instantiate new nodes during * unmount. */ return (EBUSY); } if ((mp->mnt_kern_flag & MNT_RDONLY) != 0) return (EROFS); nnode = uma_zalloc_arg(tmpfs_node_pool, tmp, M_WAITOK); /* Generic initialization. */ nnode->tn_type = type; vfs_timestamp(&nnode->tn_atime); nnode->tn_birthtime = nnode->tn_ctime = nnode->tn_mtime = nnode->tn_atime; nnode->tn_uid = uid; nnode->tn_gid = gid; nnode->tn_mode = mode; nnode->tn_id = alloc_unr64(&tmp->tm_ino_unr); nnode->tn_refcount = 1; /* Type-specific initialization. */ switch (nnode->tn_type) { case VBLK: case VCHR: nnode->tn_rdev = rdev; break; case VDIR: RB_INIT(&nnode->tn_dir.tn_dirhead); LIST_INIT(&nnode->tn_dir.tn_dupindex); MPASS(parent != nnode); MPASS(IMPLIES(parent == NULL, tmp->tm_root == NULL)); nnode->tn_dir.tn_parent = (parent == NULL) ? nnode : parent; nnode->tn_dir.tn_readdir_lastn = 0; nnode->tn_dir.tn_readdir_lastp = NULL; nnode->tn_links++; TMPFS_NODE_LOCK(nnode->tn_dir.tn_parent); nnode->tn_dir.tn_parent->tn_links++; TMPFS_NODE_UNLOCK(nnode->tn_dir.tn_parent); break; case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: MPASS(strlen(target) < MAXPATHLEN); nnode->tn_size = strlen(target); nnode->tn_link = malloc(nnode->tn_size, M_TMPFSNAME, M_WAITOK); memcpy(nnode->tn_link, target, nnode->tn_size); break; case VREG: obj = nnode->tn_reg.tn_aobj = vm_pager_allocate(OBJT_SWAP, NULL, 0, VM_PROT_DEFAULT, 0, NULL /* XXXKIB - tmpfs needs swap reservation */); VM_OBJECT_WLOCK(obj); /* OBJ_TMPFS is set together with the setting of vp->v_object */ vm_object_set_flag(obj, OBJ_TMPFS_NODE); VM_OBJECT_WUNLOCK(obj); break; default: panic("tmpfs_alloc_node: type %p %d", nnode, (int)nnode->tn_type); } TMPFS_LOCK(tmp); LIST_INSERT_HEAD(&tmp->tm_nodes_used, nnode, tn_entries); nnode->tn_attached = true; tmp->tm_nodes_inuse++; tmp->tm_refcount++; TMPFS_UNLOCK(tmp); *node = nnode; return (0); } /* * Destroys the node pointed to by node from the file system 'tmp'. * If the node references a directory, no entries are allowed. */ void tmpfs_free_node(struct tmpfs_mount *tmp, struct tmpfs_node *node) { TMPFS_LOCK(tmp); TMPFS_NODE_LOCK(node); if (!tmpfs_free_node_locked(tmp, node, false)) { TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); } } bool tmpfs_free_node_locked(struct tmpfs_mount *tmp, struct tmpfs_node *node, bool detach) { vm_object_t uobj; TMPFS_MP_ASSERT_LOCKED(tmp); TMPFS_NODE_ASSERT_LOCKED(node); KASSERT(node->tn_refcount > 0, ("node %p refcount zero", node)); node->tn_refcount--; if (node->tn_attached && (detach || node->tn_refcount == 0)) { MPASS(tmp->tm_nodes_inuse > 0); tmp->tm_nodes_inuse--; LIST_REMOVE(node, tn_entries); node->tn_attached = false; } if (node->tn_refcount > 0) return (false); #ifdef INVARIANTS MPASS(node->tn_vnode == NULL); MPASS((node->tn_vpstate & TMPFS_VNODE_ALLOCATING) == 0); #endif TMPFS_NODE_UNLOCK(node); TMPFS_UNLOCK(tmp); switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VDIR: /* FALLTHROUGH */ case VFIFO: /* FALLTHROUGH */ case VSOCK: break; case VLNK: free(node->tn_link, M_TMPFSNAME); break; case VREG: uobj = node->tn_reg.tn_aobj; if (uobj != NULL) { if (uobj->size != 0) atomic_subtract_long(&tmp->tm_pages_used, uobj->size); KASSERT((uobj->flags & OBJ_TMPFS) == 0, ("leaked OBJ_TMPFS node %p vm_obj %p", node, uobj)); vm_object_deallocate(uobj); } break; default: panic("tmpfs_free_node: type %p %d", node, (int)node->tn_type); } uma_zfree(tmpfs_node_pool, node); TMPFS_LOCK(tmp); tmpfs_free_tmp(tmp); return (true); } static __inline uint32_t tmpfs_dirent_hash(const char *name, u_int len) { uint32_t hash; hash = fnv_32_buf(name, len, FNV1_32_INIT + len) & TMPFS_DIRCOOKIE_MASK; #ifdef TMPFS_DEBUG_DIRCOOKIE_DUP hash &= 0xf; #endif if (hash < TMPFS_DIRCOOKIE_MIN) hash += TMPFS_DIRCOOKIE_MIN; return (hash); } static __inline off_t tmpfs_dirent_cookie(struct tmpfs_dirent *de) { if (de == NULL) return (TMPFS_DIRCOOKIE_EOF); MPASS(de->td_cookie >= TMPFS_DIRCOOKIE_MIN); return (de->td_cookie); } static __inline boolean_t tmpfs_dirent_dup(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUP) != 0); } static __inline boolean_t tmpfs_dirent_duphead(struct tmpfs_dirent *de) { return ((de->td_cookie & TMPFS_DIRCOOKIE_DUPHEAD) != 0); } void tmpfs_dirent_init(struct tmpfs_dirent *de, const char *name, u_int namelen) { de->td_hash = de->td_cookie = tmpfs_dirent_hash(name, namelen); memcpy(de->ud.td_name, name, namelen); de->td_namelen = namelen; } /* * Allocates a new directory entry for the node node with a name of name. * The new directory entry is returned in *de. * * The link count of node is increased by one to reflect the new object * referencing it. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_dirent(struct tmpfs_mount *tmp, struct tmpfs_node *node, const char *name, u_int len, struct tmpfs_dirent **de) { struct tmpfs_dirent *nde; nde = uma_zalloc(tmpfs_dirent_pool, M_WAITOK); nde->td_node = node; if (name != NULL) { nde->ud.td_name = malloc(len, M_TMPFSNAME, M_WAITOK); tmpfs_dirent_init(nde, name, len); } else nde->td_namelen = 0; if (node != NULL) node->tn_links++; *de = nde; return 0; } /* * Frees a directory entry. It is the caller's responsibility to destroy * the node referenced by it if needed. * * The link count of node is decreased by one to reflect the removal of an * object that referenced it. This only happens if 'node_exists' is true; * otherwise the function will not access the node referred to by the * directory entry, as it may already have been released from the outside. */ void tmpfs_free_dirent(struct tmpfs_mount *tmp, struct tmpfs_dirent *de) { struct tmpfs_node *node; node = de->td_node; if (node != NULL) { MPASS(node->tn_links > 0); node->tn_links--; } if (!tmpfs_dirent_duphead(de) && de->ud.td_name != NULL) free(de->ud.td_name, M_TMPFSNAME); uma_zfree(tmpfs_dirent_pool, de); } void tmpfs_destroy_vobject(struct vnode *vp, vm_object_t obj) { ASSERT_VOP_ELOCKED(vp, "tmpfs_destroy_vobject"); if (vp->v_type != VREG || obj == NULL) return; VM_OBJECT_WLOCK(obj); VI_LOCK(vp); vm_object_clear_flag(obj, OBJ_TMPFS); obj->un_pager.swp.swp_tmpfs = NULL; if (vp->v_writecount < 0) vp->v_writecount = 0; VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(obj); } /* * Need to clear v_object for insmntque failure. */ static void tmpfs_insmntque_dtr(struct vnode *vp, void *dtr_arg) { tmpfs_destroy_vobject(vp, vp->v_object); vp->v_object = NULL; vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } /* * Allocates a new vnode for the node node or returns a new reference to * an existing one if the node had already a vnode referencing it. The * resulting locked vnode is returned in *vpp. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_alloc_vp(struct mount *mp, struct tmpfs_node *node, int lkflag, struct vnode **vpp) { struct vnode *vp; struct tmpfs_mount *tm; vm_object_t object; int error; error = 0; tm = VFS_TO_TMPFS(mp); TMPFS_NODE_LOCK(node); tmpfs_ref_node_locked(node); loop: TMPFS_NODE_ASSERT_LOCKED(node); if ((vp = node->tn_vnode) != NULL) { MPASS((node->tn_vpstate & TMPFS_VNODE_DOOMED) == 0); VI_LOCK(vp); if ((node->tn_type == VDIR && node->tn_dir.tn_parent == NULL) || (VN_IS_DOOMED(vp) && (lkflag & LK_NOWAIT) != 0)) { VI_UNLOCK(vp); TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } if (VN_IS_DOOMED(vp)) { VI_UNLOCK(vp); node->tn_vpstate |= TMPFS_VNODE_WRECLAIM; while ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) { msleep(&node->tn_vnode, TMPFS_NODE_MTX(node), 0, "tmpfsE", 0); } goto loop; } TMPFS_NODE_UNLOCK(node); error = vget(vp, lkflag | LK_INTERLOCK, curthread); if (error == ENOENT) { TMPFS_NODE_LOCK(node); goto loop; } if (error != 0) { vp = NULL; goto out; } /* * Make sure the vnode is still there after * getting the interlock to avoid racing a free. */ if (node->tn_vnode == NULL || node->tn_vnode != vp) { vput(vp); TMPFS_NODE_LOCK(node); goto loop; } goto out; } if ((node->tn_vpstate & TMPFS_VNODE_DOOMED) || (node->tn_type == VDIR && node->tn_dir.tn_parent == NULL)) { TMPFS_NODE_UNLOCK(node); error = ENOENT; vp = NULL; goto out; } /* * otherwise lock the vp list while we call getnewvnode * since that can block. */ if (node->tn_vpstate & TMPFS_VNODE_ALLOCATING) { node->tn_vpstate |= TMPFS_VNODE_WANT; error = msleep((caddr_t) &node->tn_vpstate, TMPFS_NODE_MTX(node), 0, "tmpfs_alloc_vp", 0); if (error != 0) goto out; goto loop; } else node->tn_vpstate |= TMPFS_VNODE_ALLOCATING; TMPFS_NODE_UNLOCK(node); /* Get a new vnode and associate it with our node. */ error = getnewvnode("tmpfs", mp, VFS_TO_TMPFS(mp)->tm_nonc ? &tmpfs_vnodeop_nonc_entries : &tmpfs_vnodeop_entries, &vp); if (error != 0) goto unlock; MPASS(vp != NULL); /* lkflag is ignored, the lock is exclusive */ (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vp->v_data = node; vp->v_type = node->tn_type; /* Type-specific initialization. */ switch (node->tn_type) { case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VLNK: /* FALLTHROUGH */ case VSOCK: break; case VFIFO: vp->v_op = &tmpfs_fifoop_entries; break; case VREG: object = node->tn_reg.tn_aobj; VM_OBJECT_WLOCK(object); VI_LOCK(vp); KASSERT(vp->v_object == NULL, ("Not NULL v_object in tmpfs")); vp->v_object = object; object->un_pager.swp.swp_tmpfs = vp; vm_object_set_flag(object, OBJ_TMPFS); VI_UNLOCK(vp); VM_OBJECT_WUNLOCK(object); break; case VDIR: MPASS(node->tn_dir.tn_parent != NULL); if (node->tn_dir.tn_parent == node) vp->v_vflag |= VV_ROOT; break; default: panic("tmpfs_alloc_vp: type %p %d", node, (int)node->tn_type); } if (vp->v_type != VFIFO) VN_LOCK_ASHARE(vp); error = insmntque1(vp, mp, tmpfs_insmntque_dtr, NULL); if (error != 0) vp = NULL; unlock: TMPFS_NODE_LOCK(node); MPASS(node->tn_vpstate & TMPFS_VNODE_ALLOCATING); node->tn_vpstate &= ~TMPFS_VNODE_ALLOCATING; node->tn_vnode = vp; if (node->tn_vpstate & TMPFS_VNODE_WANT) { node->tn_vpstate &= ~TMPFS_VNODE_WANT; TMPFS_NODE_UNLOCK(node); wakeup((caddr_t) &node->tn_vpstate); } else TMPFS_NODE_UNLOCK(node); out: if (error == 0) { *vpp = vp; #ifdef INVARIANTS MPASS(*vpp != NULL && VOP_ISLOCKED(*vpp)); TMPFS_NODE_LOCK(node); MPASS(*vpp == node->tn_vnode); TMPFS_NODE_UNLOCK(node); #endif } tmpfs_free_node(tm, node); return (error); } /* * Destroys the association between the vnode vp and the node it * references. */ void tmpfs_free_vp(struct vnode *vp) { struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); TMPFS_NODE_ASSERT_LOCKED(node); node->tn_vnode = NULL; if ((node->tn_vpstate & TMPFS_VNODE_WRECLAIM) != 0) wakeup(&node->tn_vnode); node->tn_vpstate &= ~TMPFS_VNODE_WRECLAIM; vp->v_data = NULL; } /* * Allocates a new file of type 'type' and adds it to the parent directory * 'dvp'; this addition is done using the component name given in 'cnp'. * The ownership of the new file is automatically assigned based on the * credentials of the caller (through 'cnp'), the group is set based on * the parent directory and the mode is determined from the 'vap' argument. * If successful, *vpp holds a vnode to the newly created file and zero * is returned. Otherwise *vpp is NULL and the function returns an * appropriate error code. */ int tmpfs_alloc_file(struct vnode *dvp, struct vnode **vpp, struct vattr *vap, struct componentname *cnp, const char *target) { int error; struct tmpfs_dirent *de; struct tmpfs_mount *tmp; struct tmpfs_node *dnode; struct tmpfs_node *node; struct tmpfs_node *parent; ASSERT_VOP_ELOCKED(dvp, "tmpfs_alloc_file"); MPASS(cnp->cn_flags & HASBUF); tmp = VFS_TO_TMPFS(dvp->v_mount); dnode = VP_TO_TMPFS_DIR(dvp); *vpp = NULL; /* If the entry we are creating is a directory, we cannot overflow * the number of links of its parent, because it will get a new * link. */ if (vap->va_type == VDIR) { /* Ensure that we do not overflow the maximum number of links * imposed by the system. */ MPASS(dnode->tn_links <= TMPFS_LINK_MAX); if (dnode->tn_links == TMPFS_LINK_MAX) { return (EMLINK); } parent = dnode; MPASS(parent != NULL); } else parent = NULL; /* Allocate a node that represents the new file. */ error = tmpfs_alloc_node(dvp->v_mount, tmp, vap->va_type, cnp->cn_cred->cr_uid, dnode->tn_gid, vap->va_mode, parent, target, vap->va_rdev, &node); if (error != 0) return (error); /* Allocate a directory entry that points to the new file. */ error = tmpfs_alloc_dirent(tmp, node, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) { tmpfs_free_node(tmp, node); return (error); } /* Allocate a vnode for the new file. */ error = tmpfs_alloc_vp(dvp->v_mount, node, LK_EXCLUSIVE, vpp); if (error != 0) { tmpfs_free_dirent(tmp, de); tmpfs_free_node(tmp, node); return (error); } /* Now that all required items are allocated, we can proceed to * insert the new node into the directory, an operation that * cannot fail. */ if (cnp->cn_flags & ISWHITEOUT) tmpfs_dir_whiteout_remove(dvp, cnp); tmpfs_dir_attach(dvp, de); return (0); } struct tmpfs_dirent * tmpfs_dir_first(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; de = RB_MIN(tmpfs_dir, &dnode->tn_dir.tn_dirhead); dc->tdc_tree = de; if (de != NULL && tmpfs_dirent_duphead(de)) de = LIST_FIRST(&de->ud.td_duphead); dc->tdc_current = de; return (dc->tdc_current); } struct tmpfs_dirent * tmpfs_dir_next(struct tmpfs_node *dnode, struct tmpfs_dir_cursor *dc) { struct tmpfs_dirent *de; MPASS(dc->tdc_tree != NULL); if (tmpfs_dirent_dup(dc->tdc_current)) { dc->tdc_current = LIST_NEXT(dc->tdc_current, uh.td_dup.entries); if (dc->tdc_current != NULL) return (dc->tdc_current); } dc->tdc_tree = dc->tdc_current = RB_NEXT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, dc->tdc_tree); if ((de = dc->tdc_current) != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); } /* Lookup directory entry in RB-Tree. Function may return duphead entry. */ static struct tmpfs_dirent * tmpfs_dir_xlookup_hash(struct tmpfs_node *dnode, uint32_t hash) { struct tmpfs_dirent *de, dekey; dekey.td_hash = hash; de = RB_FIND(tmpfs_dir, &dnode->tn_dir.tn_dirhead, &dekey); return (de); } /* Lookup directory entry by cookie, initialize directory cursor accordingly. */ static struct tmpfs_dirent * tmpfs_dir_lookup_cookie(struct tmpfs_node *node, off_t cookie, struct tmpfs_dir_cursor *dc) { struct tmpfs_dir *dirhead = &node->tn_dir.tn_dirhead; struct tmpfs_dirent *de, dekey; MPASS(cookie >= TMPFS_DIRCOOKIE_MIN); if (cookie == node->tn_dir.tn_readdir_lastn && (de = node->tn_dir.tn_readdir_lastp) != NULL) { /* Protect against possible race, tn_readdir_last[pn] * may be updated with only shared vnode lock held. */ if (cookie == tmpfs_dirent_cookie(de)) goto out; } if ((cookie & TMPFS_DIRCOOKIE_DUP) != 0) { LIST_FOREACH(de, &node->tn_dir.tn_dupindex, uh.td_dup.index_entries) { MPASS(tmpfs_dirent_dup(de)); if (de->td_cookie == cookie) goto out; /* dupindex list is sorted. */ if (de->td_cookie < cookie) { de = NULL; goto out; } } MPASS(de == NULL); goto out; } if ((cookie & TMPFS_DIRCOOKIE_MASK) != cookie) { de = NULL; } else { dekey.td_hash = cookie; /* Recover if direntry for cookie was removed */ de = RB_NFIND(tmpfs_dir, dirhead, &dekey); } dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_duphead(de)) { dc->tdc_current = LIST_FIRST(&de->ud.td_duphead); MPASS(dc->tdc_current != NULL); } return (dc->tdc_current); out: dc->tdc_tree = de; dc->tdc_current = de; if (de != NULL && tmpfs_dirent_dup(de)) dc->tdc_tree = tmpfs_dir_xlookup_hash(node, de->td_hash); return (dc->tdc_current); } /* * Looks for a directory entry in the directory represented by node. * 'cnp' describes the name of the entry to look for. Note that the . * and .. components are not allowed as they do not physically exist * within directories. * * Returns a pointer to the entry when found, otherwise NULL. */ struct tmpfs_dirent * tmpfs_dir_lookup(struct tmpfs_node *node, struct tmpfs_node *f, struct componentname *cnp) { struct tmpfs_dir_duphead *duphead; struct tmpfs_dirent *de; uint32_t hash; MPASS(IMPLIES(cnp->cn_namelen == 1, cnp->cn_nameptr[0] != '.')); MPASS(IMPLIES(cnp->cn_namelen == 2, !(cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.'))); TMPFS_VALIDATE_DIR(node); hash = tmpfs_dirent_hash(cnp->cn_nameptr, cnp->cn_namelen); de = tmpfs_dir_xlookup_hash(node, hash); if (de != NULL && tmpfs_dirent_duphead(de)) { duphead = &de->ud.td_duphead; LIST_FOREACH(de, duphead, uh.td_dup.entries) { if (TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) break; } } else if (de != NULL) { if (!TMPFS_DIRENT_MATCHES(de, cnp->cn_nameptr, cnp->cn_namelen)) de = NULL; } if (de != NULL && f != NULL && de->td_node != f) de = NULL; return (de); } /* * Attach duplicate-cookie directory entry nde to dnode and insert to dupindex * list, allocate new cookie value. */ static void tmpfs_dir_attach_dup(struct tmpfs_node *dnode, struct tmpfs_dir_duphead *duphead, struct tmpfs_dirent *nde) { struct tmpfs_dir_duphead *dupindex; struct tmpfs_dirent *de, *pde; dupindex = &dnode->tn_dir.tn_dupindex; de = LIST_FIRST(dupindex); if (de == NULL || de->td_cookie < TMPFS_DIRCOOKIE_DUP_MAX) { if (de == NULL) nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; else nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } /* * Cookie numbers are near exhaustion. Scan dupindex list for unused * numbers. dupindex list is sorted in descending order. Keep it so * after inserting nde. */ while (1) { pde = de; de = LIST_NEXT(de, uh.td_dup.index_entries); if (de == NULL && pde->td_cookie != TMPFS_DIRCOOKIE_DUP_MIN) { /* * Last element of the index doesn't have minimal cookie * value, use it. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MIN; LIST_INSERT_AFTER(pde, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } else if (de == NULL) { /* * We are so lucky have 2^30 hash duplicates in single * directory :) Return largest possible cookie value. * It should be fine except possible issues with * VOP_READDIR restart. */ nde->td_cookie = TMPFS_DIRCOOKIE_DUP_MAX; LIST_INSERT_HEAD(dupindex, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } if (de->td_cookie + 1 == pde->td_cookie || de->td_cookie >= TMPFS_DIRCOOKIE_DUP_MAX) continue; /* No hole or invalid cookie. */ nde->td_cookie = de->td_cookie + 1; MPASS(tmpfs_dirent_dup(nde)); MPASS(pde->td_cookie > nde->td_cookie); MPASS(nde->td_cookie > de->td_cookie); LIST_INSERT_BEFORE(de, nde, uh.td_dup.index_entries); LIST_INSERT_HEAD(duphead, nde, uh.td_dup.entries); return; } } /* * Attaches the directory entry de to the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_alloc_dirent. */ void tmpfs_dir_attach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_node *dnode; struct tmpfs_dirent *xde, *nde; ASSERT_VOP_ELOCKED(vp, __func__); MPASS(de->td_namelen > 0); MPASS(de->td_hash >= TMPFS_DIRCOOKIE_MIN); MPASS(de->td_cookie == de->td_hash); dnode = VP_TO_TMPFS_DIR(vp); dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; MPASS(!tmpfs_dirent_dup(de)); xde = RB_INSERT(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); if (xde != NULL && tmpfs_dirent_duphead(xde)) tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); else if (xde != NULL) { /* * Allocate new duphead. Swap xde with duphead to avoid * adding/removing elements with the same hash. */ MPASS(!tmpfs_dirent_dup(xde)); tmpfs_alloc_dirent(VFS_TO_TMPFS(vp->v_mount), NULL, NULL, 0, &nde); /* *nde = *xde; XXX gcc 4.2.1 may generate invalid code. */ memcpy(nde, xde, sizeof(*xde)); xde->td_cookie |= TMPFS_DIRCOOKIE_DUPHEAD; LIST_INIT(&xde->ud.td_duphead); xde->td_namelen = 0; xde->td_node = NULL; tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, nde); tmpfs_dir_attach_dup(dnode, &xde->ud.td_duphead, de); } dnode->tn_size += sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } /* * Detaches the directory entry de from the directory represented by vp. * Note that this does not change the link count of the node pointed by * the directory entry, as this is done by tmpfs_free_dirent. */ void tmpfs_dir_detach(struct vnode *vp, struct tmpfs_dirent *de) { struct tmpfs_mount *tmp; struct tmpfs_dir *head; struct tmpfs_node *dnode; struct tmpfs_dirent *xde; ASSERT_VOP_ELOCKED(vp, __func__); dnode = VP_TO_TMPFS_DIR(vp); head = &dnode->tn_dir.tn_dirhead; dnode->tn_dir.tn_readdir_lastn = 0; dnode->tn_dir.tn_readdir_lastp = NULL; if (tmpfs_dirent_dup(de)) { /* Remove duphead if de was last entry. */ if (LIST_NEXT(de, uh.td_dup.entries) == NULL) { xde = tmpfs_dir_xlookup_hash(dnode, de->td_hash); MPASS(tmpfs_dirent_duphead(xde)); } else xde = NULL; LIST_REMOVE(de, uh.td_dup.entries); LIST_REMOVE(de, uh.td_dup.index_entries); if (xde != NULL) { if (LIST_EMPTY(&xde->ud.td_duphead)) { RB_REMOVE(tmpfs_dir, head, xde); tmp = VFS_TO_TMPFS(vp->v_mount); MPASS(xde->td_node == NULL); tmpfs_free_dirent(tmp, xde); } } de->td_cookie = de->td_hash; } else RB_REMOVE(tmpfs_dir, head, de); dnode->tn_size -= sizeof(struct tmpfs_dirent); dnode->tn_status |= TMPFS_NODE_ACCESSED | TMPFS_NODE_CHANGED | \ TMPFS_NODE_MODIFIED; tmpfs_update(vp); } void tmpfs_dir_destroy(struct tmpfs_mount *tmp, struct tmpfs_node *dnode) { struct tmpfs_dirent *de, *dde, *nde; RB_FOREACH_SAFE(de, tmpfs_dir, &dnode->tn_dir.tn_dirhead, nde) { RB_REMOVE(tmpfs_dir, &dnode->tn_dir.tn_dirhead, de); /* Node may already be destroyed. */ de->td_node = NULL; if (tmpfs_dirent_duphead(de)) { while ((dde = LIST_FIRST(&de->ud.td_duphead)) != NULL) { LIST_REMOVE(dde, uh.td_dup.entries); dde->td_node = NULL; tmpfs_free_dirent(tmp, dde); } } tmpfs_free_dirent(tmp, de); } } /* * Helper function for tmpfs_readdir. Creates a '.' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { int error; struct dirent dent; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOT); dent.d_fileno = node->tn_id; dent.d_type = DT_DIR; dent.d_namlen = 1; dent.d_name[0] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Creates a '..' entry for the given * directory and returns it in the uio space. The function returns 0 * on success, -1 if there was not enough space in the uio structure to * hold the directory entry or an appropriate error code if another * error happens. */ static int tmpfs_dir_getdotdotdent(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio) { struct tmpfs_node *parent; struct dirent dent; int error; TMPFS_VALIDATE_DIR(node); MPASS(uio->uio_offset == TMPFS_DIRCOOKIE_DOTDOT); /* * Return ENOENT if the current node is already removed. */ TMPFS_ASSERT_LOCKED(node); parent = node->tn_dir.tn_parent; if (parent == NULL) return (ENOENT); TMPFS_NODE_LOCK(parent); dent.d_fileno = parent->tn_id; TMPFS_NODE_UNLOCK(parent); dent.d_type = DT_DIR; dent.d_namlen = 2; dent.d_name[0] = '.'; dent.d_name[1] = '.'; dent.d_reclen = GENERIC_DIRSIZ(&dent); dirent_terminate(&dent); if (dent.d_reclen > uio->uio_resid) error = EJUSTRETURN; else error = uiomove(&dent, dent.d_reclen, uio); tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return (error); } /* * Helper function for tmpfs_readdir. Returns as much directory entries * as can fit in the uio space. The read starts at uio->uio_offset. * The function returns 0 on success, -1 if there was not enough space * in the uio structure to hold the directory entry or an appropriate * error code if another error happens. */ int tmpfs_dir_getdents(struct tmpfs_mount *tm, struct tmpfs_node *node, struct uio *uio, int maxcookies, u_long *cookies, int *ncookies) { struct tmpfs_dir_cursor dc; struct tmpfs_dirent *de; off_t off; int error; TMPFS_VALIDATE_DIR(node); off = 0; /* * Lookup the node from the current offset. The starting offset of * 0 will lookup both '.' and '..', and then the first real entry, * or EOF if there are none. Then find all entries for the dir that * fit into the buffer. Once no more entries are found (de == NULL), * the offset is set to TMPFS_DIRCOOKIE_EOF, which will cause the next * call to return 0. */ switch (uio->uio_offset) { case TMPFS_DIRCOOKIE_DOT: error = tmpfs_dir_getdotdent(tm, node, uio); if (error != 0) return (error); uio->uio_offset = TMPFS_DIRCOOKIE_DOTDOT; if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* FALLTHROUGH */ case TMPFS_DIRCOOKIE_DOTDOT: error = tmpfs_dir_getdotdotdent(tm, node, uio); if (error != 0) return (error); de = tmpfs_dir_first(node, &dc); uio->uio_offset = tmpfs_dirent_cookie(de); if (cookies != NULL) cookies[(*ncookies)++] = off = uio->uio_offset; /* EOF. */ if (de == NULL) return (0); break; case TMPFS_DIRCOOKIE_EOF: return (0); default: de = tmpfs_dir_lookup_cookie(node, uio->uio_offset, &dc); if (de == NULL) return (EINVAL); if (cookies != NULL) off = tmpfs_dirent_cookie(de); } /* Read as much entries as possible; i.e., until we reach the end of * the directory or we exhaust uio space. */ do { struct dirent d; /* Create a dirent structure representing the current * tmpfs_node and fill it. */ if (de->td_node == NULL) { d.d_fileno = 1; d.d_type = DT_WHT; } else { d.d_fileno = de->td_node->tn_id; switch (de->td_node->tn_type) { case VBLK: d.d_type = DT_BLK; break; case VCHR: d.d_type = DT_CHR; break; case VDIR: d.d_type = DT_DIR; break; case VFIFO: d.d_type = DT_FIFO; break; case VLNK: d.d_type = DT_LNK; break; case VREG: d.d_type = DT_REG; break; case VSOCK: d.d_type = DT_SOCK; break; default: panic("tmpfs_dir_getdents: type %p %d", de->td_node, (int)de->td_node->tn_type); } } d.d_namlen = de->td_namelen; MPASS(de->td_namelen < sizeof(d.d_name)); (void)memcpy(d.d_name, de->ud.td_name, de->td_namelen); d.d_reclen = GENERIC_DIRSIZ(&d); dirent_terminate(&d); /* Stop reading if the directory entry we are treating is * bigger than the amount of data that can be returned. */ if (d.d_reclen > uio->uio_resid) { error = EJUSTRETURN; break; } /* Copy the new dirent structure into the output buffer and * advance pointers. */ error = uiomove(&d, d.d_reclen, uio); if (error == 0) { de = tmpfs_dir_next(node, &dc); if (cookies != NULL) { off = tmpfs_dirent_cookie(de); MPASS(*ncookies < maxcookies); cookies[(*ncookies)++] = off; } } } while (error == 0 && uio->uio_resid > 0 && de != NULL); /* Skip setting off when using cookies as it is already done above. */ if (cookies == NULL) off = tmpfs_dirent_cookie(de); /* Update the offset and cache. */ uio->uio_offset = off; node->tn_dir.tn_readdir_lastn = off; node->tn_dir.tn_readdir_lastp = de; tmpfs_set_status(tm, node, TMPFS_NODE_ACCESSED); return error; } int tmpfs_dir_whiteout_add(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; int error; error = tmpfs_alloc_dirent(VFS_TO_TMPFS(dvp->v_mount), NULL, cnp->cn_nameptr, cnp->cn_namelen, &de); if (error != 0) return (error); tmpfs_dir_attach(dvp, de); return (0); } void tmpfs_dir_whiteout_remove(struct vnode *dvp, struct componentname *cnp) { struct tmpfs_dirent *de; de = tmpfs_dir_lookup(VP_TO_TMPFS_DIR(dvp), NULL, cnp); MPASS(de != NULL && de->td_node == NULL); tmpfs_dir_detach(dvp, de); tmpfs_free_dirent(VFS_TO_TMPFS(dvp->v_mount), de); } /* * Resizes the aobj associated with the regular file pointed to by 'vp' to the * size 'newsize'. 'vp' must point to a vnode that represents a regular file. * 'newsize' must be positive. * * Returns zero on success or an appropriate error code on failure. */ int tmpfs_reg_resize(struct vnode *vp, off_t newsize, boolean_t ignerr) { struct tmpfs_mount *tmp; struct tmpfs_node *node; vm_object_t uobj; vm_page_t m; vm_pindex_t idx, newpages, oldpages; off_t oldsize; int base, rv; MPASS(vp->v_type == VREG); MPASS(newsize >= 0); node = VP_TO_TMPFS_NODE(vp); uobj = node->tn_reg.tn_aobj; tmp = VFS_TO_TMPFS(vp->v_mount); /* * Convert the old and new sizes to the number of pages needed to * store them. It may happen that we do not need to do anything * because the last allocated page can accommodate the change on * its own. */ oldsize = node->tn_size; oldpages = OFF_TO_IDX(oldsize + PAGE_MASK); MPASS(oldpages == uobj->size); newpages = OFF_TO_IDX(newsize + PAGE_MASK); if (__predict_true(newpages == oldpages && newsize >= oldsize)) { node->tn_size = newsize; return (0); } if (newpages > oldpages && tmpfs_pages_check_avail(tmp, newpages - oldpages) == 0) return (ENOSPC); VM_OBJECT_WLOCK(uobj); if (newsize < oldsize) { /* * Zero the truncated part of the last page. */ base = newsize & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(newsize); retry: m = vm_page_grab(uobj, idx, VM_ALLOC_NOCREAT); if (m != NULL) { MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(uobj, idx, NULL, NULL)) { m = vm_page_alloc(uobj, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(uobj, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ - vm_page_lock(m); vm_page_launder(m); - vm_page_unlock(m); } else { vm_page_free(m); if (ignerr) m = NULL; else { VM_OBJECT_WUNLOCK(uobj); return (EIO); } } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); vm_page_set_dirty(m); vm_page_xunbusy(m); } } /* * Release any swap space and free any whole pages. */ if (newpages < oldpages) { swap_pager_freespace(uobj, newpages, oldpages - newpages); vm_object_page_remove(uobj, newpages, 0, 0); } } uobj->size = newpages; VM_OBJECT_WUNLOCK(uobj); atomic_add_long(&tmp->tm_pages_used, newpages - oldpages); node->tn_size = newsize; return (0); } void tmpfs_check_mtime(struct vnode *vp) { struct tmpfs_node *node; struct vm_object *obj; ASSERT_VOP_ELOCKED(vp, "check_mtime"); if (vp->v_type != VREG) return; obj = vp->v_object; KASSERT((obj->flags & (OBJ_TMPFS_NODE | OBJ_TMPFS)) == (OBJ_TMPFS_NODE | OBJ_TMPFS), ("non-tmpfs obj")); /* unlocked read */ if (obj->generation != obj->cleangeneration) { VM_OBJECT_WLOCK(obj); if (obj->generation != obj->cleangeneration) { obj->cleangeneration = obj->generation; node = VP_TO_TMPFS_NODE(vp); node->tn_status |= TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED; } VM_OBJECT_WUNLOCK(obj); } } /* * Change flags of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chflags(struct vnode *vp, u_long flags, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chflags"); node = VP_TO_TMPFS_NODE(vp); if ((flags & ~(SF_APPEND | SF_ARCHIVED | SF_IMMUTABLE | SF_NOUNLINK | UF_APPEND | UF_ARCHIVE | UF_HIDDEN | UF_IMMUTABLE | UF_NODUMP | UF_NOUNLINK | UF_OFFLINE | UF_OPAQUE | UF_READONLY | UF_REPARSE | UF_SPARSE | UF_SYSTEM)) != 0) return (EOPNOTSUPP); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* * Callers may only modify the file flags on objects they * have VADMIN rights for. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Unprivileged processes are not permitted to unset system * flags, or modify flags if any system flags are set. */ if (!priv_check_cred(cred, PRIV_VFS_SYSFLAGS)) { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND)) { error = securelevel_gt(cred, 0); if (error) return (error); } } else { if (node->tn_flags & (SF_NOUNLINK | SF_IMMUTABLE | SF_APPEND) || ((flags ^ node->tn_flags) & SF_SETTABLE)) return (EPERM); } node->tn_flags = flags; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chflags2"); return (0); } /* * Change access mode on the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chmod(struct vnode *vp, mode_t mode, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chmod"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the permissions on a file, must possess VADMIN * for that file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * Privileged processes may set the sticky bit on non-directories, * as well as set the setgid bit on a file with a group that the * process is not a member of. */ if (vp->v_type != VDIR && (mode & S_ISTXT)) { if (priv_check_cred(cred, PRIV_VFS_STICKYFILE)) return (EFTYPE); } if (!groupmember(node->tn_gid, cred) && (mode & S_ISGID)) { error = priv_check_cred(cred, PRIV_VFS_SETGID); if (error) return (error); } node->tn_mode &= ~ALLPERMS; node->tn_mode |= mode & ALLPERMS; node->tn_status |= TMPFS_NODE_CHANGED; ASSERT_VOP_ELOCKED(vp, "chmod2"); return (0); } /* * Change ownership of the given vnode. At least one of uid or gid must * be different than VNOVAL. If one is set to that value, the attribute * is unchanged. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chown(struct vnode *vp, uid_t uid, gid_t gid, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; uid_t ouid; gid_t ogid; ASSERT_VOP_ELOCKED(vp, "chown"); node = VP_TO_TMPFS_NODE(vp); /* Assign default values if they are unknown. */ MPASS(uid != VNOVAL || gid != VNOVAL); if (uid == VNOVAL) uid = node->tn_uid; if (gid == VNOVAL) gid = node->tn_gid; MPASS(uid != VNOVAL && gid != VNOVAL); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; /* * To modify the ownership of a file, must possess VADMIN for that * file. */ if ((error = VOP_ACCESS(vp, VADMIN, cred, p))) return (error); /* * To change the owner of a file, or change the group of a file to a * group of which we are not a member, the caller must have * privilege. */ if ((uid != node->tn_uid || (gid != node->tn_gid && !groupmember(gid, cred))) && (error = priv_check_cred(cred, PRIV_VFS_CHOWN))) return (error); ogid = node->tn_gid; ouid = node->tn_uid; node->tn_uid = uid; node->tn_gid = gid; node->tn_status |= TMPFS_NODE_CHANGED; if ((node->tn_mode & (S_ISUID | S_ISGID)) && (ouid != uid || ogid != gid)) { if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) node->tn_mode &= ~(S_ISUID | S_ISGID); } ASSERT_VOP_ELOCKED(vp, "chown2"); return (0); } /* * Change size of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chsize(struct vnode *vp, u_quad_t size, struct ucred *cred, struct thread *p) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chsize"); node = VP_TO_TMPFS_NODE(vp); /* Decide whether this is a valid operation based on the file type. */ error = 0; switch (vp->v_type) { case VDIR: return EISDIR; case VREG: if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; break; case VBLK: /* FALLTHROUGH */ case VCHR: /* FALLTHROUGH */ case VFIFO: /* Allow modifications of special files even if in the file * system is mounted read-only (we are not modifying the * files themselves, but the objects they represent). */ return 0; default: /* Anything else is unsupported. */ return EOPNOTSUPP; } /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = tmpfs_truncate(vp, size); /* tmpfs_truncate will raise the NOTE_EXTEND and NOTE_ATTRIB kevents * for us, as will update tn_status; no need to do that here. */ ASSERT_VOP_ELOCKED(vp, "chsize2"); return (error); } /* * Change access and modification times of the given vnode. * Caller should execute tmpfs_update on vp after a successful execution. * The vnode must be locked on entry and remain locked on exit. */ int tmpfs_chtimes(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *l) { int error; struct tmpfs_node *node; ASSERT_VOP_ELOCKED(vp, "chtimes"); node = VP_TO_TMPFS_NODE(vp); /* Disallow this operation if the file system is mounted read-only. */ if (vp->v_mount->mnt_flag & MNT_RDONLY) return EROFS; /* Immutable or append-only files cannot be modified, either. */ if (node->tn_flags & (IMMUTABLE | APPEND)) return EPERM; error = vn_utimes_perm(vp, vap, cred, l); if (error != 0) return (error); if (vap->va_atime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_ACCESSED; if (vap->va_mtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_status |= TMPFS_NODE_MODIFIED; tmpfs_itimes(vp, &vap->va_atime, &vap->va_mtime); if (vap->va_birthtime.tv_sec != VNOVAL) node->tn_birthtime = vap->va_birthtime; ASSERT_VOP_ELOCKED(vp, "chtimes2"); return (0); } void tmpfs_set_status(struct tmpfs_mount *tm, struct tmpfs_node *node, int status) { if ((node->tn_status & status) == status || tm->tm_ronly) return; TMPFS_NODE_LOCK(node); node->tn_status |= status; TMPFS_NODE_UNLOCK(node); } /* Sync timestamps */ void tmpfs_itimes(struct vnode *vp, const struct timespec *acc, const struct timespec *mod) { struct tmpfs_node *node; struct timespec now; ASSERT_VOP_LOCKED(vp, "tmpfs_itimes"); node = VP_TO_TMPFS_NODE(vp); if ((node->tn_status & (TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED)) == 0) return; vfs_timestamp(&now); TMPFS_NODE_LOCK(node); if (node->tn_status & TMPFS_NODE_ACCESSED) { if (acc == NULL) acc = &now; node->tn_atime = *acc; } if (node->tn_status & TMPFS_NODE_MODIFIED) { if (mod == NULL) mod = &now; node->tn_mtime = *mod; } if (node->tn_status & TMPFS_NODE_CHANGED) node->tn_ctime = now; node->tn_status &= ~(TMPFS_NODE_ACCESSED | TMPFS_NODE_MODIFIED | TMPFS_NODE_CHANGED); TMPFS_NODE_UNLOCK(node); /* XXX: FIX? The entropy here is desirable, but the harvesting may be expensive */ random_harvest_queue(node, sizeof(*node), RANDOM_FS_ATIME); } void tmpfs_update(struct vnode *vp) { tmpfs_itimes(vp, NULL, NULL); } int tmpfs_truncate(struct vnode *vp, off_t length) { int error; struct tmpfs_node *node; node = VP_TO_TMPFS_NODE(vp); if (length < 0) { error = EINVAL; goto out; } if (node->tn_size == length) { error = 0; goto out; } if (length > VFS_TO_TMPFS(vp->v_mount)->tm_maxfilesize) return (EFBIG); error = tmpfs_reg_resize(vp, length, FALSE); if (error == 0) node->tn_status |= TMPFS_NODE_CHANGED | TMPFS_NODE_MODIFIED; out: tmpfs_update(vp); return (error); } static __inline int tmpfs_dirtree_cmp(struct tmpfs_dirent *a, struct tmpfs_dirent *b) { if (a->td_hash > b->td_hash) return (1); else if (a->td_hash < b->td_hash) return (-1); return (0); } RB_GENERATE_STATIC(tmpfs_dir, tmpfs_dirent, uh.td_entries, tmpfs_dirtree_cmp); Index: head/sys/kern/uipc_shm.c =================================================================== --- head/sys/kern/uipc_shm.c (revision 356156) +++ head/sys/kern/uipc_shm.c (revision 356157) @@ -1,1505 +1,1503 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Support for shared swap-backed anonymous memory objects via * shm_open(2), shm_rename(2), and shm_unlink(2). * While most of the implementation is here, vm_mmap.c contains * mapping logic changes. * * posixshmcontrol(1) allows users to inspect the state of the memory * objects. Per-uid swap resource limit controls total amount of * memory that user can consume for anonymous objects, including * shared. */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct shm_mapping { char *sm_path; Fnv32_t sm_fnv; struct shmfd *sm_shmfd; LIST_ENTRY(shm_mapping) sm_link; }; static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); static LIST_HEAD(, shm_mapping) *shm_dictionary; static struct sx shm_dict_lock; static struct mtx shm_timestamp_lock; static u_long shm_hash; static struct unrhdr64 shm_ino_unr; static dev_t shm_dev_ino; #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) static void shm_init(void *arg); static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie); static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out); static fo_rdwr_t shm_read; static fo_rdwr_t shm_write; static fo_truncate_t shm_truncate; static fo_ioctl_t shm_ioctl; static fo_stat_t shm_stat; static fo_close_t shm_close; static fo_chmod_t shm_chmod; static fo_chown_t shm_chown; static fo_seek_t shm_seek; static fo_fill_kinfo_t shm_fill_kinfo; static fo_mmap_t shm_mmap; static fo_get_seals_t shm_get_seals; static fo_add_seals_t shm_add_seals; /* File descriptor operations. */ struct fileops shm_ops = { .fo_read = shm_read, .fo_write = shm_write, .fo_truncate = shm_truncate, .fo_ioctl = shm_ioctl, .fo_poll = invfo_poll, .fo_kqfilter = invfo_kqfilter, .fo_stat = shm_stat, .fo_close = shm_close, .fo_chmod = shm_chmod, .fo_chown = shm_chown, .fo_sendfile = vn_sendfile, .fo_seek = shm_seek, .fo_fill_kinfo = shm_fill_kinfo, .fo_mmap = shm_mmap, .fo_get_seals = shm_get_seals, .fo_add_seals = shm_add_seals, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; FEATURE(posix_shm, "POSIX shared memory"); static int uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) { vm_page_t m; vm_pindex_t idx; size_t tlen; int error, offset, rv; idx = OFF_TO_IDX(uio->uio_offset); offset = uio->uio_offset & PAGE_MASK; tlen = MIN(PAGE_SIZE - offset, len); VM_OBJECT_WLOCK(obj); /* * Read I/O without either a corresponding resident page or swap * page: use zero_region. This is intended to avoid instantiating * pages on read from a sparse region. */ if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL && !vm_pager_has_page(obj, idx, NULL, NULL)) { VM_OBJECT_WUNLOCK(obj); return (uiomove(__DECONST(void *, zero_region), tlen, uio)); } /* * Parallel reads of the page content from disk are prevented * by exclusive busy. * * Although the tmpfs vnode lock is held here, it is * nonetheless safe to sleep waiting for a free page. The * pageout daemon does not need to acquire the tmpfs vnode * lock to page out tobj's pages because tobj is a OBJT_SWAP * type object. */ rv = vm_page_grab_valid(&m, obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (rv != VM_PAGER_OK) { VM_OBJECT_WUNLOCK(obj); printf("uiomove_object: vm_obj %p idx %jd pager error %d\n", obj, idx, rv); return (EIO); } VM_OBJECT_WUNLOCK(obj); error = uiomove_fromphys(&m, offset, tlen, uio); if (uio->uio_rw == UIO_WRITE && error == 0) vm_page_set_dirty(m); - vm_page_lock(m); vm_page_activate(m); - vm_page_unlock(m); vm_page_sunbusy(m); return (error); } int uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) { ssize_t resid; size_t len; int error; error = 0; while ((resid = uio->uio_resid) > 0) { if (obj_size <= uio->uio_offset) break; len = MIN(obj_size - uio->uio_offset, resid); if (len == 0) break; error = uiomove_object_page(obj, len, uio); if (error != 0 || resid == uio->uio_resid) break; } return (error); } static int shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct shmfd *shmfd; off_t foffset; int error; shmfd = fp->f_data; foffset = foffset_lock(fp, 0); error = 0; switch (whence) { case L_INCR: if (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset)) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { error = EOVERFLOW; break; } offset += shmfd->shm_size; break; case L_SET: break; default: error = EINVAL; } if (error == 0) { if (offset < 0 || offset > shmfd->shm_size) error = EINVAL; else td->td_uretoff.tdu_off = offset; } foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } static int shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); if ((flags & FOF_OFFSET) == 0) { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset, uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx); } if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) error = EPERM; else error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif return (shm_dotruncate(shmfd, length)); } int shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { switch (com) { case FIONBIO: case FIOASYNC: /* * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, * just like it would on an unlinked regular file */ return (0); default: return (ENOTTY); } } static int shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif /* * Attempt to return sanish values for fstat() on a memory file * descriptor. */ bzero(sb, sizeof(*sb)); sb->st_blksize = PAGE_SIZE; sb->st_size = shmfd->shm_size; sb->st_blocks = howmany(sb->st_size, sb->st_blksize); mtx_lock(&shm_timestamp_lock); sb->st_atim = shmfd->shm_atime; sb->st_ctim = shmfd->shm_ctime; sb->st_mtim = shmfd->shm_mtime; sb->st_birthtim = shmfd->shm_birthtime; sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ sb->st_uid = shmfd->shm_uid; sb->st_gid = shmfd->shm_gid; mtx_unlock(&shm_timestamp_lock); sb->st_dev = shm_dev_ino; sb->st_ino = shmfd->shm_ino; sb->st_nlink = shmfd->shm_object->ref_count; return (0); } static int shm_close(struct file *fp, struct thread *td) { struct shmfd *shmfd; shmfd = fp->f_data; fp->f_data = NULL; shm_drop(shmfd); return (0); } static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { int error; char *path; const char *pr_path; size_t pr_pathlen; path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; /* Construct a full pathname for jailed callers. */ pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(userpath_in, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); if (error != 0) goto out; #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif /* Require paths to start with a '/' character. */ if (path[pr_pathlen] != '/') { error = EINVAL; goto out; } *path_out = path; out: if (error != 0) free(path, M_SHMFD); return (error); } static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_page_t m; vm_pindex_t idx, nobjsize; vm_ooffset_t delta; int base, rv; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); if (length == shmfd->shm_size) return (0); nobjsize = OFF_TO_IDX(length + PAGE_MASK); /* Are we shrinking? If so, trim the end. */ if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); /* * Disallow any requests to shrink the size if this * object is mapped into the kernel. */ if (shmfd->shm_kmappings > 0) return (EBUSY); /* * Zero the truncated part of the last page. */ base = length & PAGE_MASK; if (base != 0) { idx = OFF_TO_IDX(length); retry: m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); if (m != NULL) { MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(object, idx, NULL, NULL)) { m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, * and therefore not recently * accessed, immediately enqueue it * for asynchronous laundering. The * current operation is not regarded * as an access. */ vm_page_launder(m); } else { vm_page_free(m); VM_OBJECT_WUNLOCK(object); return (EIO); } } if (m != NULL) { pmap_zero_page_area(m, base, PAGE_SIZE - base); KASSERT(vm_page_all_valid(m), ("shm_dotruncate: page %p is invalid", m)); vm_page_set_dirty(m); vm_page_xunbusy(m); } } delta = IDX_TO_OFF(object->size - nobjsize); /* Toss in memory pages. */ if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* Toss pages from swap. */ if (object->type == OBJT_SWAP) swap_pager_freespace(object, nobjsize, delta); /* Free the swap accounted for shm */ swap_release_by_cred(delta, object->cred); object->charge -= delta; } else { if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); /* Try to reserve additional swap space. */ delta = IDX_TO_OFF(nobjsize - object->size); if (!swap_reserve_by_cred(delta, object->cred)) return (ENOMEM); object->charge += delta; } shmfd->shm_size = length; mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_ctime); shmfd->shm_mtime = shmfd->shm_ctime; mtx_unlock(&shm_timestamp_lock); object->size = nobjsize; return (0); } int shm_dotruncate(struct shmfd *shmfd, off_t length) { void *rl_cookie; int error; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); VM_OBJECT_WLOCK(shmfd->shm_object); error = shm_dotruncate_locked(shmfd, length, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } /* * shmfd object management including creation and reference counting * routines. */ struct shmfd * shm_alloc(struct ucred *ucred, mode_t mode) { struct shmfd *shmfd; shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); shmfd->shm_size = 0; shmfd->shm_uid = ucred->cr_uid; shmfd->shm_gid = ucred->cr_gid; shmfd->shm_mode = mode; shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL, shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); vfs_timestamp(&shmfd->shm_birthtime); shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = shmfd->shm_birthtime; shmfd->shm_ino = alloc_unr64(&shm_ino_unr); refcount_init(&shmfd->shm_refs, 1); mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); rangelock_init(&shmfd->shm_rl); #ifdef MAC mac_posixshm_init(shmfd); mac_posixshm_create(ucred, shmfd); #endif return (shmfd); } struct shmfd * shm_hold(struct shmfd *shmfd) { refcount_acquire(&shmfd->shm_refs); return (shmfd); } void shm_drop(struct shmfd *shmfd) { if (refcount_release(&shmfd->shm_refs)) { #ifdef MAC mac_posixshm_destroy(shmfd); #endif rangelock_destroy(&shmfd->shm_rl); mtx_destroy(&shmfd->shm_mtx); vm_object_deallocate(shmfd->shm_object); free(shmfd, M_SHMFD); } } /* * Determine if the credentials have sufficient permissions for a * specified combination of FREAD and FWRITE. */ int shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) { accmode_t accmode; int error; accmode = 0; if (flags & FREAD) accmode |= VREAD; if (flags & FWRITE) accmode |= VWRITE; mtx_lock(&shm_timestamp_lock); error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, accmode, ucred, NULL); mtx_unlock(&shm_timestamp_lock); return (error); } /* * Dictionary management. We maintain an in-kernel dictionary to map * paths to shmfd objects. We use the FNV hash on the path to store * the mappings in a hash table. */ static void shm_init(void *arg) { mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); sx_init(&shm_dict_lock, "shm dictionary"); shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); new_unrhdr64(&shm_ino_unr, 1); shm_dev_ino = devfs_alloc_cdp_inode(); KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); } SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); static struct shmfd * shm_lookup(char *path, Fnv32_t fnv) { struct shm_mapping *map; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) return (map->sm_shmfd); } return (NULL); } static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) { struct shm_mapping *map; map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); map->sm_path = path; map->sm_fnv = fnv; map->sm_shmfd = shm_hold(shmfd); shmfd->shm_path = path; LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); } static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) { struct shm_mapping *map; int error; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) { #ifdef MAC error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); if (error) return (error); #endif error = shm_access(map->sm_shmfd, ucred, FREAD | FWRITE); if (error) return (error); map->sm_shmfd->shm_path = NULL; LIST_REMOVE(map, sm_link); shm_drop(map->sm_shmfd); free(map->sm_path, M_SHMFD); free(map, M_SHMFD); return (0); } } return (ENOENT); } int kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode, struct filecaps *fcaps, int initial_seals) { struct filedesc *fdp; struct shmfd *shmfd; struct file *fp; char *path; void *rl_cookie; Fnv32_t fnv; mode_t cmode; int fd, error; #ifdef CAPABILITY_MODE /* * shm_open(2) is only allowed for anonymous objects. */ if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON)) return (ECAPMODE); #endif AUDIT_ARG_FFLAGS(flags); AUDIT_ARG_MODE(mode); if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) return (EINVAL); if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) return (EINVAL); /* * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. * If the decision is made later to allow additional seals, care must be * taken below to ensure that the seals are properly set if the shmfd * already existed -- this currently assumes that only F_SEAL_SEAL can * be set and doesn't take further precautions to ensure the validity of * the seals being added with respect to current mappings. */ if ((initial_seals & ~F_SEAL_SEAL) != 0) return (EINVAL); fdp = td->td_proc->p_fd; cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS; /* * shm_open(2) created shm should always have O_CLOEXEC set, as mandated * by POSIX. We allow it to be unset here so that an in-kernel * interface may be written as a thin layer around shm, optionally not * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally * in sys_shm_open() to keep this implementation compliant. */ error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); if (error) return (error); /* A SHM_ANON path pointer creates an anonymous object. */ if (userpath == SHM_ANON) { /* A read-only anonymous object is pointless. */ if ((flags & O_ACCMODE) == O_RDONLY) { fdclose(td, fp, fd); fdrop(fp, td); return (EINVAL); } shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; } else { error = shm_copyin_path(td, userpath, &path); if (error != 0) { fdclose(td, fp, fd); fdrop(fp, td); return (error); } AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); shmfd = shm_lookup(path, fnv); if (shmfd == NULL) { /* Object does not yet exist, create it if requested. */ if (flags & O_CREAT) { #ifdef MAC error = mac_posixshm_check_create(td->td_ucred, path); if (error == 0) { #endif shmfd = shm_alloc(td->td_ucred, cmode); shmfd->shm_seals = initial_seals; shm_insert(path, fnv, shmfd); #ifdef MAC } #endif } else { free(path, M_SHMFD); error = ENOENT; } } else { rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* * kern_shm_open() likely shouldn't ever error out on * trying to set a seal that already exists, unlike * F_ADD_SEALS. This would break terribly as * shm_open(2) actually sets F_SEAL_SEAL to maintain * historical behavior where the underlying file could * not be sealed. */ initial_seals &= ~shmfd->shm_seals; /* * Object already exists, obtain a new * reference if requested and permitted. */ free(path, M_SHMFD); /* * initial_seals can't set additional seals if we've * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, * then we've already removed that one from * initial_seals. This is currently redundant as we * only allow setting F_SEAL_SEAL at creation time, but * it's cheap to check and decreases the effort required * to allow additional seals. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && initial_seals != 0) error = EPERM; else if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) error = EEXIST; else { #ifdef MAC error = mac_posixshm_check_open(td->td_ucred, shmfd, FFLAGS(flags & O_ACCMODE)); if (error == 0) #endif error = shm_access(shmfd, td->td_ucred, FFLAGS(flags & O_ACCMODE)); } /* * Truncate the file back to zero length if * O_TRUNC was specified and the object was * opened with read/write. */ if (error == 0 && (flags & (O_ACCMODE | O_TRUNC)) == (O_RDWR | O_TRUNC)) { VM_OBJECT_WLOCK(shmfd->shm_object); #ifdef MAC error = mac_posixshm_check_truncate( td->td_ucred, fp->f_cred, shmfd); if (error == 0) #endif error = shm_dotruncate_locked(shmfd, 0, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); } if (error == 0) { /* * Currently we only allow F_SEAL_SEAL to be * set initially. As noted above, this would * need to be reworked should that change. */ shmfd->shm_seals |= initial_seals; shm_hold(shmfd); } rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); } sx_xunlock(&shm_dict_lock); if (error) { fdclose(td, fp, fd); fdrop(fp, td); return (error); } } finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); td->td_retval[0] = fd; fdrop(fp, td); return (0); } /* System calls. */ #ifdef COMPAT_FREEBSD12 int freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) { return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode, NULL, F_SEAL_SEAL)); } #endif int sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) { char *path; Fnv32_t fnv; int error; error = shm_copyin_path(td, uap->path, &path); if (error != 0) return (error); AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); error = shm_remove(path, fnv, td->td_ucred); sx_xunlock(&shm_dict_lock); free(path, M_TEMP); return (error); } int sys_shm_rename(struct thread *td, struct shm_rename_args *uap) { char *path_from = NULL, *path_to = NULL; Fnv32_t fnv_from, fnv_to; struct shmfd *fd_from; struct shmfd *fd_to; int error; int flags; flags = uap->flags; AUDIT_ARG_FFLAGS(flags); /* * Make sure the user passed only valid flags. * If you add a new flag, please add a new term here. */ if ((flags & ~( SHM_RENAME_NOREPLACE | SHM_RENAME_EXCHANGE )) != 0) { error = EINVAL; goto out; } /* * EXCHANGE and NOREPLACE don't quite make sense together. Let's * force the user to choose one or the other. */ if ((flags & SHM_RENAME_NOREPLACE) != 0 && (flags & SHM_RENAME_EXCHANGE) != 0) { error = EINVAL; goto out; } /* Renaming to or from anonymous makes no sense */ if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { error = EINVAL; goto out; } error = shm_copyin_path(td, uap->path_from, &path_from); if (error != 0) goto out; error = shm_copyin_path(td, uap->path_to, &path_to); if (error != 0) goto out; AUDIT_ARG_UPATH1_CANON(path_from); AUDIT_ARG_UPATH2_CANON(path_to); /* Rename with from/to equal is a no-op */ if (strcmp(path_from, path_to) == 0) goto out; fnv_from = fnv_32_str(path_from, FNV1_32_INIT); fnv_to = fnv_32_str(path_to, FNV1_32_INIT); sx_xlock(&shm_dict_lock); fd_from = shm_lookup(path_from, fnv_from); if (fd_from == NULL) { error = ENOENT; goto out_locked; } fd_to = shm_lookup(path_to, fnv_to); if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { error = EEXIST; goto out_locked; } /* * Unconditionally prevents shm_remove from invalidating the 'from' * shm's state. */ shm_hold(fd_from); error = shm_remove(path_from, fnv_from, td->td_ucred); /* * One of my assumptions failed if ENOENT (e.g. locking didn't * protect us) */ KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", path_from)); if (error != 0) { shm_drop(fd_from); goto out_locked; } /* * If we are exchanging, we need to ensure the shm_remove below * doesn't invalidate the dest shm's state. */ if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) shm_hold(fd_to); /* * NOTE: if path_to is not already in the hash, c'est la vie; * it simply means we have nothing already at path_to to unlink. * That is the ENOENT case. * * If we somehow don't have access to unlink this guy, but * did for the shm at path_from, then relink the shm to path_from * and abort with EACCES. * * All other errors: that is weird; let's relink and abort the * operation. */ error = shm_remove(path_to, fnv_to, td->td_ucred); if (error != 0 && error != ENOENT) { shm_insert(path_from, fnv_from, fd_from); shm_drop(fd_from); /* Don't free path_from now, since the hash references it */ path_from = NULL; goto out_locked; } error = 0; shm_insert(path_to, fnv_to, fd_from); /* Don't free path_to now, since the hash references it */ path_to = NULL; /* We kept a ref when we removed, and incremented again in insert */ shm_drop(fd_from); KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_from->shm_refs)); if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { shm_insert(path_from, fnv_from, fd_to); path_from = NULL; shm_drop(fd_to); KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_to->shm_refs)); } out_locked: sx_xunlock(&shm_dict_lock); out: free(path_from, M_SHMFD); free(path_to, M_SHMFD); return (error); } int shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct shmfd *shmfd; vm_prot_t maxprot; int error; bool writecnt; void *rl_cookie; shmfd = fp->f_data; maxprot = VM_PROT_NONE; rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize, &shmfd->shm_mtx); /* FREAD should always be set. */ if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0; if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; goto out; } /* Don't permit shared writable mappings on read-only descriptors. */ if (writecnt && (maxprot & VM_PROT_WRITE) == 0) { error = EACCES; goto out; } maxprot &= cap_maxprot; /* See comment in vn_mmap(). */ if ( #ifdef _LP64 objsize > OFF_MAX || #endif foff < 0 || foff > OFF_MAX - objsize) { error = EINVAL; goto out; } #ifdef MAC error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); if (error != 0) goto out; #endif mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_atime); mtx_unlock(&shm_timestamp_lock); vm_object_reference(shmfd->shm_object); if (writecnt) vm_pager_update_writecount(shmfd->shm_object, 0, objsize); error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, shmfd->shm_object, foff, writecnt, td); if (error != 0) { if (writecnt) vm_pager_release_writecount(shmfd->shm_object, 0, objsize); vm_object_deallocate(shmfd->shm_object); } out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); /* * SUSv4 says that x bits of permission need not be affected. * Be consistent with our shm_open there. */ #ifdef MAC error = mac_posixshm_check_setmode(active_cred, shmfd, mode); if (error != 0) goto out; #endif error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, VADMIN, active_cred, NULL); if (error != 0) goto out; shmfd->shm_mode = mode & ACCESSPERMS; out: mtx_unlock(&shm_timestamp_lock); return (error); } static int shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); #ifdef MAC error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); if (error != 0) goto out; #endif if (uid == (uid_t)-1) uid = shmfd->shm_uid; if (gid == (gid_t)-1) gid = shmfd->shm_gid; if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) goto out; shmfd->shm_uid = uid; shmfd->shm_gid = gid; out: mtx_unlock(&shm_timestamp_lock); return (error); } /* * Helper routines to allow the backing object of a shared memory file * descriptor to be mapped in the kernel. */ int shm_map(struct file *fp, size_t size, off_t offset, void **memp) { struct shmfd *shmfd; vm_offset_t kva, ofs; vm_object_t obj; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); /* * XXXRW: This validation is probably insufficient, and subject to * sign errors. It should be fixed. */ if (offset >= shmfd->shm_size || offset + size > round_page(shmfd->shm_size)) { VM_OBJECT_WUNLOCK(obj); return (EINVAL); } shmfd->shm_kmappings++; vm_object_reference_locked(obj); VM_OBJECT_WUNLOCK(obj); /* Map the object into the kernel_map and wire it. */ kva = vm_map_min(kernel_map); ofs = offset & PAGE_MASK; offset = trunc_page(offset); size = round_page(size + ofs); rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv == KERN_SUCCESS) { rv = vm_map_wire(kernel_map, kva, kva + size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv == KERN_SUCCESS) { *memp = (void *)(kva + ofs); return (0); } vm_map_remove(kernel_map, kva, kva + size); } else vm_object_deallocate(obj); /* On failure, drop our mapping reference. */ VM_OBJECT_WLOCK(obj); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (vm_mmap_to_errno(rv)); } /* * We require the caller to unmap the entire entry. This allows us to * safely decrement shm_kmappings when a mapping is removed. */ int shm_unmap(struct file *fp, void *mem, size_t size) { struct shmfd *shmfd; vm_map_entry_t entry; vm_offset_t kva, ofs; vm_object_t obj; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; vm_map_t map; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; kva = (vm_offset_t)mem; ofs = kva & PAGE_MASK; kva = trunc_page(kva); size = round_page(size + ofs); map = kernel_map; rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, &obj, &pindex, &prot, &wired); if (rv != KERN_SUCCESS) return (EINVAL); if (entry->start != kva || entry->end != kva + size) { vm_map_lookup_done(map, entry); return (EINVAL); } vm_map_lookup_done(map, entry); if (obj != shmfd->shm_object) return (EINVAL); vm_map_remove(map, kva, kva + size); VM_OBJECT_WLOCK(obj); KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (0); } static int shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) { const char *path, *pr_path; size_t pr_pathlen; bool visible; sx_assert(&shm_dict_lock, SA_LOCKED); kif->kf_type = KF_TYPE_SHM; kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; if (shmfd->shm_path != NULL) { if (shmfd->shm_path != NULL) { path = shmfd->shm_path; pr_path = curthread->td_ucred->cr_prison->pr_path; if (strcmp(pr_path, "/") != 0) { /* Return the jail-rooted pathname. */ pr_pathlen = strlen(pr_path); visible = strncmp(path, pr_path, pr_pathlen) == 0 && path[pr_pathlen] == '/'; if (list && !visible) return (EPERM); if (visible) path += pr_pathlen; } strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); } } return (0); } static int shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp __unused) { int res; sx_slock(&shm_dict_lock); res = shm_fill_kinfo_locked(fp->f_data, kif, false); sx_sunlock(&shm_dict_lock); return (res); } static int shm_add_seals(struct file *fp, int seals) { struct shmfd *shmfd; void *rl_cookie; vm_ooffset_t writemappings; int error, nseals; error = 0; shmfd = fp->f_data; rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX, &shmfd->shm_mtx); /* Even already-set seals should result in EPERM. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { error = EPERM; goto out; } nseals = seals & ~shmfd->shm_seals; if ((nseals & F_SEAL_WRITE) != 0) { /* * The rangelock above prevents writable mappings from being * added after we've started applying seals. The RLOCK here * is to avoid torn reads on ILP32 arches as unmapping/reducing * writemappings will be done without a rangelock. */ VM_OBJECT_RLOCK(shmfd->shm_object); writemappings = shmfd->shm_object->un_pager.swp.writemappings; VM_OBJECT_RUNLOCK(shmfd->shm_object); /* kmappings are also writable */ if (writemappings > 0) { error = EBUSY; goto out; } } shmfd->shm_seals |= nseals; out: rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx); return (error); } static int shm_get_seals(struct file *fp, int *seals) { struct shmfd *shmfd; shmfd = fp->f_data; *seals = shmfd->shm_seals; return (0); } static int sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) { struct shm_mapping *shmm; struct sbuf sb; struct kinfo_file kif; u_long i; ssize_t curlen; int error, error2; sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); sbuf_clear_flags(&sb, SBUF_INCLUDENUL); curlen = 0; error = 0; sx_slock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { error = shm_fill_kinfo_locked(shmm->sm_shmfd, &kif, true); if (error == EPERM) continue; if (error != 0) break; pack_kinfo(&kif); if (req->oldptr != NULL && kif.kf_structsize + curlen > req->oldlen) break; error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 0 : ENOMEM; if (error != 0) break; curlen += kif.kf_structsize; } } sx_sunlock(&shm_dict_lock); error2 = sbuf_finish(&sb); sbuf_delete(&sb); return (error != 0 ? error : error2); } SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, NULL, 0, sysctl_posix_shm_list, "", "POSIX SHM list"); int kern_shm_open2(struct thread *td, const char *path, int flags, mode_t mode, int shmflags, const char *name __unused) { int initial_seals; if ((shmflags & ~SHM_ALLOW_SEALING) != 0) return (EINVAL); initial_seals = F_SEAL_SEAL; if ((shmflags & SHM_ALLOW_SEALING) != 0) initial_seals &= ~F_SEAL_SEAL; return (kern_shm_open(td, path, flags, mode, NULL, initial_seals)); } /* * This version of the shm_open() interface leaves CLOEXEC behavior up to the * caller, and libc will enforce it for the traditional shm_open() call. This * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This * interface also includes a 'name' argument that is currently unused, but could * potentially be exported later via some interface for debugging purposes. * From the kernel's perspective, it is optional. Individual consumers like * memfd_create() may require it in order to be compatible with other systems * implementing the same function. */ int sys_shm_open2(struct thread *td, struct shm_open2_args *uap) { return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, uap->shmflags, uap->name)); } Index: head/sys/vm/swap_pager.c =================================================================== --- head/sys/vm/swap_pager.c (revision 356156) +++ head/sys/vm/swap_pager.c (revision 356157) @@ -1,3111 +1,3104 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1998 Matthew Dillon, * Copyright (c) 1994 John S. Dyson * Copyright (c) 1990 University of Utah. * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * New Swap System * Matthew Dillon * * Radix Bitmap 'blists'. * * - The new swapper uses the new radix bitmap code. This should scale * to arbitrarily small or arbitrarily large swap spaces and an almost * arbitrary degree of fragmentation. * * Features: * * - on the fly reallocation of swap during putpages. The new system * does not try to keep previously allocated swap blocks for dirty * pages. * * - on the fly deallocation of swap * * - No more garbage collection required. Unnecessarily allocated swap * blocks only exist for dirty vm_page_t's now and these are already * cycled (in a high-load system) by the pager. We also do on-the-fly * removal of invalidated swap blocks when a page is destroyed * or renamed. * * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ * * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 * @(#)vm_swap.c 8.5 (Berkeley) 2/17/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * MAX_PAGEOUT_CLUSTER must be a power of 2 between 1 and 64. * The 64-page limit is due to the radix code (kern/subr_blist.c). */ #ifndef MAX_PAGEOUT_CLUSTER #define MAX_PAGEOUT_CLUSTER 32 #endif #if !defined(SWB_NPAGES) #define SWB_NPAGES MAX_PAGEOUT_CLUSTER #endif #define SWAP_META_PAGES PCTRIE_COUNT /* * A swblk structure maps each page index within a * SWAP_META_PAGES-aligned and sized range to the address of an * on-disk swap block (or SWAPBLK_NONE). The collection of these * mappings for an entire vm object is implemented as a pc-trie. */ struct swblk { vm_pindex_t p; daddr_t d[SWAP_META_PAGES]; }; static MALLOC_DEFINE(M_VMPGDATA, "vm_pgdata", "swap pager private data"); static struct mtx sw_dev_mtx; static TAILQ_HEAD(, swdevt) swtailq = TAILQ_HEAD_INITIALIZER(swtailq); static struct swdevt *swdevhd; /* Allocate from here next */ static int nswapdev; /* Number of swap devices */ int swap_pager_avail; static struct sx swdev_syscall_lock; /* serialize swap(on|off) */ static u_long swap_reserved; static u_long swap_total; static int sysctl_page_shift(SYSCTL_HANDLER_ARGS); static SYSCTL_NODE(_vm_stats, OID_AUTO, swap, CTLFLAG_RD, 0, "VM swap stats"); SYSCTL_PROC(_vm, OID_AUTO, swap_reserved, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, &swap_reserved, 0, sysctl_page_shift, "A", "Amount of swap storage needed to back all allocated anonymous memory."); SYSCTL_PROC(_vm, OID_AUTO, swap_total, CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, &swap_total, 0, sysctl_page_shift, "A", "Total amount of available swap storage."); static int overcommit = 0; SYSCTL_INT(_vm, VM_OVERCOMMIT, overcommit, CTLFLAG_RW, &overcommit, 0, "Configure virtual memory overcommit behavior. See tuning(7) " "for details."); static unsigned long swzone; SYSCTL_ULONG(_vm, OID_AUTO, swzone, CTLFLAG_RD, &swzone, 0, "Actual size of swap metadata zone"); static unsigned long swap_maxpages; SYSCTL_ULONG(_vm, OID_AUTO, swap_maxpages, CTLFLAG_RD, &swap_maxpages, 0, "Maximum amount of swap supported"); static counter_u64_t swap_free_deferred; SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_deferred, CTLFLAG_RD, &swap_free_deferred, "Number of pages that deferred freeing swap space"); static counter_u64_t swap_free_completed; SYSCTL_COUNTER_U64(_vm_stats_swap, OID_AUTO, free_completed, CTLFLAG_RD, &swap_free_completed, "Number of deferred frees completed"); /* bits from overcommit */ #define SWAP_RESERVE_FORCE_ON (1 << 0) #define SWAP_RESERVE_RLIMIT_ON (1 << 1) #define SWAP_RESERVE_ALLOW_NONWIRED (1 << 2) static int sysctl_page_shift(SYSCTL_HANDLER_ARGS) { uint64_t newval; u_long value = *(u_long *)arg1; newval = ((uint64_t)value) << PAGE_SHIFT; return (sysctl_handle_64(oidp, &newval, 0, req)); } int swap_reserve(vm_ooffset_t incr) { return (swap_reserve_by_cred(incr, curthread->td_ucred)); } int swap_reserve_by_cred(vm_ooffset_t incr, struct ucred *cred) { u_long r, s, prev, pincr; int res, error; static int curfail; static struct timeval lastfail; struct uidinfo *uip; uip = cred->cr_ruidinfo; KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, (uintmax_t)incr)); #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); error = racct_add(curproc, RACCT_SWAP, incr); PROC_UNLOCK(curproc); if (error != 0) return (0); } #endif pincr = atop(incr); res = 0; prev = atomic_fetchadd_long(&swap_reserved, pincr); r = prev + pincr; if (overcommit & SWAP_RESERVE_ALLOW_NONWIRED) { s = vm_cnt.v_page_count - vm_cnt.v_free_reserved - vm_wire_count(); } else s = 0; s += swap_total; if ((overcommit & SWAP_RESERVE_FORCE_ON) == 0 || r <= s || (error = priv_check(curthread, PRIV_VM_SWAP_NOQUOTA)) == 0) { res = 1; } else { prev = atomic_fetchadd_long(&swap_reserved, -pincr); if (prev < pincr) panic("swap_reserved < incr on overcommit fail"); } if (res) { prev = atomic_fetchadd_long(&uip->ui_vmsize, pincr); if ((overcommit & SWAP_RESERVE_RLIMIT_ON) != 0 && prev + pincr > lim_cur(curthread, RLIMIT_SWAP) && priv_check(curthread, PRIV_VM_SWAP_NORLIMIT)) { res = 0; prev = atomic_fetchadd_long(&uip->ui_vmsize, -pincr); if (prev < pincr) panic("uip->ui_vmsize < incr on overcommit fail"); } } if (!res && ppsratecheck(&lastfail, &curfail, 1)) { printf("uid %d, pid %d: swap reservation for %jd bytes failed\n", uip->ui_uid, curproc->p_pid, incr); } #ifdef RACCT if (racct_enable && !res) { PROC_LOCK(curproc); racct_sub(curproc, RACCT_SWAP, incr); PROC_UNLOCK(curproc); } #endif return (res); } void swap_reserve_force(vm_ooffset_t incr) { struct uidinfo *uip; u_long pincr; KASSERT((incr & PAGE_MASK) == 0, ("%s: incr: %ju & PAGE_MASK", __func__, (uintmax_t)incr)); PROC_LOCK(curproc); #ifdef RACCT if (racct_enable) racct_add_force(curproc, RACCT_SWAP, incr); #endif pincr = atop(incr); atomic_add_long(&swap_reserved, pincr); uip = curproc->p_ucred->cr_ruidinfo; atomic_add_long(&uip->ui_vmsize, pincr); PROC_UNLOCK(curproc); } void swap_release(vm_ooffset_t decr) { struct ucred *cred; PROC_LOCK(curproc); cred = curproc->p_ucred; swap_release_by_cred(decr, cred); PROC_UNLOCK(curproc); } void swap_release_by_cred(vm_ooffset_t decr, struct ucred *cred) { u_long prev, pdecr; struct uidinfo *uip; uip = cred->cr_ruidinfo; KASSERT((decr & PAGE_MASK) == 0, ("%s: decr: %ju & PAGE_MASK", __func__, (uintmax_t)decr)); pdecr = atop(decr); prev = atomic_fetchadd_long(&swap_reserved, -pdecr); if (prev < pdecr) panic("swap_reserved < decr"); prev = atomic_fetchadd_long(&uip->ui_vmsize, -pdecr); if (prev < pdecr) printf("negative vmsize for uid = %d\n", uip->ui_uid); #ifdef RACCT if (racct_enable) racct_sub_cred(cred, RACCT_SWAP, decr); #endif } static int swap_pager_full = 2; /* swap space exhaustion (task killing) */ static int swap_pager_almost_full = 1; /* swap space exhaustion (w/hysteresis)*/ static struct mtx swbuf_mtx; /* to sync nsw_wcount_async */ static int nsw_wcount_async; /* limit async write buffers */ static int nsw_wcount_async_max;/* assigned maximum */ static int nsw_cluster_max; /* maximum VOP I/O allowed */ static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, swap_async_max, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_async_max, "I", "Maximum running async swap ops"); static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, swap_fragmentation, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_swap_fragmentation, "A", "Swap Fragmentation Info"); static struct sx sw_alloc_sx; /* * "named" and "unnamed" anon region objects. Try to reduce the overhead * of searching a named list by hashing it just a little. */ #define NOBJLISTS 8 #define NOBJLIST(handle) \ (&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)]) static struct pagerlst swap_pager_object_list[NOBJLISTS]; static uma_zone_t swwbuf_zone; static uma_zone_t swrbuf_zone; static uma_zone_t swblk_zone; static uma_zone_t swpctrie_zone; /* * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure * calls hooked from other parts of the VM system and do not appear here. * (see vm/swap_pager.h). */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *); static void swap_pager_dealloc(vm_object_t object); static int swap_pager_getpages(vm_object_t, vm_page_t *, int, int *, int *); static int swap_pager_getpages_async(vm_object_t, vm_page_t *, int, int *, int *, pgo_getpages_iodone_t, void *); static void swap_pager_putpages(vm_object_t, vm_page_t *, int, boolean_t, int *); static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after); static void swap_pager_init(void); static void swap_pager_unswapped(vm_page_t); static void swap_pager_swapoff(struct swdevt *sp); static void swap_pager_update_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end); static void swap_pager_release_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end); struct pagerops swappagerops = { .pgo_init = swap_pager_init, /* early system initialization of pager */ .pgo_alloc = swap_pager_alloc, /* allocate an OBJT_SWAP object */ .pgo_dealloc = swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ .pgo_getpages = swap_pager_getpages, /* pagein */ .pgo_getpages_async = swap_pager_getpages_async, /* pagein (async) */ .pgo_putpages = swap_pager_putpages, /* pageout */ .pgo_haspage = swap_pager_haspage, /* get backing store status for page */ .pgo_pageunswapped = swap_pager_unswapped, /* remove swap related to page */ .pgo_update_writecount = swap_pager_update_writecount, .pgo_release_writecount = swap_pager_release_writecount, }; /* * swap_*() routines are externally accessible. swp_*() routines are * internal. */ static int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ static int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ SYSCTL_INT(_vm, OID_AUTO, dmmax, CTLFLAG_RD, &nsw_cluster_max, 0, "Maximum size of a swap block in pages"); static void swp_sizecheck(void); static void swp_pager_async_iodone(struct buf *bp); static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit); static void swp_pager_free_empty_swblk(vm_object_t, struct swblk *sb); static int swapongeom(struct vnode *); static int swaponvp(struct thread *, struct vnode *, u_long); static int swapoff_one(struct swdevt *sp, struct ucred *cred); /* * Swap bitmap functions */ static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages); static daddr_t swp_pager_getswapspace(int *npages, int limit); /* * Metadata functions */ static daddr_t swp_pager_meta_build(vm_object_t, vm_pindex_t, daddr_t); static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); static void swp_pager_meta_transfer(vm_object_t src, vm_object_t dst, vm_pindex_t pindex, vm_pindex_t count); static void swp_pager_meta_free_all(vm_object_t); static daddr_t swp_pager_meta_lookup(vm_object_t, vm_pindex_t); static void swp_pager_init_freerange(daddr_t *start, daddr_t *num) { *start = SWAPBLK_NONE; *num = 0; } static void swp_pager_update_freerange(daddr_t *start, daddr_t *num, daddr_t addr) { if (*start + *num == addr) { (*num)++; } else { swp_pager_freeswapspace(*start, *num); *start = addr; *num = 1; } } static void * swblk_trie_alloc(struct pctrie *ptree) { return (uma_zalloc(swpctrie_zone, M_NOWAIT | (curproc == pageproc ? M_USE_RESERVE : 0))); } static void swblk_trie_free(struct pctrie *ptree, void *node) { uma_zfree(swpctrie_zone, node); } PCTRIE_DEFINE(SWAP, swblk, p, swblk_trie_alloc, swblk_trie_free); /* * SWP_SIZECHECK() - update swap_pager_full indication * * update the swap_pager_almost_full indication and warn when we are * about to run out of swap space, using lowat/hiwat hysteresis. * * Clear swap_pager_full ( task killing ) indication when lowat is met. * * No restrictions on call * This routine may not block. */ static void swp_sizecheck(void) { if (swap_pager_avail < nswap_lowat) { if (swap_pager_almost_full == 0) { printf("swap_pager: out of swap space\n"); swap_pager_almost_full = 1; } } else { swap_pager_full = 0; if (swap_pager_avail > nswap_hiwat) swap_pager_almost_full = 0; } } /* * SWAP_PAGER_INIT() - initialize the swap pager! * * Expected to be started from system init. NOTE: This code is run * before much else so be careful what you depend on. Most of the VM * system has yet to be initialized at this point. */ static void swap_pager_init(void) { /* * Initialize object lists */ int i; for (i = 0; i < NOBJLISTS; ++i) TAILQ_INIT(&swap_pager_object_list[i]); mtx_init(&sw_dev_mtx, "swapdev", NULL, MTX_DEF); sx_init(&sw_alloc_sx, "swspsx"); sx_init(&swdev_syscall_lock, "swsysc"); } static void swap_pager_counters(void) { swap_free_deferred = counter_u64_alloc(M_WAITOK); swap_free_completed = counter_u64_alloc(M_WAITOK); } SYSINIT(swap_counters, SI_SUB_CPU, SI_ORDER_ANY, swap_pager_counters, NULL); /* * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process * * Expected to be started from pageout process once, prior to entering * its main loop. */ void swap_pager_swap_init(void) { unsigned long n, n2; /* * Number of in-transit swap bp operations. Don't * exhaust the pbufs completely. Make sure we * initialize workable values (0 will work for hysteresis * but it isn't very efficient). * * The nsw_cluster_max is constrained by the bp->b_pages[] * array, which has MAXPHYS / PAGE_SIZE entries, and our locally * defined MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are * constrained by the swap device interleave stripe size. * * Currently we hardwire nsw_wcount_async to 4. This limit is * designed to prevent other I/O from having high latencies due to * our pageout I/O. The value 4 works well for one or two active swap * devices but is probably a little low if you have more. Even so, * a higher value would probably generate only a limited improvement * with three or four active swap devices since the system does not * typically have to pageout at extreme bandwidths. We will want * at least 2 per swap devices, and 4 is a pretty good value if you * have one NFS swap device due to the command/ack latency over NFS. * So it all works out pretty well. */ nsw_cluster_max = min(MAXPHYS / PAGE_SIZE, MAX_PAGEOUT_CLUSTER); nsw_wcount_async = 4; nsw_wcount_async_max = nsw_wcount_async; mtx_init(&swbuf_mtx, "async swbuf mutex", NULL, MTX_DEF); swwbuf_zone = pbuf_zsecond_create("swwbuf", nswbuf / 4); swrbuf_zone = pbuf_zsecond_create("swrbuf", nswbuf / 2); /* * Initialize our zone, taking the user's requested size or * estimating the number we need based on the number of pages * in the system. */ n = maxswzone != 0 ? maxswzone / sizeof(struct swblk) : vm_cnt.v_page_count / 2; swpctrie_zone = uma_zcreate("swpctrie", pctrie_node_size(), NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM); if (swpctrie_zone == NULL) panic("failed to create swap pctrie zone."); swblk_zone = uma_zcreate("swblk", sizeof(struct swblk), NULL, NULL, NULL, NULL, _Alignof(struct swblk) - 1, UMA_ZONE_VM); if (swblk_zone == NULL) panic("failed to create swap blk zone."); n2 = n; do { if (uma_zone_reserve_kva(swblk_zone, n)) break; /* * if the allocation failed, try a zone two thirds the * size of the previous attempt. */ n -= ((n + 2) / 3); } while (n > 0); /* * Often uma_zone_reserve_kva() cannot reserve exactly the * requested size. Account for the difference when * calculating swap_maxpages. */ n = uma_zone_get_max(swblk_zone); if (n < n2) printf("Swap blk zone entries changed from %lu to %lu.\n", n2, n); /* absolute maximum we can handle assuming 100% efficiency */ swap_maxpages = n * SWAP_META_PAGES; swzone = n * sizeof(struct swblk); if (!uma_zone_reserve_kva(swpctrie_zone, n)) printf("Cannot reserve swap pctrie zone, " "reduce kern.maxswzone.\n"); } static vm_object_t swap_pager_alloc_init(void *handle, struct ucred *cred, vm_ooffset_t size, vm_ooffset_t offset) { vm_object_t object; if (cred != NULL) { if (!swap_reserve_by_cred(size, cred)) return (NULL); crhold(cred); } /* * The un_pager.swp.swp_blks trie is initialized by * vm_object_allocate() to ensure the correct order of * visibility to other threads. */ object = vm_object_allocate(OBJT_SWAP, OFF_TO_IDX(offset + PAGE_MASK + size)); object->un_pager.swp.writemappings = 0; object->handle = handle; if (cred != NULL) { object->cred = cred; object->charge = size; } return (object); } /* * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate * its metadata structures. * * This routine is called from the mmap and fork code to create a new * OBJT_SWAP object. * * This routine must ensure that no live duplicate is created for * the named object request, which is protected against by * holding the sw_alloc_sx lock in case handle != NULL. */ static vm_object_t swap_pager_alloc(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t offset, struct ucred *cred) { vm_object_t object; if (handle != NULL) { /* * Reference existing named region or allocate new one. There * should not be a race here against swp_pager_meta_build() * as called from vm_page_remove() in regards to the lookup * of the handle. */ sx_xlock(&sw_alloc_sx); object = vm_pager_object_lookup(NOBJLIST(handle), handle); if (object == NULL) { object = swap_pager_alloc_init(handle, cred, size, offset); if (object != NULL) { TAILQ_INSERT_TAIL(NOBJLIST(object->handle), object, pager_object_list); } } sx_xunlock(&sw_alloc_sx); } else { object = swap_pager_alloc_init(handle, cred, size, offset); } return (object); } /* * SWAP_PAGER_DEALLOC() - remove swap metadata from object * * The swap backing for the object is destroyed. The code is * designed such that we can reinstantiate it later, but this * routine is typically called only when the entire object is * about to be destroyed. * * The object must be locked. */ static void swap_pager_dealloc(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_DEAD) != 0, ("dealloc of reachable obj")); /* * Remove from list right away so lookups will fail if we block for * pageout completion. */ if ((object->flags & OBJ_ANON) == 0 && object->handle != NULL) { VM_OBJECT_WUNLOCK(object); sx_xlock(&sw_alloc_sx); TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list); sx_xunlock(&sw_alloc_sx); VM_OBJECT_WLOCK(object); } vm_object_pip_wait(object, "swpdea"); /* * Free all remaining metadata. We only bother to free it from * the swap meta data. We do not attempt to free swapblk's still * associated with vm_page_t's for this object. We do not care * if paging is still in progress on some objects. */ swp_pager_meta_free_all(object); object->handle = NULL; object->type = OBJT_DEAD; } /************************************************************************ * SWAP PAGER BITMAP ROUTINES * ************************************************************************/ /* * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space * * Allocate swap for up to the requested number of pages, and at * least a minimum number of pages. The starting swap block number * (a page index) is returned or SWAPBLK_NONE if the allocation * failed. * * Also has the side effect of advising that somebody made a mistake * when they configured swap and didn't configure enough. * * This routine may not sleep. * * We allocate in round-robin fashion from the configured devices. */ static daddr_t swp_pager_getswapspace(int *io_npages, int limit) { daddr_t blk; struct swdevt *sp; int mpages, npages; blk = SWAPBLK_NONE; mpages = *io_npages; npages = imin(BLIST_MAX_ALLOC, mpages); mtx_lock(&sw_dev_mtx); sp = swdevhd; while (!TAILQ_EMPTY(&swtailq)) { if (sp == NULL) sp = TAILQ_FIRST(&swtailq); if ((sp->sw_flags & SW_CLOSING) == 0) blk = blist_alloc(sp->sw_blist, &npages, mpages); if (blk != SWAPBLK_NONE) break; sp = TAILQ_NEXT(sp, sw_list); if (swdevhd == sp) { if (npages <= limit) break; mpages = npages - 1; npages >>= 1; } } if (blk != SWAPBLK_NONE) { *io_npages = npages; blk += sp->sw_first; sp->sw_used += npages; swap_pager_avail -= npages; swp_sizecheck(); swdevhd = TAILQ_NEXT(sp, sw_list); } else { if (swap_pager_full != 2) { printf("swp_pager_getswapspace(%d): failed\n", *io_npages); swap_pager_full = 2; swap_pager_almost_full = 1; } swdevhd = NULL; } mtx_unlock(&sw_dev_mtx); return (blk); } static bool swp_pager_isondev(daddr_t blk, struct swdevt *sp) { return (blk >= sp->sw_first && blk < sp->sw_end); } static void swp_pager_strategy(struct buf *bp) { struct swdevt *sp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (swp_pager_isondev(bp->b_blkno, sp)) { mtx_unlock(&sw_dev_mtx); if ((sp->sw_flags & SW_UNMAPPED) != 0 && unmapped_buf_allowed) { bp->b_data = unmapped_buf; bp->b_offset = 0; } else { pmap_qenter((vm_offset_t)bp->b_data, &bp->b_pages[0], bp->b_bcount / PAGE_SIZE); } sp->sw_strategy(bp, sp); return; } } panic("Swapdev not found"); } /* * SWP_PAGER_FREESWAPSPACE() - free raw swap space * * This routine returns the specified swap blocks back to the bitmap. * * This routine may not sleep. */ static void swp_pager_freeswapspace(daddr_t blk, daddr_t npages) { struct swdevt *sp; if (npages == 0) return; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (swp_pager_isondev(blk, sp)) { sp->sw_used -= npages; /* * If we are attempting to stop swapping on * this device, we don't want to mark any * blocks free lest they be reused. */ if ((sp->sw_flags & SW_CLOSING) == 0) { blist_free(sp->sw_blist, blk - sp->sw_first, npages); swap_pager_avail += npages; swp_sizecheck(); } mtx_unlock(&sw_dev_mtx); return; } } panic("Swapdev not found"); } /* * SYSCTL_SWAP_FRAGMENTATION() - produce raw swap space stats */ static int sysctl_swap_fragmentation(SYSCTL_HANDLER_ARGS) { struct sbuf sbuf; struct swdevt *sp; const char *devname; int error; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (vn_isdisk(sp->sw_vp, NULL)) devname = devtoname(sp->sw_vp->v_rdev); else devname = "[file]"; sbuf_printf(&sbuf, "\nFree space on device %s:\n", devname); blist_stats(sp->sw_blist, &sbuf); } mtx_unlock(&sw_dev_mtx); error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page * range within an object. * * This is a globally accessible routine. * * This routine removes swapblk assignments from swap metadata. * * The external callers of this routine typically have already destroyed * or renamed vm_page_t's associated with this range in the object so * we should be ok. * * The object must be locked. */ void swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size) { swp_pager_meta_free(object, start, size); } /* * SWAP_PAGER_RESERVE() - reserve swap blocks in object * * Assigns swap blocks to the specified range within the object. The * swap blocks are not zeroed. Any previous swap assignment is destroyed. * * Returns 0 on success, -1 on failure. */ int swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) { daddr_t addr, blk, n_free, s_free; int i, j, n; swp_pager_init_freerange(&s_free, &n_free); VM_OBJECT_WLOCK(object); for (i = 0; i < size; i += n) { n = size - i; blk = swp_pager_getswapspace(&n, 1); if (blk == SWAPBLK_NONE) { swp_pager_meta_free(object, start, i); VM_OBJECT_WUNLOCK(object); return (-1); } for (j = 0; j < n; ++j) { addr = swp_pager_meta_build(object, start + i + j, blk + j); if (addr != SWAPBLK_NONE) swp_pager_update_freerange(&s_free, &n_free, addr); } } swp_pager_freeswapspace(s_free, n_free); VM_OBJECT_WUNLOCK(object); return (0); } static bool swp_pager_xfer_source(vm_object_t srcobject, vm_object_t dstobject, vm_pindex_t pindex, daddr_t addr) { daddr_t dstaddr; KASSERT(srcobject->type == OBJT_SWAP, ("%s: Srcobject not swappable", __func__)); if (dstobject->type == OBJT_SWAP && swp_pager_meta_lookup(dstobject, pindex) != SWAPBLK_NONE) { /* Caller should destroy the source block. */ return (false); } /* * Destination has no swapblk and is not resident, transfer source. * swp_pager_meta_build() can sleep. */ vm_object_pip_add(srcobject, 1); VM_OBJECT_WUNLOCK(srcobject); vm_object_pip_add(dstobject, 1); dstaddr = swp_pager_meta_build(dstobject, pindex, addr); KASSERT(dstaddr == SWAPBLK_NONE, ("Unexpected destination swapblk")); vm_object_pip_wakeup(dstobject); VM_OBJECT_WLOCK(srcobject); vm_object_pip_wakeup(srcobject); return (true); } /* * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager * and destroy the source. * * Copy any valid swapblks from the source to the destination. In * cases where both the source and destination have a valid swapblk, * we keep the destination's. * * This routine is allowed to sleep. It may sleep allocating metadata * indirectly through swp_pager_meta_build() or if paging is still in * progress on the source. * * The source object contains no vm_page_t's (which is just as well) * * The source object is of type OBJT_SWAP. * * The source and destination objects must be locked. * Both object locks may temporarily be released. */ void swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, vm_pindex_t offset, int destroysource) { VM_OBJECT_ASSERT_WLOCKED(srcobject); VM_OBJECT_ASSERT_WLOCKED(dstobject); /* * If destroysource is set, we remove the source object from the * swap_pager internal queue now. */ if (destroysource && (srcobject->flags & OBJ_ANON) == 0 && srcobject->handle != NULL) { vm_object_pip_add(srcobject, 1); VM_OBJECT_WUNLOCK(srcobject); vm_object_pip_add(dstobject, 1); VM_OBJECT_WUNLOCK(dstobject); sx_xlock(&sw_alloc_sx); TAILQ_REMOVE(NOBJLIST(srcobject->handle), srcobject, pager_object_list); sx_xunlock(&sw_alloc_sx); VM_OBJECT_WLOCK(dstobject); vm_object_pip_wakeup(dstobject); VM_OBJECT_WLOCK(srcobject); vm_object_pip_wakeup(srcobject); } /* * Transfer source to destination. */ swp_pager_meta_transfer(srcobject, dstobject, offset, dstobject->size); /* * Free left over swap blocks in source. * * We have to revert the type to OBJT_DEFAULT so we do not accidentally * double-remove the object from the swap queues. */ if (destroysource) { swp_pager_meta_free_all(srcobject); /* * Reverting the type is not necessary, the caller is going * to destroy srcobject directly, but I'm doing it here * for consistency since we've removed the object from its * queues. */ srcobject->type = OBJT_DEFAULT; } } /* * SWAP_PAGER_HASPAGE() - determine if we have good backing store for * the requested page. * * We determine whether good backing store exists for the requested * page and return TRUE if it does, FALSE if it doesn't. * * If TRUE, we also try to determine how much valid, contiguous backing * store exists before and after the requested page. */ static boolean_t swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { daddr_t blk, blk0; int i; VM_OBJECT_ASSERT_LOCKED(object); KASSERT(object->type == OBJT_SWAP, ("%s: object not swappable", __func__)); /* * do we have good backing store at the requested index ? */ blk0 = swp_pager_meta_lookup(object, pindex); if (blk0 == SWAPBLK_NONE) { if (before) *before = 0; if (after) *after = 0; return (FALSE); } /* * find backwards-looking contiguous good backing store */ if (before != NULL) { for (i = 1; i < SWB_NPAGES; i++) { if (i > pindex) break; blk = swp_pager_meta_lookup(object, pindex - i); if (blk != blk0 - i) break; } *before = i - 1; } /* * find forward-looking contiguous good backing store */ if (after != NULL) { for (i = 1; i < SWB_NPAGES; i++) { blk = swp_pager_meta_lookup(object, pindex + i); if (blk != blk0 + i) break; } *after = i - 1; } return (TRUE); } /* * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page * * This removes any associated swap backing store, whether valid or * not, from the page. * * This routine is typically called when a page is made dirty, at * which point any associated swap can be freed. MADV_FREE also * calls us in a special-case situation * * NOTE!!! If the page is clean and the swap was valid, the caller * should make the page dirty before calling this routine. This routine * does NOT change the m->dirty status of the page. Also: MADV_FREE * depends on it. * * This routine may not sleep. * * The object containing the page may be locked. */ static void swap_pager_unswapped(vm_page_t m) { struct swblk *sb; vm_object_t obj; /* * Handle enqueing deferred frees first. If we do not have the * object lock we wait for the page daemon to clear the space. */ obj = m->object; if (!VM_OBJECT_WOWNED(obj)) { VM_PAGE_OBJECT_BUSY_ASSERT(m); /* * The caller is responsible for synchronization but we * will harmlessly handle races. This is typically provided * by only calling unswapped() when a page transitions from * clean to dirty. */ if ((m->a.flags & (PGA_SWAP_SPACE | PGA_SWAP_FREE)) == PGA_SWAP_SPACE) { vm_page_aflag_set(m, PGA_SWAP_FREE); counter_u64_add(swap_free_deferred, 1); } return; } if ((m->a.flags & PGA_SWAP_FREE) != 0) counter_u64_add(swap_free_completed, 1); vm_page_aflag_clear(m, PGA_SWAP_FREE | PGA_SWAP_SPACE); /* * The meta data only exists if the object is OBJT_SWAP * and even then might not be allocated yet. */ KASSERT(m->object->type == OBJT_SWAP, ("Free object not swappable")); sb = SWAP_PCTRIE_LOOKUP(&m->object->un_pager.swp.swp_blks, rounddown(m->pindex, SWAP_META_PAGES)); if (sb == NULL) return; if (sb->d[m->pindex % SWAP_META_PAGES] == SWAPBLK_NONE) return; swp_pager_freeswapspace(sb->d[m->pindex % SWAP_META_PAGES], 1); sb->d[m->pindex % SWAP_META_PAGES] = SWAPBLK_NONE; swp_pager_free_empty_swblk(m->object, sb); } /* * swap_pager_getpages() - bring pages in from swap * * Attempt to page in the pages in array "ma" of length "count". The * caller may optionally specify that additional pages preceding and * succeeding the specified range be paged in. The number of such pages * is returned in the "rbehind" and "rahead" parameters, and they will * be in the inactive queue upon return. * * The pages in "ma" must be busied and will remain busied upon return. */ static int swap_pager_getpages(vm_object_t object, vm_page_t *ma, int count, int *rbehind, int *rahead) { struct buf *bp; vm_page_t bm, mpred, msucc, p; vm_pindex_t pindex; daddr_t blk; int i, maxahead, maxbehind, reqcount; reqcount = count; /* * Determine the final number of read-behind pages and * allocate them BEFORE releasing the object lock. Otherwise, * there can be a problematic race with vm_object_split(). * Specifically, vm_object_split() might first transfer pages * that precede ma[0] in the current object to a new object, * and then this function incorrectly recreates those pages as * read-behind pages in the current object. */ KASSERT(object->type == OBJT_SWAP, ("%s: object not swappable", __func__)); if (!swap_pager_haspage(object, ma[0]->pindex, &maxbehind, &maxahead)) return (VM_PAGER_FAIL); /* * Clip the readahead and readbehind ranges to exclude resident pages. */ if (rahead != NULL) { KASSERT(reqcount - 1 <= maxahead, ("page count %d extends beyond swap block", reqcount)); *rahead = imin(*rahead, maxahead - (reqcount - 1)); pindex = ma[reqcount - 1]->pindex; msucc = TAILQ_NEXT(ma[reqcount - 1], listq); if (msucc != NULL && msucc->pindex - pindex - 1 < *rahead) *rahead = msucc->pindex - pindex - 1; } if (rbehind != NULL) { *rbehind = imin(*rbehind, maxbehind); pindex = ma[0]->pindex; mpred = TAILQ_PREV(ma[0], pglist, listq); if (mpred != NULL && pindex - mpred->pindex - 1 < *rbehind) *rbehind = pindex - mpred->pindex - 1; } bm = ma[0]; for (i = 0; i < count; i++) ma[i]->oflags |= VPO_SWAPINPROG; /* * Allocate readahead and readbehind pages. */ if (rbehind != NULL) { for (i = 1; i <= *rbehind; i++) { p = vm_page_alloc(object, ma[0]->pindex - i, VM_ALLOC_NORMAL); if (p == NULL) break; p->oflags |= VPO_SWAPINPROG; bm = p; } *rbehind = i - 1; } if (rahead != NULL) { for (i = 0; i < *rahead; i++) { p = vm_page_alloc(object, ma[reqcount - 1]->pindex + i + 1, VM_ALLOC_NORMAL); if (p == NULL) break; p->oflags |= VPO_SWAPINPROG; } *rahead = i; } if (rbehind != NULL) count += *rbehind; if (rahead != NULL) count += *rahead; vm_object_pip_add(object, count); pindex = bm->pindex; blk = swp_pager_meta_lookup(object, pindex); KASSERT(blk != SWAPBLK_NONE, ("no swap blocking containing %p(%jx)", object, (uintmax_t)pindex)); VM_OBJECT_WUNLOCK(object); bp = uma_zalloc(swrbuf_zone, M_WAITOK); /* Pages cannot leave the object while busy. */ for (i = 0, p = bm; i < count; i++, p = TAILQ_NEXT(p, listq)) { MPASS(p->pindex == bm->pindex + i); bp->b_pages[i] = p; } bp->b_flags |= B_PAGING; bp->b_iocmd = BIO_READ; bp->b_iodone = swp_pager_async_iodone; bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_blkno = blk; bp->b_bcount = PAGE_SIZE * count; bp->b_bufsize = PAGE_SIZE * count; bp->b_npages = count; bp->b_pgbefore = rbehind != NULL ? *rbehind : 0; bp->b_pgafter = rahead != NULL ? *rahead : 0; VM_CNT_INC(v_swapin); VM_CNT_ADD(v_swappgsin, count); /* * perform the I/O. NOTE!!! bp cannot be considered valid after * this point because we automatically release it on completion. * Instead, we look at the one page we are interested in which we * still hold a lock on even through the I/O completion. * * The other pages in our ma[] array are also released on completion, * so we cannot assume they are valid anymore either. * * NOTE: b_blkno is destroyed by the call to swapdev_strategy */ BUF_KERNPROC(bp); swp_pager_strategy(bp); /* * Wait for the pages we want to complete. VPO_SWAPINPROG is always * cleared on completion. If an I/O error occurs, SWAPBLK_NONE * is set in the metadata for each page in the request. */ VM_OBJECT_WLOCK(object); while ((ma[0]->oflags & VPO_SWAPINPROG) != 0) { ma[0]->oflags |= VPO_SWAPSLEEP; VM_CNT_INC(v_intrans); if (VM_OBJECT_SLEEP(object, &object->handle, PSWP, "swread", hz * 20)) { printf( "swap_pager: indefinite wait buffer: bufobj: %p, blkno: %jd, size: %ld\n", bp->b_bufobj, (intmax_t)bp->b_blkno, bp->b_bcount); } } /* * If we had an unrecoverable read error pages will not be valid. */ for (i = 0; i < reqcount; i++) if (ma[i]->valid != VM_PAGE_BITS_ALL) return (VM_PAGER_ERROR); return (VM_PAGER_OK); /* * A final note: in a low swap situation, we cannot deallocate swap * and mark a page dirty here because the caller is likely to mark * the page clean when we return, causing the page to possibly revert * to all-zero's later. */ } /* * swap_pager_getpages_async(): * * Right now this is emulation of asynchronous operation on top of * swap_pager_getpages(). */ static int swap_pager_getpages_async(vm_object_t object, vm_page_t *ma, int count, int *rbehind, int *rahead, pgo_getpages_iodone_t iodone, void *arg) { int r, error; r = swap_pager_getpages(object, ma, count, rbehind, rahead); VM_OBJECT_WUNLOCK(object); switch (r) { case VM_PAGER_OK: error = 0; break; case VM_PAGER_ERROR: error = EIO; break; case VM_PAGER_FAIL: error = EINVAL; break; default: panic("unhandled swap_pager_getpages() error %d", r); } (iodone)(arg, ma, count, error); VM_OBJECT_WLOCK(object); return (r); } /* * swap_pager_putpages: * * Assign swap (if necessary) and initiate I/O on the specified pages. * * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects * are automatically converted to SWAP objects. * * In a low memory situation we may block in VOP_STRATEGY(), but the new * vm_page reservation system coupled with properly written VFS devices * should ensure that no low-memory deadlock occurs. This is an area * which needs work. * * The parent has N vm_object_pip_add() references prior to * calling us and will remove references for rtvals[] that are * not set to VM_PAGER_PEND. We need to remove the rest on I/O * completion. * * The parent has soft-busy'd the pages it passes us and will unbusy * those whose rtvals[] entry is not set to VM_PAGER_PEND on return. * We need to unbusy the rest on I/O completion. */ static void swap_pager_putpages(vm_object_t object, vm_page_t *ma, int count, int flags, int *rtvals) { struct buf *bp; daddr_t addr, blk, n_free, s_free; vm_page_t mreq; int i, j, n; bool async; KASSERT(count == 0 || ma[0]->object == object, ("%s: object mismatch %p/%p", __func__, object, ma[0]->object)); /* * Step 1 * * Turn object into OBJT_SWAP. Force sync if not a pageout process. */ if (object->type != OBJT_SWAP) { addr = swp_pager_meta_build(object, 0, SWAPBLK_NONE); KASSERT(addr == SWAPBLK_NONE, ("unexpected object swap block")); } VM_OBJECT_WUNLOCK(object); async = curproc == pageproc && (flags & VM_PAGER_PUT_SYNC) == 0; swp_pager_init_freerange(&s_free, &n_free); /* * Step 2 * * Assign swap blocks and issue I/O. We reallocate swap on the fly. * The page is left dirty until the pageout operation completes * successfully. */ for (i = 0; i < count; i += n) { /* Maximum I/O size is limited by maximum swap block size. */ n = min(count - i, nsw_cluster_max); /* Get a block of swap of size up to size n. */ blk = swp_pager_getswapspace(&n, 4); if (blk == SWAPBLK_NONE) { for (j = 0; j < n; ++j) rtvals[i + j] = VM_PAGER_FAIL; continue; } /* * All I/O parameters have been satisfied. Build the I/O * request and assign the swap space. */ if (async) { mtx_lock(&swbuf_mtx); while (nsw_wcount_async == 0) msleep(&nsw_wcount_async, &swbuf_mtx, PVM, "swbufa", 0); nsw_wcount_async--; mtx_unlock(&swbuf_mtx); } bp = uma_zalloc(swwbuf_zone, M_WAITOK); if (async) bp->b_flags = B_ASYNC; bp->b_flags |= B_PAGING; bp->b_iocmd = BIO_WRITE; bp->b_rcred = crhold(thread0.td_ucred); bp->b_wcred = crhold(thread0.td_ucred); bp->b_bcount = PAGE_SIZE * n; bp->b_bufsize = PAGE_SIZE * n; bp->b_blkno = blk; VM_OBJECT_WLOCK(object); for (j = 0; j < n; ++j) { mreq = ma[i + j]; vm_page_aflag_clear(mreq, PGA_SWAP_FREE); addr = swp_pager_meta_build(mreq->object, mreq->pindex, blk + j); if (addr != SWAPBLK_NONE) swp_pager_update_freerange(&s_free, &n_free, addr); MPASS(mreq->dirty == VM_PAGE_BITS_ALL); mreq->oflags |= VPO_SWAPINPROG; bp->b_pages[j] = mreq; } VM_OBJECT_WUNLOCK(object); bp->b_npages = n; /* * Must set dirty range for NFS to work. */ bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bcount; VM_CNT_INC(v_swapout); VM_CNT_ADD(v_swappgsout, bp->b_npages); /* * We unconditionally set rtvals[] to VM_PAGER_PEND so that we * can call the async completion routine at the end of a * synchronous I/O operation. Otherwise, our caller would * perform duplicate unbusy and wakeup operations on the page * and object, respectively. */ for (j = 0; j < n; j++) rtvals[i + j] = VM_PAGER_PEND; /* * asynchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy. */ if (async) { bp->b_iodone = swp_pager_async_iodone; BUF_KERNPROC(bp); swp_pager_strategy(bp); continue; } /* * synchronous * * NOTE: b_blkno is destroyed by the call to swapdev_strategy. */ bp->b_iodone = bdone; swp_pager_strategy(bp); /* * Wait for the sync I/O to complete. */ bwait(bp, PVM, "swwrt"); /* * Now that we are through with the bp, we can call the * normal async completion, which frees everything up. */ swp_pager_async_iodone(bp); } swp_pager_freeswapspace(s_free, n_free); VM_OBJECT_WLOCK(object); } /* * swp_pager_async_iodone: * * Completion routine for asynchronous reads and writes from/to swap. * Also called manually by synchronous code to finish up a bp. * * This routine may not sleep. */ static void swp_pager_async_iodone(struct buf *bp) { int i; vm_object_t object = NULL; /* * Report error - unless we ran out of memory, in which case * we've already logged it in swapgeom_strategy(). */ if (bp->b_ioflags & BIO_ERROR && bp->b_error != ENOMEM) { printf( "swap_pager: I/O error - %s failed; blkno %ld," "size %ld, error %d\n", ((bp->b_iocmd == BIO_READ) ? "pagein" : "pageout"), (long)bp->b_blkno, (long)bp->b_bcount, bp->b_error ); } /* * remove the mapping for kernel virtual */ if (buf_mapped(bp)) pmap_qremove((vm_offset_t)bp->b_data, bp->b_npages); else bp->b_data = bp->b_kvabase; if (bp->b_npages) { object = bp->b_pages[0]->object; VM_OBJECT_WLOCK(object); } /* * cleanup pages. If an error occurs writing to swap, we are in * very serious trouble. If it happens to be a disk error, though, * we may be able to recover by reassigning the swap later on. So * in this case we remove the m->swapblk assignment for the page * but do not free it in the rlist. The errornous block(s) are thus * never reallocated as swap. Redirty the page and continue. */ for (i = 0; i < bp->b_npages; ++i) { vm_page_t m = bp->b_pages[i]; m->oflags &= ~VPO_SWAPINPROG; if (m->oflags & VPO_SWAPSLEEP) { m->oflags &= ~VPO_SWAPSLEEP; wakeup(&object->handle); } /* We always have space after I/O, successful or not. */ vm_page_aflag_set(m, PGA_SWAP_SPACE); if (bp->b_ioflags & BIO_ERROR) { /* * If an error occurs I'd love to throw the swapblk * away without freeing it back to swapspace, so it * can never be used again. But I can't from an * interrupt. */ if (bp->b_iocmd == BIO_READ) { /* * NOTE: for reads, m->dirty will probably * be overridden by the original caller of * getpages so don't play cute tricks here. */ vm_page_invalid(m); } else { /* * If a write error occurs, reactivate page * so it doesn't clog the inactive list, * then finish the I/O. */ MPASS(m->dirty == VM_PAGE_BITS_ALL); + /* PQ_UNSWAPPABLE? */ - vm_page_lock(m); vm_page_activate(m); - vm_page_unlock(m); vm_page_sunbusy(m); } } else if (bp->b_iocmd == BIO_READ) { /* * NOTE: for reads, m->dirty will probably be * overridden by the original caller of getpages so * we cannot set them in order to free the underlying * swap in a low-swap situation. I don't think we'd * want to do that anyway, but it was an optimization * that existed in the old swapper for a time before * it got ripped out due to precisely this problem. */ KASSERT(!pmap_page_is_mapped(m), ("swp_pager_async_iodone: page %p is mapped", m)); KASSERT(m->dirty == 0, ("swp_pager_async_iodone: page %p is dirty", m)); vm_page_valid(m); if (i < bp->b_pgbefore || i >= bp->b_npages - bp->b_pgafter) vm_page_readahead_finish(m); } else { /* * For write success, clear the dirty * status, then finish the I/O ( which decrements the * busy count and possibly wakes waiter's up ). * A page is only written to swap after a period of * inactivity. Therefore, we do not expect it to be * reused. */ KASSERT(!pmap_page_is_write_mapped(m), ("swp_pager_async_iodone: page %p is not write" " protected", m)); vm_page_undirty(m); - vm_page_lock(m); vm_page_deactivate_noreuse(m); - vm_page_unlock(m); vm_page_sunbusy(m); } } /* * adjust pip. NOTE: the original parent may still have its own * pip refs on the object. */ if (object != NULL) { vm_object_pip_wakeupn(object, bp->b_npages); VM_OBJECT_WUNLOCK(object); } /* * swapdev_strategy() manually sets b_vp and b_bufobj before calling * bstrategy(). Set them back to NULL now we're done with it, or we'll * trigger a KASSERT in relpbuf(). */ if (bp->b_vp) { bp->b_vp = NULL; bp->b_bufobj = NULL; } /* * release the physical I/O buffer */ if (bp->b_flags & B_ASYNC) { mtx_lock(&swbuf_mtx); if (++nsw_wcount_async == 1) wakeup(&nsw_wcount_async); mtx_unlock(&swbuf_mtx); } uma_zfree((bp->b_iocmd == BIO_READ) ? swrbuf_zone : swwbuf_zone, bp); } int swap_pager_nswapdev(void) { return (nswapdev); } static void swp_pager_force_dirty(vm_page_t m) { vm_page_dirty(m); #ifdef INVARIANTS - vm_page_lock(m); if (!vm_page_wired(m) && m->a.queue == PQ_NONE) panic("page %p is neither wired nor queued", m); - vm_page_unlock(m); #endif vm_page_xunbusy(m); swap_pager_unswapped(m); } static void swp_pager_force_launder(vm_page_t m) { vm_page_dirty(m); - vm_page_lock(m); vm_page_launder(m); - vm_page_unlock(m); vm_page_xunbusy(m); swap_pager_unswapped(m); } /* * SWP_PAGER_FORCE_PAGEIN() - force swap blocks to be paged in * * This routine dissociates pages starting at the given index within an * object from their backing store, paging them in if they do not reside * in memory. Pages that are paged in are marked dirty and placed in the * laundry queue. Pages are marked dirty because they no longer have * backing store. They are placed in the laundry queue because they have * not been accessed recently. Otherwise, they would already reside in * memory. */ static void swp_pager_force_pagein(vm_object_t object, vm_pindex_t pindex, int npages) { vm_page_t ma[npages]; int i, j; KASSERT(npages > 0, ("%s: No pages", __func__)); KASSERT(npages <= MAXPHYS / PAGE_SIZE, ("%s: Too many pages: %d", __func__, npages)); KASSERT(object->type == OBJT_SWAP, ("%s: Object not swappable", __func__)); vm_object_pip_add(object, npages); vm_page_grab_pages(object, pindex, VM_ALLOC_NORMAL, ma, npages); for (i = j = 0;; i++) { /* Count nonresident pages, to page-in all at once. */ if (i < npages && ma[i]->valid != VM_PAGE_BITS_ALL) continue; if (j < i) { /* Page-in nonresident pages. Mark for laundering. */ if (swap_pager_getpages(object, &ma[j], i - j, NULL, NULL) != VM_PAGER_OK) panic("%s: read from swap failed", __func__); do { swp_pager_force_launder(ma[j]); } while (++j < i); } if (i == npages) break; /* Mark dirty a resident page. */ swp_pager_force_dirty(ma[j++]); } vm_object_pip_wakeupn(object, npages); } /* * swap_pager_swapoff_object: * * Page in all of the pages that have been paged out for an object * to a swap device. */ static void swap_pager_swapoff_object(struct swdevt *sp, vm_object_t object) { struct swblk *sb; vm_pindex_t pi, s_pindex; daddr_t blk, n_blks, s_blk; int i; KASSERT(object->type == OBJT_SWAP, ("%s: Object not swappable", __func__)); n_blks = 0; for (pi = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pi)) != NULL; ) { for (i = 0; i < SWAP_META_PAGES; i++) { blk = sb->d[i]; if (!swp_pager_isondev(blk, sp)) blk = SWAPBLK_NONE; /* * If there are no blocks/pages accumulated, start a new * accumulation here. */ if (n_blks == 0) { if (blk != SWAPBLK_NONE) { s_blk = blk; s_pindex = sb->p + i; n_blks = 1; } continue; } /* * If the accumulation can be extended without breaking * the sequence of consecutive blocks and pages that * swp_pager_force_pagein() depends on, do so. */ if (n_blks < MAXPHYS / PAGE_SIZE && s_blk + n_blks == blk && s_pindex + n_blks == sb->p + i) { ++n_blks; continue; } /* * The sequence of consecutive blocks and pages cannot * be extended, so page them all in here. Then, * because doing so involves releasing and reacquiring * a lock that protects the swap block pctrie, do not * rely on the current swap block. Break this loop and * re-fetch the same pindex from the pctrie again. */ swp_pager_force_pagein(object, s_pindex, n_blks); n_blks = 0; break; } if (i == SWAP_META_PAGES) pi = sb->p + SWAP_META_PAGES; } if (n_blks > 0) swp_pager_force_pagein(object, s_pindex, n_blks); } /* * swap_pager_swapoff: * * Page in all of the pages that have been paged out to the * given device. The corresponding blocks in the bitmap must be * marked as allocated and the device must be flagged SW_CLOSING. * There may be no processes swapped out to the device. * * This routine may block. */ static void swap_pager_swapoff(struct swdevt *sp) { vm_object_t object; int retries; sx_assert(&swdev_syscall_lock, SA_XLOCKED); retries = 0; full_rescan: mtx_lock(&vm_object_list_mtx); TAILQ_FOREACH(object, &vm_object_list, object_list) { if (object->type != OBJT_SWAP) continue; mtx_unlock(&vm_object_list_mtx); /* Depends on type-stability. */ VM_OBJECT_WLOCK(object); /* * Dead objects are eventually terminated on their own. */ if ((object->flags & OBJ_DEAD) != 0) goto next_obj; /* * Sync with fences placed after pctrie * initialization. We must not access pctrie below * unless we checked that our object is swap and not * dead. */ atomic_thread_fence_acq(); if (object->type != OBJT_SWAP) goto next_obj; swap_pager_swapoff_object(sp, object); next_obj: VM_OBJECT_WUNLOCK(object); mtx_lock(&vm_object_list_mtx); } mtx_unlock(&vm_object_list_mtx); if (sp->sw_used) { /* * Objects may be locked or paging to the device being * removed, so we will miss their pages and need to * make another pass. We have marked this device as * SW_CLOSING, so the activity should finish soon. */ retries++; if (retries > 100) { panic("swapoff: failed to locate %d swap blocks", sp->sw_used); } pause("swpoff", hz / 20); goto full_rescan; } EVENTHANDLER_INVOKE(swapoff, sp); } /************************************************************************ * SWAP META DATA * ************************************************************************ * * These routines manipulate the swap metadata stored in the * OBJT_SWAP object. * * Swap metadata is implemented with a global hash and not directly * linked into the object. Instead the object simply contains * appropriate tracking counters. */ /* * SWP_PAGER_SWBLK_EMPTY() - is a range of blocks free? */ static bool swp_pager_swblk_empty(struct swblk *sb, int start, int limit) { int i; MPASS(0 <= start && start <= limit && limit <= SWAP_META_PAGES); for (i = start; i < limit; i++) { if (sb->d[i] != SWAPBLK_NONE) return (false); } return (true); } /* * SWP_PAGER_FREE_EMPTY_SWBLK() - frees if a block is free * * Nothing is done if the block is still in use. */ static void swp_pager_free_empty_swblk(vm_object_t object, struct swblk *sb) { if (swp_pager_swblk_empty(sb, 0, SWAP_META_PAGES)) { SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); uma_zfree(swblk_zone, sb); } } /* * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object * * We first convert the object to a swap object if it is a default * object. * * The specified swapblk is added to the object's swap metadata. If * the swapblk is not valid, it is freed instead. Any previously * assigned swapblk is returned. */ static daddr_t swp_pager_meta_build(vm_object_t object, vm_pindex_t pindex, daddr_t swapblk) { static volatile int swblk_zone_exhausted, swpctrie_zone_exhausted; struct swblk *sb, *sb1; vm_pindex_t modpi, rdpi; daddr_t prev_swapblk; int error, i; VM_OBJECT_ASSERT_WLOCKED(object); /* * Convert default object to swap object if necessary */ if (object->type != OBJT_SWAP) { pctrie_init(&object->un_pager.swp.swp_blks); /* * Ensure that swap_pager_swapoff()'s iteration over * object_list does not see a garbage pctrie. */ atomic_thread_fence_rel(); object->type = OBJT_SWAP; object->un_pager.swp.writemappings = 0; KASSERT((object->flags & OBJ_ANON) != 0 || object->handle == NULL, ("default pager %p with handle %p", object, object->handle)); } rdpi = rounddown(pindex, SWAP_META_PAGES); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb == NULL) { if (swapblk == SWAPBLK_NONE) return (SWAPBLK_NONE); for (;;) { sb = uma_zalloc(swblk_zone, M_NOWAIT | (curproc == pageproc ? M_USE_RESERVE : 0)); if (sb != NULL) { sb->p = rdpi; for (i = 0; i < SWAP_META_PAGES; i++) sb->d[i] = SWAPBLK_NONE; if (atomic_cmpset_int(&swblk_zone_exhausted, 1, 0)) printf("swblk zone ok\n"); break; } VM_OBJECT_WUNLOCK(object); if (uma_zone_exhausted(swblk_zone)) { if (atomic_cmpset_int(&swblk_zone_exhausted, 0, 1)) printf("swap blk zone exhausted, " "increase kern.maxswzone\n"); vm_pageout_oom(VM_OOM_SWAPZ); pause("swzonxb", 10); } else uma_zwait(swblk_zone); VM_OBJECT_WLOCK(object); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb != NULL) /* * Somebody swapped out a nearby page, * allocating swblk at the rdpi index, * while we dropped the object lock. */ goto allocated; } for (;;) { error = SWAP_PCTRIE_INSERT( &object->un_pager.swp.swp_blks, sb); if (error == 0) { if (atomic_cmpset_int(&swpctrie_zone_exhausted, 1, 0)) printf("swpctrie zone ok\n"); break; } VM_OBJECT_WUNLOCK(object); if (uma_zone_exhausted(swpctrie_zone)) { if (atomic_cmpset_int(&swpctrie_zone_exhausted, 0, 1)) printf("swap pctrie zone exhausted, " "increase kern.maxswzone\n"); vm_pageout_oom(VM_OOM_SWAPZ); pause("swzonxp", 10); } else uma_zwait(swpctrie_zone); VM_OBJECT_WLOCK(object); sb1 = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rdpi); if (sb1 != NULL) { uma_zfree(swblk_zone, sb); sb = sb1; goto allocated; } } } allocated: MPASS(sb->p == rdpi); modpi = pindex % SWAP_META_PAGES; /* Return prior contents of metadata. */ prev_swapblk = sb->d[modpi]; /* Enter block into metadata. */ sb->d[modpi] = swapblk; /* * Free the swblk if we end up with the empty page run. */ if (swapblk == SWAPBLK_NONE) swp_pager_free_empty_swblk(object, sb); return (prev_swapblk); } /* * SWP_PAGER_META_TRANSFER() - free a range of blocks in the srcobject's swap * metadata, or transfer it into dstobject. * * This routine will free swap metadata structures as they are cleaned * out. */ static void swp_pager_meta_transfer(vm_object_t srcobject, vm_object_t dstobject, vm_pindex_t pindex, vm_pindex_t count) { struct swblk *sb; daddr_t n_free, s_free; vm_pindex_t offset, last; int i, limit, start; VM_OBJECT_ASSERT_WLOCKED(srcobject); if (srcobject->type != OBJT_SWAP || count == 0) return; swp_pager_init_freerange(&s_free, &n_free); offset = pindex; last = pindex + count; for (;;) { sb = SWAP_PCTRIE_LOOKUP_GE(&srcobject->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL || sb->p >= last) break; start = pindex > sb->p ? pindex - sb->p : 0; limit = last - sb->p < SWAP_META_PAGES ? last - sb->p : SWAP_META_PAGES; for (i = start; i < limit; i++) { if (sb->d[i] == SWAPBLK_NONE) continue; if (dstobject == NULL || !swp_pager_xfer_source(srcobject, dstobject, sb->p + i - offset, sb->d[i])) { swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); } sb->d[i] = SWAPBLK_NONE; } pindex = sb->p + SWAP_META_PAGES; if (swp_pager_swblk_empty(sb, 0, start) && swp_pager_swblk_empty(sb, limit, SWAP_META_PAGES)) { SWAP_PCTRIE_REMOVE(&srcobject->un_pager.swp.swp_blks, sb->p); uma_zfree(swblk_zone, sb); } } swp_pager_freeswapspace(s_free, n_free); } /* * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata * * The requested range of blocks is freed, with any associated swap * returned to the swap bitmap. * * This routine will free swap metadata structures as they are cleaned * out. This routine does *NOT* operate on swap metadata associated * with resident pages. */ static void swp_pager_meta_free(vm_object_t object, vm_pindex_t pindex, vm_pindex_t count) { swp_pager_meta_transfer(object, NULL, pindex, count); } /* * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object * * This routine locates and destroys all swap metadata associated with * an object. */ static void swp_pager_meta_free_all(vm_object_t object) { struct swblk *sb; daddr_t n_free, s_free; vm_pindex_t pindex; int i; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_SWAP) return; swp_pager_init_freerange(&s_free, &n_free); for (pindex = 0; (sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pindex)) != NULL;) { pindex = sb->p + SWAP_META_PAGES; for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->d[i] == SWAPBLK_NONE) continue; swp_pager_update_freerange(&s_free, &n_free, sb->d[i]); } SWAP_PCTRIE_REMOVE(&object->un_pager.swp.swp_blks, sb->p); uma_zfree(swblk_zone, sb); } swp_pager_freeswapspace(s_free, n_free); } /* * SWP_PAGER_METACTL() - misc control of swap meta data. * * This routine is capable of looking up, or removing swapblk * assignments in the swap meta data. It returns the swapblk being * looked-up, popped, or SWAPBLK_NONE if the block was invalid. * * When acting on a busy resident page and paging is in progress, we * have to wait until paging is complete but otherwise can act on the * busy page. */ static daddr_t swp_pager_meta_lookup(vm_object_t object, vm_pindex_t pindex) { struct swblk *sb; VM_OBJECT_ASSERT_LOCKED(object); /* * The meta data only exists if the object is OBJT_SWAP * and even then might not be allocated yet. */ KASSERT(object->type == OBJT_SWAP, ("Lookup object not swappable")); sb = SWAP_PCTRIE_LOOKUP(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL) return (SWAPBLK_NONE); return (sb->d[pindex % SWAP_META_PAGES]); } /* * Returns the least page index which is greater than or equal to the * parameter pindex and for which there is a swap block allocated. * Returns object's size if the object's type is not swap or if there * are no allocated swap blocks for the object after the requested * pindex. */ vm_pindex_t swap_pager_find_least(vm_object_t object, vm_pindex_t pindex) { struct swblk *sb; int i; VM_OBJECT_ASSERT_LOCKED(object); if (object->type != OBJT_SWAP) return (object->size); sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, rounddown(pindex, SWAP_META_PAGES)); if (sb == NULL) return (object->size); if (sb->p < pindex) { for (i = pindex % SWAP_META_PAGES; i < SWAP_META_PAGES; i++) { if (sb->d[i] != SWAPBLK_NONE) return (sb->p + i); } sb = SWAP_PCTRIE_LOOKUP_GE(&object->un_pager.swp.swp_blks, roundup(pindex, SWAP_META_PAGES)); if (sb == NULL) return (object->size); } for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->d[i] != SWAPBLK_NONE) return (sb->p + i); } /* * We get here if a swblk is present in the trie but it * doesn't map any blocks. */ MPASS(0); return (object->size); } /* * System call swapon(name) enables swapping on device name, * which must be in the swdevsw. Return EBUSY * if already swapping on this device. */ #ifndef _SYS_SYSPROTO_H_ struct swapon_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int sys_swapon(struct thread *td, struct swapon_args *uap) { struct vattr attr; struct vnode *vp; struct nameidata nd; int error; error = priv_check(td, PRIV_SWAPON); if (error) return (error); sx_xlock(&swdev_syscall_lock); /* * Swap metadata may not fit in the KVM if we have physical * memory of >1GB. */ if (swblk_zone == NULL) { error = ENOMEM; goto done; } NDINIT(&nd, LOOKUP, ISOPEN | FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; if (vn_isdisk(vp, &error)) { error = swapongeom(vp); } else if (vp->v_type == VREG && (vp->v_mount->mnt_vfc->vfc_flags & VFCF_NETWORK) != 0 && (error = VOP_GETATTR(vp, &attr, td->td_ucred)) == 0) { /* * Allow direct swapping to NFS regular files in the same * way that nfs_mountroot() sets up diskless swapping. */ error = swaponvp(td, vp, attr.va_size / DEV_BSIZE); } if (error) vrele(vp); done: sx_xunlock(&swdev_syscall_lock); return (error); } /* * Check that the total amount of swap currently configured does not * exceed half the theoretical maximum. If it does, print a warning * message. */ static void swapon_check_swzone(void) { /* recommend using no more than half that amount */ if (swap_total > swap_maxpages / 2) { printf("warning: total configured swap (%lu pages) " "exceeds maximum recommended amount (%lu pages).\n", swap_total, swap_maxpages / 2); printf("warning: increase kern.maxswzone " "or reduce amount of swap.\n"); } } static void swaponsomething(struct vnode *vp, void *id, u_long nblks, sw_strategy_t *strategy, sw_close_t *close, dev_t dev, int flags) { struct swdevt *sp, *tsp; swblk_t dvbase; u_long mblocks; /* * nblks is in DEV_BSIZE'd chunks, convert to PAGE_SIZE'd chunks. * First chop nblks off to page-align it, then convert. * * sw->sw_nblks is in page-sized chunks now too. */ nblks &= ~(ctodb(1) - 1); nblks = dbtoc(nblks); /* * If we go beyond this, we get overflows in the radix * tree bitmap code. */ mblocks = 0x40000000 / BLIST_META_RADIX; if (nblks > mblocks) { printf( "WARNING: reducing swap size to maximum of %luMB per unit\n", mblocks / 1024 / 1024 * PAGE_SIZE); nblks = mblocks; } sp = malloc(sizeof *sp, M_VMPGDATA, M_WAITOK | M_ZERO); sp->sw_vp = vp; sp->sw_id = id; sp->sw_dev = dev; sp->sw_nblks = nblks; sp->sw_used = 0; sp->sw_strategy = strategy; sp->sw_close = close; sp->sw_flags = flags; sp->sw_blist = blist_create(nblks, M_WAITOK); /* * Do not free the first blocks in order to avoid overwriting * any bsd label at the front of the partition */ blist_free(sp->sw_blist, howmany(BBSIZE, PAGE_SIZE), nblks - howmany(BBSIZE, PAGE_SIZE)); dvbase = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(tsp, &swtailq, sw_list) { if (tsp->sw_end >= dvbase) { /* * We put one uncovered page between the devices * in order to definitively prevent any cross-device * I/O requests */ dvbase = tsp->sw_end + 1; } } sp->sw_first = dvbase; sp->sw_end = dvbase + nblks; TAILQ_INSERT_TAIL(&swtailq, sp, sw_list); nswapdev++; swap_pager_avail += nblks - howmany(BBSIZE, PAGE_SIZE); swap_total += nblks; swapon_check_swzone(); swp_sizecheck(); mtx_unlock(&sw_dev_mtx); EVENTHANDLER_INVOKE(swapon, sp); } /* * SYSCALL: swapoff(devname) * * Disable swapping on the given device. * * XXX: Badly designed system call: it should use a device index * rather than filename as specification. We keep sw_vp around * only to make this work. */ #ifndef _SYS_SYSPROTO_H_ struct swapoff_args { char *name; }; #endif /* * MPSAFE */ /* ARGSUSED */ int sys_swapoff(struct thread *td, struct swapoff_args *uap) { struct vnode *vp; struct nameidata nd; struct swdevt *sp; int error; error = priv_check(td, PRIV_SWAPOFF); if (error) return (error); sx_xlock(&swdev_syscall_lock); NDINIT(&nd, LOOKUP, FOLLOW | AUDITVNODE1, UIO_USERSPACE, uap->name, td); error = namei(&nd); if (error) goto done; NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_vp == vp) break; } mtx_unlock(&sw_dev_mtx); if (sp == NULL) { error = EINVAL; goto done; } error = swapoff_one(sp, td->td_ucred); done: sx_xunlock(&swdev_syscall_lock); return (error); } static int swapoff_one(struct swdevt *sp, struct ucred *cred) { u_long nblks; #ifdef MAC int error; #endif sx_assert(&swdev_syscall_lock, SA_XLOCKED); #ifdef MAC (void) vn_lock(sp->sw_vp, LK_EXCLUSIVE | LK_RETRY); error = mac_system_check_swapoff(cred, sp->sw_vp); (void) VOP_UNLOCK(sp->sw_vp, 0); if (error != 0) return (error); #endif nblks = sp->sw_nblks; /* * We can turn off this swap device safely only if the * available virtual memory in the system will fit the amount * of data we will have to page back in, plus an epsilon so * the system doesn't become critically low on swap space. */ if (vm_free_count() + swap_pager_avail < nblks + nswap_lowat) return (ENOMEM); /* * Prevent further allocations on this device. */ mtx_lock(&sw_dev_mtx); sp->sw_flags |= SW_CLOSING; swap_pager_avail -= blist_fill(sp->sw_blist, 0, nblks); swap_total -= nblks; mtx_unlock(&sw_dev_mtx); /* * Page in the contents of the device and close it. */ swap_pager_swapoff(sp); sp->sw_close(curthread, sp); mtx_lock(&sw_dev_mtx); sp->sw_id = NULL; TAILQ_REMOVE(&swtailq, sp, sw_list); nswapdev--; if (nswapdev == 0) { swap_pager_full = 2; swap_pager_almost_full = 1; } if (swdevhd == sp) swdevhd = NULL; mtx_unlock(&sw_dev_mtx); blist_destroy(sp->sw_blist); free(sp, M_VMPGDATA); return (0); } void swapoff_all(void) { struct swdevt *sp, *spt; const char *devname; int error; sx_xlock(&swdev_syscall_lock); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH_SAFE(sp, &swtailq, sw_list, spt) { mtx_unlock(&sw_dev_mtx); if (vn_isdisk(sp->sw_vp, NULL)) devname = devtoname(sp->sw_vp->v_rdev); else devname = "[file]"; error = swapoff_one(sp, thread0.td_ucred); if (error != 0) { printf("Cannot remove swap device %s (error=%d), " "skipping.\n", devname, error); } else if (bootverbose) { printf("Swap device %s removed.\n", devname); } mtx_lock(&sw_dev_mtx); } mtx_unlock(&sw_dev_mtx); sx_xunlock(&swdev_syscall_lock); } void swap_pager_status(int *total, int *used) { struct swdevt *sp; *total = 0; *used = 0; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { *total += sp->sw_nblks; *used += sp->sw_used; } mtx_unlock(&sw_dev_mtx); } int swap_dev_info(int name, struct xswdev *xs, char *devname, size_t len) { struct swdevt *sp; const char *tmp_devname; int error, n; n = 0; error = ENOENT; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (n != name) { n++; continue; } xs->xsw_version = XSWDEV_VERSION; xs->xsw_dev = sp->sw_dev; xs->xsw_flags = sp->sw_flags; xs->xsw_nblks = sp->sw_nblks; xs->xsw_used = sp->sw_used; if (devname != NULL) { if (vn_isdisk(sp->sw_vp, NULL)) tmp_devname = devtoname(sp->sw_vp->v_rdev); else tmp_devname = "[file]"; strncpy(devname, tmp_devname, len); } error = 0; break; } mtx_unlock(&sw_dev_mtx); return (error); } #if defined(COMPAT_FREEBSD11) #define XSWDEV_VERSION_11 1 struct xswdev11 { u_int xsw_version; uint32_t xsw_dev; int xsw_flags; int xsw_nblks; int xsw_used; }; #endif #if defined(__amd64__) && defined(COMPAT_FREEBSD32) struct xswdev32 { u_int xsw_version; u_int xsw_dev1, xsw_dev2; int xsw_flags; int xsw_nblks; int xsw_used; }; #endif static int sysctl_vm_swap_info(SYSCTL_HANDLER_ARGS) { struct xswdev xs; #if defined(__amd64__) && defined(COMPAT_FREEBSD32) struct xswdev32 xs32; #endif #if defined(COMPAT_FREEBSD11) struct xswdev11 xs11; #endif int error; if (arg2 != 1) /* name length */ return (EINVAL); error = swap_dev_info(*(int *)arg1, &xs, NULL, 0); if (error != 0) return (error); #if defined(__amd64__) && defined(COMPAT_FREEBSD32) if (req->oldlen == sizeof(xs32)) { xs32.xsw_version = XSWDEV_VERSION; xs32.xsw_dev1 = xs.xsw_dev; xs32.xsw_dev2 = xs.xsw_dev >> 32; xs32.xsw_flags = xs.xsw_flags; xs32.xsw_nblks = xs.xsw_nblks; xs32.xsw_used = xs.xsw_used; error = SYSCTL_OUT(req, &xs32, sizeof(xs32)); return (error); } #endif #if defined(COMPAT_FREEBSD11) if (req->oldlen == sizeof(xs11)) { xs11.xsw_version = XSWDEV_VERSION_11; xs11.xsw_dev = xs.xsw_dev; /* truncation */ xs11.xsw_flags = xs.xsw_flags; xs11.xsw_nblks = xs.xsw_nblks; xs11.xsw_used = xs.xsw_used; error = SYSCTL_OUT(req, &xs11, sizeof(xs11)); return (error); } #endif error = SYSCTL_OUT(req, &xs, sizeof(xs)); return (error); } SYSCTL_INT(_vm, OID_AUTO, nswapdev, CTLFLAG_RD, &nswapdev, 0, "Number of swap devices"); SYSCTL_NODE(_vm, OID_AUTO, swap_info, CTLFLAG_RD | CTLFLAG_MPSAFE, sysctl_vm_swap_info, "Swap statistics by device"); /* * Count the approximate swap usage in pages for a vmspace. The * shadowed or not yet copied on write swap blocks are not accounted. * The map must be locked. */ long vmspace_swap_count(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t cur; vm_object_t object; struct swblk *sb; vm_pindex_t e, pi; long count; int i; map = &vmspace->vm_map; count = 0; VM_MAP_ENTRY_FOREACH(cur, map) { if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; object = cur->object.vm_object; if (object == NULL || object->type != OBJT_SWAP) continue; VM_OBJECT_RLOCK(object); if (object->type != OBJT_SWAP) goto unlock; pi = OFF_TO_IDX(cur->offset); e = pi + OFF_TO_IDX(cur->end - cur->start); for (;; pi = sb->p + SWAP_META_PAGES) { sb = SWAP_PCTRIE_LOOKUP_GE( &object->un_pager.swp.swp_blks, pi); if (sb == NULL || sb->p >= e) break; for (i = 0; i < SWAP_META_PAGES; i++) { if (sb->p + i < e && sb->d[i] != SWAPBLK_NONE) count++; } } unlock: VM_OBJECT_RUNLOCK(object); } return (count); } /* * GEOM backend * * Swapping onto disk devices. * */ static g_orphan_t swapgeom_orphan; static struct g_class g_swap_class = { .name = "SWAP", .version = G_VERSION, .orphan = swapgeom_orphan, }; DECLARE_GEOM_CLASS(g_swap_class, g_class); static void swapgeom_close_ev(void *arg, int flags) { struct g_consumer *cp; cp = arg; g_access(cp, -1, -1, 0); g_detach(cp); g_destroy_consumer(cp); } /* * Add a reference to the g_consumer for an inflight transaction. */ static void swapgeom_acquire(struct g_consumer *cp) { mtx_assert(&sw_dev_mtx, MA_OWNED); cp->index++; } /* * Remove a reference from the g_consumer. Post a close event if all * references go away, since the function might be called from the * biodone context. */ static void swapgeom_release(struct g_consumer *cp, struct swdevt *sp) { mtx_assert(&sw_dev_mtx, MA_OWNED); cp->index--; if (cp->index == 0) { if (g_post_event(swapgeom_close_ev, cp, M_NOWAIT, NULL) == 0) sp->sw_id = NULL; } } static void swapgeom_done(struct bio *bp2) { struct swdevt *sp; struct buf *bp; struct g_consumer *cp; bp = bp2->bio_caller2; cp = bp2->bio_from; bp->b_ioflags = bp2->bio_flags; if (bp2->bio_error) bp->b_ioflags |= BIO_ERROR; bp->b_resid = bp->b_bcount - bp2->bio_completed; bp->b_error = bp2->bio_error; bp->b_caller1 = NULL; bufdone(bp); sp = bp2->bio_caller1; mtx_lock(&sw_dev_mtx); swapgeom_release(cp, sp); mtx_unlock(&sw_dev_mtx); g_destroy_bio(bp2); } static void swapgeom_strategy(struct buf *bp, struct swdevt *sp) { struct bio *bio; struct g_consumer *cp; mtx_lock(&sw_dev_mtx); cp = sp->sw_id; if (cp == NULL) { mtx_unlock(&sw_dev_mtx); bp->b_error = ENXIO; bp->b_ioflags |= BIO_ERROR; bufdone(bp); return; } swapgeom_acquire(cp); mtx_unlock(&sw_dev_mtx); if (bp->b_iocmd == BIO_WRITE) bio = g_new_bio(); else bio = g_alloc_bio(); if (bio == NULL) { mtx_lock(&sw_dev_mtx); swapgeom_release(cp, sp); mtx_unlock(&sw_dev_mtx); bp->b_error = ENOMEM; bp->b_ioflags |= BIO_ERROR; printf("swap_pager: cannot allocate bio\n"); bufdone(bp); return; } bp->b_caller1 = bio; bio->bio_caller1 = sp; bio->bio_caller2 = bp; bio->bio_cmd = bp->b_iocmd; bio->bio_offset = (bp->b_blkno - sp->sw_first) * PAGE_SIZE; bio->bio_length = bp->b_bcount; bio->bio_done = swapgeom_done; if (!buf_mapped(bp)) { bio->bio_ma = bp->b_pages; bio->bio_data = unmapped_buf; bio->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; bio->bio_ma_n = bp->b_npages; bio->bio_flags |= BIO_UNMAPPED; } else { bio->bio_data = bp->b_data; bio->bio_ma = NULL; } g_io_request(bio, cp); return; } static void swapgeom_orphan(struct g_consumer *cp) { struct swdevt *sp; int destroy; mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_id == cp) { sp->sw_flags |= SW_CLOSING; break; } } /* * Drop reference we were created with. Do directly since we're in a * special context where we don't have to queue the call to * swapgeom_close_ev(). */ cp->index--; destroy = ((sp != NULL) && (cp->index == 0)); if (destroy) sp->sw_id = NULL; mtx_unlock(&sw_dev_mtx); if (destroy) swapgeom_close_ev(cp, 0); } static void swapgeom_close(struct thread *td, struct swdevt *sw) { struct g_consumer *cp; mtx_lock(&sw_dev_mtx); cp = sw->sw_id; sw->sw_id = NULL; mtx_unlock(&sw_dev_mtx); /* * swapgeom_close() may be called from the biodone context, * where we cannot perform topology changes. Delegate the * work to the events thread. */ if (cp != NULL) g_waitfor_event(swapgeom_close_ev, cp, M_WAITOK, NULL); } static int swapongeom_locked(struct cdev *dev, struct vnode *vp) { struct g_provider *pp; struct g_consumer *cp; static struct g_geom *gp; struct swdevt *sp; u_long nblks; int error; pp = g_dev_getprovider(dev); if (pp == NULL) return (ENODEV); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { cp = sp->sw_id; if (cp != NULL && cp->provider == pp) { mtx_unlock(&sw_dev_mtx); return (EBUSY); } } mtx_unlock(&sw_dev_mtx); if (gp == NULL) gp = g_new_geomf(&g_swap_class, "swap"); cp = g_new_consumer(gp); cp->index = 1; /* Number of active I/Os, plus one for being active. */ cp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE; g_attach(cp, pp); /* * XXX: Every time you think you can improve the margin for * footshooting, somebody depends on the ability to do so: * savecore(8) wants to write to our swapdev so we cannot * set an exclusive count :-( */ error = g_access(cp, 1, 1, 0); if (error != 0) { g_detach(cp); g_destroy_consumer(cp); return (error); } nblks = pp->mediasize / DEV_BSIZE; swaponsomething(vp, cp, nblks, swapgeom_strategy, swapgeom_close, dev2udev(dev), (pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ? SW_UNMAPPED : 0); return (0); } static int swapongeom(struct vnode *vp) { int error; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (vp->v_type != VCHR || VN_IS_DOOMED(vp)) { error = ENOENT; } else { g_topology_lock(); error = swapongeom_locked(vp->v_rdev, vp); g_topology_unlock(); } VOP_UNLOCK(vp, 0); return (error); } /* * VNODE backend * * This is used mainly for network filesystem (read: probably only tested * with NFS) swapfiles. * */ static void swapdev_strategy(struct buf *bp, struct swdevt *sp) { struct vnode *vp2; bp->b_blkno = ctodb(bp->b_blkno - sp->sw_first); vp2 = sp->sw_id; vhold(vp2); if (bp->b_iocmd == BIO_WRITE) { if (bp->b_bufobj) bufobj_wdrop(bp->b_bufobj); bufobj_wref(&vp2->v_bufobj); } if (bp->b_bufobj != &vp2->v_bufobj) bp->b_bufobj = &vp2->v_bufobj; bp->b_vp = vp2; bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); return; } static void swapdev_close(struct thread *td, struct swdevt *sp) { VOP_CLOSE(sp->sw_vp, FREAD | FWRITE, td->td_ucred, td); vrele(sp->sw_vp); } static int swaponvp(struct thread *td, struct vnode *vp, u_long nblks) { struct swdevt *sp; int error; if (nblks == 0) return (ENXIO); mtx_lock(&sw_dev_mtx); TAILQ_FOREACH(sp, &swtailq, sw_list) { if (sp->sw_id == vp) { mtx_unlock(&sw_dev_mtx); return (EBUSY); } } mtx_unlock(&sw_dev_mtx); (void) vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); #ifdef MAC error = mac_system_check_swapon(td->td_ucred, vp); if (error == 0) #endif error = VOP_OPEN(vp, FREAD | FWRITE, td->td_ucred, td, NULL); (void) VOP_UNLOCK(vp, 0); if (error) return (error); swaponsomething(vp, vp, nblks, swapdev_strategy, swapdev_close, NODEV, 0); return (0); } static int sysctl_swap_async_max(SYSCTL_HANDLER_ARGS) { int error, new, n; new = nsw_wcount_async_max; error = sysctl_handle_int(oidp, &new, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (new > nswbuf / 2 || new < 1) return (EINVAL); mtx_lock(&swbuf_mtx); while (nsw_wcount_async_max != new) { /* * Adjust difference. If the current async count is too low, * we will need to sqeeze our update slowly in. Sleep with a * higher priority than getpbuf() to finish faster. */ n = new - nsw_wcount_async_max; if (nsw_wcount_async + n >= 0) { nsw_wcount_async += n; nsw_wcount_async_max += n; wakeup(&nsw_wcount_async); } else { nsw_wcount_async_max -= nsw_wcount_async; nsw_wcount_async = 0; msleep(&nsw_wcount_async, &swbuf_mtx, PSWP, "swpsysctl", 0); } } mtx_unlock(&swbuf_mtx); return (0); } static void swap_pager_update_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { VM_OBJECT_WLOCK(object); KASSERT((object->flags & OBJ_ANON) == 0, ("Splittable object with writecount")); object->un_pager.swp.writemappings += (vm_ooffset_t)end - start; VM_OBJECT_WUNLOCK(object); } static void swap_pager_release_writecount(vm_object_t object, vm_offset_t start, vm_offset_t end) { VM_OBJECT_WLOCK(object); KASSERT((object->flags & OBJ_ANON) == 0, ("Splittable object with writecount")); object->un_pager.swp.writemappings -= (vm_ooffset_t)end - start; VM_OBJECT_WUNLOCK(object); } Index: head/sys/vm/vm_fault.c =================================================================== --- head/sys/vm/vm_fault.c (revision 356156) +++ head/sys/vm/vm_fault.c (revision 356157) @@ -1,1913 +1,1899 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Page fault handling module. */ #include __FBSDID("$FreeBSD$"); #include "opt_ktrace.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #include #include #define PFBAK 4 #define PFFOR 4 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) #define VM_FAULT_DONTNEED_MIN 1048576 struct faultstate { vm_page_t m; vm_object_t object; vm_pindex_t pindex; vm_page_t first_m; vm_object_t first_object; vm_pindex_t first_pindex; vm_map_t map; vm_map_entry_t entry; int map_generation; bool lookup_still_valid; struct vnode *vp; }; static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead); static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked); static int vm_pfault_oom_attempts = 3; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, &vm_pfault_oom_attempts, 0, "Number of page allocation attempts in page fault handler before it " "triggers OOM handling"); static int vm_pfault_oom_wait = 10; SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, &vm_pfault_oom_wait, 0, "Number of seconds to wait for free pages before retrying " "the page fault handler"); static inline void fault_page_release(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { /* * We are likely to loop around again and attempt to busy * this page. Deactivating it leaves it available for * pageout while optimizing fault restarts. */ - vm_page_lock(m); vm_page_deactivate(m); - vm_page_unlock(m); vm_page_xunbusy(m); *mp = NULL; } } static inline void fault_page_free(vm_page_t *mp) { vm_page_t m; m = *mp; if (m != NULL) { VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_wired(m)) vm_page_free(m); else vm_page_xunbusy(m); *mp = NULL; } } static inline void unlock_map(struct faultstate *fs) { if (fs->lookup_still_valid) { vm_map_lookup_done(fs->map, fs->entry); fs->lookup_still_valid = false; } } static void unlock_vp(struct faultstate *fs) { if (fs->vp != NULL) { vput(fs->vp); fs->vp = NULL; } } static void fault_deallocate(struct faultstate *fs) { fault_page_release(&fs->m); vm_object_pip_wakeup(fs->object); if (fs->object != fs->first_object) { VM_OBJECT_WLOCK(fs->first_object); fault_page_free(&fs->first_m); VM_OBJECT_WUNLOCK(fs->first_object); vm_object_pip_wakeup(fs->first_object); } vm_object_deallocate(fs->first_object); unlock_map(fs); unlock_vp(fs); } static void unlock_and_deallocate(struct faultstate *fs) { VM_OBJECT_WUNLOCK(fs->object); fault_deallocate(fs); } static void vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, vm_prot_t fault_type, int fault_flags) { bool need_dirty; if (((prot & VM_PROT_WRITE) == 0 && (fault_flags & VM_FAULT_DIRTY) == 0) || (m->oflags & VPO_UNMANAGED) != 0) return; VM_PAGE_OBJECT_BUSY_ASSERT(m); need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && (fault_flags & VM_FAULT_WIRE) == 0) || (fault_flags & VM_FAULT_DIRTY) != 0; vm_object_set_writeable_dirty(m->object); /* * If the fault is a write, we know that this page is being * written NOW so dirty it explicitly to save on * pmap_is_modified() calls later. * * Also, since the page is now dirty, we can possibly tell * the pager to release any swap backing the page. */ if (need_dirty && vm_page_set_dirty(m) == 0) { /* * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC * if the page is already dirty to prevent data written with * the expectation of being synced from not being synced. * Likewise if this entry does not request NOSYNC then make * sure the page isn't marked NOSYNC. Applications sharing * data should use the same flags to avoid ping ponging. */ if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) vm_page_aflag_set(m, PGA_NOSYNC); else vm_page_aflag_clear(m, PGA_NOSYNC); } } /* * Unlocks fs.first_object and fs.map on success. */ static int vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { vm_page_t m, m_map; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 vm_page_t m_super; int flags; #endif int psind, rv; MPASS(fs->vp == NULL); vm_object_busy(fs->first_object); m = vm_page_lookup(fs->first_object, fs->first_pindex); /* A busy page can be mapped for read|execute access. */ if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)) || !vm_page_all_valid(m)) { rv = KERN_FAILURE; goto out; } m_map = m; psind = 0; #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ VM_NRESERVLEVEL > 0 if ((m->flags & PG_FICTITIOUS) == 0 && (m_super = vm_reserv_to_superpage(m)) != NULL && rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & (pagesizes[m_super->psind] - 1)) && !wired && pmap_ps_enabled(fs->map->pmap)) { flags = PS_ALL_VALID; if ((prot & VM_PROT_WRITE) != 0) { /* * Create a superpage mapping allowing write access * only if none of the constituent pages are busy and * all of them are already dirty (except possibly for * the page that was faulted on). */ flags |= PS_NONE_BUSY; if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) flags |= PS_ALL_DIRTY; } if (vm_page_ps_test(m_super, flags, m)) { m_map = m_super; psind = m_super->psind; vaddr = rounddown2(vaddr, pagesizes[psind]); /* Preset the modified bit for dirty superpages. */ if ((flags & PS_ALL_DIRTY) != 0) fault_type |= VM_PROT_WRITE; } } #endif rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); if (rv != KERN_SUCCESS) goto out; if (m_hold != NULL) { *m_hold = m; vm_page_wire(m); } vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags); if (psind == 0 && !wired) vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); VM_OBJECT_RUNLOCK(fs->first_object); vm_map_lookup_done(fs->map, fs->entry); curthread->td_ru.ru_minflt++; out: vm_object_unbusy(fs->first_object); return (rv); } static void vm_fault_restore_map_lock(struct faultstate *fs) { VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); if (!vm_map_trylock_read(fs->map)) { VM_OBJECT_WUNLOCK(fs->first_object); vm_map_lock_read(fs->map); VM_OBJECT_WLOCK(fs->first_object); } fs->lookup_still_valid = true; } static void vm_fault_populate_check_page(vm_page_t m) { /* * Check each page to ensure that the pager is obeying the * interface: the page must be installed in the object, fully * valid, and exclusively busied. */ MPASS(m != NULL); MPASS(vm_page_all_valid(m)); MPASS(vm_page_xbusied(m)); } static void vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, vm_pindex_t last) { vm_page_t m; vm_pindex_t pidx; VM_OBJECT_ASSERT_WLOCKED(object); MPASS(first <= last); for (pidx = first, m = vm_page_lookup(object, pidx); pidx <= last; pidx++, m = vm_page_next(m)) { vm_fault_populate_check_page(m); - vm_page_lock(m); vm_page_deactivate(m); - vm_page_unlock(m); vm_page_xunbusy(m); } } static int vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) { - struct mtx *m_mtx; vm_offset_t vaddr; vm_page_t m; vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; int i, npages, psind, rv; MPASS(fs->object == fs->first_object); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); MPASS(fs->first_object->backing_object == NULL); MPASS(fs->lookup_still_valid); pager_first = OFF_TO_IDX(fs->entry->offset); pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; unlock_map(fs); unlock_vp(fs); /* * Call the pager (driver) populate() method. * * There is no guarantee that the method will be called again * if the current fault is for read, and a future fault is * for write. Report the entry's maximum allowed protection * to the driver. */ rv = vm_pager_populate(fs->first_object, fs->first_pindex, fault_type, fs->entry->max_protection, &pager_first, &pager_last); VM_OBJECT_ASSERT_WLOCKED(fs->first_object); if (rv == VM_PAGER_BAD) { /* * VM_PAGER_BAD is the backdoor for a pager to request * normal fault handling. */ vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ return (KERN_NOT_RECEIVER); } if (rv != VM_PAGER_OK) return (KERN_FAILURE); /* AKA SIGSEGV */ /* Ensure that the driver is obeying the interface. */ MPASS(pager_first <= pager_last); MPASS(fs->first_pindex <= pager_last); MPASS(fs->first_pindex >= pager_first); MPASS(pager_last < fs->first_object->size); vm_fault_restore_map_lock(fs); if (fs->map->timestamp != fs->map_generation) { vm_fault_populate_cleanup(fs->first_object, pager_first, pager_last); return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ } /* * The map is unchanged after our last unlock. Process the fault. * * The range [pager_first, pager_last] that is given to the * pager is only a hint. The pager may populate any range * within the object that includes the requested page index. * In case the pager expanded the range, clip it to fit into * the map entry. */ map_first = OFF_TO_IDX(fs->entry->offset); if (map_first > pager_first) { vm_fault_populate_cleanup(fs->first_object, pager_first, map_first - 1); pager_first = map_first; } map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; if (map_last < pager_last) { vm_fault_populate_cleanup(fs->first_object, map_last + 1, pager_last); pager_last = map_last; } for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); pidx <= pager_last; pidx += npages, m = vm_page_next(&m[npages - 1])) { vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) psind = m->psind; if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || !pmap_ps_enabled(fs->map->pmap) || wired)) psind = 0; #else psind = 0; #endif npages = atop(pagesizes[psind]); for (i = 0; i < npages; i++) { vm_fault_populate_check_page(&m[i]); vm_fault_dirty(fs->entry, &m[i], prot, fault_type, fault_flags); } VM_OBJECT_WUNLOCK(fs->first_object); rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), psind); #if defined(__amd64__) if (psind > 0 && rv == KERN_FAILURE) { for (i = 0; i < npages; i++) { rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), &m[i], prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); MPASS(rv == KERN_SUCCESS); } } #else MPASS(rv == KERN_SUCCESS); #endif VM_OBJECT_WLOCK(fs->first_object); - m_mtx = NULL; for (i = 0; i < npages; i++) { - if ((fault_flags & VM_FAULT_WIRE) != 0) { + if ((fault_flags & VM_FAULT_WIRE) != 0) vm_page_wire(&m[i]); - } else { - vm_page_change_lock(&m[i], &m_mtx); + else vm_page_activate(&m[i]); - } if (m_hold != NULL && m[i].pindex == fs->first_pindex) { *m_hold = &m[i]; vm_page_wire(&m[i]); } vm_page_xunbusy(&m[i]); } - if (m_mtx != NULL) - mtx_unlock(m_mtx); } curthread->td_ru.ru_majflt++; return (KERN_SUCCESS); } static int prot_fault_translation; SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, &prot_fault_translation, 0, "Control signal to deliver on protection fault"); /* compat definition to keep common code for signal translation */ #define UCODE_PAGEFLT 12 #ifdef T_PAGEFLT _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); #endif /* * vm_fault_trap: * * Handle a page fault occurring at the given address, * requiring the given permissions, in the map specified. * If successful, the page is inserted into the * associated physical map. * * NOTE: the given address should be truncated to the * proper page address. * * KERN_SUCCESS is returned if the page fault is handled; otherwise, * a standard error specifying why the fault is fatal is returned. * * The map in question must be referenced, and remains so. * Caller may hold no locks. */ int vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, int *signo, int *ucode) { int result; MPASS(signo == NULL || ucode != NULL); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) ktrfault(vaddr, fault_type); #endif result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, NULL); KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || result == KERN_INVALID_ADDRESS || result == KERN_RESOURCE_SHORTAGE || result == KERN_PROTECTION_FAILURE || result == KERN_OUT_OF_BOUNDS, ("Unexpected Mach error %d from vm_fault()", result)); #ifdef KTRACE if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) ktrfaultend(result); #endif if (result != KERN_SUCCESS && signo != NULL) { switch (result) { case KERN_FAILURE: case KERN_INVALID_ADDRESS: *signo = SIGSEGV; *ucode = SEGV_MAPERR; break; case KERN_RESOURCE_SHORTAGE: *signo = SIGBUS; *ucode = BUS_OOMERR; break; case KERN_OUT_OF_BOUNDS: *signo = SIGBUS; *ucode = BUS_OBJERR; break; case KERN_PROTECTION_FAILURE: if (prot_fault_translation == 0) { /* * Autodetect. This check also covers * the images without the ABI-tag ELF * note. */ if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && curproc->p_osrel >= P_OSREL_SIGSEGV) { *signo = SIGSEGV; *ucode = SEGV_ACCERR; } else { *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } } else if (prot_fault_translation == 1) { /* Always compat mode. */ *signo = SIGBUS; *ucode = UCODE_PAGEFLT; } else { /* Always SIGSEGV mode. */ *signo = SIGSEGV; *ucode = SEGV_ACCERR; } break; default: KASSERT(0, ("Unexpected Mach error %d from vm_fault()", result)); break; } } return (result); } static int vm_fault_lock_vnode(struct faultstate *fs) { struct vnode *vp; int error, locked; if (fs->object->type != OBJT_VNODE) return (KERN_SUCCESS); vp = fs->object->handle; if (vp == fs->vp) { ASSERT_VOP_LOCKED(vp, "saved vnode is not locked"); return (KERN_SUCCESS); } /* * Perform an unlock in case the desired vnode changed while * the map was unlocked during a retry. */ unlock_vp(fs); locked = VOP_ISLOCKED(vp); if (locked != LK_EXCLUSIVE) locked = LK_SHARED; /* * We must not sleep acquiring the vnode lock while we have * the page exclusive busied or the object's * paging-in-progress count incremented. Otherwise, we could * deadlock. */ error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread); if (error == 0) { fs->vp = vp; return (KERN_SUCCESS); } vhold(vp); unlock_and_deallocate(fs); error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread); vdrop(vp); fs->vp = vp; KASSERT(error == 0, ("vm_fault: vget failed %d", error)); return (KERN_RESOURCE_SHORTAGE); } /* * Wait/Retry if the page is busy. We have to do this if the page is * either exclusive or shared busy because the vm_pager may be using * read busy for pageouts (and even pageins if it is the vnode pager), * and we could end up trying to pagein and pageout the same page * simultaneously. * * We can theoretically allow the busy case on a read fault if the page * is marked valid, but since such pages are typically already pmap'd, * putting that special case in might be more effort then it is worth. * We cannot under any circumstances mess around with a shared busied * page except, perhaps, to pmap it. */ static void vm_fault_busy_sleep(struct faultstate *fs) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(fs->m, PGA_REFERENCED); if (fs->object != fs->first_object) { fault_page_release(&fs->first_m); vm_object_pip_wakeup(fs->first_object); } vm_object_pip_wakeup(fs->object); unlock_map(fs); if (fs->m == vm_page_lookup(fs->object, fs->pindex)) vm_page_busy_sleep(fs->m, "vmpfw", false); else VM_OBJECT_WUNLOCK(fs->object); VM_CNT_INC(v_intrans); vm_object_deallocate(fs->first_object); } int vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags, vm_page_t *m_hold) { struct faultstate fs; struct domainset *dset; vm_object_t next_object, retry_object; vm_offset_t e_end, e_start; vm_pindex_t retry_pindex; vm_prot_t prot, retry_prot; int ahead, alloc_req, behind, cluster_offset, era, faultcount; int nera, oom, result, rv; u_char behavior; boolean_t wired; /* Passed by reference. */ bool dead, hardfault, is_first_object_locked; VM_CNT_INC(v_vm_faults); if ((curthread->td_pflags & TDP_NOFAULTING) != 0) return (KERN_PROTECTION_FAILURE); fs.vp = NULL; faultcount = 0; nera = -1; hardfault = false; RetryFault: oom = 0; RetryFault_oom: /* * Find the backing store object and offset into it to begin the * search. */ fs.map = map; result = vm_map_lookup(&fs.map, vaddr, fault_type | VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); if (result != KERN_SUCCESS) { unlock_vp(&fs); return (result); } fs.map_generation = fs.map->timestamp; if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { panic("%s: fault on nofault entry, addr: %#lx", __func__, (u_long)vaddr); } if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && fs.entry->wiring_thread != curthread) { vm_map_unlock_read(fs.map); vm_map_lock(fs.map); if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { unlock_vp(&fs); fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; vm_map_unlock_and_wait(fs.map, 0); } else vm_map_unlock(fs.map); goto RetryFault; } MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); if (wired) fault_type = prot | (fault_type & VM_PROT_COPY); else KASSERT((fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); /* * Try to avoid lock contention on the top-level object through * special-case handling of some types of page faults, specifically, * those that are mapping an existing page from the top-level object. * Under this condition, a read lock on the object suffices, allowing * multiple page faults of a similar type to run in parallel. */ if (fs.vp == NULL /* avoid locked vnode leak */ && (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) { VM_OBJECT_RLOCK(fs.first_object); rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, fault_flags, wired, m_hold); if (rv == KERN_SUCCESS) return (rv); if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { VM_OBJECT_RUNLOCK(fs.first_object); VM_OBJECT_WLOCK(fs.first_object); } } else { VM_OBJECT_WLOCK(fs.first_object); } /* * Make a reference to this object to prevent its disposal while we * are messing with it. Once we have the reference, the map is free * to be diddled. Since objects reference their shadows (and copies), * they will stay around as well. * * Bump the paging-in-progress count to prevent size changes (e.g. * truncation operations) during I/O. */ vm_object_reference_locked(fs.first_object); vm_object_pip_add(fs.first_object, 1); fs.lookup_still_valid = true; fs.m = fs.first_m = NULL; /* * Search for the page at object/offset. */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; while (TRUE) { KASSERT(fs.m == NULL, ("page still set %p at loop start", fs.m)); /* * If the object is marked for imminent termination, * we retry here, since the collapse pass has raced * with us. Otherwise, if we see terminally dead * object, return fail. */ if ((fs.object->flags & OBJ_DEAD) != 0) { dead = fs.object->type == OBJT_DEAD; unlock_and_deallocate(&fs); if (dead) return (KERN_PROTECTION_FAILURE); pause("vmf_de", 1); goto RetryFault; } /* * See if page is resident */ fs.m = vm_page_lookup(fs.object, fs.pindex); if (fs.m != NULL) { if (vm_page_tryxbusy(fs.m) == 0) { vm_fault_busy_sleep(&fs); goto RetryFault; } /* * The page is marked busy for other processes and the * pagedaemon. If it still isn't completely valid * (readable), jump to readrest, else break-out ( we * found the page ). */ if (!vm_page_all_valid(fs.m)) goto readrest; break; /* break to PAGE HAS BEEN FOUND */ } KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); /* * Page is not resident. If the pager might contain the page * or this is the beginning of the search, allocate a new * page. (Default objects are zero-fill, so there is no real * pager for them.) */ if (fs.object->type != OBJT_DEFAULT || fs.object == fs.first_object) { if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) { rv = vm_fault_lock_vnode(&fs); MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE); if (rv == KERN_RESOURCE_SHORTAGE) goto RetryFault; } if (fs.pindex >= fs.object->size) { unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } if (fs.object == fs.first_object && (fs.first_object->flags & OBJ_POPULATE) != 0 && fs.first_object->shadow_count == 0) { rv = vm_fault_populate(&fs, prot, fault_type, fault_flags, wired, m_hold); switch (rv) { case KERN_SUCCESS: case KERN_FAILURE: unlock_and_deallocate(&fs); return (rv); case KERN_RESOURCE_SHORTAGE: unlock_and_deallocate(&fs); goto RetryFault; case KERN_NOT_RECEIVER: /* * Pager's populate() method * returned VM_PAGER_BAD. */ break; default: panic("inconsistent return codes"); } } /* * Allocate a new page for this object/offset pair. * * Unlocked read of the p_flag is harmless. At * worst, the P_KILLED might be not observed * there, and allocation can fail, causing * restart and new reading of the p_flag. */ dset = fs.object->domain.dr_policy; if (dset == NULL) dset = curthread->td_domain.dr_policy; if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { #if VM_NRESERVLEVEL > 0 vm_object_color(fs.object, atop(vaddr) - fs.pindex); #endif alloc_req = P_KILLED(curproc) ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; if (fs.object->type != OBJT_VNODE && fs.object->backing_object == NULL) alloc_req |= VM_ALLOC_ZERO; fs.m = vm_page_alloc(fs.object, fs.pindex, alloc_req); } if (fs.m == NULL) { unlock_and_deallocate(&fs); if (vm_pfault_oom_attempts < 0 || oom < vm_pfault_oom_attempts) { oom++; vm_waitpfault(dset, vm_pfault_oom_wait * hz); goto RetryFault_oom; } if (bootverbose) printf( "proc %d (%s) failed to alloc page on fault, starting OOM\n", curproc->p_pid, curproc->p_comm); vm_pageout_oom(VM_OOM_MEM_PF); goto RetryFault; } } readrest: /* * At this point, we have either allocated a new page or found * an existing page that is only partially valid. * * We hold a reference on the current object and the page is * exclusive busied. */ /* * If the pager for the current object might have the page, * then determine the number of additional pages to read and * potentially reprioritize previously read pages for earlier * reclamation. These operations should only be performed * once per page fault. Even if the current pager doesn't * have the page, the number of additional pages to read will * apply to subsequent objects in the shadow chain. */ if (fs.object->type != OBJT_DEFAULT && nera == -1 && !P_KILLED(curproc)) { KASSERT(fs.lookup_still_valid, ("map unlocked")); era = fs.entry->read_ahead; behavior = vm_map_entry_behavior(fs.entry); if (behavior == MAP_ENTRY_BEHAV_RANDOM) { nera = 0; } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { nera = VM_FAULT_READ_AHEAD_MAX; if (vaddr == fs.entry->next_read) vm_fault_dontneed(&fs, vaddr, nera); } else if (vaddr == fs.entry->next_read) { /* * This is a sequential fault. Arithmetically * increase the requested number of pages in * the read-ahead window. The requested * number of pages is "# of sequential faults * x (read ahead min + 1) + read ahead min" */ nera = VM_FAULT_READ_AHEAD_MIN; if (era > 0) { nera += era + 1; if (nera > VM_FAULT_READ_AHEAD_MAX) nera = VM_FAULT_READ_AHEAD_MAX; } if (era == VM_FAULT_READ_AHEAD_MAX) vm_fault_dontneed(&fs, vaddr, nera); } else { /* * This is a non-sequential fault. */ nera = 0; } if (era != nera) { /* * A read lock on the map suffices to update * the read ahead count safely. */ fs.entry->read_ahead = nera; } /* * Prepare for unlocking the map. Save the map * entry's start and end addresses, which are used to * optimize the size of the pager operation below. * Even if the map entry's addresses change after * unlocking the map, using the saved addresses is * safe. */ e_start = fs.entry->start; e_end = fs.entry->end; } /* * Call the pager to retrieve the page if there is a chance * that the pager has it, and potentially retrieve additional * pages at the same time. */ if (fs.object->type != OBJT_DEFAULT) { /* * Release the map lock before locking the vnode or * sleeping in the pager. (If the current object has * a shadow, then an earlier iteration of this loop * may have already unlocked the map.) */ unlock_map(&fs); rv = vm_fault_lock_vnode(&fs); MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE); if (rv == KERN_RESOURCE_SHORTAGE) goto RetryFault; KASSERT(fs.vp == NULL || !fs.map->system_map, ("vm_fault: vnode-backed object mapped by system map")); /* * Page in the requested page and hint the pager, * that it may bring up surrounding pages. */ if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || P_KILLED(curproc)) { behind = 0; ahead = 0; } else { /* Is this a sequential fault? */ if (nera > 0) { behind = 0; ahead = nera; } else { /* * Request a cluster of pages that is * aligned to a VM_FAULT_READ_DEFAULT * page offset boundary within the * object. Alignment to a page offset * boundary is more likely to coincide * with the underlying file system * block than alignment to a virtual * address boundary. */ cluster_offset = fs.pindex % VM_FAULT_READ_DEFAULT; behind = ulmin(cluster_offset, atop(vaddr - e_start)); ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; } ahead = ulmin(ahead, atop(e_end - vaddr) - 1); } rv = vm_pager_get_pages(fs.object, &fs.m, 1, &behind, &ahead); if (rv == VM_PAGER_OK) { faultcount = behind + 1 + ahead; hardfault = true; break; /* break to PAGE HAS BEEN FOUND */ } if (rv == VM_PAGER_ERROR) printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); /* * If an I/O error occurred or the requested page was * outside the range of the pager, clean up and return * an error. */ if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { fault_page_free(&fs.m); unlock_and_deallocate(&fs); return (KERN_OUT_OF_BOUNDS); } } /* * The requested page does not exist at this object/ * offset. Remove the invalid page from the object, * waking up anyone waiting for it, and continue on to * the next object. However, if this is the top-level * object, we must leave the busy page in place to * prevent another process from rushing past us, and * inserting the page in that object at the same time * that we are. */ if (fs.object == fs.first_object) { fs.first_m = fs.m; fs.m = NULL; } else fault_page_free(&fs.m); /* * Move on to the next object. Lock the next object before * unlocking the current one. */ next_object = fs.object->backing_object; if (next_object == NULL) { /* * If there's no object left, fill the page in the top * object with zeros. */ if (fs.object != fs.first_object) { vm_object_pip_wakeup(fs.object); VM_OBJECT_WUNLOCK(fs.object); fs.object = fs.first_object; fs.pindex = fs.first_pindex; VM_OBJECT_WLOCK(fs.object); } MPASS(fs.first_m != NULL); MPASS(fs.m == NULL); fs.m = fs.first_m; fs.first_m = NULL; /* * Zero the page if necessary and mark it valid. */ if ((fs.m->flags & PG_ZERO) == 0) { pmap_zero_page(fs.m); } else { VM_CNT_INC(v_ozfod); } VM_CNT_INC(v_zfod); vm_page_valid(fs.m); /* Don't try to prefault neighboring pages. */ faultcount = 1; break; /* break to PAGE HAS BEEN FOUND */ } else { MPASS(fs.first_m != NULL); KASSERT(fs.object != next_object, ("object loop %p", next_object)); VM_OBJECT_WLOCK(next_object); vm_object_pip_add(next_object, 1); if (fs.object != fs.first_object) vm_object_pip_wakeup(fs.object); fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); VM_OBJECT_WUNLOCK(fs.object); fs.object = next_object; } } vm_page_assert_xbusied(fs.m); /* * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock * is held.] */ /* * If the page is being written, but isn't already owned by the * top-level object, we have to copy it into a new page owned by the * top-level object. */ if (fs.object != fs.first_object) { /* * We only really need to copy if we want to write it. */ if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { /* * This allows pages to be virtually copied from a * backing_object into the first_object, where the * backing object has no other refs to it, and cannot * gain any more refs. Instead of a bcopy, we just * move the page from the backing object to the * first object. Note that we must mark the page * dirty in the first object so that it will go out * to swap when needed. */ is_first_object_locked = false; if ( /* * Only one shadow object */ (fs.object->shadow_count == 1) && /* * No COW refs, except us */ (fs.object->ref_count == 1) && /* * No one else can look this object up */ (fs.object->handle == NULL) && /* * No other ways to look the object up */ ((fs.object->flags & OBJ_ANON) != 0) && (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && /* * We don't chase down the shadow chain */ fs.object == fs.first_object->backing_object) { /* * Remove but keep xbusy for replace. fs.m is * moved into fs.first_object and left busy * while fs.first_m is conditionally freed. */ vm_page_remove_xbusy(fs.m); vm_page_replace(fs.m, fs.first_object, fs.first_pindex, fs.first_m); vm_page_dirty(fs.m); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(fs.m, fs.first_object, fs.object, OFF_TO_IDX( fs.first_object->backing_object_offset)); #endif VM_OBJECT_WUNLOCK(fs.object); fs.first_m = fs.m; fs.m = NULL; VM_CNT_INC(v_cow_optim); } else { VM_OBJECT_WUNLOCK(fs.object); /* * Oh, well, lets copy it. */ pmap_copy_page(fs.m, fs.first_m); vm_page_valid(fs.first_m); if (wired && (fault_flags & VM_FAULT_WIRE) == 0) { vm_page_wire(fs.first_m); vm_page_unwire(fs.m, PQ_INACTIVE); } /* * We no longer need the old page or object. */ fault_page_release(&fs.m); } /* * fs.object != fs.first_object due to above * conditional */ vm_object_pip_wakeup(fs.object); /* * We only try to prefault read-only mappings to the * neighboring pages when this copy-on-write fault is * a hard fault. In other cases, trying to prefault * is typically wasted effort. */ if (faultcount == 0) faultcount = 1; /* * Only use the new page below... */ fs.object = fs.first_object; fs.pindex = fs.first_pindex; fs.m = fs.first_m; if (!is_first_object_locked) VM_OBJECT_WLOCK(fs.object); VM_CNT_INC(v_cow_faults); curthread->td_cow++; } else { prot &= ~VM_PROT_WRITE; } } /* * We must verify that the maps have not changed since our last * lookup. */ if (!fs.lookup_still_valid) { if (!vm_map_trylock_read(fs.map)) { unlock_and_deallocate(&fs); goto RetryFault; } fs.lookup_still_valid = true; if (fs.map->timestamp != fs.map_generation) { result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); /* * If we don't need the page any longer, put it on the inactive * list (the easiest thing to do here). If no one needs it, * pageout will grab it eventually. */ if (result != KERN_SUCCESS) { unlock_and_deallocate(&fs); /* * If retry of map lookup would have blocked then * retry fault from start. */ if (result == KERN_FAILURE) goto RetryFault; return (result); } if ((retry_object != fs.first_object) || (retry_pindex != fs.first_pindex)) { unlock_and_deallocate(&fs); goto RetryFault; } /* * Check whether the protection has changed or the object has * been copied while we left the map unlocked. Changing from * read to write permission is OK - we leave the page * write-protected, and catch the write fault. Changing from * write to read permission means that we can't mark the page * write-enabled after all. */ prot &= retry_prot; fault_type &= retry_prot; if (prot == 0) { unlock_and_deallocate(&fs); goto RetryFault; } /* Reassert because wired may have changed. */ KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, ("!wired && VM_FAULT_WIRE")); } } /* * If the page was filled by a pager, save the virtual address that * should be faulted on next under a sequential access pattern to the * map entry. A read lock on the map suffices to update this address * safely. */ if (hardfault) fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; vm_page_assert_xbusied(fs.m); vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags); /* * Page must be completely valid or it is not fit to * map into user space. vm_pager_get_pages() ensures this. */ KASSERT(vm_page_all_valid(fs.m), ("vm_fault: page %p partially invalid", fs.m)); VM_OBJECT_WUNLOCK(fs.object); /* * Put this page into the physical map. We had to do the unlock above * because pmap_enter() may sleep. We don't put the page * back on the active queue until later so that the pageout daemon * won't find it (yet). */ pmap_enter(fs.map->pmap, vaddr, fs.m, prot, fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && wired == 0) vm_fault_prefault(&fs, vaddr, faultcount > 0 ? behind : PFBAK, faultcount > 0 ? ahead : PFFOR, false); /* * If the page is not wired down, then put it where the pageout daemon * can find it. */ - if ((fault_flags & VM_FAULT_WIRE) != 0) { + if ((fault_flags & VM_FAULT_WIRE) != 0) vm_page_wire(fs.m); - } else { - vm_page_lock(fs.m); + else vm_page_activate(fs.m); - vm_page_unlock(fs.m); - } if (m_hold != NULL) { *m_hold = fs.m; vm_page_wire(fs.m); } vm_page_xunbusy(fs.m); fs.m = NULL; /* * Unlock everything, and return */ fault_deallocate(&fs); if (hardfault) { VM_CNT_INC(v_io_faults); curthread->td_ru.ru_majflt++; #ifdef RACCT if (racct_enable && fs.object->type == OBJT_VNODE) { PROC_LOCK(curproc); if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { racct_add_force(curproc, RACCT_WRITEBPS, PAGE_SIZE + behind * PAGE_SIZE); racct_add_force(curproc, RACCT_WRITEIOPS, 1); } else { racct_add_force(curproc, RACCT_READBPS, PAGE_SIZE + ahead * PAGE_SIZE); racct_add_force(curproc, RACCT_READIOPS, 1); } PROC_UNLOCK(curproc); } #endif } else curthread->td_ru.ru_minflt++; return (KERN_SUCCESS); } /* * Speed up the reclamation of pages that precede the faulting pindex within * the first object of the shadow chain. Essentially, perform the equivalent * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes * the faulting pindex by the cluster size when the pages read by vm_fault() * cross a cluster-size boundary. The cluster size is the greater of the * smallest superpage size and VM_FAULT_DONTNEED_MIN. * * When "fs->first_object" is a shadow object, the pages in the backing object * that precede the faulting pindex are deactivated by vm_fault(). So, this * function must only be concerned with pages in the first object. */ static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) { vm_map_entry_t entry; vm_object_t first_object, object; vm_offset_t end, start; vm_page_t m, m_next; vm_pindex_t pend, pstart; vm_size_t size; object = fs->object; VM_OBJECT_ASSERT_WLOCKED(object); first_object = fs->first_object; if (first_object != object) { if (!VM_OBJECT_TRYWLOCK(first_object)) { VM_OBJECT_WUNLOCK(object); VM_OBJECT_WLOCK(first_object); VM_OBJECT_WLOCK(object); } } /* Neither fictitious nor unmanaged pages can be reclaimed. */ if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { size = VM_FAULT_DONTNEED_MIN; if (MAXPAGESIZES > 1 && size < pagesizes[1]) size = pagesizes[1]; end = rounddown2(vaddr, size); if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && (entry = fs->entry)->start < end) { if (end - entry->start < size) start = entry->start; else start = end - size; pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); pstart = OFF_TO_IDX(entry->offset) + atop(start - entry->start); m_next = vm_page_find_least(first_object, pstart); pend = OFF_TO_IDX(entry->offset) + atop(end - entry->start); while ((m = m_next) != NULL && m->pindex < pend) { m_next = TAILQ_NEXT(m, listq); if (!vm_page_all_valid(m) || vm_page_busied(m)) continue; /* * Don't clear PGA_REFERENCED, since it would * likely represent a reference by a different * process. * * Typically, at this point, prefetched pages * are still in the inactive queue. Only * pages that triggered page faults are in the - * active queue. + * active queue. The test for whether the page + * is in the inactive queue is racy; in the + * worst case we will requeue the page + * unnecessarily. */ - vm_page_lock(m); if (!vm_page_inactive(m)) vm_page_deactivate(m); - vm_page_unlock(m); } } } if (first_object != object) VM_OBJECT_WUNLOCK(first_object); } /* * vm_fault_prefault provides a quick way of clustering * pagefaults into a processes address space. It is a "cousin" * of vm_map_pmap_enter, except it runs at page fault time instead * of mmap time. */ static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, int backward, int forward, bool obj_locked) { pmap_t pmap; vm_map_entry_t entry; vm_object_t backing_object, lobject; vm_offset_t addr, starta; vm_pindex_t pindex; vm_page_t m; int i; pmap = fs->map->pmap; if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) return; entry = fs->entry; if (addra < backward * PAGE_SIZE) { starta = entry->start; } else { starta = addra - backward * PAGE_SIZE; if (starta < entry->start) starta = entry->start; } /* * Generate the sequence of virtual addresses that are candidates for * prefaulting in an outward spiral from the faulting virtual address, * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... * If the candidate address doesn't have a backing physical page, then * the loop immediately terminates. */ for (i = 0; i < 2 * imax(backward, forward); i++) { addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : PAGE_SIZE); if (addr > addra + forward * PAGE_SIZE) addr = 0; if (addr < starta || addr >= entry->end) continue; if (!pmap_is_prefaultable(pmap, addr)) continue; pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; lobject = entry->object.vm_object; if (!obj_locked) VM_OBJECT_RLOCK(lobject); while ((m = vm_page_lookup(lobject, pindex)) == NULL && lobject->type == OBJT_DEFAULT && (backing_object = lobject->backing_object) != NULL) { KASSERT((lobject->backing_object_offset & PAGE_MASK) == 0, ("vm_fault_prefault: unaligned object offset")); pindex += lobject->backing_object_offset >> PAGE_SHIFT; VM_OBJECT_RLOCK(backing_object); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); lobject = backing_object; } if (m == NULL) { if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); break; } if (vm_page_all_valid(m) && (m->flags & PG_FICTITIOUS) == 0) pmap_enter_quick(pmap, addr, m, entry->protection); if (!obj_locked || lobject != entry->object.vm_object) VM_OBJECT_RUNLOCK(lobject); } } /* * Hold each of the physical pages that are mapped by the specified range of * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid * and allow the specified types of access, "prot". If all of the implied * pages are successfully held, then the number of held pages is returned * together with pointers to those pages in the array "ma". However, if any * of the pages cannot be held, -1 is returned. */ int vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, vm_prot_t prot, vm_page_t *ma, int max_count) { vm_offset_t end, va; vm_page_t *mp; int count; boolean_t pmap_failed; if (len == 0) return (0); end = round_page(addr + len); addr = trunc_page(addr); /* * Check for illegal addresses. */ if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) return (-1); if (atop(end - addr) > max_count) panic("vm_fault_quick_hold_pages: count > max_count"); count = atop(end - addr); /* * Most likely, the physical pages are resident in the pmap, so it is * faster to try pmap_extract_and_hold() first. */ pmap_failed = FALSE; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { *mp = pmap_extract_and_hold(map->pmap, va, prot); if (*mp == NULL) pmap_failed = TRUE; else if ((prot & VM_PROT_WRITE) != 0 && (*mp)->dirty != VM_PAGE_BITS_ALL) { /* * Explicitly dirty the physical page. Otherwise, the * caller's changes may go unnoticed because they are * performed through an unmanaged mapping or by a DMA * operation. * * The object lock is not held here. * See vm_page_clear_dirty_mask(). */ vm_page_dirty(*mp); } } if (pmap_failed) { /* * One or more pages could not be held by the pmap. Either no * page was mapped at the specified virtual address or that * mapping had insufficient permissions. Attempt to fault in * and hold these pages. * * If vm_fault_disable_pagefaults() was called, * i.e., TDP_NOFAULTING is set, we must not sleep nor * acquire MD VM locks, which means we must not call * vm_fault(). Some (out of tree) callers mark * too wide a code area with vm_fault_disable_pagefaults() * already, use the VM_PROT_QUICK_NOFAULT flag to request * the proper behaviour explicitly. */ if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && (curthread->td_pflags & TDP_NOFAULTING) != 0) goto error; for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) if (*mp == NULL && vm_fault(map, va, prot, VM_FAULT_NORMAL, mp) != KERN_SUCCESS) goto error; } return (count); error: for (mp = ma; mp < ma + count; mp++) if (*mp != NULL) vm_page_unwire(*mp, PQ_INACTIVE); return (-1); } /* * Routine: * vm_fault_copy_entry * Function: * Create new shadow object backing dst_entry with private copy of * all underlying pages. When src_entry is equal to dst_entry, * function implements COW for wired-down map entry. Otherwise, * it forks wired entry into dst_map. * * In/out conditions: * The source and destination maps must be locked for write. * The source map entry must be wired down (or be a sharing map * entry corresponding to a main map entry that is wired down). */ void vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, vm_map_entry_t dst_entry, vm_map_entry_t src_entry, vm_ooffset_t *fork_charge) { vm_object_t backing_object, dst_object, object, src_object; vm_pindex_t dst_pindex, pindex, src_pindex; vm_prot_t access, prot; vm_offset_t vaddr; vm_page_t dst_m; vm_page_t src_m; boolean_t upgrade; #ifdef lint src_map++; #endif /* lint */ upgrade = src_entry == dst_entry; access = prot = dst_entry->protection; src_object = src_entry->object.vm_object; src_pindex = OFF_TO_IDX(src_entry->offset); if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { dst_object = src_object; vm_object_reference(dst_object); } else { /* * Create the top-level object for the destination entry. * Doesn't actually shadow anything - we copy the pages * directly. */ dst_object = vm_object_allocate_anon(atop(dst_entry->end - dst_entry->start), NULL, NULL, 0); #if VM_NRESERVLEVEL > 0 dst_object->flags |= OBJ_COLORED; dst_object->pg_color = atop(dst_entry->start); #endif dst_object->domain = src_object->domain; dst_object->charge = dst_entry->end - dst_entry->start; } VM_OBJECT_WLOCK(dst_object); KASSERT(upgrade || dst_entry->object.vm_object == NULL, ("vm_fault_copy_entry: vm_object not NULL")); if (src_object != dst_object) { dst_entry->object.vm_object = dst_object; dst_entry->offset = 0; dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; } if (fork_charge != NULL) { KASSERT(dst_entry->cred == NULL, ("vm_fault_copy_entry: leaked swp charge")); dst_object->cred = curthread->td_ucred; crhold(dst_object->cred); *fork_charge += dst_object->charge; } else if ((dst_object->type == OBJT_DEFAULT || dst_object->type == OBJT_SWAP) && dst_object->cred == NULL) { KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", dst_entry)); dst_object->cred = dst_entry->cred; dst_entry->cred = NULL; } /* * If not an upgrade, then enter the mappings in the pmap as * read and/or execute accesses. Otherwise, enter them as * write accesses. * * A writeable large page mapping is only created if all of * the constituent small page mappings are modified. Marking * PTEs as modified on inception allows promotion to happen * without taking potentially large number of soft faults. */ if (!upgrade) access &= ~VM_PROT_WRITE; /* * Loop through all of the virtual pages within the entry's * range, copying each page from the source object to the * destination object. Since the source is wired, those pages * must exist. In contrast, the destination is pageable. * Since the destination object doesn't share any backing storage * with the source object, all of its pages must be dirtied, * regardless of whether they can be written. */ for (vaddr = dst_entry->start, dst_pindex = 0; vaddr < dst_entry->end; vaddr += PAGE_SIZE, dst_pindex++) { again: /* * Find the page in the source object, and copy it in. * Because the source is wired down, the page will be * in memory. */ if (src_object != dst_object) VM_OBJECT_RLOCK(src_object); object = src_object; pindex = src_pindex + dst_pindex; while ((src_m = vm_page_lookup(object, pindex)) == NULL && (backing_object = object->backing_object) != NULL) { /* * Unless the source mapping is read-only or * it is presently being upgraded from * read-only, the first object in the shadow * chain should provide all of the pages. In * other words, this loop body should never be * executed when the source mapping is already * read/write. */ KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || upgrade, ("vm_fault_copy_entry: main object missing page")); VM_OBJECT_RLOCK(backing_object); pindex += OFF_TO_IDX(object->backing_object_offset); if (object != dst_object) VM_OBJECT_RUNLOCK(object); object = backing_object; } KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); if (object != dst_object) { /* * Allocate a page in the destination object. */ dst_m = vm_page_alloc(dst_object, (src_object == dst_object ? src_pindex : 0) + dst_pindex, VM_ALLOC_NORMAL); if (dst_m == NULL) { VM_OBJECT_WUNLOCK(dst_object); VM_OBJECT_RUNLOCK(object); vm_wait(dst_object); VM_OBJECT_WLOCK(dst_object); goto again; } pmap_copy_page(src_m, dst_m); VM_OBJECT_RUNLOCK(object); dst_m->dirty = dst_m->valid = src_m->valid; } else { dst_m = src_m; if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) goto again; if (dst_m->pindex >= dst_object->size) { /* * We are upgrading. Index can occur * out of bounds if the object type is * vnode and the file was truncated. */ vm_page_xunbusy(dst_m); break; } } VM_OBJECT_WUNLOCK(dst_object); /* * Enter it in the pmap. If a wired, copy-on-write * mapping is being replaced by a write-enabled * mapping, then wire that new mapping. * * The page can be invalid if the user called * msync(MS_INVALIDATE) or truncated the backing vnode * or shared memory object. In this case, do not * insert it into pmap, but still do the copy so that * all copies of the wired map entry have similar * backing pages. */ if (vm_page_all_valid(dst_m)) { pmap_enter(dst_map->pmap, vaddr, dst_m, prot, access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); } /* * Mark it no longer busy, and put it on the active list. */ VM_OBJECT_WLOCK(dst_object); if (upgrade) { if (src_m != dst_m) { vm_page_unwire(src_m, PQ_INACTIVE); vm_page_wire(dst_m); } else { KASSERT(vm_page_wired(dst_m), ("dst_m %p is not wired", dst_m)); } } else { - vm_page_lock(dst_m); vm_page_activate(dst_m); - vm_page_unlock(dst_m); } vm_page_xunbusy(dst_m); } VM_OBJECT_WUNLOCK(dst_object); if (upgrade) { dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); vm_object_deallocate(src_object); } } /* * Block entry into the machine-independent layer's page fault handler by * the calling thread. Subsequent calls to vm_fault() by that thread will * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of * spurious page faults. */ int vm_fault_disable_pagefaults(void) { return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); } void vm_fault_enable_pagefaults(int save) { curthread_pflags_restore(save); } Index: head/sys/vm/vm_object.c =================================================================== --- head/sys/vm/vm_object.c (revision 356156) +++ head/sys/vm/vm_object.c (revision 356157) @@ -1,2763 +1,2756 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Virtual memory object module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include /* for curproc, pageproc */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int old_msync; SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0, "Use old (insecure) msync behavior"); static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *allclean, boolean_t *eio); static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean); static void vm_object_qcollapse(vm_object_t object); static void vm_object_vndeallocate(vm_object_t object); static void vm_object_backing_remove(vm_object_t object); /* * Virtual memory objects maintain the actual data * associated with allocated virtual memory. A given * page of memory exists within exactly one object. * * An object is only deallocated when all "references" * are given up. Only one "reference" to a given * region of an object should be writeable. * * Associated with each object is a list of all resident * memory pages belonging to that object; this list is * maintained by the "vm_page" module, and locked by the object's * lock. * * Each object also records a "pager" routine which is * used to retrieve (and store) pages to the proper backing * storage. In addition, objects may be backed by other * objects from which they were virtual-copied. * * The only items within the object structure which are * modified after time of creation are: * reference count locked by object's lock * pager routine locked by object's lock * */ struct object_q vm_object_list; struct mtx vm_object_list_mtx; /* lock for object list and count */ struct vm_object kernel_object_store; static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats"); static counter_u64_t object_collapses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD, &object_collapses, "VM object collapses"); static counter_u64_t object_bypasses = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD, &object_bypasses, "VM object bypasses"); static void counter_startup(void) { object_collapses = counter_u64_alloc(M_WAITOK); object_bypasses = counter_u64_alloc(M_WAITOK); } SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); static uma_zone_t obj_zone; static int vm_object_zinit(void *mem, int size, int flags); #ifdef INVARIANTS static void vm_object_zdtor(void *mem, int size, void *arg); static void vm_object_zdtor(void *mem, int size, void *arg) { vm_object_t object; object = (vm_object_t)mem; KASSERT(object->ref_count == 0, ("object %p ref_count = %d", object, object->ref_count)); KASSERT(TAILQ_EMPTY(&object->memq), ("object %p has resident pages in its memq", object)); KASSERT(vm_radix_is_empty(&object->rtree), ("object %p has resident pages in its trie", object)); #if VM_NRESERVLEVEL > 0 KASSERT(LIST_EMPTY(&object->rvq), ("object %p has reservations", object)); #endif KASSERT(REFCOUNT_COUNT(object->paging_in_progress) == 0, ("object %p paging_in_progress = %d", object, REFCOUNT_COUNT(object->paging_in_progress))); KASSERT(object->busy == 0, ("object %p busy = %d", object, object->busy)); KASSERT(object->resident_page_count == 0, ("object %p resident_page_count = %d", object, object->resident_page_count)); KASSERT(object->shadow_count == 0, ("object %p shadow_count = %d", object, object->shadow_count)); KASSERT(object->type == OBJT_DEAD, ("object %p has non-dead type %d", object, object->type)); } #endif static int vm_object_zinit(void *mem, int size, int flags) { vm_object_t object; object = (vm_object_t)mem; rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW); /* These are true for any object that has been freed */ object->type = OBJT_DEAD; vm_radix_init(&object->rtree); refcount_init(&object->ref_count, 0); refcount_init(&object->paging_in_progress, 0); refcount_init(&object->busy, 0); object->resident_page_count = 0; object->shadow_count = 0; object->flags = OBJ_DEAD; mtx_lock(&vm_object_list_mtx); TAILQ_INSERT_TAIL(&vm_object_list, object, object_list); mtx_unlock(&vm_object_list_mtx); return (0); } static void _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags, vm_object_t object, void *handle) { TAILQ_INIT(&object->memq); LIST_INIT(&object->shadow_head); object->type = type; if (type == OBJT_SWAP) pctrie_init(&object->un_pager.swp.swp_blks); /* * Ensure that swap_pager_swapoff() iteration over object_list * sees up to date type and pctrie head if it observed * non-dead object. */ atomic_thread_fence_rel(); object->pg_color = 0; object->flags = flags; object->size = size; object->domain.dr_policy = NULL; object->generation = 1; object->cleangeneration = 1; refcount_init(&object->ref_count, 1); object->memattr = VM_MEMATTR_DEFAULT; object->cred = NULL; object->charge = 0; object->handle = handle; object->backing_object = NULL; object->backing_object_offset = (vm_ooffset_t) 0; #if VM_NRESERVLEVEL > 0 LIST_INIT(&object->rvq); #endif umtx_shm_object_init(object); } /* * vm_object_init: * * Initialize the VM objects module. */ void vm_object_init(void) { TAILQ_INIT(&vm_object_list); mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF); rw_init(&kernel_object->lock, "kernel vm object"); _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL); #if VM_NRESERVLEVEL > 0 kernel_object->flags |= OBJ_COLORED; kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS); #endif /* * The lock portion of struct vm_object must be type stable due * to vm_pageout_fallback_object_lock locking a vm object * without holding any references to it. */ obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL, #ifdef INVARIANTS vm_object_zdtor, #else NULL, #endif vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); vm_radix_zinit(); } void vm_object_clear_flag(vm_object_t object, u_short bits) { VM_OBJECT_ASSERT_WLOCKED(object); object->flags &= ~bits; } /* * Sets the default memory attribute for the specified object. Pages * that are allocated to this object are by default assigned this memory * attribute. * * Presently, this function must be called before any pages are allocated * to the object. In the future, this requirement may be relaxed for * "default" and "swap" objects. */ int vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr) { VM_OBJECT_ASSERT_WLOCKED(object); switch (object->type) { case OBJT_DEFAULT: case OBJT_DEVICE: case OBJT_MGTDEVICE: case OBJT_PHYS: case OBJT_SG: case OBJT_SWAP: case OBJT_VNODE: if (!TAILQ_EMPTY(&object->memq)) return (KERN_FAILURE); break; case OBJT_DEAD: return (KERN_INVALID_ARGUMENT); default: panic("vm_object_set_memattr: object %p is of undefined type", object); } object->memattr = memattr; return (KERN_SUCCESS); } void vm_object_pip_add(vm_object_t object, short i) { refcount_acquiren(&object->paging_in_progress, i); } void vm_object_pip_wakeup(vm_object_t object) { refcount_release(&object->paging_in_progress); } void vm_object_pip_wakeupn(vm_object_t object, short i) { refcount_releasen(&object->paging_in_progress, i); } void vm_object_pip_wait(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_WLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); refcount_wait(&object->paging_in_progress, waitid, PVM); VM_OBJECT_WLOCK(object); } } void vm_object_pip_wait_unlocked(vm_object_t object, char *waitid) { VM_OBJECT_ASSERT_UNLOCKED(object); while (REFCOUNT_COUNT(object->paging_in_progress) > 0) refcount_wait(&object->paging_in_progress, waitid, PVM); } /* * vm_object_allocate: * * Returns a new object with the given size. */ vm_object_t vm_object_allocate(objtype_t type, vm_pindex_t size) { vm_object_t object; u_short flags; switch (type) { case OBJT_DEAD: panic("vm_object_allocate: can't create OBJT_DEAD"); case OBJT_DEFAULT: case OBJT_SWAP: flags = OBJ_COLORED; break; case OBJT_DEVICE: case OBJT_SG: flags = OBJ_FICTITIOUS | OBJ_UNMANAGED; break; case OBJT_MGTDEVICE: flags = OBJ_FICTITIOUS; break; case OBJT_PHYS: flags = OBJ_UNMANAGED; break; case OBJT_VNODE: flags = 0; break; default: panic("vm_object_allocate: type %d is undefined", type); } object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK); _vm_object_allocate(type, size, flags, object, NULL); return (object); } /* * vm_object_allocate_anon: * * Returns a new default object of the given size and marked as * anonymous memory for special split/collapse handling. Color * to be initialized by the caller. */ vm_object_t vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object, struct ucred *cred, vm_size_t charge) { vm_object_t handle, object; if (backing_object == NULL) handle = NULL; else if ((backing_object->flags & OBJ_ANON) != 0) handle = backing_object->handle; else handle = backing_object; object = uma_zalloc(obj_zone, M_WAITOK); _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING, object, handle); object->cred = cred; object->charge = cred != NULL ? charge : 0; return (object); } /* * vm_object_reference: * * Gets another reference to the given object. Note: OBJ_DEAD * objects can be referenced during final cleaning. */ void vm_object_reference(vm_object_t object) { struct vnode *vp; u_int old; if (object == NULL) return; /* * Many places assume exclusive access to objects with a single * ref. vm_object_collapse() in particular will directly mainpulate * references for objects in this state. vnode objects only need * the lock for the first ref to reference the vnode. */ if (!refcount_acquire_if_gt(&object->ref_count, object->type == OBJT_VNODE ? 0 : 1)) { VM_OBJECT_RLOCK(object); old = refcount_acquire(&object->ref_count); if (object->type == OBJT_VNODE && old == 0) { vp = object->handle; vref(vp); } VM_OBJECT_RUNLOCK(object); } } /* * vm_object_reference_locked: * * Gets another reference to the given object. * * The object must be locked. */ void vm_object_reference_locked(vm_object_t object) { struct vnode *vp; u_int old; VM_OBJECT_ASSERT_LOCKED(object); old = refcount_acquire(&object->ref_count); if (object->type == OBJT_VNODE && old == 0) { vp = object->handle; vref(vp); } } /* * Handle deallocating an object of type OBJT_VNODE. */ static void vm_object_vndeallocate(vm_object_t object) { struct vnode *vp = (struct vnode *) object->handle; bool last; KASSERT(object->type == OBJT_VNODE, ("vm_object_vndeallocate: not a vnode object")); KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp")); /* Object lock to protect handle lookup. */ last = refcount_release(&object->ref_count); VM_OBJECT_RUNLOCK(object); if (!last) return; if (!umtx_shm_vnobj_persistent) umtx_shm_object_terminated(object); /* vrele may need the vnode lock. */ vrele(vp); } /* * vm_object_deallocate: * * Release a reference to the specified object, * gained either through a vm_object_allocate * or a vm_object_reference call. When all references * are gone, storage associated with this object * may be relinquished. * * No object may be locked. */ void vm_object_deallocate(vm_object_t object) { vm_object_t robject, temp; bool released; while (object != NULL) { /* * If the reference count goes to 0 we start calling * vm_object_terminate() on the object chain. A ref count * of 1 may be a special case depending on the shadow count * being 0 or 1. These cases require a write lock on the * object. */ if ((object->flags & OBJ_ANON) == 0) released = refcount_release_if_gt(&object->ref_count, 1); else released = refcount_release_if_gt(&object->ref_count, 2); if (released) return; if (object->type == OBJT_VNODE) { VM_OBJECT_RLOCK(object); if (object->type == OBJT_VNODE) { vm_object_vndeallocate(object); return; } VM_OBJECT_RUNLOCK(object); } VM_OBJECT_WLOCK(object); KASSERT(object->ref_count > 0, ("vm_object_deallocate: object deallocated too many times: %d", object->type)); if (refcount_release(&object->ref_count)) goto doterm; if (object->ref_count > 1) { VM_OBJECT_WUNLOCK(object); return; } else if (object->ref_count == 1) { if (object->shadow_count == 0 && (object->flags & OBJ_ANON) != 0) { vm_object_set_flag(object, OBJ_ONEMAPPING); } else if (object->shadow_count == 1) { KASSERT((object->flags & OBJ_ANON) != 0, ("obj %p with shadow_count > 0 is not anon", object)); robject = LIST_FIRST(&object->shadow_head); KASSERT(robject != NULL, ("vm_object_deallocate: ref_count: %d, " "shadow_count: %d", object->ref_count, object->shadow_count)); KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object %p", object)); if (!VM_OBJECT_TRYWLOCK(robject)) { /* * Avoid a potential deadlock. */ refcount_acquire(&object->ref_count); VM_OBJECT_WUNLOCK(object); /* * More likely than not the thread * holding robject's lock has lower * priority than the current thread. * Let the lower priority thread run. */ pause("vmo_de", 1); continue; } /* * Collapse object into its shadow unless its * shadow is dead. In that case, object will * be deallocated by the thread that is * deallocating its shadow. */ if ((robject->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON) { refcount_acquire(&robject->ref_count); retry: if (REFCOUNT_COUNT(robject->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(object); vm_object_pip_wait(robject, "objde1"); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else if (REFCOUNT_COUNT(object->paging_in_progress) > 0) { VM_OBJECT_WUNLOCK(robject); VM_OBJECT_WUNLOCK(object); refcount_wait( &object->paging_in_progress, "objde2", PVM); VM_OBJECT_WLOCK(robject); temp = robject->backing_object; if (object == temp) { VM_OBJECT_WLOCK(object); goto retry; } } else VM_OBJECT_WUNLOCK(object); if (robject->ref_count == 1) { refcount_release(&robject->ref_count); object = robject; goto doterm; } object = robject; vm_object_collapse(object); VM_OBJECT_WUNLOCK(object); continue; } VM_OBJECT_WUNLOCK(robject); } VM_OBJECT_WUNLOCK(object); return; } doterm: umtx_shm_object_terminated(object); temp = object->backing_object; if (temp != NULL) { KASSERT((object->flags & OBJ_TMPFS_NODE) == 0, ("shadowed tmpfs v_object 2 %p", object)); vm_object_backing_remove(object); } /* * Don't double-terminate, we could be in a termination * recursion due to the terminate having to sync data * to disk. */ if ((object->flags & OBJ_DEAD) == 0) { vm_object_set_flag(object, OBJ_DEAD); vm_object_terminate(object); } else VM_OBJECT_WUNLOCK(object); object = temp; } } /* * vm_object_destroy removes the object from the global object list * and frees the space for the object. */ void vm_object_destroy(vm_object_t object) { /* * Release the allocation charge. */ if (object->cred != NULL) { swap_release_by_cred(object->charge, object->cred); object->charge = 0; crfree(object->cred); object->cred = NULL; } /* * Free the space for the object. */ uma_zfree(obj_zone, object); } static void vm_object_backing_remove_locked(vm_object_t object) { vm_object_t backing_object; backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(backing_object); if ((object->flags & OBJ_SHADOWLIST) != 0) { LIST_REMOVE(object, shadow_list); backing_object->shadow_count--; object->flags &= ~OBJ_SHADOWLIST; } object->backing_object = NULL; } static void vm_object_backing_remove(vm_object_t object) { vm_object_t backing_object; VM_OBJECT_ASSERT_WLOCKED(object); if ((object->flags & OBJ_SHADOWLIST) != 0) { backing_object = object->backing_object; VM_OBJECT_WLOCK(backing_object); vm_object_backing_remove_locked(object); VM_OBJECT_WUNLOCK(backing_object); } else object->backing_object = NULL; } static void vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object) { VM_OBJECT_ASSERT_WLOCKED(object); if ((backing_object->flags & OBJ_ANON) != 0) { VM_OBJECT_ASSERT_WLOCKED(backing_object); LIST_INSERT_HEAD(&backing_object->shadow_head, object, shadow_list); backing_object->shadow_count++; object->flags |= OBJ_SHADOWLIST; } object->backing_object = backing_object; } static void vm_object_backing_insert(vm_object_t object, vm_object_t backing_object) { VM_OBJECT_ASSERT_WLOCKED(object); if ((backing_object->flags & OBJ_ANON) != 0) { VM_OBJECT_WLOCK(backing_object); vm_object_backing_insert_locked(object, backing_object); VM_OBJECT_WUNLOCK(backing_object); } else object->backing_object = backing_object; } /* * vm_object_terminate_pages removes any remaining pageable pages * from the object and resets the object to an empty state. */ static void vm_object_terminate_pages(vm_object_t object) { vm_page_t p, p_next; VM_OBJECT_ASSERT_WLOCKED(object); /* * Free any remaining pageable pages. This also removes them from the * paging queues. However, don't free wired pages, just remove them * from the object. Rather than incrementally removing each page from * the object, the page and object are reset to any empty state. */ TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) { vm_page_assert_unbusied(p); KASSERT(p->object == object && (p->ref_count & VPRC_OBJREF) != 0, ("vm_object_terminate_pages: page %p is inconsistent", p)); p->object = NULL; if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) { VM_CNT_INC(v_pfree); vm_page_free(p); } } /* * If the object contained any pages, then reset it to an empty state. * None of the object's fields, including "resident_page_count", were * modified by the preceding loop. */ if (object->resident_page_count != 0) { vm_radix_reclaim_allnodes(&object->rtree); TAILQ_INIT(&object->memq); object->resident_page_count = 0; if (object->type == OBJT_VNODE) vdrop(object->handle); } } /* * vm_object_terminate actually destroys the specified object, freeing * up all previously used resources. * * The object must be locked. * This routine may block. */ void vm_object_terminate(vm_object_t object) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_DEAD) != 0, ("terminating non-dead obj %p", object)); /* * wait for the pageout daemon to be done with the object */ vm_object_pip_wait(object, "objtrm"); KASSERT(!REFCOUNT_COUNT(object->paging_in_progress), ("vm_object_terminate: pageout in progress")); KASSERT(object->ref_count == 0, ("vm_object_terminate: object with references, ref_count=%d", object->ref_count)); if ((object->flags & OBJ_PG_DTOR) == 0) vm_object_terminate_pages(object); #if VM_NRESERVLEVEL > 0 if (__predict_false(!LIST_EMPTY(&object->rvq))) vm_reserv_break_all(object); #endif KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT || object->type == OBJT_SWAP, ("%s: non-swap obj %p has cred", __func__, object)); /* * Let the pager know object is dead. */ vm_pager_deallocate(object); VM_OBJECT_WUNLOCK(object); vm_object_destroy(object); } /* * Make the page read-only so that we can clear the object flags. However, if * this is a nosync mmap then the object is likely to stay dirty so do not * mess with the page and do not clear the object flags. Returns TRUE if the * page should be flushed, and FALSE otherwise. */ static boolean_t vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean) { vm_page_assert_busied(p); /* * If we have been asked to skip nosync pages and this is a * nosync page, skip it. Note that the object flags were not * cleared in this case so we do not have to set them. */ if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) { *allclean = FALSE; return (FALSE); } else { pmap_remove_write(p); return (p->dirty != 0); } } /* * vm_object_page_clean * * Clean all dirty pages in the specified range of object. Leaves page * on whatever queue it is currently on. If NOSYNC is set then do not * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC), * leaving the object dirty. * * When stuffing pages asynchronously, allow clustering. XXX we need a * synchronous clustering mode implementation. * * Odd semantics: if start == end, we clean everything. * * The object must be locked. * * Returns FALSE if some page from the range was not written, as * reported by the pager, and TRUE otherwise. */ boolean_t vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end, int flags) { vm_page_t np, p; vm_pindex_t pi, tend, tstart; int curgeneration, n, pagerflags; boolean_t eio, res, allclean; VM_OBJECT_ASSERT_WLOCKED(object); if (object->type != OBJT_VNODE || !vm_object_mightbedirty(object) || object->resident_page_count == 0) return (TRUE); pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK; pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0; tstart = OFF_TO_IDX(start); tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK); allclean = tstart == 0 && tend >= object->size; res = TRUE; rescan: curgeneration = object->generation; for (p = vm_page_find_least(object, tstart); p != NULL; p = np) { pi = p->pindex; if (pi >= tend) break; np = TAILQ_NEXT(p, listq); if (vm_page_none_valid(p)) continue; if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) { if (object->generation != curgeneration && (flags & OBJPC_SYNC) != 0) goto rescan; np = vm_page_find_least(object, pi); continue; } if (!vm_object_page_remove_write(p, flags, &allclean)) { vm_page_xunbusy(p); continue; } n = vm_object_page_collect_flush(object, p, pagerflags, flags, &allclean, &eio); if (eio) { res = FALSE; allclean = FALSE; } if (object->generation != curgeneration && (flags & OBJPC_SYNC) != 0) goto rescan; /* * If the VOP_PUTPAGES() did a truncated write, so * that even the first page of the run is not fully * written, vm_pageout_flush() returns 0 as the run * length. Since the condition that caused truncated * write may be permanent, e.g. exhausted free space, * accepting n == 0 would cause an infinite loop. * * Forwarding the iterator leaves the unwritten page * behind, but there is not much we can do there if * filesystem refuses to write it. */ if (n == 0) { n = 1; allclean = FALSE; } np = vm_page_find_least(object, pi + n); } #if 0 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0); #endif if (allclean) object->cleangeneration = curgeneration; return (res); } static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags, int flags, boolean_t *allclean, boolean_t *eio) { vm_page_t ma[vm_pageout_page_count], p_first, tp; int count, i, mreq, runlen; vm_page_lock_assert(p, MA_NOTOWNED); vm_page_assert_xbusied(p); VM_OBJECT_ASSERT_WLOCKED(object); count = 1; mreq = 0; for (tp = p; count < vm_pageout_page_count; count++) { tp = vm_page_next(tp); if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; if (!vm_object_page_remove_write(tp, flags, allclean)) { vm_page_xunbusy(tp); break; } } for (p_first = p; count < vm_pageout_page_count; count++) { tp = vm_page_prev(p_first); if (tp == NULL || vm_page_tryxbusy(tp) == 0) break; if (!vm_object_page_remove_write(tp, flags, allclean)) { vm_page_xunbusy(tp); break; } p_first = tp; mreq++; } for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++) ma[i] = tp; vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio); return (runlen); } /* * Note that there is absolutely no sense in writing out * anonymous objects, so we track down the vnode object * to write out. * We invalidate (remove) all pages from the address space * for semantic correctness. * * If the backing object is a device object with unmanaged pages, then any * mappings to the specified range of pages must be removed before this * function is called. * * Note: certain anonymous maps, such as MAP_NOSYNC maps, * may start out with a NULL object. */ boolean_t vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size, boolean_t syncio, boolean_t invalidate) { vm_object_t backing_object; struct vnode *vp; struct mount *mp; int error, flags, fsync_after; boolean_t res; if (object == NULL) return (TRUE); res = TRUE; error = 0; VM_OBJECT_WLOCK(object); while ((backing_object = object->backing_object) != NULL) { VM_OBJECT_WLOCK(backing_object); offset += object->backing_object_offset; VM_OBJECT_WUNLOCK(object); object = backing_object; if (object->size < OFF_TO_IDX(offset + size)) size = IDX_TO_OFF(object->size) - offset; } /* * Flush pages if writing is allowed, invalidate them * if invalidation requested. Pages undergoing I/O * will be ignored by vm_object_page_remove(). * * We cannot lock the vnode and then wait for paging * to complete without deadlocking against vm_fault. * Instead we simply call vm_object_page_remove() and * allow it to block internally on a page-by-page * basis when it encounters pages undergoing async * I/O. */ if (object->type == OBJT_VNODE && vm_object_mightbedirty(object) != 0 && ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) { VM_OBJECT_WUNLOCK(object); (void) vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (syncio && !invalidate && offset == 0 && atop(size) == object->size) { /* * If syncing the whole mapping of the file, * it is faster to schedule all the writes in * async mode, also allowing the clustering, * and then wait for i/o to complete. */ flags = 0; fsync_after = TRUE; } else { flags = (syncio || invalidate) ? OBJPC_SYNC : 0; flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0; fsync_after = FALSE; } VM_OBJECT_WLOCK(object); res = vm_object_page_clean(object, offset, offset + size, flags); VM_OBJECT_WUNLOCK(object); if (fsync_after) error = VOP_FSYNC(vp, MNT_WAIT, curthread); VOP_UNLOCK(vp, 0); vn_finished_write(mp); if (error != 0) res = FALSE; VM_OBJECT_WLOCK(object); } if ((object->type == OBJT_VNODE || object->type == OBJT_DEVICE) && invalidate) { if (object->type == OBJT_DEVICE) /* * The option OBJPR_NOTMAPPED must be passed here * because vm_object_page_remove() cannot remove * unmanaged mappings. */ flags = OBJPR_NOTMAPPED; else if (old_msync) flags = 0; else flags = OBJPR_CLEANONLY; vm_object_page_remove(object, OFF_TO_IDX(offset), OFF_TO_IDX(offset + size + PAGE_MASK), flags); } VM_OBJECT_WUNLOCK(object); return (res); } /* * Determine whether the given advice can be applied to the object. Advice is * not applied to unmanaged pages since they never belong to page queues, and * since MADV_FREE is destructive, it can apply only to anonymous pages that * have been mapped at most once. */ static bool vm_object_advice_applies(vm_object_t object, int advice) { if ((object->flags & OBJ_UNMANAGED) != 0) return (false); if (advice != MADV_FREE) return (true); return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) == (OBJ_ONEMAPPING | OBJ_ANON)); } static void vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex, vm_size_t size) { if (advice == MADV_FREE && object->type == OBJT_SWAP) swap_pager_freespace(object, pindex, size); } /* * vm_object_madvise: * * Implements the madvise function at the object/page level. * * MADV_WILLNEED (any object) * * Activate the specified pages if they are resident. * * MADV_DONTNEED (any object) * * Deactivate the specified pages if they are resident. * * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, * OBJ_ONEMAPPING only) * * Deactivate and clean the specified pages if they are * resident. This permits the process to reuse the pages * without faulting or the kernel to reclaim the pages * without I/O. */ void vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end, int advice) { vm_pindex_t tpindex; vm_object_t backing_object, tobject; vm_page_t m, tm; if (object == NULL) return; relookup: VM_OBJECT_WLOCK(object); if (!vm_object_advice_applies(object, advice)) { VM_OBJECT_WUNLOCK(object); return; } for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) { tobject = object; /* * If the next page isn't resident in the top-level object, we * need to search the shadow chain. When applying MADV_FREE, we * take care to release any swap space used to store * non-resident pages. */ if (m == NULL || pindex < m->pindex) { /* * Optimize a common case: if the top-level object has * no backing object, we can skip over the non-resident * range in constant time. */ if (object->backing_object == NULL) { tpindex = (m != NULL && m->pindex < end) ? m->pindex : end; vm_object_madvise_freespace(object, advice, pindex, tpindex - pindex); if ((pindex = tpindex) == end) break; goto next_page; } tpindex = pindex; do { vm_object_madvise_freespace(tobject, advice, tpindex, 1); /* * Prepare to search the next object in the * chain. */ backing_object = tobject->backing_object; if (backing_object == NULL) goto next_pindex; VM_OBJECT_WLOCK(backing_object); tpindex += OFF_TO_IDX(tobject->backing_object_offset); if (tobject != object) VM_OBJECT_WUNLOCK(tobject); tobject = backing_object; if (!vm_object_advice_applies(tobject, advice)) goto next_pindex; } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { next_page: tm = m; m = TAILQ_NEXT(m, listq); } /* * If the page is not in a normal state, skip it. The page * can not be invalidated while the object lock is held. */ if (!vm_page_all_valid(tm) || vm_page_wired(tm)) goto next_pindex; KASSERT((tm->flags & PG_FICTITIOUS) == 0, ("vm_object_madvise: page %p is fictitious", tm)); KASSERT((tm->oflags & VPO_UNMANAGED) == 0, ("vm_object_madvise: page %p is not managed", tm)); if (vm_page_tryxbusy(tm) == 0) { if (object != tobject) VM_OBJECT_WUNLOCK(object); if (advice == MADV_WILLNEED) { /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ vm_page_aflag_set(tm, PGA_REFERENCED); } vm_page_busy_sleep(tm, "madvpo", false); goto relookup; } - vm_page_lock(tm); vm_page_advise(tm, advice); - vm_page_unlock(tm); vm_page_xunbusy(tm); vm_object_madvise_freespace(tobject, advice, tm->pindex, 1); next_pindex: if (tobject != object) VM_OBJECT_WUNLOCK(tobject); } VM_OBJECT_WUNLOCK(object); } /* * vm_object_shadow: * * Create a new object which is backed by the * specified existing object range. The source * object reference is deallocated. * * The new object and offset into that object * are returned in the source parameters. */ void vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length, struct ucred *cred, bool shared) { vm_object_t source; vm_object_t result; source = *object; /* * Don't create the new object if the old object isn't shared. * * If we hold the only reference we can guarantee that it won't * increase while we have the map locked. Otherwise the race is * harmless and we will end up with an extra shadow object that * will be collapsed later. */ if (source != NULL && source->ref_count == 1 && (source->flags & OBJ_ANON) != 0) return; /* * Allocate a new object with the given length. */ result = vm_object_allocate_anon(atop(length), source, cred, length); /* * Store the offset into the source object, and fix up the offset into * the new object. */ result->backing_object_offset = *offset; if (shared || source != NULL) { VM_OBJECT_WLOCK(result); /* * The new object shadows the source object, adding a * reference to it. Our caller changes his reference * to point to the new object, removing a reference to * the source object. Net result: no change of * reference count, unless the caller needs to add one * more reference due to forking a shared map entry. */ if (shared) { vm_object_reference_locked(result); vm_object_clear_flag(result, OBJ_ONEMAPPING); } /* * Try to optimize the result object's page color when * shadowing in order to maintain page coloring * consistency in the combined shadowed object. */ if (source != NULL) { vm_object_backing_insert(result, source); result->domain = source->domain; #if VM_NRESERVLEVEL > 0 result->flags |= source->flags & OBJ_COLORED; result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER - 1)) - 1); #endif } VM_OBJECT_WUNLOCK(result); } /* * Return the new things */ *offset = 0; *object = result; } /* * vm_object_split: * * Split the pages in a map entry into a new object. This affords * easier removal of unused pages, and keeps object inheritance from * being a negative impact on memory usage. */ void vm_object_split(vm_map_entry_t entry) { vm_page_t m, m_next; vm_object_t orig_object, new_object, source; vm_pindex_t idx, offidxstart; vm_size_t size; orig_object = entry->object.vm_object; if ((orig_object->flags & OBJ_ANON) == 0) return; if (orig_object->ref_count <= 1) return; VM_OBJECT_WUNLOCK(orig_object); offidxstart = OFF_TO_IDX(entry->offset); size = atop(entry->end - entry->start); /* * If swap_pager_copy() is later called, it will convert new_object * into a swap object. */ new_object = vm_object_allocate_anon(size, orig_object, orig_object->cred, ptoa(size)); /* * At this point, the new object is still private, so the order in * which the original and new objects are locked does not matter. */ VM_OBJECT_WLOCK(new_object); VM_OBJECT_WLOCK(orig_object); new_object->domain = orig_object->domain; source = orig_object->backing_object; if (source != NULL) { if ((source->flags & (OBJ_ANON | OBJ_DEAD)) != 0) { VM_OBJECT_WLOCK(source); if ((source->flags & OBJ_DEAD) != 0) { VM_OBJECT_WUNLOCK(source); VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); new_object->cred = NULL; vm_object_deallocate(new_object); VM_OBJECT_WLOCK(orig_object); return; } vm_object_backing_insert_locked(new_object, source); vm_object_reference_locked(source); /* for new_object */ vm_object_clear_flag(source, OBJ_ONEMAPPING); VM_OBJECT_WUNLOCK(source); } else { vm_object_backing_insert(new_object, source); vm_object_reference(source); } new_object->backing_object_offset = orig_object->backing_object_offset + entry->offset; } if (orig_object->cred != NULL) { crhold(orig_object->cred); KASSERT(orig_object->charge >= ptoa(size), ("orig_object->charge < 0")); orig_object->charge -= ptoa(size); } retry: m = vm_page_find_least(orig_object, offidxstart); for (; m != NULL && (idx = m->pindex - offidxstart) < size; m = m_next) { m_next = TAILQ_NEXT(m, listq); /* * We must wait for pending I/O to complete before we can * rename the page. * * We do not have to VM_PROT_NONE the page as mappings should * not be changed by this operation. */ if (vm_page_tryxbusy(m) == 0) { VM_OBJECT_WUNLOCK(new_object); vm_page_sleep_if_busy(m, "spltwt"); VM_OBJECT_WLOCK(new_object); goto retry; } /* * The page was left invalid. Likely placed there by * an incomplete fault. Just remove and ignore. */ if (vm_page_none_valid(m)) { if (vm_page_remove(m)) vm_page_free(m); continue; } /* vm_page_rename() will dirty the page. */ if (vm_page_rename(m, new_object, idx)) { vm_page_xunbusy(m); VM_OBJECT_WUNLOCK(new_object); VM_OBJECT_WUNLOCK(orig_object); vm_radix_wait(); VM_OBJECT_WLOCK(orig_object); VM_OBJECT_WLOCK(new_object); goto retry; } #if VM_NRESERVLEVEL > 0 /* * If some of the reservation's allocated pages remain with * the original object, then transferring the reservation to * the new object is neither particularly beneficial nor * particularly harmful as compared to leaving the reservation * with the original object. If, however, all of the * reservation's allocated pages are transferred to the new * object, then transferring the reservation is typically * beneficial. Determining which of these two cases applies * would be more costly than unconditionally renaming the * reservation. */ vm_reserv_rename(m, new_object, orig_object, offidxstart); #endif if (orig_object->type != OBJT_SWAP) vm_page_xunbusy(m); } if (orig_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case the orig_object's * and new_object's locks are released and reacquired. */ swap_pager_copy(orig_object, new_object, offidxstart, 0); TAILQ_FOREACH(m, &new_object->memq, listq) vm_page_xunbusy(m); } VM_OBJECT_WUNLOCK(orig_object); VM_OBJECT_WUNLOCK(new_object); entry->object.vm_object = new_object; entry->offset = 0LL; vm_object_deallocate(orig_object); VM_OBJECT_WLOCK(new_object); } #define OBSC_COLLAPSE_NOWAIT 0x0002 #define OBSC_COLLAPSE_WAIT 0x0004 static vm_page_t vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next, int op) { vm_object_t backing_object; VM_OBJECT_ASSERT_WLOCKED(object); backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(backing_object); KASSERT(p == NULL || p->object == object || p->object == backing_object, ("invalid ownership %p %p %p", p, object, backing_object)); if ((op & OBSC_COLLAPSE_NOWAIT) != 0) return (next); /* The page is only NULL when rename fails. */ if (p == NULL) { VM_OBJECT_WUNLOCK(object); VM_OBJECT_WUNLOCK(backing_object); vm_radix_wait(); } else { if (p->object == object) VM_OBJECT_WUNLOCK(backing_object); else VM_OBJECT_WUNLOCK(object); vm_page_busy_sleep(p, "vmocol", false); } VM_OBJECT_WLOCK(object); VM_OBJECT_WLOCK(backing_object); return (TAILQ_FIRST(&backing_object->memq)); } static bool vm_object_scan_all_shadowed(vm_object_t object) { vm_object_t backing_object; vm_page_t p, pp; vm_pindex_t backing_offset_index, new_pindex, pi, ps; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; if ((backing_object->flags & OBJ_ANON) == 0) return (false); pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset); p = vm_page_find_least(backing_object, pi); ps = swap_pager_find_least(backing_object, pi); /* * Only check pages inside the parent object's range and * inside the parent object's mapping of the backing object. */ for (;; pi++) { if (p != NULL && p->pindex < pi) p = TAILQ_NEXT(p, listq); if (ps < pi) ps = swap_pager_find_least(backing_object, pi); if (p == NULL && ps >= backing_object->size) break; else if (p == NULL) pi = ps; else pi = MIN(p->pindex, ps); new_pindex = pi - backing_offset_index; if (new_pindex >= object->size) break; /* * See if the parent has the page or if the parent's object * pager has the page. If the parent has the page but the page * is not valid, the parent's object pager must have the page. * * If this fails, the parent does not completely shadow the * object and we might as well give up now. */ pp = vm_page_lookup(object, new_pindex); /* * The valid check here is stable due to object lock being * required to clear valid and initiate paging. */ if ((pp == NULL || vm_page_none_valid(pp)) && !vm_pager_has_page(object, new_pindex, NULL, NULL)) return (false); } return (true); } static bool vm_object_collapse_scan(vm_object_t object, int op) { vm_object_t backing_object; vm_page_t next, p, pp; vm_pindex_t backing_offset_index, new_pindex; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(object->backing_object); backing_object = object->backing_object; backing_offset_index = OFF_TO_IDX(object->backing_object_offset); /* * Initial conditions */ if ((op & OBSC_COLLAPSE_WAIT) != 0) vm_object_set_flag(backing_object, OBJ_DEAD); /* * Our scan */ for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) { next = TAILQ_NEXT(p, listq); new_pindex = p->pindex - backing_offset_index; /* * Check for busy page */ if (vm_page_tryxbusy(p) == 0) { next = vm_object_collapse_scan_wait(object, p, next, op); continue; } KASSERT(p->object == backing_object, ("vm_object_collapse_scan: object mismatch")); if (p->pindex < backing_offset_index || new_pindex >= object->size) { if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); continue; } pp = vm_page_lookup(object, new_pindex); if (pp != NULL && vm_page_tryxbusy(pp) == 0) { vm_page_xunbusy(p); /* * The page in the parent is busy and possibly not * (yet) valid. Until its state is finalized by the * busy bit owner, we can't tell whether it shadows the * original page. Therefore, we must either skip it * and the original (backing_object) page or wait for * its state to be finalized. * * This is due to a race with vm_fault() where we must * unbusy the original (backing_obj) page before we can * (re)lock the parent. Hence we can get here. */ next = vm_object_collapse_scan_wait(object, pp, next, op); continue; } if (pp != NULL && vm_page_none_valid(pp)) { /* * The page was invalid in the parent. Likely placed * there by an incomplete fault. Just remove and * ignore. p can replace it. */ if (vm_page_remove(pp)) vm_page_free(pp); pp = NULL; } if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL, NULL)) { /* * The page already exists in the parent OR swap exists * for this location in the parent. Leave the parent's * page alone. Destroy the original page from the * backing object. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, p->pindex, 1); KASSERT(!pmap_page_is_mapped(p), ("freeing mapped page %p", p)); if (vm_page_remove(p)) vm_page_free(p); if (pp != NULL) vm_page_xunbusy(pp); continue; } /* * Page does not exist in parent, rename the page from the * backing object to the main object. * * If the page was mapped to a process, it can remain mapped * through the rename. vm_page_rename() will dirty the page. */ if (vm_page_rename(p, object, new_pindex)) { vm_page_xunbusy(p); if (pp != NULL) vm_page_xunbusy(pp); next = vm_object_collapse_scan_wait(object, NULL, next, op); continue; } /* Use the old pindex to free the right page. */ if (backing_object->type == OBJT_SWAP) swap_pager_freespace(backing_object, new_pindex + backing_offset_index, 1); #if VM_NRESERVLEVEL > 0 /* * Rename the reservation. */ vm_reserv_rename(p, object, backing_object, backing_offset_index); #endif vm_page_xunbusy(p); } return (true); } /* * this version of collapse allows the operation to occur earlier and * when paging_in_progress is true for an object... This is not a complete * operation, but should plug 99.9% of the rest of the leaks. */ static void vm_object_qcollapse(vm_object_t object) { vm_object_t backing_object = object->backing_object; VM_OBJECT_ASSERT_WLOCKED(object); VM_OBJECT_ASSERT_WLOCKED(backing_object); if (backing_object->ref_count != 1) return; vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT); } /* * vm_object_collapse: * * Collapse an object with the object backing it. * Pages in the backing object are moved into the * parent, and the backing object is deallocated. */ void vm_object_collapse(vm_object_t object) { vm_object_t backing_object, new_backing_object; VM_OBJECT_ASSERT_WLOCKED(object); while (TRUE) { /* * Verify that the conditions are right for collapse: * * The object exists and the backing object exists. */ if ((backing_object = object->backing_object) == NULL) break; /* * we check the backing object first, because it is most likely * not collapsable. */ if ((backing_object->flags & OBJ_ANON) == 0) break; VM_OBJECT_WLOCK(backing_object); if ((backing_object->flags & OBJ_DEAD) != 0 || (object->flags & (OBJ_DEAD | OBJ_ANON)) != OBJ_ANON) { VM_OBJECT_WUNLOCK(backing_object); break; } if (REFCOUNT_COUNT(object->paging_in_progress) > 0 || REFCOUNT_COUNT(backing_object->paging_in_progress) > 0) { vm_object_qcollapse(object); VM_OBJECT_WUNLOCK(backing_object); break; } /* * We know that we can either collapse the backing object (if * the parent is the only reference to it) or (perhaps) have * the parent bypass the object if the parent happens to shadow * all the resident pages in the entire backing object. * * This is ignoring pager-backed pages such as swap pages. * vm_object_collapse_scan fails the shadowing test in this * case. */ if (backing_object->ref_count == 1) { vm_object_pip_add(object, 1); vm_object_pip_add(backing_object, 1); /* * If there is exactly one reference to the backing * object, we can collapse it into the parent. */ vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT); #if VM_NRESERVLEVEL > 0 /* * Break any reservations from backing_object. */ if (__predict_false(!LIST_EMPTY(&backing_object->rvq))) vm_reserv_break_all(backing_object); #endif /* * Move the pager from backing_object to object. */ if (backing_object->type == OBJT_SWAP) { /* * swap_pager_copy() can sleep, in which case * the backing_object's and object's locks are * released and reacquired. * Since swap_pager_copy() is being asked to * destroy the source, it will change the * backing_object's type to OBJT_DEFAULT. */ swap_pager_copy( backing_object, object, OFF_TO_IDX(object->backing_object_offset), TRUE); } /* * Object now shadows whatever backing_object did. * Note that the reference to * backing_object->backing_object moves from within * backing_object to within object. */ vm_object_backing_remove_locked(object); new_backing_object = backing_object->backing_object; if (new_backing_object != NULL) { VM_OBJECT_WLOCK(new_backing_object); vm_object_backing_remove_locked(backing_object); vm_object_backing_insert_locked(object, new_backing_object); VM_OBJECT_WUNLOCK(new_backing_object); } object->backing_object_offset += backing_object->backing_object_offset; /* * Discard backing_object. * * Since the backing object has no pages, no pager left, * and no object references within it, all that is * necessary is to dispose of it. */ KASSERT(backing_object->ref_count == 1, ( "backing_object %p was somehow re-referenced during collapse!", backing_object)); vm_object_pip_wakeup(backing_object); backing_object->type = OBJT_DEAD; refcount_release(&backing_object->ref_count); VM_OBJECT_WUNLOCK(backing_object); vm_object_destroy(backing_object); vm_object_pip_wakeup(object); counter_u64_add(object_collapses, 1); } else { /* * If we do not entirely shadow the backing object, * there is nothing we can do so we give up. */ if (object->resident_page_count != object->size && !vm_object_scan_all_shadowed(object)) { VM_OBJECT_WUNLOCK(backing_object); break; } /* * Make the parent shadow the next object in the * chain. Deallocating backing_object will not remove * it, since its reference count is at least 2. */ vm_object_backing_remove_locked(object); new_backing_object = backing_object->backing_object; if (new_backing_object != NULL) { vm_object_backing_insert(object, new_backing_object); vm_object_reference(new_backing_object); object->backing_object_offset += backing_object->backing_object_offset; } /* * Drop the reference count on backing_object. Since * its ref_count was at least 2, it will not vanish. */ refcount_release(&backing_object->ref_count); VM_OBJECT_WUNLOCK(backing_object); counter_u64_add(object_bypasses, 1); } /* * Try again with this object's new backing object. */ } } /* * vm_object_page_remove: * * For the given object, either frees or invalidates each of the * specified pages. In general, a page is freed. However, if a page is * wired for any reason other than the existence of a managed, wired * mapping, then it may be invalidated but not removed from the object. * Pages are specified by the given range ["start", "end") and the option * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range * extends from "start" to the end of the object. If the option * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the * specified range are affected. If the option OBJPR_NOTMAPPED is * specified, then the pages within the specified range must have no * mappings. Otherwise, if this option is not specified, any mappings to * the specified pages are removed before the pages are freed or * invalidated. * * In general, this operation should only be performed on objects that * contain managed pages. There are, however, two exceptions. First, it * is performed on the kernel and kmem objects by vm_map_entry_delete(). * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device- * backed pages. In both of these cases, the option OBJPR_CLEANONLY must * not be specified and the option OBJPR_NOTMAPPED must be specified. * * The object must be locked. */ void vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int options) { vm_page_t p, next; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_UNMANAGED) == 0 || (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED, ("vm_object_page_remove: illegal options for object %p", object)); if (object->resident_page_count == 0) return; vm_object_pip_add(object, 1); again: p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); /* * If the page is wired for any reason besides the existence * of managed, wired mappings, then it cannot be freed. For * example, fictitious pages, which represent device memory, * are inherently wired and cannot be freed. They can, * however, be invalidated if the option OBJPR_CLEANONLY is * not specified. */ if (vm_page_tryxbusy(p) == 0) { vm_page_sleep_if_busy(p, "vmopar"); goto again; } if (vm_page_wired(p)) { wired: if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0) pmap_remove_all(p); if ((options & OBJPR_CLEANONLY) == 0) { vm_page_invalid(p); vm_page_undirty(p); } vm_page_xunbusy(p); continue; } KASSERT((p->flags & PG_FICTITIOUS) == 0, ("vm_object_page_remove: page %p is fictitious", p)); if ((options & OBJPR_CLEANONLY) != 0 && !vm_page_none_valid(p)) { if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_write(p)) goto wired; if (p->dirty != 0) { vm_page_xunbusy(p); continue; } } if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0 && !vm_page_try_remove_all(p)) goto wired; vm_page_free(p); } vm_object_pip_wakeup(object); } /* * vm_object_page_noreuse: * * For the given object, attempt to move the specified pages to * the head of the inactive queue. This bypasses regular LRU * operation and allows the pages to be reused quickly under memory * pressure. If a page is wired for any reason, then it will not * be queued. Pages are specified by the range ["start", "end"). * As a special case, if "end" is zero, then the range extends from * "start" to the end of the object. * * This operation should only be performed on objects that * contain non-fictitious, managed pages. * * The object must be locked. */ void vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { - struct mtx *mtx; vm_page_t p, next; VM_OBJECT_ASSERT_LOCKED(object); KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0, ("vm_object_page_noreuse: illegal object %p", object)); if (object->resident_page_count == 0) return; p = vm_page_find_least(object, start); /* * Here, the variable "p" is either (1) the page with the least pindex * greater than or equal to the parameter "start" or (2) NULL. */ - mtx = NULL; for (; p != NULL && (p->pindex < end || end == 0); p = next) { next = TAILQ_NEXT(p, listq); - vm_page_change_lock(p, &mtx); vm_page_deactivate_noreuse(p); } - if (mtx != NULL) - mtx_unlock(mtx); } /* * Populate the specified range of the object with valid pages. Returns * TRUE if the range is successfully populated and FALSE otherwise. * * Note: This function should be optimized to pass a larger array of * pages to vm_pager_get_pages() before it is applied to a non- * OBJT_DEVICE object. * * The object must be locked. */ boolean_t vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end) { vm_page_t m; vm_pindex_t pindex; int rv; VM_OBJECT_ASSERT_WLOCKED(object); for (pindex = start; pindex < end; pindex++) { rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL); if (rv != VM_PAGER_OK) break; /* * Keep "m" busy because a subsequent iteration may unlock * the object. */ } if (pindex > start) { m = vm_page_lookup(object, start); while (m != NULL && m->pindex < pindex) { vm_page_xunbusy(m); m = TAILQ_NEXT(m, listq); } } return (pindex == end); } /* * Routine: vm_object_coalesce * Function: Coalesces two objects backing up adjoining * regions of memory into a single object. * * returns TRUE if objects were combined. * * NOTE: Only works at the moment if the second object is NULL - * if it's not, which object do we lock first? * * Parameters: * prev_object First object to coalesce * prev_offset Offset into prev_object * prev_size Size of reference to prev_object * next_size Size of reference to the second object * reserved Indicator that extension region has * swap accounted for * * Conditions: * The object must *not* be locked. */ boolean_t vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset, vm_size_t prev_size, vm_size_t next_size, boolean_t reserved) { vm_pindex_t next_pindex; if (prev_object == NULL) return (TRUE); if ((prev_object->flags & OBJ_ANON) == 0) return (FALSE); VM_OBJECT_WLOCK(prev_object); /* * Try to collapse the object first */ vm_object_collapse(prev_object); /* * Can't coalesce if: . more than one reference . paged out . shadows * another object . has a copy elsewhere (any of which mean that the * pages not mapped to prev_entry may be in use anyway) */ if (prev_object->backing_object != NULL) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_size >>= PAGE_SHIFT; next_size >>= PAGE_SHIFT; next_pindex = OFF_TO_IDX(prev_offset) + prev_size; if (prev_object->ref_count > 1 && prev_object->size != next_pindex && (prev_object->flags & OBJ_ONEMAPPING) == 0) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } /* * Account for the charge. */ if (prev_object->cred != NULL) { /* * If prev_object was charged, then this mapping, * although not charged now, may become writable * later. Non-NULL cred in the object would prevent * swap reservation during enabling of the write * access, so reserve swap now. Failed reservation * cause allocation of the separate object for the map * entry, and swap reservation for this entry is * managed in appropriate time. */ if (!reserved && !swap_reserve_by_cred(ptoa(next_size), prev_object->cred)) { VM_OBJECT_WUNLOCK(prev_object); return (FALSE); } prev_object->charge += ptoa(next_size); } /* * Remove any pages that may still be in the object from a previous * deallocation. */ if (next_pindex < prev_object->size) { vm_object_page_remove(prev_object, next_pindex, next_pindex + next_size, 0); if (prev_object->type == OBJT_SWAP) swap_pager_freespace(prev_object, next_pindex, next_size); #if 0 if (prev_object->cred != NULL) { KASSERT(prev_object->charge >= ptoa(prev_object->size - next_pindex), ("object %p overcharged 1 %jx %jx", prev_object, (uintmax_t)next_pindex, (uintmax_t)next_size)); prev_object->charge -= ptoa(prev_object->size - next_pindex); } #endif } /* * Extend the object if necessary. */ if (next_pindex + next_size > prev_object->size) prev_object->size = next_pindex + next_size; VM_OBJECT_WUNLOCK(prev_object); return (TRUE); } void vm_object_set_writeable_dirty(vm_object_t object) { /* Only set for vnodes & tmpfs */ if (object->type != OBJT_VNODE && (object->flags & OBJ_TMPFS_NODE) == 0) return; atomic_add_int(&object->generation, 1); } /* * vm_object_unwire: * * For each page offset within the specified range of the given object, * find the highest-level page in the shadow chain and unwire it. A page * must exist at every page offset, and the highest-level page must be * wired. */ void vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length, uint8_t queue) { vm_object_t tobject, t1object; vm_page_t m, tm; vm_pindex_t end_pindex, pindex, tpindex; int depth, locked_depth; KASSERT((offset & PAGE_MASK) == 0, ("vm_object_unwire: offset is not page aligned")); KASSERT((length & PAGE_MASK) == 0, ("vm_object_unwire: length is not a multiple of PAGE_SIZE")); /* The wired count of a fictitious page never changes. */ if ((object->flags & OBJ_FICTITIOUS) != 0) return; pindex = OFF_TO_IDX(offset); end_pindex = pindex + atop(length); again: locked_depth = 1; VM_OBJECT_RLOCK(object); m = vm_page_find_least(object, pindex); while (pindex < end_pindex) { if (m == NULL || pindex < m->pindex) { /* * The first object in the shadow chain doesn't * contain a page at the current index. Therefore, * the page must exist in a backing object. */ tobject = object; tpindex = pindex; depth = 0; do { tpindex += OFF_TO_IDX(tobject->backing_object_offset); tobject = tobject->backing_object; KASSERT(tobject != NULL, ("vm_object_unwire: missing page")); if ((tobject->flags & OBJ_FICTITIOUS) != 0) goto next_page; depth++; if (depth == locked_depth) { locked_depth++; VM_OBJECT_RLOCK(tobject); } } while ((tm = vm_page_lookup(tobject, tpindex)) == NULL); } else { tm = m; m = TAILQ_NEXT(m, listq); } if (vm_page_trysbusy(tm) == 0) { for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; if (tm->object != tobject) VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } vm_page_busy_sleep(tm, "unwbo", true); goto again; } vm_page_unwire(tm, queue); vm_page_sunbusy(tm); next_page: pindex++; } /* Release the accumulated object locks. */ for (tobject = object; locked_depth >= 1; locked_depth--) { t1object = tobject->backing_object; VM_OBJECT_RUNLOCK(tobject); tobject = t1object; } } /* * Return the vnode for the given object, or NULL if none exists. * For tmpfs objects, the function may return NULL if there is * no vnode allocated at the time of the call. */ struct vnode * vm_object_vnode(vm_object_t object) { struct vnode *vp; VM_OBJECT_ASSERT_LOCKED(object); if (object->type == OBJT_VNODE) { vp = object->handle; KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__)); } else if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0) { vp = object->un_pager.swp.swp_tmpfs; KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__)); } else { vp = NULL; } return (vp); } /* * Busy the vm object. This prevents new pages belonging to the object from * becoming busy. Existing pages persist as busy. Callers are responsible * for checking page state before proceeding. */ void vm_object_busy(vm_object_t obj) { VM_OBJECT_ASSERT_LOCKED(obj); refcount_acquire(&obj->busy); /* The fence is required to order loads of page busy. */ atomic_thread_fence_acq_rel(); } void vm_object_unbusy(vm_object_t obj) { refcount_release(&obj->busy); } void vm_object_busy_wait(vm_object_t obj, const char *wmesg) { VM_OBJECT_ASSERT_UNLOCKED(obj); if (obj->busy) refcount_sleep(&obj->busy, wmesg, PVM); } /* * Return the kvme type of the given object. * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL. */ int vm_object_kvme_type(vm_object_t object, struct vnode **vpp) { VM_OBJECT_ASSERT_LOCKED(object); if (vpp != NULL) *vpp = vm_object_vnode(object); switch (object->type) { case OBJT_DEFAULT: return (KVME_TYPE_DEFAULT); case OBJT_VNODE: return (KVME_TYPE_VNODE); case OBJT_SWAP: if ((object->flags & OBJ_TMPFS_NODE) != 0) return (KVME_TYPE_VNODE); return (KVME_TYPE_SWAP); case OBJT_DEVICE: return (KVME_TYPE_DEVICE); case OBJT_PHYS: return (KVME_TYPE_PHYS); case OBJT_DEAD: return (KVME_TYPE_DEAD); case OBJT_SG: return (KVME_TYPE_SG); case OBJT_MGTDEVICE: return (KVME_TYPE_MGTDEVICE); default: return (KVME_TYPE_UNKNOWN); } } static int sysctl_vm_object_list(SYSCTL_HANDLER_ARGS) { struct kinfo_vmobject *kvo; char *fullpath, *freepath; struct vnode *vp; struct vattr va; vm_object_t obj; vm_page_t m; int count, error; if (req->oldptr == NULL) { /* * If an old buffer has not been provided, generate an * estimate of the space needed for a subsequent call. */ mtx_lock(&vm_object_list_mtx); count = 0; TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; count++; } mtx_unlock(&vm_object_list_mtx); return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) * count * 11 / 10)); } kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK); error = 0; /* * VM objects are type stable and are never removed from the * list once added. This allows us to safely read obj->object_list * after reacquiring the VM object lock. */ mtx_lock(&vm_object_list_mtx); TAILQ_FOREACH(obj, &vm_object_list, object_list) { if (obj->type == OBJT_DEAD) continue; VM_OBJECT_RLOCK(obj); if (obj->type == OBJT_DEAD) { VM_OBJECT_RUNLOCK(obj); continue; } mtx_unlock(&vm_object_list_mtx); kvo->kvo_size = ptoa(obj->size); kvo->kvo_resident = obj->resident_page_count; kvo->kvo_ref_count = obj->ref_count; kvo->kvo_shadow_count = obj->shadow_count; kvo->kvo_memattr = obj->memattr; kvo->kvo_active = 0; kvo->kvo_inactive = 0; TAILQ_FOREACH(m, &obj->memq, listq) { /* * A page may belong to the object but be * dequeued and set to PQ_NONE while the * object lock is not held. This makes the * reads of m->queue below racy, and we do not * count pages set to PQ_NONE. However, this * sysctl is only meant to give an * approximation of the system anyway. */ if (m->a.queue == PQ_ACTIVE) kvo->kvo_active++; else if (m->a.queue == PQ_INACTIVE) kvo->kvo_inactive++; } kvo->kvo_vn_fileid = 0; kvo->kvo_vn_fsid = 0; kvo->kvo_vn_fsid_freebsd11 = 0; freepath = NULL; fullpath = ""; kvo->kvo_type = vm_object_kvme_type(obj, &vp); if (vp != NULL) vref(vp); VM_OBJECT_RUNLOCK(obj); if (vp != NULL) { vn_fullpath(curthread, vp, &fullpath, &freepath); vn_lock(vp, LK_SHARED | LK_RETRY); if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) { kvo->kvo_vn_fileid = va.va_fileid; kvo->kvo_vn_fsid = va.va_fsid; kvo->kvo_vn_fsid_freebsd11 = va.va_fsid; /* truncate */ } vput(vp); } strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path)); if (freepath != NULL) free(freepath, M_TEMP); /* Pack record size down */ kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) + strlen(kvo->kvo_path) + 1; kvo->kvo_structsize = roundup(kvo->kvo_structsize, sizeof(uint64_t)); error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize); mtx_lock(&vm_object_list_mtx); if (error) break; } mtx_unlock(&vm_object_list_mtx); free(kvo, M_TEMP); return (error); } SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject", "List of VM objects"); #include "opt_ddb.h" #ifdef DDB #include #include #include static int _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry) { vm_map_t tmpm; vm_map_entry_t tmpe; vm_object_t obj; if (map == 0) return 0; if (entry == 0) { VM_MAP_ENTRY_FOREACH(tmpe, map) { if (_vm_object_in_map(map, object, tmpe)) { return 1; } } } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { tmpm = entry->object.sub_map; VM_MAP_ENTRY_FOREACH(tmpe, tmpm) { if (_vm_object_in_map(tmpm, object, tmpe)) { return 1; } } } else if ((obj = entry->object.vm_object) != NULL) { for (; obj; obj = obj->backing_object) if (obj == object) { return 1; } } return 0; } static int vm_object_in_map(vm_object_t object) { struct proc *p; /* sx_slock(&allproc_lock); */ FOREACH_PROC_IN_SYSTEM(p) { if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */) continue; if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) { /* sx_sunlock(&allproc_lock); */ return 1; } } /* sx_sunlock(&allproc_lock); */ if (_vm_object_in_map(kernel_map, object, 0)) return 1; return 0; } DB_SHOW_COMMAND(vmochk, vm_object_check) { vm_object_t object; /* * make sure that internal objs are in a map somewhere * and none have zero ref counts. */ TAILQ_FOREACH(object, &vm_object_list, object_list) { if ((object->flags & OBJ_ANON) != 0) { if (object->ref_count == 0) { db_printf("vmochk: internal obj has zero ref count: %ld\n", (long)object->size); } if (!vm_object_in_map(object)) { db_printf( "vmochk: internal obj is not in a map: " "ref: %d, size: %lu: 0x%lx, backing_object: %p\n", object->ref_count, (u_long)object->size, (u_long)object->size, (void *)object->backing_object); } } } } /* * vm_object_print: [ debug ] */ DB_SHOW_COMMAND(object, vm_object_print_static) { /* XXX convert args. */ vm_object_t object = (vm_object_t)addr; boolean_t full = have_addr; vm_page_t p; /* XXX count is an (unused) arg. Avoid shadowing it. */ #define count was_count int count; if (object == NULL) return; db_iprintf( "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n", object, (int)object->type, (uintmax_t)object->size, object->resident_page_count, object->ref_count, object->flags, object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge); db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n", object->shadow_count, object->backing_object ? object->backing_object->ref_count : 0, object->backing_object, (uintmax_t)object->backing_object_offset); if (!full) return; db_indent += 2; count = 0; TAILQ_FOREACH(p, &object->memq, listq) { if (count == 0) db_iprintf("memory:="); else if (count == 6) { db_printf("\n"); db_iprintf(" ..."); count = 0; } else db_printf(","); count++; db_printf("(off=0x%jx,page=0x%jx)", (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p)); } if (count != 0) db_printf("\n"); db_indent -= 2; } /* XXX. */ #undef count /* XXX need this non-static entry for calling from vm_map_print. */ void vm_object_print( /* db_expr_t */ long addr, boolean_t have_addr, /* db_expr_t */ long count, char *modif) { vm_object_print_static(addr, have_addr, count, modif); } DB_SHOW_COMMAND(vmopag, vm_object_print_pages) { vm_object_t object; vm_pindex_t fidx; vm_paddr_t pa; vm_page_t m, prev_m; int rcount, nl, c; nl = 0; TAILQ_FOREACH(object, &vm_object_list, object_list) { db_printf("new object: %p\n", (void *)object); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; fidx = 0; pa = -1; TAILQ_FOREACH(m, &object->memq, listq) { if (m->pindex > 128) break; if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL && prev_m->pindex + 1 != m->pindex) { if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; rcount = 0; } } if (rcount && (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) { ++rcount; continue; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } fidx = m->pindex; pa = VM_PAGE_TO_PHYS(m); rcount = 1; } if (rcount) { db_printf(" index(%ld)run(%d)pa(0x%lx)\n", (long)fidx, rcount, (long)pa); if (nl > 18) { c = cngetc(); if (c != ' ') return; nl = 0; } nl++; } } } #endif /* DDB */ Index: head/sys/vm/vm_page.c =================================================================== --- head/sys/vm/vm_page.c (revision 356156) +++ head/sys/vm/vm_page.c (revision 356157) @@ -1,5288 +1,5281 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1998 Matthew Dillon. All Rights Reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91 */ /*- * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * Resident memory management module. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include extern int uma_startup_count(int); extern void uma_startup(void *, int); extern int vmem_startup_count(void); struct vm_domain vm_dom[MAXMEMDOM]; DPCPU_DEFINE_STATIC(struct vm_batchqueue, pqbatch[MAXMEMDOM][PQ_COUNT]); struct mtx_padalign __exclusive_cache_line pa_lock[PA_LOCK_COUNT]; struct mtx_padalign __exclusive_cache_line vm_domainset_lock; /* The following fields are protected by the domainset lock. */ domainset_t __exclusive_cache_line vm_min_domains; domainset_t __exclusive_cache_line vm_severe_domains; static int vm_min_waiters; static int vm_severe_waiters; static int vm_pageproc_waiters; static SYSCTL_NODE(_vm_stats, OID_AUTO, page, CTLFLAG_RD, 0, "VM page statistics"); static counter_u64_t pqstate_commit_retries = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, pqstate_commit_retries, CTLFLAG_RD, &pqstate_commit_retries, "Number of failed per-page atomic queue state updates"); static counter_u64_t queue_ops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_ops, CTLFLAG_RD, &queue_ops, "Number of batched queue operations"); static counter_u64_t queue_nops = EARLY_COUNTER; SYSCTL_COUNTER_U64(_vm_stats_page, OID_AUTO, queue_nops, CTLFLAG_RD, &queue_nops, "Number of batched queue operations with no effects"); static void counter_startup(void) { pqstate_commit_retries = counter_u64_alloc(M_WAITOK); queue_ops = counter_u64_alloc(M_WAITOK); queue_nops = counter_u64_alloc(M_WAITOK); } SYSINIT(page_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL); /* * bogus page -- for I/O to/from partially complete buffers, * or for paging into sparsely invalid regions. */ vm_page_t bogus_page; vm_page_t vm_page_array; long vm_page_array_size; long first_page; static int boot_pages; SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &boot_pages, 0, "number of pages allocated for bootstrapping the VM system"); static TAILQ_HEAD(, vm_page) blacklist_head; static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages"); static uma_zone_t fakepg_zone; static void vm_page_alloc_check(vm_page_t m); static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, bool nonshared, bool locked); static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits); static void vm_page_enqueue(vm_page_t m, uint8_t queue); static bool vm_page_free_prep(vm_page_t m); static void vm_page_free_toq(vm_page_t m); static void vm_page_init(void *dummy); static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred); static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred); static void vm_page_mvqueue(vm_page_t m, const uint8_t queue, const uint16_t nflag); static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high); static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, bool noreuse); static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req); static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags); static void vm_page_zone_release(void *arg, void **store, int cnt); SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init, NULL); static void vm_page_init(void *dummy) { fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM); bogus_page = vm_page_alloc(NULL, 0, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED); } /* * The cache page zone is initialized later since we need to be able to allocate * pages before UMA is fully initialized. */ static void vm_page_init_cache_zones(void *dummy __unused) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int cache, domain, maxcache, pool; maxcache = 0; TUNABLE_INT_FETCH("vm.pgcache_zone_max", &maxcache); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (pool = 0; pool < VM_NFREEPOOL; pool++) { pgcache = &vmd->vmd_pgcache[pool]; pgcache->domain = domain; pgcache->pool = pool; pgcache->zone = uma_zcache_create("vm pgcache", PAGE_SIZE, NULL, NULL, NULL, NULL, vm_page_zone_import, vm_page_zone_release, pgcache, UMA_ZONE_VM); /* * Limit each pool's zone to 0.1% of the pages in the * domain. */ cache = maxcache != 0 ? maxcache : vmd->vmd_page_count / 1000; uma_zone_set_maxcache(pgcache->zone, cache); } } } SYSINIT(vm_page2, SI_SUB_VM_CONF, SI_ORDER_ANY, vm_page_init_cache_zones, NULL); /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */ #if PAGE_SIZE == 32768 #ifdef CTASSERT CTASSERT(sizeof(u_long) >= 8); #endif #endif /* * vm_set_page_size: * * Sets the page size, perhaps based upon the memory * size. Must be called before any use of page-size * dependent functions. */ void vm_set_page_size(void) { if (vm_cnt.v_page_size == 0) vm_cnt.v_page_size = PAGE_SIZE; if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0) panic("vm_set_page_size: page size not a power of two"); } /* * vm_page_blacklist_next: * * Find the next entry in the provided string of blacklist * addresses. Entries are separated by space, comma, or newline. * If an invalid integer is encountered then the rest of the * string is skipped. Updates the list pointer to the next * character, or NULL if the string is exhausted or invalid. */ static vm_paddr_t vm_page_blacklist_next(char **list, char *end) { vm_paddr_t bad; char *cp, *pos; if (list == NULL || *list == NULL) return (0); if (**list =='\0') { *list = NULL; return (0); } /* * If there's no end pointer then the buffer is coming from * the kenv and we know it's null-terminated. */ if (end == NULL) end = *list + strlen(*list); /* Ensure that strtoq() won't walk off the end */ if (*end != '\0') { if (*end == '\n' || *end == ' ' || *end == ',') *end = '\0'; else { printf("Blacklist not terminated, skipping\n"); *list = NULL; return (0); } } for (pos = *list; *pos != '\0'; pos = cp) { bad = strtoq(pos, &cp, 0); if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') { if (bad == 0) { if (++cp < end) continue; else break; } } else break; if (*cp == '\0' || ++cp >= end) *list = NULL; else *list = cp; return (trunc_page(bad)); } printf("Garbage in RAM blacklist, skipping\n"); *list = NULL; return (0); } bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose) { struct vm_domain *vmd; vm_page_t m; int ret; m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) return (true); /* page does not exist, no failure */ vmd = vm_pagequeue_domain(m); vm_domain_free_lock(vmd); ret = vm_phys_unfree_page(m); vm_domain_free_unlock(vmd); if (ret != 0) { vm_domain_freecnt_inc(vmd, -1); TAILQ_INSERT_TAIL(&blacklist_head, m, listq); if (verbose) printf("Skipping page with pa 0x%jx\n", (uintmax_t)pa); } return (ret); } /* * vm_page_blacklist_check: * * Iterate through the provided string of blacklist addresses, pulling * each entry out of the physical allocator free list and putting it * onto a list for reporting via the vm.page_blacklist sysctl. */ static void vm_page_blacklist_check(char *list, char *end) { vm_paddr_t pa; char *next; next = list; while (next != NULL) { if ((pa = vm_page_blacklist_next(&next, end)) == 0) continue; vm_page_blacklist_add(pa, bootverbose); } } /* * vm_page_blacklist_load: * * Search for a special module named "ram_blacklist". It'll be a * plain text file provided by the user via the loader directive * of the same name. */ static void vm_page_blacklist_load(char **list, char **end) { void *mod; u_char *ptr; u_int len; mod = NULL; ptr = NULL; mod = preload_search_by_type("ram_blacklist"); if (mod != NULL) { ptr = preload_fetch_addr(mod); len = preload_fetch_size(mod); } *list = ptr; if (ptr != NULL) *end = ptr + len; else *end = NULL; return; } static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS) { vm_page_t m; struct sbuf sbuf; int error, first; first = 1; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); TAILQ_FOREACH(m, &blacklist_head, listq) { sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",", (uintmax_t)m->phys_addr); first = 0; } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } /* * Initialize a dummy page for use in scans of the specified paging queue. * In principle, this function only needs to set the flag PG_MARKER. * Nonetheless, it write busies the page as a safety precaution. */ static void vm_page_init_marker(vm_page_t marker, int queue, uint16_t aflags) { bzero(marker, sizeof(*marker)); marker->flags = PG_MARKER; marker->a.flags = aflags; marker->busy_lock = VPB_CURTHREAD_EXCLUSIVE; marker->a.queue = queue; } static void vm_page_domain_init(int domain) { struct vm_domain *vmd; struct vm_pagequeue *pq; int i; vmd = VM_DOMAIN(domain); bzero(vmd, sizeof(*vmd)); *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) = "vm inactive pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) = "vm active pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_LAUNDRY].pq_name) = "vm laundry pagequeue"; *__DECONST(char **, &vmd->vmd_pagequeues[PQ_UNSWAPPABLE].pq_name) = "vm unswappable pagequeue"; vmd->vmd_domain = domain; vmd->vmd_page_count = 0; vmd->vmd_free_count = 0; vmd->vmd_segs = 0; vmd->vmd_oom = FALSE; for (i = 0; i < PQ_COUNT; i++) { pq = &vmd->vmd_pagequeues[i]; TAILQ_INIT(&pq->pq_pl); mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue", MTX_DEF | MTX_DUPOK); pq->pq_pdpages = 0; vm_page_init_marker(&vmd->vmd_markers[i], i, 0); } mtx_init(&vmd->vmd_free_mtx, "vm page free queue", NULL, MTX_DEF); mtx_init(&vmd->vmd_pageout_mtx, "vm pageout lock", NULL, MTX_DEF); snprintf(vmd->vmd_name, sizeof(vmd->vmd_name), "%d", domain); /* * inacthead is used to provide FIFO ordering for LRU-bypassing * insertions. */ vm_page_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl, &vmd->vmd_inacthead, plinks.q); /* * The clock pages are used to implement active queue scanning without * requeues. Scans start at clock[0], which is advanced after the scan * ends. When the two clock hands meet, they are reset and scanning * resumes from the head of the queue. */ vm_page_init_marker(&vmd->vmd_clock[0], PQ_ACTIVE, PGA_ENQUEUED); vm_page_init_marker(&vmd->vmd_clock[1], PQ_ACTIVE, PGA_ENQUEUED); TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&vmd->vmd_pagequeues[PQ_ACTIVE].pq_pl, &vmd->vmd_clock[1], plinks.q); } /* * Initialize a physical page in preparation for adding it to the free * lists. */ static void vm_page_init_page(vm_page_t m, vm_paddr_t pa, int segind) { m->object = NULL; m->ref_count = 0; m->busy_lock = VPB_UNBUSIED; m->flags = m->a.flags = 0; m->phys_addr = pa; m->a.queue = PQ_NONE; m->psind = 0; m->segind = segind; m->order = VM_NFREEORDER; m->pool = VM_FREEPOOL_DEFAULT; m->valid = m->dirty = 0; pmap_page_init(m); } #ifndef PMAP_HAS_PAGE_ARRAY static vm_paddr_t vm_page_array_alloc(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t page_range) { vm_paddr_t new_end; /* * Reserve an unmapped guard page to trap access to vm_page_array[-1]. * However, because this page is allocated from KVM, out-of-bounds * accesses using the direct map will not be trapped. */ *vaddr += PAGE_SIZE; /* * Allocate physical memory for the page structures, and map it. */ new_end = trunc_page(end - page_range * sizeof(struct vm_page)); vm_page_array = (vm_page_t)pmap_map(vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); vm_page_array_size = page_range; return (new_end); } #endif /* * vm_page_startup: * * Initializes the resident memory module. Allocates physical memory for * bootstrapping UMA and some data structures that are used to manage * physical pages. Initializes these structures, and populates the free * page queues. */ vm_offset_t vm_page_startup(vm_offset_t vaddr) { struct vm_phys_seg *seg; vm_page_t m; char *list, *listend; vm_offset_t mapped; vm_paddr_t end, high_avail, low_avail, new_end, size; vm_paddr_t page_range __unused; vm_paddr_t last_pa, pa; u_long pagecount; int biggestone, i, segind; #ifdef WITNESS int witness_size; #endif #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) long ii; #endif vaddr = round_page(vaddr); vm_phys_early_startup(); biggestone = vm_phys_avail_largest(); end = phys_avail[biggestone+1]; /* * Initialize the page and queue locks. */ mtx_init(&vm_domainset_lock, "vm domainset lock", NULL, MTX_DEF); for (i = 0; i < PA_LOCK_COUNT; i++) mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF); for (i = 0; i < vm_ndomains; i++) vm_page_domain_init(i); /* * Allocate memory for use when boot strapping the kernel memory * allocator. Tell UMA how many zones we are going to create * before going fully functional. UMA will add its zones. * * VM startup zones: vmem, vmem_btag, VM OBJECT, RADIX NODE, MAP, * KMAP ENTRY, MAP ENTRY, VMSPACE. */ boot_pages = uma_startup_count(8); #ifndef UMA_MD_SMALL_ALLOC /* vmem_startup() calls uma_prealloc(). */ boot_pages += vmem_startup_count(); /* vm_map_startup() calls uma_prealloc(). */ boot_pages += howmany(MAX_KMAP, slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR)); /* * Before going fully functional kmem_init() does allocation * from "KMAP ENTRY" and vmem_create() does allocation from "vmem". */ boot_pages += 2; #endif /* * CTFLAG_RDTUN doesn't work during the early boot process, so we must * manually fetch the value. */ TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages); new_end = end - (boot_pages * UMA_SLAB_SIZE); new_end = trunc_page(new_end); mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, end - new_end); uma_startup((void *)mapped, boot_pages); #ifdef WITNESS witness_size = round_page(witness_startup_count()); new_end -= witness_size; mapped = pmap_map(&vaddr, new_end, new_end + witness_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)mapped, witness_size); witness_startup((void *)mapped); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \ defined(__i386__) || defined(__mips__) || defined(__riscv) || \ defined(__powerpc64__) /* * Allocate a bitmap to indicate that a random physical page * needs to be included in a minidump. * * The amd64 port needs this to indicate which direct map pages * need to be dumped, via calls to dump_add_page()/dump_drop_page(). * * However, i386 still needs this workspace internally within the * minidump code. In theory, they are not needed on i386, but are * included should the sf_buf code decide to use them. */ last_pa = 0; for (i = 0; dump_avail[i + 1] != 0; i += 2) if (dump_avail[i + 1] > last_pa) last_pa = dump_avail[i + 1]; page_range = last_pa / PAGE_SIZE; vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY); new_end -= vm_page_dump_size; vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end, new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE); bzero((void *)vm_page_dump, vm_page_dump_size); #else (void)last_pa; #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include the UMA bootstrap pages, witness pages and vm_page_dump * in a crash dump. When pmap_map() uses the direct map, they are * not automatically included. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; #ifdef __amd64__ /* * Request that the physical pages underlying the message buffer be * included in a crash dump. Since the message buffer is accessed * through the direct map, they are not automatically included. */ pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr); last_pa = pa + round_page(msgbufsize); while (pa < last_pa) { dump_add_page(pa); pa += PAGE_SIZE; } #endif /* * Compute the number of pages of memory that will be available for * use, taking into account the overhead of a page structure per page. * In other words, solve * "available physical memory" - round_page(page_range * * sizeof(struct vm_page)) = page_range * PAGE_SIZE * for page_range. */ low_avail = phys_avail[0]; high_avail = phys_avail[1]; for (i = 0; i < vm_phys_nsegs; i++) { if (vm_phys_segs[i].start < low_avail) low_avail = vm_phys_segs[i].start; if (vm_phys_segs[i].end > high_avail) high_avail = vm_phys_segs[i].end; } /* Skip the first chunk. It is already accounted for. */ for (i = 2; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i] < low_avail) low_avail = phys_avail[i]; if (phys_avail[i + 1] > high_avail) high_avail = phys_avail[i + 1]; } first_page = low_avail / PAGE_SIZE; #ifdef VM_PHYSSEG_SPARSE size = 0; for (i = 0; i < vm_phys_nsegs; i++) size += vm_phys_segs[i].end - vm_phys_segs[i].start; for (i = 0; phys_avail[i + 1] != 0; i += 2) size += phys_avail[i + 1] - phys_avail[i]; #elif defined(VM_PHYSSEG_DENSE) size = high_avail - low_avail; #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif #ifdef PMAP_HAS_PAGE_ARRAY pmap_page_array_startup(size / PAGE_SIZE); biggestone = vm_phys_avail_largest(); end = new_end = phys_avail[biggestone + 1]; #else #ifdef VM_PHYSSEG_DENSE /* * In the VM_PHYSSEG_DENSE case, the number of pages can account for * the overhead of a page structure per page only if vm_page_array is * allocated from the last physical memory chunk. Otherwise, we must * allocate page structures representing the physical memory * underlying vm_page_array, even though they will not be used. */ if (new_end != high_avail) page_range = size / PAGE_SIZE; else #endif { page_range = size / (PAGE_SIZE + sizeof(struct vm_page)); /* * If the partial bytes remaining are large enough for * a page (PAGE_SIZE) without a corresponding * 'struct vm_page', then new_end will contain an * extra page after subtracting the length of the VM * page array. Compensate by subtracting an extra * page from new_end. */ if (size % (PAGE_SIZE + sizeof(struct vm_page)) >= PAGE_SIZE) { if (new_end == high_avail) high_avail -= PAGE_SIZE; new_end -= PAGE_SIZE; } } end = new_end; new_end = vm_page_array_alloc(&vaddr, end, page_range); #endif #if VM_NRESERVLEVEL > 0 /* * Allocate physical memory for the reservation management system's * data structures, and map it. */ new_end = vm_reserv_startup(&vaddr, new_end); #endif #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \ defined(__riscv) || defined(__powerpc64__) /* * Include vm_page_array and vm_reserv_array in a crash dump. */ for (pa = new_end; pa < end; pa += PAGE_SIZE) dump_add_page(pa); #endif phys_avail[biggestone + 1] = new_end; /* * Add physical memory segments corresponding to the available * physical pages. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) if (vm_phys_avail_size(i) != 0) vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]); /* * Initialize the physical memory allocator. */ vm_phys_init(); /* * Initialize the page structures and add every available page to the * physical memory allocator's free lists. */ #if defined(__i386__) && defined(VM_PHYSSEG_DENSE) for (ii = 0; ii < vm_page_array_size; ii++) { m = &vm_page_array[ii]; vm_page_init_page(m, (first_page + ii) << PAGE_SHIFT, 0); m->flags = PG_FICTITIOUS; } #endif vm_cnt.v_page_count = 0; for (segind = 0; segind < vm_phys_nsegs; segind++) { seg = &vm_phys_segs[segind]; for (m = seg->first_page, pa = seg->start; pa < seg->end; m++, pa += PAGE_SIZE) vm_page_init_page(m, pa, segind); /* * Add the segment to the free lists only if it is covered by * one of the ranges in phys_avail. Because we've added the * ranges to the vm_phys_segs array, we can assume that each * segment is either entirely contained in one of the ranges, * or doesn't overlap any of them. */ for (i = 0; phys_avail[i + 1] != 0; i += 2) { struct vm_domain *vmd; if (seg->start < phys_avail[i] || seg->end > phys_avail[i + 1]) continue; m = seg->first_page; pagecount = (u_long)atop(seg->end - seg->start); vmd = VM_DOMAIN(seg->domain); vm_domain_free_lock(vmd); vm_phys_enqueue_contig(m, pagecount); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, pagecount); vm_cnt.v_page_count += (u_int)pagecount; vmd = VM_DOMAIN(seg->domain); vmd->vmd_page_count += (u_int)pagecount; vmd->vmd_segs |= 1UL << m->segind; break; } } /* * Remove blacklisted pages from the physical memory allocator. */ TAILQ_INIT(&blacklist_head); vm_page_blacklist_load(&list, &listend); vm_page_blacklist_check(list, listend); list = kern_getenv("vm.blacklist"); vm_page_blacklist_check(list, NULL); freeenv(list); #if VM_NRESERVLEVEL > 0 /* * Initialize the reservation management system. */ vm_reserv_init(); #endif return (vaddr); } void vm_page_reference(vm_page_t m) { vm_page_aflag_set(m, PGA_REFERENCED); } static bool vm_page_acquire_flags(vm_page_t m, int allocflags) { bool locked; if ((allocflags & (VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY)) != 0) locked = vm_page_trysbusy(m); else locked = vm_page_tryxbusy(m); if (locked && (allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); return (locked); } /* * vm_page_busy_sleep_flags * * Sleep for busy according to VM_ALLOC_ parameters. */ static bool vm_page_busy_sleep_flags(vm_object_t object, vm_page_t m, const char *wmesg, int allocflags) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); /* * Reference the page before unlocking and * sleeping so that the page daemon is less * likely to reclaim it. */ if ((allocflags & VM_ALLOC_NOCREAT) == 0) vm_page_aflag_set(m, PGA_REFERENCED); if (_vm_page_busy_sleep(object, m, wmesg, (allocflags & VM_ALLOC_IGN_SBUSY) != 0, true)) VM_OBJECT_WLOCK(object); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); return (true); } /* * vm_page_busy_acquire: * * Acquire the busy lock as described by VM_ALLOC_* flags. Will loop * and drop the object lock if necessary. */ bool vm_page_busy_acquire(vm_page_t m, int allocflags) { vm_object_t obj; bool locked; /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; for (;;) { if (vm_page_acquire_flags(m, allocflags)) return (true); if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (false); if (obj != NULL) locked = VM_OBJECT_WOWNED(obj); else locked = false; MPASS(locked || vm_page_wired(m)); if (_vm_page_busy_sleep(obj, m, "vmpba", (allocflags & VM_ALLOC_SBUSY) != 0, locked)) VM_OBJECT_WLOCK(obj); if ((allocflags & VM_ALLOC_WAITFAIL) != 0) return (false); KASSERT(m->object == obj || m->object == NULL, ("vm_page_busy_acquire: page %p does not belong to %p", m, obj)); } } /* * vm_page_busy_downgrade: * * Downgrade an exclusive busy page into a single shared busy page. */ void vm_page_busy_downgrade(vm_page_t m) { u_int x; vm_page_assert_xbusied(m); x = m->busy_lock; for (;;) { if (atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_SHARERS_WORD(1))) break; } if ((x & VPB_BIT_WAITERS) != 0) wakeup(m); } /* * * vm_page_busy_tryupgrade: * * Attempt to upgrade a single shared busy into an exclusive busy. */ int vm_page_busy_tryupgrade(vm_page_t m) { u_int ce, x; vm_page_assert_sbusied(m); x = m->busy_lock; ce = VPB_CURTHREAD_EXCLUSIVE; for (;;) { if (VPB_SHARERS(x) > 1) return (0); KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_busy_tryupgrade: invalid lock state")); if (!atomic_fcmpset_acq_int(&m->busy_lock, &x, ce | (x & VPB_BIT_WAITERS))) continue; return (1); } } /* * vm_page_sbusied: * * Return a positive value if the page is shared busied, 0 otherwise. */ int vm_page_sbusied(vm_page_t m) { u_int x; x = m->busy_lock; return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED); } /* * vm_page_sunbusy: * * Shared unbusy a page. */ void vm_page_sunbusy(vm_page_t m) { u_int x; vm_page_assert_sbusied(m); x = m->busy_lock; for (;;) { if (VPB_SHARERS(x) > 1) { if (atomic_fcmpset_int(&m->busy_lock, &x, x - VPB_ONE_SHARER)) break; continue; } KASSERT((x & ~VPB_BIT_WAITERS) == VPB_SHARERS_WORD(1), ("vm_page_sunbusy: invalid lock state")); if (!atomic_fcmpset_rel_int(&m->busy_lock, &x, VPB_UNBUSIED)) continue; if ((x & VPB_BIT_WAITERS) == 0) break; wakeup(m); break; } } /* * vm_page_busy_sleep: * * Sleep if the page is busy, using the page pointer as wchan. * This is used to implement the hard-path of busying mechanism. * * If nonshared is true, sleep only if the page is xbusy. * * The object lock must be held on entry and will be released on exit. */ void vm_page_busy_sleep(vm_page_t m, const char *wmesg, bool nonshared) { vm_object_t obj; obj = m->object; VM_OBJECT_ASSERT_LOCKED(obj); vm_page_lock_assert(m, MA_NOTOWNED); if (!_vm_page_busy_sleep(obj, m, wmesg, nonshared, true)) VM_OBJECT_DROP(obj); } /* * _vm_page_busy_sleep: * * Internal busy sleep function. */ static bool _vm_page_busy_sleep(vm_object_t obj, vm_page_t m, const char *wmesg, bool nonshared, bool locked) { u_int x; /* * If the object is busy we must wait for that to drain to zero * before trying the page again. */ if (obj != NULL && vm_object_busied(obj)) { if (locked) VM_OBJECT_DROP(obj); vm_object_busy_wait(obj, wmesg); return (locked); } sleepq_lock(m); x = m->busy_lock; if (x == VPB_UNBUSIED || (nonshared && (x & VPB_BIT_SHARED) != 0) || ((x & VPB_BIT_WAITERS) == 0 && !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS))) { sleepq_release(m); return (false); } if (locked) VM_OBJECT_DROP(obj); DROP_GIANT(); sleepq_add(m, NULL, wmesg, 0, 0); sleepq_wait(m, PVM); PICKUP_GIANT(); return (locked); } /* * vm_page_trysbusy: * * Try to shared busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_trysbusy(vm_page_t m) { vm_object_t obj; u_int x; obj = m->object; x = m->busy_lock; for (;;) { if ((x & VPB_BIT_SHARED) == 0) return (0); /* * Reduce the window for transient busies that will trigger * false negatives in vm_page_ps_test(). */ if (obj != NULL && vm_object_busied(obj)) return (0); if (atomic_fcmpset_acq_int(&m->busy_lock, &x, x + VPB_ONE_SHARER)) break; } /* Refetch the object now that we're guaranteed that it is stable. */ obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_sunbusy(m); return (0); } return (1); } /* * vm_page_tryxbusy: * * Try to exclusive busy a page. * If the operation succeeds 1 is returned otherwise 0. * The operation never sleeps. */ int vm_page_tryxbusy(vm_page_t m) { vm_object_t obj; if (atomic_cmpset_acq_int(&(m)->busy_lock, VPB_UNBUSIED, VPB_CURTHREAD_EXCLUSIVE) == 0) return (0); obj = m->object; if (obj != NULL && vm_object_busied(obj)) { vm_page_xunbusy(m); return (0); } return (1); } static void vm_page_xunbusy_hard_tail(vm_page_t m) { atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED); /* Wake the waiter. */ wakeup(m); } /* * vm_page_xunbusy_hard: * * Called when unbusy has failed because there is a waiter. */ void vm_page_xunbusy_hard(vm_page_t m) { vm_page_assert_xbusied(m); vm_page_xunbusy_hard_tail(m); } void vm_page_xunbusy_hard_unchecked(vm_page_t m) { vm_page_assert_xbusied_unchecked(m); vm_page_xunbusy_hard_tail(m); } /* * Avoid releasing and reacquiring the same page lock. */ void vm_page_change_lock(vm_page_t m, struct mtx **mtx) { struct mtx *mtx1; mtx1 = vm_page_lockptr(m); if (*mtx == mtx1) return; if (*mtx != NULL) mtx_unlock(*mtx); *mtx = mtx1; mtx_lock(mtx1); } /* * vm_page_unhold_pages: * * Unhold each of the pages that is referenced by the given array. */ void vm_page_unhold_pages(vm_page_t *ma, int count) { for (; count != 0; count--) { vm_page_unwire(*ma, PQ_ACTIVE); ma++; } } vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa) { vm_page_t m; #ifdef VM_PHYSSEG_SPARSE m = vm_phys_paddr_to_vm_page(pa); if (m == NULL) m = vm_phys_fictitious_to_vm_page(pa); return (m); #elif defined(VM_PHYSSEG_DENSE) long pi; pi = atop(pa); if (pi >= first_page && (pi - first_page) < vm_page_array_size) { m = &vm_page_array[pi - first_page]; return (m); } return (vm_phys_fictitious_to_vm_page(pa)); #else #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined." #endif } /* * vm_page_getfake: * * Create a fictitious page with the specified physical address and * memory attribute. The memory attribute is the only the machine- * dependent aspect of a fictitious page that must be initialized. */ vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr) { vm_page_t m; m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO); vm_page_initfake(m, paddr, memattr); return (m); } void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { if ((m->flags & PG_FICTITIOUS) != 0) { /* * The page's memattr might have changed since the * previous initialization. Update the pmap to the * new memattr. */ goto memattr; } m->phys_addr = paddr; m->a.queue = PQ_NONE; /* Fictitious pages don't use "segind". */ m->flags = PG_FICTITIOUS; /* Fictitious pages don't use "order" or "pool". */ m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; /* Fictitious pages are unevictable. */ m->ref_count = 1; pmap_page_init(m); memattr: pmap_page_set_memattr(m, memattr); } /* * vm_page_putfake: * * Release a fictitious page. */ void vm_page_putfake(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m)); KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_putfake: bad page %p", m)); vm_page_xunbusy(m); uma_zfree(fakepg_zone, m); } /* * vm_page_updatefake: * * Update the given fictitious page to the specified physical address and * memory attribute. */ void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr) { KASSERT((m->flags & PG_FICTITIOUS) != 0, ("vm_page_updatefake: bad page %p", m)); m->phys_addr = paddr; pmap_page_set_memattr(m, memattr); } /* * vm_page_free: * * Free a page. */ void vm_page_free(vm_page_t m) { m->flags &= ~PG_ZERO; vm_page_free_toq(m); } /* * vm_page_free_zero: * * Free a page to the zerod-pages queue */ void vm_page_free_zero(vm_page_t m) { m->flags |= PG_ZERO; vm_page_free_toq(m); } /* * Unbusy and handle the page queueing for a page from a getpages request that * was optionally read ahead or behind. */ void vm_page_readahead_finish(vm_page_t m) { /* We shouldn't put invalid pages on queues. */ KASSERT(!vm_page_none_valid(m), ("%s: %p is invalid", __func__, m)); /* * Since the page is not the actually needed one, whether it should * be activated or deactivated is not obvious. Empirical results * have shown that deactivating the page is usually the best choice, * unless the page is wanted by another thread. */ - vm_page_lock(m); if ((m->busy_lock & VPB_BIT_WAITERS) != 0) vm_page_activate(m); else vm_page_deactivate(m); - vm_page_unlock(m); vm_page_xunbusy_unchecked(m); } /* * vm_page_sleep_if_busy: * * Sleep and release the object lock if the page is busied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_busy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; if (vm_page_busied(m) || (obj != NULL && obj->busy)) { vm_page_busy_sleep(m, msg, false); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_sleep_if_xbusy: * * Sleep and release the object lock if the page is xbusied. * Returns TRUE if the thread slept. * * The given page must be unlocked and object containing it must * be locked. */ int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg) { vm_object_t obj; vm_page_lock_assert(m, MA_NOTOWNED); VM_OBJECT_ASSERT_WLOCKED(m->object); /* * The page-specific object must be cached because page * identity can change during the sleep, causing the * re-lock of a different object. * It is assumed that a reference to the object is already * held by the callers. */ obj = m->object; if (vm_page_xbusied(m) || (obj != NULL && obj->busy)) { vm_page_busy_sleep(m, msg, true); VM_OBJECT_WLOCK(obj); return (TRUE); } return (FALSE); } /* * vm_page_dirty_KBI: [ internal use only ] * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). * * This function should only be called by vm_page_dirty(). */ void vm_page_dirty_KBI(vm_page_t m) { /* Refer to this operation by its public name. */ KASSERT(vm_page_all_valid(m), ("vm_page_dirty: page is invalid!")); m->dirty = VM_PAGE_BITS_ALL; } /* * vm_page_insert: [ internal use only ] * * Inserts the given mem entry into the object and object list. * * The object must be locked. */ int vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) { vm_page_t mpred; VM_OBJECT_ASSERT_WLOCKED(object); mpred = vm_radix_lookup_le(&object->rtree, pindex); return (vm_page_insert_after(m, object, pindex, mpred)); } /* * vm_page_insert_after: * * Inserts the page "m" into the specified object at offset "pindex". * * The page "mpred" must immediately precede the offset "pindex" within * the specified object. * * The object must be locked. */ static int vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex, vm_page_t mpred) { vm_page_t msucc; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(m->object == NULL, ("vm_page_insert_after: page already inserted")); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_after: object doesn't contain mpred")); KASSERT(mpred->pindex < pindex, ("vm_page_insert_after: mpred doesn't precede pindex")); msucc = TAILQ_NEXT(mpred, listq); } else msucc = TAILQ_FIRST(&object->memq); if (msucc != NULL) KASSERT(msucc->pindex > pindex, ("vm_page_insert_after: msucc doesn't succeed pindex")); /* * Record the object/offset pair in this page. */ m->object = object; m->pindex = pindex; m->ref_count |= VPRC_OBJREF; /* * Now link into the object's ordered list of backed pages. */ if (vm_radix_insert(&object->rtree, m)) { m->object = NULL; m->pindex = 0; m->ref_count &= ~VPRC_OBJREF; return (1); } vm_page_insert_radixdone(m, object, mpred); return (0); } /* * vm_page_insert_radixdone: * * Complete page "m" insertion into the specified object after the * radix trie hooking. * * The page "mpred" must precede the offset "m->pindex" within the * specified object. * * The object must be locked. */ static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object != NULL && m->object == object, ("vm_page_insert_radixdone: page %p has inconsistent object", m)); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("vm_page_insert_radixdone: page %p is missing object ref", m)); if (mpred != NULL) { KASSERT(mpred->object == object, ("vm_page_insert_radixdone: object doesn't contain mpred")); KASSERT(mpred->pindex < m->pindex, ("vm_page_insert_radixdone: mpred doesn't precede pindex")); } if (mpred != NULL) TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq); else TAILQ_INSERT_HEAD(&object->memq, m, listq); /* * Show that the object has one more resident page. */ object->resident_page_count++; /* * Hold the vnode until the last page is released. */ if (object->resident_page_count == 1 && object->type == OBJT_VNODE) vhold(object->handle); /* * Since we are inserting a new and possibly dirty page, * update the object's generation count. */ if (pmap_page_is_write_mapped(m)) vm_object_set_writeable_dirty(object); } /* * Do the work to remove a page from its object. The caller is responsible for * updating the page's fields to reflect this removal. */ static void vm_page_object_remove(vm_page_t m) { vm_object_t object; vm_page_t mrem; vm_page_assert_xbusied(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((m->ref_count & VPRC_OBJREF) != 0, ("page %p is missing its object ref", m)); /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0) vm_pager_page_unswapped(m); mrem = vm_radix_remove(&object->rtree, m->pindex); KASSERT(mrem == m, ("removed page %p, expected page %p", mrem, m)); /* * Now remove from the object's list of backed pages. */ TAILQ_REMOVE(&object->memq, m, listq); /* * And show that the object has one fewer resident page. */ object->resident_page_count--; /* * The vnode may now be recycled. */ if (object->resident_page_count == 0 && object->type == OBJT_VNODE) vdrop(object->handle); } /* * vm_page_remove: * * Removes the specified page from its containing object, but does not * invalidate any backing storage. Returns true if the object's reference * was the last reference to the page, and false otherwise. * * The object must be locked and the page must be exclusively busied. * The exclusive busy will be released on return. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ bool vm_page_remove(vm_page_t m) { bool dropped; dropped = vm_page_remove_xbusy(m); vm_page_xunbusy(m); return (dropped); } /* * vm_page_remove_xbusy * * Removes the page but leaves the xbusy held. Returns true if this * removed the final ref and false otherwise. */ bool vm_page_remove_xbusy(vm_page_t m) { vm_page_object_remove(m); m->object = NULL; return (vm_page_drop(m, VPRC_OBJREF) == VPRC_OBJREF); } /* * vm_page_lookup: * * Returns the page associated with the object/offset * pair specified; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_lookup(vm_object_t object, vm_pindex_t pindex) { VM_OBJECT_ASSERT_LOCKED(object); return (vm_radix_lookup(&object->rtree, pindex)); } /* * vm_page_find_least: * * Returns the page associated with the object with least pindex * greater than or equal to the parameter pindex, or NULL. * * The object must be locked. */ vm_page_t vm_page_find_least(vm_object_t object, vm_pindex_t pindex) { vm_page_t m; VM_OBJECT_ASSERT_LOCKED(object); if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex) m = vm_radix_lookup_ge(&object->rtree, pindex); return (m); } /* * Returns the given page's successor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_next(vm_page_t m) { vm_page_t next; VM_OBJECT_ASSERT_LOCKED(m->object); if ((next = TAILQ_NEXT(m, listq)) != NULL) { MPASS(next->object == m->object); if (next->pindex != m->pindex + 1) next = NULL; } return (next); } /* * Returns the given page's predecessor (by pindex) within the object if it is * resident; if none is found, NULL is returned. * * The object must be locked. */ vm_page_t vm_page_prev(vm_page_t m) { vm_page_t prev; VM_OBJECT_ASSERT_LOCKED(m->object); if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL) { MPASS(prev->object == m->object); if (prev->pindex != m->pindex - 1) prev = NULL; } return (prev); } /* * Uses the page mnew as a replacement for an existing page at index * pindex which must be already present in the object. * * Both pages must be exclusively busied on enter. The old page is * unbusied on exit. * * A return value of true means mold is now free. If this is not the * final ref and the caller does not hold a wire reference it may not * continue to access the page. */ static bool vm_page_replace_hold(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_t mret; bool dropped; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_xbusied(mold); KASSERT(mnew->object == NULL && (mnew->ref_count & VPRC_OBJREF) == 0, ("vm_page_replace: page %p already in object", mnew)); /* * This function mostly follows vm_page_insert() and * vm_page_remove() without the radix, object count and vnode * dance. Double check such functions for more comments. */ mnew->object = object; mnew->pindex = pindex; atomic_set_int(&mnew->ref_count, VPRC_OBJREF); mret = vm_radix_replace(&object->rtree, mnew); KASSERT(mret == mold, ("invalid page replacement, mold=%p, mret=%p", mold, mret)); KASSERT((mold->oflags & VPO_UNMANAGED) == (mnew->oflags & VPO_UNMANAGED), ("vm_page_replace: mismatched VPO_UNMANAGED")); /* Keep the resident page list in sorted order. */ TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq); TAILQ_REMOVE(&object->memq, mold, listq); mold->object = NULL; /* * The object's resident_page_count does not change because we have * swapped one page for another, but the generation count should * change if the page is dirty. */ if (pmap_page_is_write_mapped(mnew)) vm_object_set_writeable_dirty(object); dropped = vm_page_drop(mold, VPRC_OBJREF) == VPRC_OBJREF; vm_page_xunbusy(mold); return (dropped); } void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold) { vm_page_assert_xbusied(mnew); if (vm_page_replace_hold(mnew, object, pindex, mold)) vm_page_free(mold); } /* * vm_page_rename: * * Move the given memory entry from its * current object to the specified target object/offset. * * Note: swap associated with the page must be invalidated by the move. We * have to do this for several reasons: (1) we aren't freeing the * page, (2) we are dirtying the page, (3) the VM system is probably * moving the page from object A to B, and will then later move * the backing store from A to B and we can't have a conflict. * * Note: we *always* dirty the page. It is necessary both for the * fact that we moved it, and because we may be invalidating * swap. * * The objects must be locked. */ int vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) { vm_page_t mpred; vm_pindex_t opidx; VM_OBJECT_ASSERT_WLOCKED(new_object); KASSERT(m->ref_count != 0, ("vm_page_rename: page %p has no refs", m)); mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex); KASSERT(mpred == NULL || mpred->pindex != new_pindex, ("vm_page_rename: pindex already renamed")); /* * Create a custom version of vm_page_insert() which does not depend * by m_prev and can cheat on the implementation aspects of the * function. */ opidx = m->pindex; m->pindex = new_pindex; if (vm_radix_insert(&new_object->rtree, m)) { m->pindex = opidx; return (1); } /* * The operation cannot fail anymore. The removal must happen before * the listq iterator is tainted. */ m->pindex = opidx; vm_page_object_remove(m); /* Return back to the new pindex to complete vm_page_insert(). */ m->pindex = new_pindex; m->object = new_object; vm_page_insert_radixdone(m, new_object, mpred); vm_page_dirty(m); return (0); } /* * vm_page_alloc: * * Allocate and return a page that is associated with the specified * object and offset pair. By default, this page is exclusive busied. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req) { return (vm_page_alloc_after(object, pindex, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } vm_page_t vm_page_alloc_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req) { return (vm_page_alloc_domain_after(object, pindex, domain, req, object != NULL ? vm_radix_lookup_le(&object->rtree, pindex) : NULL)); } /* * Allocate a page in the specified object with the given page index. To * optimize insertion of the page into the object, the caller must also specifiy * the resident page in the object with largest index smaller than the given * page index, or NULL if no such page exists. */ vm_page_t vm_page_alloc_after(vm_object_t object, vm_pindex_t pindex, int req, vm_page_t mpred) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_domain_after(object, pindex, domain, req, mpred); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } /* * Returns true if the number of free pages exceeds the minimum * for the request class and false otherwise. */ static int _vm_domain_allocate(struct vm_domain *vmd, int req_class, int npages) { u_int limit, old, new; if (req_class == VM_ALLOC_INTERRUPT) limit = 0; else if (req_class == VM_ALLOC_SYSTEM) limit = vmd->vmd_interrupt_free_min; else limit = vmd->vmd_free_reserved; /* * Attempt to reserve the pages. Fail if we're below the limit. */ limit += npages; old = vmd->vmd_free_count; do { if (old < limit) return (0); new = old - npages; } while (atomic_fcmpset_int(&vmd->vmd_free_count, &old, new) == 0); /* Wake the page daemon if we've crossed the threshold. */ if (vm_paging_needed(vmd, new) && !vm_paging_needed(vmd, old)) pagedaemon_wakeup(vmd->vmd_domain); /* Only update bitsets on transitions. */ if ((old >= vmd->vmd_free_min && new < vmd->vmd_free_min) || (old >= vmd->vmd_free_severe && new < vmd->vmd_free_severe)) vm_domain_set(vmd); return (1); } int vm_domain_allocate(struct vm_domain *vmd, int req, int npages) { int req_class; /* * The page daemon is allowed to dig deeper into the free page list. */ req_class = req & VM_ALLOC_CLASS_MASK; if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; return (_vm_domain_allocate(vmd, req_class, npages)); } vm_page_t vm_page_alloc_domain_after(vm_object_t object, vm_pindex_t pindex, int domain, int req, vm_page_t mpred) { struct vm_domain *vmd; vm_page_t m; int flags, pool; KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); KASSERT(mpred == NULL || mpred->pindex < pindex, ("mpred %p doesn't precede pindex 0x%jx", mpred, (uintmax_t)pindex)); if (object != NULL) VM_OBJECT_ASSERT_WLOCKED(object); flags = 0; m = NULL; pool = object != NULL ? VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the page from a reservation? */ if (vm_object_reserv(object) && (m = vm_reserv_alloc_page(object, pindex, domain, req, mpred)) != NULL) { domain = vm_phys_domain(m); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vmd->vmd_pgcache[pool].zone != NULL) { m = uma_zalloc(vmd->vmd_pgcache[pool].zone, M_NOWAIT); if (m != NULL) { flags |= PG_PCPU_CACHE; goto found; } } if (vm_domain_allocate(vmd, req, 1)) { /* * If not, allocate it from the free page queues. */ vm_domain_free_lock(vmd); m = vm_phys_alloc_pages(domain, pool, 0); vm_domain_free_unlock(vmd); if (m == NULL) { vm_domain_freecnt_inc(vmd, 1); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_inactive(domain)) goto again; #endif } } if (m == NULL) { /* * Not allocatable, give up. */ if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } /* * At this point we had better have found a good page. */ found: vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ if ((req & VM_ALLOC_ZERO) != 0) flags |= (m->flags & PG_ZERO); if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; m->flags = flags; m->a.flags = 0; m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; m->busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) m->busy_lock = VPB_CURTHREAD_EXCLUSIVE; if ((req & VM_ALLOC_SBUSY) != 0) m->busy_lock = VPB_SHARERS_WORD(1); if (req & VM_ALLOC_WIRED) { vm_wire_add(1); m->ref_count = 1; } m->a.act_count = 0; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if (req & VM_ALLOC_WIRED) { vm_wire_sub(1); m->ref_count = 0; } KASSERT(m->object == NULL, ("page %p has object", m)); m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } /* Ignore device objects; the pager sets "memattr" for them. */ if (object->memattr != VM_MEMATTR_DEFAULT && (object->flags & OBJ_FICTITIOUS) == 0) pmap_page_set_memattr(m, object->memattr); } else m->pindex = pindex; return (m); } /* * vm_page_alloc_contig: * * Allocate a contiguous set of physical pages of the given size "npages" * from the free lists. All of the physical pages must be at or above * the given physical address "low" and below the given physical address * "high". The given value "alignment" determines the alignment of the * first physical page in the set. If the given value "boundary" is * non-zero, then the set of physical pages cannot cross any physical * address boundary that is a multiple of that value. Both "alignment" * and "boundary" must be a power of two. * * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT, * then the memory attribute setting for the physical pages is configured * to the object's memory attribute setting. Otherwise, the memory * attribute setting for the physical pages is configured to "memattr", * overriding the object's memory attribute setting. However, if the * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the * memory attribute setting for the physical pages cannot be configured * to VM_MEMATTR_DEFAULT. * * The specified object may not contain fictitious pages. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NODUMP do not include the page in a kernel core dump * VM_ALLOC_NOOBJ page is not associated with an object and * should not be exclusive busy * VM_ALLOC_SBUSY shared busy the allocated page * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, object, pindex, &domain, &req); do { m = vm_page_alloc_contig_domain(object, pindex, domain, req, npages, low, high, alignment, boundary, memattr); if (m != NULL) break; } while (vm_domainset_iter_page(&di, object, &domain) == 0); return (m); } vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr) { struct vm_domain *vmd; vm_page_t m, m_ret, mpred; u_int busy_lock, flags, oflags; mpred = NULL; /* XXX: pacify gcc */ KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) && (object != NULL || (req & VM_ALLOC_SBUSY) == 0) && ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) != (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)), ("vm_page_alloc_contig: inconsistent object(%p)/req(%x)", object, req)); KASSERT(object == NULL || (req & VM_ALLOC_WAITOK) == 0, ("Can't sleep and retry object insertion.")); if (object != NULL) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT((object->flags & OBJ_FICTITIOUS) == 0, ("vm_page_alloc_contig: object %p has fictitious pages", object)); } KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero")); if (object != NULL) { mpred = vm_radix_lookup_le(&object->rtree, pindex); KASSERT(mpred == NULL || mpred->pindex != pindex, ("vm_page_alloc_contig: pindex already allocated")); } /* * Can we allocate the pages without the number of free pages falling * below the lower bound for the allocation class? */ m_ret = NULL; again: #if VM_NRESERVLEVEL > 0 /* * Can we allocate the pages from a reservation? */ if (vm_object_reserv(object) && (m_ret = vm_reserv_alloc_contig(object, pindex, domain, req, mpred, npages, low, high, alignment, boundary)) != NULL) { domain = vm_phys_domain(m_ret); vmd = VM_DOMAIN(domain); goto found; } #endif vmd = VM_DOMAIN(domain); if (vm_domain_allocate(vmd, req, npages)) { /* * allocate them from the free page queues. */ vm_domain_free_lock(vmd); m_ret = vm_phys_alloc_contig(domain, npages, low, high, alignment, boundary); vm_domain_free_unlock(vmd); if (m_ret == NULL) { vm_domain_freecnt_inc(vmd, npages); #if VM_NRESERVLEVEL > 0 if (vm_reserv_reclaim_contig(domain, npages, low, high, alignment, boundary)) goto again; #endif } } if (m_ret == NULL) { if (vm_domain_alloc_fail(vmd, object, req)) goto again; return (NULL); } #if VM_NRESERVLEVEL > 0 found: #endif for (m = m_ret; m < &m_ret[npages]; m++) { vm_page_dequeue(m); vm_page_alloc_check(m); } /* * Initialize the pages. Only the PG_ZERO flag is inherited. */ flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; if ((req & VM_ALLOC_NODUMP) != 0) flags |= PG_NODUMP; oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ? VPO_UNMANAGED : 0; busy_lock = VPB_UNBUSIED; if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0) busy_lock = VPB_CURTHREAD_EXCLUSIVE; if ((req & VM_ALLOC_SBUSY) != 0) busy_lock = VPB_SHARERS_WORD(1); if ((req & VM_ALLOC_WIRED) != 0) vm_wire_add(npages); if (object != NULL) { if (object->memattr != VM_MEMATTR_DEFAULT && memattr == VM_MEMATTR_DEFAULT) memattr = object->memattr; } for (m = m_ret; m < &m_ret[npages]; m++) { m->a.flags = 0; m->flags = (m->flags | PG_NODUMP) & flags; m->busy_lock = busy_lock; if ((req & VM_ALLOC_WIRED) != 0) m->ref_count = 1; m->a.act_count = 0; m->oflags = oflags; if (object != NULL) { if (vm_page_insert_after(m, object, pindex, mpred)) { if ((req & VM_ALLOC_WIRED) != 0) vm_wire_sub(npages); KASSERT(m->object == NULL, ("page %p has object", m)); mpred = m; for (m = m_ret; m < &m_ret[npages]; m++) { if (m <= mpred && (req & VM_ALLOC_WIRED) != 0) m->ref_count = 0; m->oflags = VPO_UNMANAGED; m->busy_lock = VPB_UNBUSIED; /* Don't change PG_ZERO. */ vm_page_free_toq(m); } if (req & VM_ALLOC_WAITFAIL) { VM_OBJECT_WUNLOCK(object); vm_radix_wait(); VM_OBJECT_WLOCK(object); } return (NULL); } mpred = m; } else m->pindex = pindex; if (memattr != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, memattr); pindex++; } return (m_ret); } /* * Check a page that has been freshly dequeued from a freelist. */ static void vm_page_alloc_check(vm_page_t m) { KASSERT(m->object == NULL, ("page %p has object", m)); KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("page %p has unexpected queue %d, flags %#x", m, m->a.queue, (m->a.flags & PGA_QUEUE_STATE_MASK))); KASSERT(m->ref_count == 0, ("page %p has references", m)); KASSERT(!vm_page_busied(m), ("page %p is busy", m)); KASSERT(m->dirty == 0, ("page %p is dirty", m)); KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has unexpected memattr %d", m, pmap_page_get_memattr(m))); KASSERT(m->valid == 0, ("free page %p is valid", m)); } /* * vm_page_alloc_freelist: * * Allocate a physical page from the specified free page list. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * optional allocation flags: * VM_ALLOC_COUNT(number) the number of additional pages that the caller * intends to allocate * VM_ALLOC_WIRED wire the allocated page * VM_ALLOC_ZERO prefer a zeroed page */ vm_page_t vm_page_alloc_freelist(int freelist, int req) { struct vm_domainset_iter di; vm_page_t m; int domain; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { m = vm_page_alloc_freelist_domain(domain, freelist, req); if (m != NULL) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (m); } vm_page_t vm_page_alloc_freelist_domain(int domain, int freelist, int req) { struct vm_domain *vmd; vm_page_t m; u_int flags; m = NULL; vmd = VM_DOMAIN(domain); again: if (vm_domain_allocate(vmd, req, 1)) { vm_domain_free_lock(vmd); m = vm_phys_alloc_freelist_pages(domain, freelist, VM_FREEPOOL_DIRECT, 0); vm_domain_free_unlock(vmd); if (m == NULL) vm_domain_freecnt_inc(vmd, 1); } if (m == NULL) { if (vm_domain_alloc_fail(vmd, NULL, req)) goto again; return (NULL); } vm_page_dequeue(m); vm_page_alloc_check(m); /* * Initialize the page. Only the PG_ZERO flag is inherited. */ m->a.flags = 0; flags = 0; if ((req & VM_ALLOC_ZERO) != 0) flags = PG_ZERO; m->flags &= flags; if ((req & VM_ALLOC_WIRED) != 0) { vm_wire_add(1); m->ref_count = 1; } /* Unmanaged pages don't use "act_count". */ m->oflags = VPO_UNMANAGED; return (m); } static int vm_page_zone_import(void *arg, void **store, int cnt, int domain, int flags) { struct vm_domain *vmd; struct vm_pgcache *pgcache; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); /* * The page daemon should avoid creating extra memory pressure since its * main purpose is to replenish the store of free pages. */ if (vmd->vmd_severeset || curproc == pageproc || !_vm_domain_allocate(vmd, VM_ALLOC_NORMAL, cnt)) return (0); domain = vmd->vmd_domain; vm_domain_free_lock(vmd); i = vm_phys_alloc_npages(domain, pgcache->pool, cnt, (vm_page_t *)store); vm_domain_free_unlock(vmd); if (cnt != i) vm_domain_freecnt_inc(vmd, cnt - i); return (i); } static void vm_page_zone_release(void *arg, void **store, int cnt) { struct vm_domain *vmd; struct vm_pgcache *pgcache; vm_page_t m; int i; pgcache = arg; vmd = VM_DOMAIN(pgcache->domain); vm_domain_free_lock(vmd); for (i = 0; i < cnt; i++) { m = (vm_page_t)store[i]; vm_phys_free_pages(m, 0); } vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } #define VPSC_ANY 0 /* No restrictions. */ #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */ #define VPSC_NOSUPER 2 /* Skip superpages. */ /* * vm_page_scan_contig: * * Scan vm_page_array[] between the specified entries "m_start" and * "m_end" for a run of contiguous physical pages that satisfy the * specified conditions, and return the lowest page in the run. The * specified "alignment" determines the alignment of the lowest physical * page in the run. If the specified "boundary" is non-zero, then the * run of physical pages cannot span a physical address that is a * multiple of "boundary". * * "m_end" is never dereferenced, so it need not point to a vm_page * structure within vm_page_array[]. * * "npages" must be greater than zero. "m_start" and "m_end" must not * span a hole (or discontiguity) in the physical address space. Both * "alignment" and "boundary" must be a power of two. */ vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options) { struct mtx *m_mtx; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_run; #if VM_NRESERVLEVEL > 0 int level; #endif int m_inc, order, run_ext, run_len; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); m_run = NULL; run_len = 0; m_mtx = NULL; for (m = m_start; m < m_end && run_len < npages; m += m_inc) { KASSERT((m->flags & PG_MARKER) == 0, ("page %p is PG_MARKER", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || m->ref_count >= 1, ("fictitious page %p has invalid ref count", m)); /* * If the current page would be the start of a run, check its * physical address against the end, alignment, and boundary * conditions. If it doesn't satisfy these conditions, either * terminate the scan or advance to the next page that * satisfies the failed condition. */ if (run_len == 0) { KASSERT(m_run == NULL, ("m_run != NULL")); if (m + npages > m_end) break; pa = VM_PAGE_TO_PHYS(m); if ((pa & (alignment - 1)) != 0) { m_inc = atop(roundup2(pa, alignment) - pa); continue; } if (rounddown2(pa ^ (pa + ptoa(npages) - 1), boundary) != 0) { m_inc = atop(roundup2(pa, boundary) - pa); continue; } } else KASSERT(m_run != NULL, ("m_run == NULL")); vm_page_change_lock(m, &m_mtx); m_inc = 1; retry: if (vm_page_wired(m)) run_ext = 0; #if VM_NRESERVLEVEL > 0 else if ((level = vm_reserv_level(m)) >= 0 && (options & VPSC_NORESERV) != 0) { run_ext = 0; /* Advance to the end of the reservation. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); } #endif else if ((object = m->object) != NULL) { /* * The page is considered eligible for relocation if * and only if it could be laundered or reclaimed by * the page daemon. */ if (!VM_OBJECT_TRYRLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_RLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_RUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) { run_ext = 0; #if VM_NRESERVLEVEL > 0 } else if ((options & VPSC_NOSUPER) != 0 && (level = vm_reserv_level_iffullpop(m)) >= 0) { run_ext = 0; /* Advance to the end of the superpage. */ pa = VM_PAGE_TO_PHYS(m); m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) - pa); #endif } else if (object->memattr == VM_MEMATTR_DEFAULT && vm_page_queue(m) != PQ_NONE && !vm_page_busied(m) && !vm_page_wired(m)) { /* * The page is allocated but eligible for * relocation. Extend the current run by one * page. */ KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT((m->oflags & (VPO_SWAPINPROG | VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ run_ext = 1; } else run_ext = 0; VM_OBJECT_RUNLOCK(object); #if VM_NRESERVLEVEL > 0 } else if (level >= 0) { /* * The page is reserved but not yet allocated. In * other words, it is still free. Extend the current * run by one page. */ run_ext = 1; #endif } else if ((order = m->order) < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it is the * first page in a power-of-two-sized run of * contiguous free pages. Add these pages to the end * of the current run, and jump ahead. */ run_ext = 1 << order; m_inc = 1 << order; } else { /* * Skip the page for one of the following reasons: (1) * It is enqueued in the physical memory allocator's * free page queues. However, it is not the first * page in a run of contiguous free pages. (This case * rarely occurs because the scan is performed in * ascending order.) (2) It is not reserved, and it is * transitioning from free to allocated. (Conversely, * the transition from allocated to free for managed * pages is blocked by the page lock.) (3) It is * allocated but not contained by an object and not * wired, e.g., allocated by Xen's balloon driver. */ run_ext = 0; } /* * Extend or reset the current run of pages. */ if (run_ext > 0) { if (run_len == 0) m_run = m; run_len += run_ext; } else { if (run_len > 0) { m_run = NULL; run_len = 0; } } } if (m_mtx != NULL) mtx_unlock(m_mtx); if (run_len >= npages) return (m_run); return (NULL); } /* * vm_page_reclaim_run: * * Try to relocate each of the allocated virtual pages within the * specified run of physical pages to a new physical address. Free the * physical pages underlying the relocated virtual pages. A virtual page * is relocatable if and only if it could be laundered or reclaimed by * the page daemon. Whenever possible, a virtual page is relocated to a * physical address above "high". * * Returns 0 if every physical page within the run was already free or * just freed by a successful relocation. Otherwise, returns a non-zero * value indicating why the last attempt to relocate a virtual page was * unsuccessful. * * "req_class" must be an allocation class. */ static int vm_page_reclaim_run(int req_class, int domain, u_long npages, vm_page_t m_run, vm_paddr_t high) { struct vm_domain *vmd; struct mtx *m_mtx; struct spglist free; vm_object_t object; vm_paddr_t pa; vm_page_t m, m_end, m_new; int error, order, req; KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class, ("req_class is not an allocation class")); SLIST_INIT(&free); error = 0; m = m_run; m_end = m_run + npages; m_mtx = NULL; for (; error == 0 && m < m_end; m++) { KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0, ("page %p is PG_FICTITIOUS or PG_MARKER", m)); /* * Avoid releasing and reacquiring the same page lock. */ vm_page_change_lock(m, &m_mtx); retry: /* * Racily check for wirings. Races are handled below. */ if (vm_page_wired(m)) error = EBUSY; else if ((object = m->object) != NULL) { /* * The page is relocated if and only if it could be * laundered or reclaimed by the page daemon. */ if (!VM_OBJECT_TRYWLOCK(object)) { mtx_unlock(m_mtx); VM_OBJECT_WLOCK(object); mtx_lock(m_mtx); if (m->object != object) { /* * The page may have been freed. */ VM_OBJECT_WUNLOCK(object); goto retry; } } /* Don't care: PG_NODUMP, PG_ZERO. */ if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP && object->type != OBJT_VNODE) error = EINVAL; else if (object->memattr != VM_MEMATTR_DEFAULT) error = EINVAL; else if (vm_page_queue(m) != PQ_NONE && vm_page_tryxbusy(m) != 0) { if (vm_page_wired(m)) { vm_page_xunbusy(m); error = EBUSY; goto unlock; } KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT, ("page %p has an unexpected memattr", m)); KASSERT(m->oflags == 0, ("page %p has unexpected oflags", m)); /* Don't care: PGA_NOSYNC. */ if (!vm_page_none_valid(m)) { /* * First, try to allocate a new page * that is above "high". Failing * that, try to allocate a new page * that is below "m_run". Allocate * the new page between the end of * "m_run" and "high" only as a last * resort. */ req = req_class | VM_ALLOC_NOOBJ; if ((m->flags & PG_NODUMP) != 0) req |= VM_ALLOC_NODUMP; if (trunc_page(high) != ~(vm_paddr_t)PAGE_MASK) { m_new = vm_page_alloc_contig( NULL, 0, req, 1, round_page(high), ~(vm_paddr_t)0, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } else m_new = NULL; if (m_new == NULL) { pa = VM_PAGE_TO_PHYS(m_run); m_new = vm_page_alloc_contig( NULL, 0, req, 1, 0, pa - 1, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { pa += ptoa(npages); m_new = vm_page_alloc_contig( NULL, 0, req, 1, pa, high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); } if (m_new == NULL) { vm_page_xunbusy(m); error = ENOMEM; goto unlock; } /* * Unmap the page and check for new * wirings that may have been acquired * through a pmap lookup. */ if (object->ref_count != 0 && !vm_page_try_remove_all(m)) { vm_page_xunbusy(m); vm_page_free(m_new); error = EBUSY; goto unlock; } /* * Replace "m" with the new page. For * vm_page_replace(), "m" must be busy * and dequeued. Finally, change "m" * as if vm_page_free() was called. */ m_new->a.flags = m->a.flags & ~PGA_QUEUE_STATE_MASK; KASSERT(m_new->oflags == VPO_UNMANAGED, ("page %p is managed", m_new)); m_new->oflags = 0; pmap_copy_page(m, m_new); m_new->valid = m->valid; m_new->dirty = m->dirty; m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_replace_hold(m_new, object, m->pindex, m) && vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); /* * The new page must be deactivated * before the object is unlocked. */ vm_page_change_lock(m_new, &m_mtx); vm_page_deactivate(m_new); } else { m->flags &= ~PG_ZERO; vm_page_dequeue(m); if (vm_page_free_prep(m)) SLIST_INSERT_HEAD(&free, m, plinks.s.ss); KASSERT(m->dirty == 0, ("page %p is dirty", m)); } } else error = EBUSY; unlock: VM_OBJECT_WUNLOCK(object); } else { MPASS(vm_phys_domain(m) == domain); vmd = VM_DOMAIN(domain); vm_domain_free_lock(vmd); order = m->order; if (order < VM_NFREEORDER) { /* * The page is enqueued in the physical memory * allocator's free page queues. Moreover, it * is the first page in a power-of-two-sized * run of contiguous free pages. Jump ahead * to the last page within that run, and * continue from there. */ m += (1 << order) - 1; } #if VM_NRESERVLEVEL > 0 else if (vm_reserv_is_page_free(m)) order = 0; #endif vm_domain_free_unlock(vmd); if (order == VM_NFREEORDER) error = EINVAL; } } if (m_mtx != NULL) mtx_unlock(m_mtx); if ((m = SLIST_FIRST(&free)) != NULL) { int cnt; vmd = VM_DOMAIN(domain); cnt = 0; vm_domain_free_lock(vmd); do { MPASS(vm_phys_domain(m) == domain); SLIST_REMOVE_HEAD(&free, plinks.s.ss); vm_phys_free_pages(m, 0); cnt++; } while ((m = SLIST_FIRST(&free)) != NULL); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, cnt); } return (error); } #define NRUNS 16 CTASSERT(powerof2(NRUNS)); #define RUN_INDEX(count) ((count) & (NRUNS - 1)) #define MIN_RECLAIM 8 /* * vm_page_reclaim_contig: * * Reclaim allocated, contiguous physical memory satisfying the specified * conditions by relocating the virtual pages using that physical memory. * Returns true if reclamation is successful and false otherwise. Since * relocation requires the allocation of physical pages, reclamation may * fail due to a shortage of free pages. When reclamation fails, callers * are expected to perform vm_wait() before retrying a failed allocation * operation, e.g., vm_page_alloc_contig(). * * The caller must always specify an allocation class through "req". * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs a page * VM_ALLOC_INTERRUPT interrupt time request * * The optional allocation flags are ignored. * * "npages" must be greater than zero. Both "alignment" and "boundary" * must be a power of two. */ bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domain *vmd; vm_paddr_t curr_low; vm_page_t m_run, m_runs[NRUNS]; u_long count, reclaimed; int error, i, options, req_class; KASSERT(npages > 0, ("npages is 0")); KASSERT(powerof2(alignment), ("alignment is not a power of 2")); KASSERT(powerof2(boundary), ("boundary is not a power of 2")); req_class = req & VM_ALLOC_CLASS_MASK; /* * The page daemon is allowed to dig deeper into the free page list. */ if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT) req_class = VM_ALLOC_SYSTEM; /* * Return if the number of free pages cannot satisfy the requested * allocation. */ vmd = VM_DOMAIN(domain); count = vmd->vmd_free_count; if (count < npages + vmd->vmd_free_reserved || (count < npages + vmd->vmd_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) || (count < npages && req_class == VM_ALLOC_INTERRUPT)) return (false); /* * Scan up to three times, relaxing the restrictions ("options") on * the reclamation of reservations and superpages each time. */ for (options = VPSC_NORESERV;;) { /* * Find the highest runs that satisfy the given constraints * and restrictions, and record them in "m_runs". */ curr_low = low; count = 0; for (;;) { m_run = vm_phys_scan_contig(domain, npages, curr_low, high, alignment, boundary, options); if (m_run == NULL) break; curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages); m_runs[RUN_INDEX(count)] = m_run; count++; } /* * Reclaim the highest runs in LIFO (descending) order until * the number of reclaimed pages, "reclaimed", is at least * MIN_RECLAIM. Reset "reclaimed" each time because each * reclamation is idempotent, and runs will (likely) recur * from one scan to the next as restrictions are relaxed. */ reclaimed = 0; for (i = 0; count > 0 && i < NRUNS; i++) { count--; m_run = m_runs[RUN_INDEX(count)]; error = vm_page_reclaim_run(req_class, domain, npages, m_run, high); if (error == 0) { reclaimed += npages; if (reclaimed >= MIN_RECLAIM) return (true); } } /* * Either relax the restrictions on the next scan or return if * the last scan had no restrictions. */ if (options == VPSC_NORESERV) options = VPSC_NOSUPER; else if (options == VPSC_NOSUPER) options = VPSC_ANY; else if (options == VPSC_ANY) return (reclaimed != 0); } } bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary) { struct vm_domainset_iter di; int domain; bool ret; vm_domainset_iter_page_init(&di, NULL, 0, &domain, &req); do { ret = vm_page_reclaim_contig_domain(domain, req, npages, low, high, alignment, boundary); if (ret) break; } while (vm_domainset_iter_page(&di, NULL, &domain) == 0); return (ret); } /* * Set the domain in the appropriate page level domainset. */ void vm_domain_set(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (!vmd->vmd_minset && vm_paging_min(vmd)) { vmd->vmd_minset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_min_domains); } if (!vmd->vmd_severeset && vm_paging_severe(vmd)) { vmd->vmd_severeset = 1; DOMAINSET_SET(vmd->vmd_domain, &vm_severe_domains); } mtx_unlock(&vm_domainset_lock); } /* * Clear the domain from the appropriate page level domainset. */ void vm_domain_clear(struct vm_domain *vmd) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_minset && !vm_paging_min(vmd)) { vmd->vmd_minset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_min_domains); if (vm_min_waiters != 0) { vm_min_waiters = 0; wakeup(&vm_min_domains); } } if (vmd->vmd_severeset && !vm_paging_severe(vmd)) { vmd->vmd_severeset = 0; DOMAINSET_CLR(vmd->vmd_domain, &vm_severe_domains); if (vm_severe_waiters != 0) { vm_severe_waiters = 0; wakeup(&vm_severe_domains); } } /* * If pageout daemon needs pages, then tell it that there are * some free. */ if (vmd->vmd_pageout_pages_needed && vmd->vmd_free_count >= vmd->vmd_pageout_free_min) { wakeup(&vmd->vmd_pageout_pages_needed); vmd->vmd_pageout_pages_needed = 0; } /* See comments in vm_wait_doms(). */ if (vm_pageproc_waiters) { vm_pageproc_waiters = 0; wakeup(&vm_pageproc_waiters); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the min threshold globally. */ void vm_wait_min(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_min()) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } /* * Wait for free pages to exceed the severe threshold globally. */ void vm_wait_severe(void) { mtx_lock(&vm_domainset_lock); while (vm_page_count_severe()) { vm_severe_waiters++; msleep(&vm_severe_domains, &vm_domainset_lock, PVM, "vmwait", 0); } mtx_unlock(&vm_domainset_lock); } u_int vm_wait_count(void) { return (vm_severe_waiters + vm_min_waiters + vm_pageproc_waiters); } void vm_wait_doms(const domainset_t *wdoms) { /* * We use racey wakeup synchronization to avoid expensive global * locking for the pageproc when sleeping with a non-specific vm_wait. * To handle this, we only sleep for one tick in this instance. It * is expected that most allocations for the pageproc will come from * kmem or vm_page_grab* which will use the more specific and * race-free vm_wait_domain(). */ if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); vm_pageproc_waiters++; msleep(&vm_pageproc_waiters, &vm_domainset_lock, PVM | PDROP, "pageprocwait", 1); } else { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(wdoms)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PVM | PDROP, "vmwait", 0); } else mtx_unlock(&vm_domainset_lock); } } /* * vm_wait_domain: * * Sleep until free pages are available for allocation. * - Called in various places after failed memory allocations. */ void vm_wait_domain(int domain) { struct vm_domain *vmd; domainset_t wdom; vmd = VM_DOMAIN(domain); vm_domain_free_assert_unlocked(vmd); if (curproc == pageproc) { mtx_lock(&vm_domainset_lock); if (vmd->vmd_free_count < vmd->vmd_pageout_free_min) { vmd->vmd_pageout_pages_needed = 1; msleep(&vmd->vmd_pageout_pages_needed, &vm_domainset_lock, PDROP | PSWP, "VMWait", 0); } else mtx_unlock(&vm_domainset_lock); } else { if (pageproc == NULL) panic("vm_wait in early boot"); DOMAINSET_ZERO(&wdom); DOMAINSET_SET(vmd->vmd_domain, &wdom); vm_wait_doms(&wdom); } } /* * vm_wait: * * Sleep until free pages are available for allocation in the * affinity domains of the obj. If obj is NULL, the domain set * for the calling thread is used. * Called in various places after failed memory allocations. */ void vm_wait(vm_object_t obj) { struct domainset *d; d = NULL; /* * Carefully fetch pointers only once: the struct domainset * itself is ummutable but the pointer might change. */ if (obj != NULL) d = obj->domain.dr_policy; if (d == NULL) d = curthread->td_domain.dr_policy; vm_wait_doms(&d->ds_mask); } /* * vm_domain_alloc_fail: * * Called when a page allocation function fails. Informs the * pagedaemon and performs the requested wait. Requires the * domain_free and object lock on entry. Returns with the * object lock held and free lock released. Returns an error when * retry is necessary. * */ static int vm_domain_alloc_fail(struct vm_domain *vmd, vm_object_t object, int req) { vm_domain_free_assert_unlocked(vmd); atomic_add_int(&vmd->vmd_pageout_deficit, max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1)); if (req & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) { if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_wait_domain(vmd->vmd_domain); if (object != NULL) VM_OBJECT_WLOCK(object); if (req & VM_ALLOC_WAITOK) return (EAGAIN); } return (0); } /* * vm_waitpfault: * * Sleep until free pages are available for allocation. * - Called only in vm_fault so that processes page faulting * can be easily tracked. * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing * processes will be able to grab memory first. Do not change * this balance without careful testing first. */ void vm_waitpfault(struct domainset *dset, int timo) { /* * XXX Ideally we would wait only until the allocation could * be satisfied. This condition can cause new allocators to * consume all freed pages while old allocators wait. */ mtx_lock(&vm_domainset_lock); if (vm_page_count_min_set(&dset->ds_mask)) { vm_min_waiters++; msleep(&vm_min_domains, &vm_domainset_lock, PUSER | PDROP, "pfault", timo); } else mtx_unlock(&vm_domainset_lock); } static struct vm_pagequeue * _vm_page_pagequeue(vm_page_t m, uint8_t queue) { return (&vm_pagequeue_domain(m)->vmd_pagequeues[queue]); } #ifdef INVARIANTS static struct vm_pagequeue * vm_page_pagequeue(vm_page_t m) { return (_vm_page_pagequeue(m, vm_page_astate_load(m).queue)); } #endif static __always_inline bool vm_page_pqstate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_astate_t tmp; tmp = *old; do { if (__predict_true(vm_page_astate_fcmpset(m, old, new))) return (true); counter_u64_add(pqstate_commit_retries, 1); } while (old->_bits == tmp._bits); return (false); } /* * Do the work of committing a queue state update that moves the page out of * its current queue. */ static bool _vm_page_pqstate_commit_dequeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { vm_page_t next; vm_pagequeue_assert_locked(pq); KASSERT(vm_page_pagequeue(m) == pq, ("%s: queue %p does not match page %p", __func__, pq, m)); KASSERT(old->queue != PQ_NONE && new.queue != old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); /* * Once the queue index of the page changes there is nothing * synchronizing with further updates to the page's physical * queue state. Therefore we must speculatively remove the page * from the queue now and be prepared to roll back if the queue * state update fails. If the page is not physically enqueued then * we just update its queue index. */ if ((old->flags & PGA_ENQUEUED) != 0) { new.flags &= ~PGA_ENQUEUED; next = TAILQ_NEXT(m, plinks.q); TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_pagequeue_cnt_dec(pq); if (!vm_page_pqstate_fcmpset(m, old, new)) { if (next == NULL) TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); else TAILQ_INSERT_BEFORE(next, m, plinks.q); vm_pagequeue_cnt_inc(pq); return (false); } else { return (true); } } else { return (vm_page_pqstate_fcmpset(m, old, new)); } } static bool vm_page_pqstate_commit_dequeue(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_pagequeue *pq; vm_page_astate_t as; bool ret; pq = _vm_page_pagequeue(m, old->queue); /* * The queue field and PGA_ENQUEUED flag are stable only so long as the * corresponding page queue lock is held. */ vm_pagequeue_lock(pq); as = vm_page_astate_load(m); if (__predict_false(as._bits != old->_bits)) { *old = as; ret = false; } else { ret = _vm_page_pqstate_commit_dequeue(pq, m, old, new); } vm_pagequeue_unlock(pq); return (ret); } /* * Commit a queue state update that enqueues or requeues a page. */ static bool _vm_page_pqstate_commit_requeue(struct vm_pagequeue *pq, vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { struct vm_domain *vmd; vm_pagequeue_assert_locked(pq); KASSERT(old->queue != PQ_NONE && new.queue == old->queue, ("%s: invalid queue indices %d %d", __func__, old->queue, new.queue)); new.flags |= PGA_ENQUEUED; if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if ((old->flags & PGA_ENQUEUED) != 0) TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); else vm_pagequeue_cnt_inc(pq); /* * Give PGA_REQUEUE_HEAD precedence over PGA_REQUEUE. In particular, if * both flags are set in close succession, only PGA_REQUEUE_HEAD will be * applied, even if it was set first. */ if ((old->flags & PGA_REQUEUE_HEAD) != 0) { vmd = vm_pagequeue_domain(m); KASSERT(pq == &vmd->vmd_pagequeues[PQ_INACTIVE], ("%s: invalid page queue for page %p", __func__, m)); TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q); } else { TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q); } return (true); } /* * Commit a queue state update that encodes a request for a deferred queue * operation. */ static bool vm_page_pqstate_commit_request(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(old->queue == new.queue || new.queue != PQ_NONE, ("%s: invalid state, queue %d flags %x", __func__, new.queue, new.flags)); if (old->_bits != new._bits && !vm_page_pqstate_fcmpset(m, old, new)) return (false); vm_page_pqbatch_submit(m, new.queue); return (true); } /* * A generic queue state update function. This handles more cases than the * specialized functions above. */ bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { if (old->_bits == new._bits) return (true); if (old->queue != PQ_NONE && new.queue != old->queue) { if (!vm_page_pqstate_commit_dequeue(m, old, new)) return (false); if (new.queue != PQ_NONE) vm_page_pqbatch_submit(m, new.queue); } else { if (!vm_page_pqstate_fcmpset(m, old, new)) return (false); if (new.queue != PQ_NONE && ((new.flags & ~old->flags) & PGA_QUEUE_OP_MASK) != 0) vm_page_pqbatch_submit(m, new.queue); } return (true); } /* * Apply deferred queue state updates to a page. */ static inline void vm_pqbatch_process_page(struct vm_pagequeue *pq, vm_page_t m, uint8_t queue) { vm_page_astate_t new, old; CRITICAL_ASSERT(curthread); vm_pagequeue_assert_locked(pq); KASSERT(queue < PQ_COUNT, ("%s: invalid queue index %d", __func__, queue)); KASSERT(pq == _vm_page_pagequeue(m, queue), ("%s: page %p does not belong to queue %p", __func__, m, pq)); for (old = vm_page_astate_load(m);;) { if (__predict_false(old.queue != queue || (old.flags & PGA_QUEUE_OP_MASK) == 0)) { counter_u64_add(queue_nops, 1); break; } KASSERT(old.queue != PQ_NONE || (old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); new = old; if ((old.flags & PGA_DEQUEUE) != 0) { new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; if (__predict_true(_vm_page_pqstate_commit_dequeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } else { new.flags &= ~(PGA_REQUEUE | PGA_REQUEUE_HEAD); if (__predict_true(_vm_page_pqstate_commit_requeue(pq, m, &old, new))) { counter_u64_add(queue_ops, 1); break; } } } } static void vm_pqbatch_process(struct vm_pagequeue *pq, struct vm_batchqueue *bq, uint8_t queue) { int i; for (i = 0; i < bq->bq_cnt; i++) vm_pqbatch_process_page(pq, bq->bq_pa[i], queue); vm_batchqueue_init(bq); } /* * vm_page_pqbatch_submit: [ internal use only ] * * Enqueue a page in the specified page queue's batched work queue. * The caller must have encoded the requested operation in the page * structure's a.flags field. */ void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue) { struct vm_batchqueue *bq; struct vm_pagequeue *pq; int domain; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("page %p is unmanaged", m)); KASSERT(queue < PQ_COUNT, ("invalid queue %d", queue)); domain = vm_phys_domain(m); pq = &vm_pagequeue_domain(m)->vmd_pagequeues[queue]; critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); if (vm_batchqueue_insert(bq, m)) { critical_exit(); return; } critical_exit(); vm_pagequeue_lock(pq); critical_enter(); bq = DPCPU_PTR(pqbatch[domain][queue]); vm_pqbatch_process(pq, bq, queue); vm_pqbatch_process_page(pq, m, queue); vm_pagequeue_unlock(pq); critical_exit(); } /* * vm_page_pqbatch_drain: [ internal use only ] * * Force all per-CPU page queue batch queues to be drained. This is * intended for use in severe memory shortages, to ensure that pages * do not remain stuck in the batch queues. */ void vm_page_pqbatch_drain(void) { struct thread *td; struct vm_domain *vmd; struct vm_pagequeue *pq; int cpu, domain, queue; td = curthread; CPU_FOREACH(cpu) { thread_lock(td); sched_bind(td, cpu); thread_unlock(td); for (domain = 0; domain < vm_ndomains; domain++) { vmd = VM_DOMAIN(domain); for (queue = 0; queue < PQ_COUNT; queue++) { pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); critical_enter(); vm_pqbatch_process(pq, DPCPU_PTR(pqbatch[domain][queue]), queue); critical_exit(); vm_pagequeue_unlock(pq); } } } thread_lock(td); sched_unbind(td); thread_unlock(td); } /* * vm_page_dequeue_deferred: [ internal use only ] * * Request removal of the given page from its current page * queue. Physical removal from the queue may be deferred * indefinitely. * * The page must be locked. */ void vm_page_dequeue_deferred(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags |= PGA_DEQUEUE; } while (!vm_page_pqstate_commit_request(m, &old, new)); } /* * vm_page_dequeue: * * Remove the page from whichever page queue it's in, if any, before * returning. */ void vm_page_dequeue(vm_page_t m) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue == PQ_NONE) { KASSERT((old.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p has unexpected queue state", __func__, m)); break; } new = old; new.flags &= ~PGA_QUEUE_OP_MASK; new.queue = PQ_NONE; } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); } /* * Schedule the given page for insertion into the specified page queue. * Physical insertion of the page may be deferred indefinitely. */ static void vm_page_enqueue(vm_page_t m, uint8_t queue) { - vm_page_assert_locked(m); KASSERT(m->a.queue == PQ_NONE && (m->a.flags & PGA_QUEUE_STATE_MASK) == 0, ("%s: page %p is already enqueued", __func__, m)); KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); m->a.queue = queue; if ((m->a.flags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, queue); } /* * vm_page_requeue: [ internal use only ] * * Schedule a requeue of the given page. * * The page must be locked. */ void vm_page_requeue(vm_page_t m) { vm_page_assert_locked(m); KASSERT(vm_page_queue(m) != PQ_NONE, ("%s: page %p is not logically enqueued", __func__, m)); KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); if ((m->a.flags & PGA_REQUEUE) == 0) vm_page_aflag_set(m, PGA_REQUEUE); vm_page_pqbatch_submit(m, atomic_load_8(&m->a.queue)); } /* * vm_page_swapqueue: [ internal use only ] * * Move the page from one queue to another, or to the tail of its * current queue, in the face of a possible concurrent free of the * page. */ void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if (old.queue != oldq || (old.flags & PGA_DEQUEUE) != 0) return; new = old; new.flags |= PGA_REQUEUE; new.queue = newq; } while (!vm_page_pqstate_commit_dequeue(m, &old, new)); vm_page_pqbatch_submit(m, newq); } /* * vm_page_free_prep: * * Prepares the given page to be put on the free list, * disassociating it from any VM object. The caller may return * the page to the free list only if this function returns true. * * The object must be locked. The page must be locked if it is * managed. */ static bool vm_page_free_prep(vm_page_t m) { /* * Synchronize with threads that have dropped a reference to this * page. */ atomic_thread_fence_acq(); if (vm_page_sbusied(m)) panic("vm_page_free_prep: freeing shared busy page %p", m); #if defined(DIAGNOSTIC) && defined(PHYS_TO_DMAP) if (PMAP_HAS_DMAP && (m->flags & PG_ZERO) != 0) { uint64_t *p; int i; p = (uint64_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); for (i = 0; i < PAGE_SIZE / sizeof(uint64_t); i++, p++) KASSERT(*p == 0, ("vm_page_free_prep %p PG_ZERO %d %jx", m, i, (uintmax_t)*p)); } #endif if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(!pmap_page_is_mapped(m), ("vm_page_free_prep: freeing mapped page %p", m)); KASSERT((m->a.flags & (PGA_EXECUTABLE | PGA_WRITEABLE)) == 0, ("vm_page_free_prep: mapping flags set in page %p", m)); } else { KASSERT(m->a.queue == PQ_NONE, ("vm_page_free_prep: unmanaged page %p is queued", m)); } VM_CNT_INC(v_tfree); if (m->object != NULL) { KASSERT(((m->oflags & VPO_UNMANAGED) != 0) == ((m->object->flags & OBJ_UNMANAGED) != 0), ("vm_page_free_prep: managed flag mismatch for page %p", m)); vm_page_object_remove(m); /* * The object reference can be released without an atomic * operation. */ KASSERT((m->flags & PG_FICTITIOUS) != 0 || m->ref_count == VPRC_OBJREF, ("vm_page_free_prep: page %p has unexpected ref_count %u", m, m->ref_count)); m->object = NULL; m->ref_count -= VPRC_OBJREF; vm_page_xunbusy(m); } if (vm_page_xbusied(m)) panic("vm_page_free_prep: freeing exclusive busy page %p", m); /* * If fictitious remove object association and * return. */ if ((m->flags & PG_FICTITIOUS) != 0) { KASSERT(m->ref_count == 1, ("fictitious page %p is referenced", m)); KASSERT(m->a.queue == PQ_NONE, ("fictitious page %p is queued", m)); return (false); } /* * Pages need not be dequeued before they are returned to the physical * memory allocator, but they must at least be marked for a deferred * dequeue. */ if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_dequeue_deferred(m); m->valid = 0; vm_page_undirty(m); if (m->ref_count != 0) panic("vm_page_free_prep: page %p has references", m); /* * Restore the default memory attribute to the page. */ if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT) pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT); #if VM_NRESERVLEVEL > 0 /* * Determine whether the page belongs to a reservation. If the page was * allocated from a per-CPU cache, it cannot belong to a reservation, so * as an optimization, we avoid the check in that case. */ if ((m->flags & PG_PCPU_CACHE) == 0 && vm_reserv_free_page(m)) return (false); #endif return (true); } /* * vm_page_free_toq: * * Returns the given page to the free list, disassociating it * from any VM object. * * The object must be locked. The page must be locked if it is * managed. */ static void vm_page_free_toq(vm_page_t m) { struct vm_domain *vmd; uma_zone_t zone; if (!vm_page_free_prep(m)) return; vmd = vm_pagequeue_domain(m); zone = vmd->vmd_pgcache[m->pool].zone; if ((m->flags & PG_PCPU_CACHE) != 0 && zone != NULL) { uma_zfree(zone, m); return; } vm_domain_free_lock(vmd); vm_phys_free_pages(m, 0); vm_domain_free_unlock(vmd); vm_domain_freecnt_inc(vmd, 1); } /* * vm_page_free_pages_toq: * * Returns a list of pages to the free list, disassociating it * from any VM object. In other words, this is equivalent to * calling vm_page_free_toq() for each page of a list of VM objects. * * The objects must be locked. The pages must be locked if it is * managed. */ void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count) { vm_page_t m; int count; if (SLIST_EMPTY(free)) return; count = 0; while ((m = SLIST_FIRST(free)) != NULL) { count++; SLIST_REMOVE_HEAD(free, plinks.s.ss); vm_page_free_toq(m); } if (update_wire_count) vm_wire_sub(count); } /* * Mark this page as wired down, preventing reclamation by the page daemon * or when the containing object is destroyed. */ void vm_page_wire(vm_page_t m) { u_int old; KASSERT(m->object != NULL, ("vm_page_wire: page %p does not belong to an object", m)); if (!vm_page_busied(m) && !vm_object_busied(m->object)) VM_OBJECT_ASSERT_LOCKED(m->object); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(m->ref_count) >= 1, ("vm_page_wire: fictitious page %p has zero wirings", m)); old = atomic_fetchadd_int(&m->ref_count, 1); KASSERT(VPRC_WIRE_COUNT(old) != VPRC_WIRE_COUNT_MAX, ("vm_page_wire: counter overflow for page %p", m)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } } /* * Attempt to wire a mapped page following a pmap lookup of that page. * This may fail if a thread is concurrently tearing down mappings of the page. * The transient failure is acceptable because it translates to the * failure of the caller pmap_extract_and_hold(), which should be then * followed by the vm_fault() fallback, see e.g. vm_fault_quick_hold_pages(). */ bool vm_page_wire_mapped(vm_page_t m) { u_int old; old = m->ref_count; do { KASSERT(old > 0, ("vm_page_wire_mapped: wiring unreferenced page %p", m)); if ((old & VPRC_BLOCKED) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old + 1)); if (VPRC_WIRE_COUNT(old) == 0) { if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); vm_wire_add(1); } return (true); } /* * Release a wiring reference to a managed page. If the page still belongs to * an object, update its position in the page queues to reflect the reference. * If the wiring was the last reference to the page, free the page. */ static void vm_page_unwire_managed(vm_page_t m, uint8_t nqueue, bool noreuse) { u_int old; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is unmanaged", __func__, m)); /* * Update LRU state before releasing the wiring reference. * Use a release store when updating the reference count to * synchronize with vm_page_free_prep(). */ old = m->ref_count; do { KASSERT(VPRC_WIRE_COUNT(old) > 0, ("vm_page_unwire: wire count underflow for page %p", m)); if (old > VPRC_OBJREF + 1) { /* * The page has at least one other wiring reference. An * earlier iteration of this loop may have called * vm_page_release_toq() and cleared PGA_DEQUEUE, so * re-set it if necessary. */ if ((vm_page_astate_load(m).flags & PGA_DEQUEUE) == 0) vm_page_aflag_set(m, PGA_DEQUEUE); } else if (old == VPRC_OBJREF + 1) { /* * This is the last wiring. Clear PGA_DEQUEUE and * update the page's queue state to reflect the * reference. If the page does not belong to an object * (i.e., the VPRC_OBJREF bit is clear), we only need to * clear leftover queue state. */ vm_page_release_toq(m, nqueue, false); } else if (old == 1) { vm_page_aflag_clear(m, PGA_DEQUEUE); } } while (!atomic_fcmpset_rel_int(&m->ref_count, &old, old - 1)); if (VPRC_WIRE_COUNT(old) == 1) { vm_wire_sub(1); if (old == 1) vm_page_free(m); } } /* * Release one wiring of the specified page, potentially allowing it to be * paged out. * * Only managed pages belonging to an object can be paged out. If the number * of wirings transitions to zero and the page is eligible for page out, then * the page is added to the specified paging queue. If the released wiring * represented the last reference to the page, the page is freed. * * A managed page must be locked. */ void vm_page_unwire(vm_page_t m, uint8_t nqueue) { KASSERT(nqueue < PQ_COUNT, ("vm_page_unwire: invalid queue %u request for page %p", nqueue, m)); if ((m->oflags & VPO_UNMANAGED) != 0) { if (vm_page_unwire_noq(m) && m->ref_count == 0) vm_page_free(m); return; } vm_page_unwire_managed(m, nqueue, false); } /* * Unwire a page without (re-)inserting it into a page queue. It is up * to the caller to enqueue, requeue, or free the page as appropriate. * In most cases involving managed pages, vm_page_unwire() should be used * instead. */ bool vm_page_unwire_noq(vm_page_t m) { u_int old; old = vm_page_drop(m, 1); KASSERT(VPRC_WIRE_COUNT(old) != 0, ("vm_page_unref: counter underflow for page %p", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0 || VPRC_WIRE_COUNT(old) > 1, ("vm_page_unref: missing ref on fictitious page %p", m)); if (VPRC_WIRE_COUNT(old) > 1) return (false); if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_clear(m, PGA_DEQUEUE); vm_wire_sub(1); return (true); } /* * Ensure that the page ends up in the specified page queue. If the page is * active or being moved to the active queue, ensure that its act_count is * at least ACT_INIT but do not otherwise mess with it. * * A managed page must be locked. */ static __always_inline void vm_page_mvqueue(vm_page_t m, const uint8_t nqueue, const uint16_t nflag) { vm_page_astate_t old, new; KASSERT(m->ref_count > 0, ("%s: page %p does not carry any references", __func__, m)); KASSERT(nflag == PGA_REQUEUE || nflag == PGA_REQUEUE_HEAD, ("%s: invalid flags %x", __func__, nflag)); if ((m->oflags & VPO_UNMANAGED) != 0) return; old = vm_page_astate_load(m); do { if ((old.flags & PGA_DEQUEUE) != 0) break; new = old; if (nqueue == PQ_ACTIVE) new.act_count = max(old.act_count, ACT_INIT); if (old.queue == nqueue) { if (nqueue != PQ_ACTIVE) new.flags |= nflag; } else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Put the specified page on the active list (if appropriate). */ void vm_page_activate(vm_page_t m) { vm_page_mvqueue(m, PQ_ACTIVE, PGA_REQUEUE); } /* * Move the specified page to the tail of the inactive queue, or requeue * the page if it is already in the inactive queue. */ void vm_page_deactivate(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE); } void vm_page_deactivate_noreuse(vm_page_t m) { vm_page_mvqueue(m, PQ_INACTIVE, PGA_REQUEUE_HEAD); } /* * Put a page in the laundry, or requeue it if it is already there. */ void vm_page_launder(vm_page_t m) { vm_page_mvqueue(m, PQ_LAUNDRY, PGA_REQUEUE); } /* * Put a page in the PQ_UNSWAPPABLE holding queue. */ void vm_page_unswappable(vm_page_t m) { - vm_page_assert_locked(m); KASSERT(!vm_page_wired(m) && (m->oflags & VPO_UNMANAGED) == 0, ("page %p already unswappable", m)); vm_page_dequeue(m); vm_page_enqueue(m, PQ_UNSWAPPABLE); } /* * Release a page back to the page queues in preparation for unwiring. */ static void vm_page_release_toq(vm_page_t m, uint8_t nqueue, const bool noreuse) { vm_page_astate_t old, new; uint16_t nflag; /* * Use a check of the valid bits to determine whether we should * accelerate reclamation of the page. The object lock might not be * held here, in which case the check is racy. At worst we will either * accelerate reclamation of a valid page and violate LRU, or * unnecessarily defer reclamation of an invalid page. * * If we were asked to not cache the page, place it near the head of the * inactive queue so that is reclaimed sooner. */ if (noreuse || m->valid == 0) { nqueue = PQ_INACTIVE; nflag = PGA_REQUEUE_HEAD; } else { nflag = PGA_REQUEUE; } old = vm_page_astate_load(m); do { new = old; /* * If the page is already in the active queue and we are not * trying to accelerate reclamation, simply mark it as * referenced and avoid any queue operations. */ new.flags &= ~PGA_QUEUE_OP_MASK; if (nflag != PGA_REQUEUE_HEAD && old.queue == PQ_ACTIVE) new.flags |= PGA_REFERENCED; else { new.flags |= nflag; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); } /* * Unwire a page and either attempt to free it or re-add it to the page queues. */ void vm_page_release(vm_page_t m, int flags) { vm_object_t object; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release: page %p is unmanaged", m)); if ((flags & VPR_TRYFREE) != 0) { for (;;) { object = (vm_object_t)atomic_load_ptr(&m->object); if (object == NULL) break; /* Depends on type-stability. */ if (vm_page_busied(m) || !VM_OBJECT_TRYWLOCK(object)) break; if (object == m->object) { vm_page_release_locked(m, flags); VM_OBJECT_WUNLOCK(object); return; } VM_OBJECT_WUNLOCK(object); } } vm_page_unwire_managed(m, PQ_INACTIVE, flags != 0); } /* See vm_page_release(). */ void vm_page_release_locked(vm_page_t m, int flags) { VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("vm_page_release_locked: page %p is unmanaged", m)); if (vm_page_unwire_noq(m)) { if ((flags & VPR_TRYFREE) != 0 && (m->object->ref_count == 0 || !pmap_page_is_mapped(m)) && m->dirty == 0 && vm_page_tryxbusy(m)) { vm_page_free(m); } else { - vm_page_lock(m); vm_page_release_toq(m, PQ_INACTIVE, flags != 0); - vm_page_unlock(m); } } } static bool vm_page_try_blocked_op(vm_page_t m, void (*op)(vm_page_t)) { u_int old; KASSERT(m->object != NULL && (m->oflags & VPO_UNMANAGED) == 0, ("vm_page_try_blocked_op: page %p has no object", m)); KASSERT(vm_page_busied(m), ("vm_page_try_blocked_op: page %p is not busy", m)); VM_OBJECT_ASSERT_LOCKED(m->object); old = m->ref_count; do { KASSERT(old != 0, ("vm_page_try_blocked_op: page %p has no references", m)); if (VPRC_WIRE_COUNT(old) != 0) return (false); } while (!atomic_fcmpset_int(&m->ref_count, &old, old | VPRC_BLOCKED)); (op)(m); /* * If the object is read-locked, new wirings may be created via an * object lookup. */ old = vm_page_drop(m, VPRC_BLOCKED); KASSERT(!VM_OBJECT_WOWNED(m->object) || old == (VPRC_BLOCKED | VPRC_OBJREF), ("vm_page_try_blocked_op: unexpected refcount value %u for %p", old, m)); return (true); } /* * Atomically check for wirings and remove all mappings of the page. */ bool vm_page_try_remove_all(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_all)); } /* * Atomically check for wirings and remove all writeable mappings of the page. */ bool vm_page_try_remove_write(vm_page_t m) { return (vm_page_try_blocked_op(m, pmap_remove_write)); } /* * vm_page_advise * * Apply the specified advice to the given page. * * The object and page must be locked. */ void vm_page_advise(vm_page_t m, int advice) { - vm_page_assert_locked(m); VM_OBJECT_ASSERT_WLOCKED(m->object); if (advice == MADV_FREE) /* * Mark the page clean. This will allow the page to be freed * without first paging it out. MADV_FREE pages are often * quickly reused by malloc(3), so we do not do anything that * would result in a page fault on a later access. */ vm_page_undirty(m); else if (advice != MADV_DONTNEED) { if (advice == MADV_WILLNEED) vm_page_activate(m); return; } + if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) + vm_page_dirty(m); + /* * Clear any references to the page. Otherwise, the page daemon will * immediately reactivate the page. */ vm_page_aflag_clear(m, PGA_REFERENCED); - - if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m)) - vm_page_dirty(m); /* * Place clean pages near the head of the inactive queue rather than * the tail, thus defeating the queue's LRU operation and ensuring that * the page will be reused quickly. Dirty pages not already in the * laundry are moved there. */ if (m->dirty == 0) vm_page_deactivate_noreuse(m); else if (!vm_page_in_laundry(m)) vm_page_launder(m); } static inline int vm_page_grab_pflags(int allocflags) { int pflags; KASSERT((allocflags & VM_ALLOC_NOBUSY) == 0 || (allocflags & VM_ALLOC_WIRED) != 0, ("vm_page_grab_pflags: the pages must be busied or wired")); KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_pflags: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY " "mismatch")); pflags = allocflags & ~(VM_ALLOC_NOWAIT | VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL | VM_ALLOC_NOBUSY); if ((allocflags & VM_ALLOC_NOWAIT) == 0) pflags |= VM_ALLOC_WAITFAIL; if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) pflags |= VM_ALLOC_SBUSY; return (pflags); } /* * Grab a page, waiting until we are waken up due to the page * changing state. We keep on waiting, if the page continues * to be in the object. If the page doesn't exist, first allocate it * and then conditionally zero it. * * This routine may sleep. * * The object must be locked on entry. The lock will, however, be released * and reacquired if the routine sleeps. */ vm_page_t vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; int pflags; VM_OBJECT_ASSERT_WLOCKED(object); pflags = vm_page_grab_pflags(allocflags); retrylookup: if ((m = vm_page_lookup(object, pindex)) != NULL) { if (!vm_page_acquire_flags(m, allocflags)) { if (vm_page_busy_sleep_flags(object, m, "pgrbwt", allocflags)) goto retrylookup; return (NULL); } goto out; } if ((allocflags & VM_ALLOC_NOCREAT) != 0) return (NULL); m = vm_page_alloc(object, pindex, pflags); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) return (NULL); goto retrylookup; } if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0) pmap_zero_page(m); out: if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } return (m); } /* * Grab a page and make it valid, paging in if necessary. Pages missing from * their pager are zero filled and validated. If a VM_ALLOC_COUNT is supplied * and the page is not valid as many as VM_INITIAL_PAGEIN pages can be brought * in simultaneously. Additional pages will be left on a paging queue but * will neither be wired nor busy regardless of allocflags. */ int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags) { vm_page_t m; vm_page_t ma[VM_INITIAL_PAGEIN]; bool sleep, xbusy; int after, i, pflags, rv; KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 || (allocflags & VM_ALLOC_IGN_SBUSY) != 0, ("vm_page_grab_valid: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch")); KASSERT((allocflags & (VM_ALLOC_NOWAIT | VM_ALLOC_WAITFAIL | VM_ALLOC_ZERO)) == 0, ("vm_page_grab_valid: Invalid flags 0x%X", allocflags)); VM_OBJECT_ASSERT_WLOCKED(object); pflags = allocflags & ~(VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY); pflags |= VM_ALLOC_WAITFAIL; retrylookup: xbusy = false; if ((m = vm_page_lookup(object, pindex)) != NULL) { /* * If the page is fully valid it can only become invalid * with the object lock held. If it is not valid it can * become valid with the busy lock held. Therefore, we * may unnecessarily lock the exclusive busy here if we * race with I/O completion not using the object lock. * However, we will not end up with an invalid page and a * shared lock. */ if (!vm_page_all_valid(m) || (allocflags & (VM_ALLOC_IGN_SBUSY | VM_ALLOC_SBUSY)) == 0) { sleep = !vm_page_tryxbusy(m); xbusy = true; } else sleep = !vm_page_trysbusy(m); if (sleep) { (void)vm_page_busy_sleep_flags(object, m, "pgrbwt", allocflags); goto retrylookup; } if ((allocflags & VM_ALLOC_NOCREAT) != 0 && !vm_page_all_valid(m)) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); *mp = NULL; return (VM_PAGER_FAIL); } if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_wire(m); if (vm_page_all_valid(m)) goto out; } else if ((allocflags & VM_ALLOC_NOCREAT) != 0) { *mp = NULL; return (VM_PAGER_FAIL); } else if ((m = vm_page_alloc(object, pindex, pflags)) != NULL) { xbusy = true; } else { goto retrylookup; } vm_page_assert_xbusied(m); MPASS(xbusy); if (vm_pager_has_page(object, pindex, NULL, &after)) { after = MIN(after, VM_INITIAL_PAGEIN); after = MIN(after, allocflags >> VM_ALLOC_COUNT_SHIFT); after = MAX(after, 1); ma[0] = m; for (i = 1; i < after; i++) { if ((ma[i] = vm_page_next(ma[i - 1])) != NULL) { if (ma[i]->valid || !vm_page_tryxbusy(ma[i])) break; } else { ma[i] = vm_page_alloc(object, m->pindex + i, VM_ALLOC_NORMAL); if (ma[i] == NULL) break; } } after = i; rv = vm_pager_get_pages(object, ma, after, NULL, NULL); /* Pager may have replaced a page. */ m = ma[0]; if (rv != VM_PAGER_OK) { if ((allocflags & VM_ALLOC_WIRED) != 0) vm_page_unwire_noq(m); for (i = 0; i < after; i++) { if (!vm_page_wired(ma[i])) vm_page_free(ma[i]); else vm_page_xunbusy(ma[i]); } *mp = NULL; return (rv); } for (i = 1; i < after; i++) vm_page_readahead_finish(ma[i]); MPASS(vm_page_all_valid(m)); } else { vm_page_zero_invalid(m, TRUE); } out: if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if (xbusy) vm_page_xunbusy(m); else vm_page_sunbusy(m); } if ((allocflags & VM_ALLOC_SBUSY) != 0 && xbusy) vm_page_busy_downgrade(m); *mp = m; return (VM_PAGER_OK); } /* * Return the specified range of pages from the given object. For each * page offset within the range, if a page already exists within the object * at that offset and it is busy, then wait for it to change state. If, * instead, the page doesn't exist, then allocate it. * * The caller must always specify an allocation class. * * allocation classes: * VM_ALLOC_NORMAL normal process request * VM_ALLOC_SYSTEM system *really* needs the pages * * The caller must always specify that the pages are to be busied and/or * wired. * * optional allocation flags: * VM_ALLOC_IGN_SBUSY do not sleep on soft busy pages * VM_ALLOC_NOBUSY do not exclusive busy the page * VM_ALLOC_NOWAIT do not sleep * VM_ALLOC_SBUSY set page to sbusy state * VM_ALLOC_WIRED wire the pages * VM_ALLOC_ZERO zero and validate any invalid pages * * If VM_ALLOC_NOWAIT is not specified, this routine may sleep. Otherwise, it * may return a partial prefix of the requested range. */ int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count) { vm_page_t m, mpred; int pflags; int i; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(((u_int)allocflags >> VM_ALLOC_COUNT_SHIFT) == 0, ("vm_page_grap_pages: VM_ALLOC_COUNT() is not allowed")); pflags = vm_page_grab_pflags(allocflags); if (count == 0) return (0); i = 0; retrylookup: m = vm_radix_lookup_le(&object->rtree, pindex + i); if (m == NULL || m->pindex != pindex + i) { mpred = m; m = NULL; } else mpred = TAILQ_PREV(m, pglist, listq); for (; i < count; i++) { if (m != NULL) { if (!vm_page_acquire_flags(m, allocflags)) { if (vm_page_busy_sleep_flags(object, m, "grbmaw", allocflags)) goto retrylookup; break; } } else { if ((allocflags & VM_ALLOC_NOCREAT) != 0) break; m = vm_page_alloc_after(object, pindex + i, pflags | VM_ALLOC_COUNT(count - i), mpred); if (m == NULL) { if ((allocflags & VM_ALLOC_NOWAIT) != 0) break; goto retrylookup; } } if (vm_page_none_valid(m) && (allocflags & VM_ALLOC_ZERO) != 0) { if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); vm_page_valid(m); } if ((allocflags & VM_ALLOC_NOBUSY) != 0) { if ((allocflags & VM_ALLOC_IGN_SBUSY) != 0) vm_page_sunbusy(m); else vm_page_xunbusy(m); } ma[i] = mpred = m; m = vm_page_next(m); } return (i); } /* * Mapping function for valid or dirty bits in a page. * * Inputs are required to range within a page. */ vm_page_bits_t vm_page_bits(int base, int size) { int first_bit; int last_bit; KASSERT( base + size <= PAGE_SIZE, ("vm_page_bits: illegal base/size %d/%d", base, size) ); if (size == 0) /* handle degenerate case */ return (0); first_bit = base >> DEV_BSHIFT; last_bit = (base + size - 1) >> DEV_BSHIFT; return (((vm_page_bits_t)2 << last_bit) - ((vm_page_bits_t)1 << first_bit)); } void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set) { #if PAGE_SIZE == 32768 atomic_set_64((uint64_t *)bits, set); #elif PAGE_SIZE == 16384 atomic_set_32((uint32_t *)bits, set); #elif (PAGE_SIZE == 8192) && defined(atomic_set_16) atomic_set_16((uint16_t *)bits, set); #elif (PAGE_SIZE == 4096) && defined(atomic_set_8) atomic_set_8((uint8_t *)bits, set); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_set_32((uint32_t *)addr, set << shift); #endif /* PAGE_SIZE */ } static inline void vm_page_bits_clear(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t clear) { #if PAGE_SIZE == 32768 atomic_clear_64((uint64_t *)bits, clear); #elif PAGE_SIZE == 16384 atomic_clear_32((uint32_t *)bits, clear); #elif (PAGE_SIZE == 8192) && defined(atomic_clear_16) atomic_clear_16((uint16_t *)bits, clear); #elif (PAGE_SIZE == 4096) && defined(atomic_clear_8) atomic_clear_8((uint8_t *)bits, clear); #else /* PAGE_SIZE <= 8192 */ uintptr_t addr; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); atomic_clear_32((uint32_t *)addr, clear << shift); #endif /* PAGE_SIZE */ } static inline vm_page_bits_t vm_page_bits_swap(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t newbits) { #if PAGE_SIZE == 32768 uint64_t old; old = *bits; while (atomic_fcmpset_64(bits, &old, newbits) == 0); return (old); #elif PAGE_SIZE == 16384 uint32_t old; old = *bits; while (atomic_fcmpset_32(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 8192) && defined(atomic_fcmpset_16) uint16_t old; old = *bits; while (atomic_fcmpset_16(bits, &old, newbits) == 0); return (old); #elif (PAGE_SIZE == 4096) && defined(atomic_fcmpset_8) uint8_t old; old = *bits; while (atomic_fcmpset_8(bits, &old, newbits) == 0); return (old); #else /* PAGE_SIZE <= 4096*/ uintptr_t addr; uint32_t old, new, mask; int shift; addr = (uintptr_t)bits; /* * Use a trick to perform a 32-bit atomic on the * containing aligned word, to not depend on the existence * of atomic_{set, swap, clear}_{8, 16}. */ shift = addr & (sizeof(uint32_t) - 1); #if BYTE_ORDER == BIG_ENDIAN shift = (sizeof(uint32_t) - sizeof(vm_page_bits_t) - shift) * NBBY; #else shift *= NBBY; #endif addr &= ~(sizeof(uint32_t) - 1); mask = VM_PAGE_BITS_ALL << shift; old = *bits; do { new = old & ~mask; new |= newbits << shift; } while (atomic_fcmpset_32((uint32_t *)addr, &old, new) == 0); return (old >> shift); #endif /* PAGE_SIZE */ } /* * vm_page_set_valid_range: * * Sets portions of a page valid. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zeroed. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_valid_range(vm_page_t m, int base, int size) { int endoff, frag; vm_page_bits_t pagebits; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & (1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Assert that no previously invalid block that is now being validated * is already dirty. */ KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0, ("vm_page_set_valid_range: page %p is dirty", m)); /* * Set valid bits inclusive of any overlap. */ pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); } /* * Set the page dirty bits and free the invalid swap space if * present. Returns the previous dirty bits. */ vm_page_bits_t vm_page_set_dirty(vm_page_t m) { vm_page_bits_t old; VM_PAGE_OBJECT_BUSY_ASSERT(m); if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) { old = m->dirty; m->dirty = VM_PAGE_BITS_ALL; } else old = vm_page_bits_swap(m, &m->dirty, VM_PAGE_BITS_ALL); if (old == 0 && (m->a.flags & PGA_SWAP_SPACE) != 0) vm_pager_page_unswapped(m); return (old); } /* * Clear the given bits from the specified page's dirty field. */ static __inline void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits) { vm_page_assert_busied(m); /* * If the page is xbusied and not write mapped we are the * only thread that can modify dirty bits. Otherwise, The pmap * layer can call vm_page_dirty() without holding a distinguished * lock. The combination of page busy and atomic operations * suffice to guarantee consistency of the page dirty field. */ if (vm_page_xbusied(m) && !pmap_page_is_write_mapped(m)) m->dirty &= ~pagebits; else vm_page_bits_clear(m, &m->dirty, pagebits); } /* * vm_page_set_validclean: * * Sets portions of a page valid and clean. The arguments are expected * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive * of any partial chunks touched by the range. The invalid portion of * such chunks will be zero'd. * * (base + size) must be less then or equal to PAGE_SIZE. */ void vm_page_set_validclean(vm_page_t m, int base, int size) { vm_page_bits_t oldvalid, pagebits; int endoff, frag; vm_page_assert_busied(m); if (size == 0) /* handle degenerate case */ return; /* * If the base is not DEV_BSIZE aligned and the valid * bit is clear, we have to zero out a portion of the * first block. */ if ((frag = rounddown2(base, DEV_BSIZE)) != base && (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, frag, base - frag); /* * If the ending offset is not DEV_BSIZE aligned and the * valid bit is clear, we have to zero out a portion of * the last block. */ endoff = base + size; if ((frag = rounddown2(endoff, DEV_BSIZE)) != endoff && (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0) pmap_zero_page_area(m, endoff, DEV_BSIZE - (endoff & (DEV_BSIZE - 1))); /* * Set valid, clear dirty bits. If validating the entire * page we can safely clear the pmap modify bit. We also * use this opportunity to clear the PGA_NOSYNC flag. If a process * takes a write fault on a MAP_NOSYNC memory area the flag will * be set again. * * We set valid bits inclusive of any overlap, but we can only * clear dirty bits for DEV_BSIZE chunks that are fully within * the range. */ oldvalid = m->valid; pagebits = vm_page_bits(base, size); if (vm_page_xbusied(m)) m->valid |= pagebits; else vm_page_bits_set(m, &m->valid, pagebits); #if 0 /* NOT YET */ if ((frag = base & (DEV_BSIZE - 1)) != 0) { frag = DEV_BSIZE - frag; base += frag; size -= frag; if (size < 0) size = 0; } pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); #endif if (base == 0 && size == PAGE_SIZE) { /* * The page can only be modified within the pmap if it is * mapped, and it can only be mapped if it was previously * fully valid. */ if (oldvalid == VM_PAGE_BITS_ALL) /* * Perform the pmap_clear_modify() first. Otherwise, * a concurrent pmap operation, such as * pmap_protect(), could clear a modification in the * pmap and set the dirty field on the page before * pmap_clear_modify() had begun and after the dirty * field was cleared here. */ pmap_clear_modify(m); m->dirty = 0; vm_page_aflag_clear(m, PGA_NOSYNC); } else if (oldvalid != VM_PAGE_BITS_ALL && vm_page_xbusied(m)) m->dirty &= ~pagebits; else vm_page_clear_dirty_mask(m, pagebits); } void vm_page_clear_dirty(vm_page_t m, int base, int size) { vm_page_clear_dirty_mask(m, vm_page_bits(base, size)); } /* * vm_page_set_invalid: * * Invalidates DEV_BSIZE'd chunks within a page. Both the * valid and dirty bits for the effected areas are cleared. */ void vm_page_set_invalid(vm_page_t m, int base, int size) { vm_page_bits_t bits; vm_object_t object; /* * The object lock is required so that pages can't be mapped * read-only while we're in the process of invalidating them. */ object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); vm_page_assert_busied(m); if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) + size >= object->un_pager.vnp.vnp_size) bits = VM_PAGE_BITS_ALL; else bits = vm_page_bits(base, size); if (object->ref_count != 0 && vm_page_all_valid(m) && bits != 0) pmap_remove_all(m); KASSERT((bits == 0 && vm_page_all_valid(m)) || !pmap_page_is_mapped(m), ("vm_page_set_invalid: page %p is mapped", m)); if (vm_page_xbusied(m)) { m->valid &= ~bits; m->dirty &= ~bits; } else { vm_page_bits_clear(m, &m->valid, bits); vm_page_bits_clear(m, &m->dirty, bits); } } /* * vm_page_invalid: * * Invalidates the entire page. The page must be busy, unmapped, and * the enclosing object must be locked. The object locks protects * against concurrent read-only pmap enter which is done without * busy. */ void vm_page_invalid(vm_page_t m) { vm_page_assert_busied(m); VM_OBJECT_ASSERT_LOCKED(m->object); MPASS(!pmap_page_is_mapped(m)); if (vm_page_xbusied(m)) m->valid = 0; else vm_page_bits_clear(m, &m->valid, VM_PAGE_BITS_ALL); } /* * vm_page_zero_invalid() * * The kernel assumes that the invalid portions of a page contain * garbage, but such pages can be mapped into memory by user code. * When this occurs, we must zero out the non-valid portions of the * page so user code sees what it expects. * * Pages are most often semi-valid when the end of a file is mapped * into memory and the file's size is not page aligned. */ void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) { int b; int i; /* * Scan the valid bits looking for invalid sections that * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the * valid bit may be set ) have already been zeroed by * vm_page_set_validclean(). */ for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { if (i == (PAGE_SIZE / DEV_BSIZE) || (m->valid & ((vm_page_bits_t)1 << i))) { if (i > b) { pmap_zero_page_area(m, b << DEV_BSHIFT, (i - b) << DEV_BSHIFT); } b = i + 1; } } /* * setvalid is TRUE when we can safely set the zero'd areas * as being valid. We can do this if there are no cache consistancy * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. */ if (setvalid) vm_page_valid(m); } /* * vm_page_is_valid: * * Is (partial) page valid? Note that the case where size == 0 * will return FALSE in the degenerate case where the page is * entirely invalid, and TRUE otherwise. * * Some callers envoke this routine without the busy lock held and * handle races via higher level locks. Typical callers should * hold a busy lock to prevent invalidation. */ int vm_page_is_valid(vm_page_t m, int base, int size) { vm_page_bits_t bits; bits = vm_page_bits(base, size); return (m->valid != 0 && (m->valid & bits) == bits); } /* * Returns true if all of the specified predicates are true for the entire * (super)page and false otherwise. */ bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m) { vm_object_t object; int i, npages; object = m->object; if (skip_m != NULL && skip_m->object != object) return (false); VM_OBJECT_ASSERT_LOCKED(object); npages = atop(pagesizes[m->psind]); /* * The physically contiguous pages that make up a superpage, i.e., a * page with a page size index ("psind") greater than zero, will * occupy adjacent entries in vm_page_array[]. */ for (i = 0; i < npages; i++) { /* Always test object consistency, including "skip_m". */ if (m[i].object != object) return (false); if (&m[i] == skip_m) continue; if ((flags & PS_NONE_BUSY) != 0 && vm_page_busied(&m[i])) return (false); if ((flags & PS_ALL_DIRTY) != 0) { /* * Calling vm_page_test_dirty() or pmap_is_modified() * might stop this case from spuriously returning * "false". However, that would require a write lock * on the object containing "m[i]". */ if (m[i].dirty != VM_PAGE_BITS_ALL) return (false); } if ((flags & PS_ALL_VALID) != 0 && m[i].valid != VM_PAGE_BITS_ALL) return (false); } return (true); } /* * Set the page's dirty bits if the page is modified. */ void vm_page_test_dirty(vm_page_t m) { vm_page_assert_busied(m); if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m)) vm_page_dirty(m); } void vm_page_valid(vm_page_t m) { vm_page_assert_busied(m); if (vm_page_xbusied(m)) m->valid = VM_PAGE_BITS_ALL; else vm_page_bits_set(m, &m->valid, VM_PAGE_BITS_ALL); } void vm_page_lock_KBI(vm_page_t m, const char *file, int line) { mtx_lock_flags_(vm_page_lockptr(m), 0, file, line); } void vm_page_unlock_KBI(vm_page_t m, const char *file, int line) { mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line); } int vm_page_trylock_KBI(vm_page_t m, const char *file, int line) { return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line)); } #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line) { vm_page_lock_assert_KBI(m, MA_OWNED, file, line); } void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line) { mtx_assert_(vm_page_lockptr(m), a, file, line); } #endif #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m) { /* * Certain of the page's fields may only be modified by the * holder of a page or object busy. */ if (m->object != NULL && !vm_page_busied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits) { if ((bits & PGA_WRITEABLE) == 0) return; /* * The PGA_WRITEABLE flag can only be set if the page is * managed, is exclusively busied or the object is locked. * Currently, this flag is only set by pmap_enter(). */ KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("PGA_WRITEABLE on unmanaged page")); if (!vm_page_xbusied(m)) VM_OBJECT_ASSERT_BUSY(m->object); } #endif #include "opt_ddb.h" #ifdef DDB #include #include DB_SHOW_COMMAND(page, vm_page_print_page_info) { db_printf("vm_cnt.v_free_count: %d\n", vm_free_count()); db_printf("vm_cnt.v_inactive_count: %d\n", vm_inactive_count()); db_printf("vm_cnt.v_active_count: %d\n", vm_active_count()); db_printf("vm_cnt.v_laundry_count: %d\n", vm_laundry_count()); db_printf("vm_cnt.v_wire_count: %d\n", vm_wire_count()); db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved); db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min); db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target); db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target); } DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) { int dom; db_printf("pq_free %d\n", vm_free_count()); for (dom = 0; dom < vm_ndomains; dom++) { db_printf( "dom %d page_cnt %d free %d pq_act %d pq_inact %d pq_laund %d pq_unsw %d\n", dom, vm_dom[dom].vmd_page_count, vm_dom[dom].vmd_free_count, vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_LAUNDRY].pq_cnt, vm_dom[dom].vmd_pagequeues[PQ_UNSWAPPABLE].pq_cnt); } } DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo) { vm_page_t m; boolean_t phys, virt; if (!have_addr) { db_printf("show pginfo addr\n"); return; } phys = strchr(modif, 'p') != NULL; virt = strchr(modif, 'v') != NULL; if (virt) m = PHYS_TO_VM_PAGE(pmap_kextract(addr)); else if (phys) m = PHYS_TO_VM_PAGE(addr); else m = (vm_page_t)addr; db_printf( "page %p obj %p pidx 0x%jx phys 0x%jx q %d ref %u\n" " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n", m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr, m->a.queue, m->ref_count, m->a.flags, m->oflags, m->flags, m->a.act_count, m->busy_lock, m->valid, m->dirty); } #endif /* DDB */ Index: head/sys/vm/vm_page.h =================================================================== --- head/sys/vm/vm_page.h (revision 356156) +++ head/sys/vm/vm_page.h (revision 356157) @@ -1,1003 +1,995 @@ /*- * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) * * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_page.h 8.2 (Berkeley) 12/13/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. * * $FreeBSD$ */ /* * Resident memory system definitions. */ #ifndef _VM_PAGE_ #define _VM_PAGE_ #include /* * Management of resident (logical) pages. * * A small structure is kept for each resident * page, indexed by page number. Each structure * is an element of several collections: * * A radix tree used to quickly * perform object/offset lookups * * A list of all pages for a given object, * so they can be quickly deactivated at * time of deallocation. * * An ordered list of pages due for pageout. * * In addition, the structure contains the object * and offset to which this page belongs (for pageout), * and sundry status bits. * * In general, operations on this structure's mutable fields are * synchronized using either one of or a combination of locks. If a * field is annotated with two of these locks then holding either is * sufficient for read access but both are required for write access. * The physical address of a page is used to select its page lock from * a pool. The queue lock for a page depends on the value of its queue * field and is described in detail below. * * The following annotations are possible: * (A) the field is atomic and may require additional synchronization. * (B) the page busy lock. * (C) the field is immutable. * (F) the per-domain lock for the free queues * (M) Machine dependent, defined by pmap layer. * (O) the object that the page belongs to. * (P) the page lock. * (Q) the page's queue lock. * * The busy lock is an embedded reader-writer lock that protects the * page's contents and identity (i.e., its tuple) as * well as certain valid/dirty modifications. To avoid bloating the * the page structure, the busy lock lacks some of the features available * the kernel's general-purpose synchronization primitives. As a result, * busy lock ordering rules are not verified, lock recursion is not * detected, and an attempt to xbusy a busy page or sbusy an xbusy page * results will trigger a panic rather than causing the thread to block. * vm_page_sleep_if_busy() can be used to sleep until the page's busy * state changes, after which the caller must re-lookup the page and * re-evaluate its state. vm_page_busy_acquire() will block until * the lock is acquired. * * The valid field is protected by the page busy lock (B) and object * lock (O). Transitions from invalid to valid are generally done * via I/O or zero filling and do not require the object lock. * These must be protected with the busy lock to prevent page-in or * creation races. Page invalidation generally happens as a result * of truncate or msync. When invalidated, pages must not be present * in pmap and must hold the object lock to prevent concurrent * speculative read-only mappings that do not require busy. I/O * routines may check for validity without a lock if they are prepared * to handle invalidation races with higher level locks (vnode) or are * unconcerned with races so long as they hold a reference to prevent * recycling. When a valid bit is set while holding a shared busy * lock (A) atomic operations are used to protect against concurrent * modification. * * In contrast, the synchronization of accesses to the page's * dirty field is a mix of machine dependent (M) and busy (B). In * the machine-independent layer, the page busy must be held to * operate on the field. However, the pmap layer is permitted to * set all bits within the field without holding that lock. If the * underlying architecture does not support atomic read-modify-write * operations on the field's type, then the machine-independent * layer uses a 32-bit atomic on the aligned 32-bit word that * contains the dirty field. In the machine-independent layer, * the implementation of read-modify-write operations on the * field is encapsulated in vm_page_clear_dirty_mask(). An * exclusive busy lock combined with pmap_remove_{write/all}() is the * only way to ensure a page can not become dirty. I/O generally * removes the page from pmap to ensure exclusive access and atomic * writes. * * The ref_count field tracks references to the page. References that * prevent the page from being reclaimable are called wirings and are * counted in the low bits of ref_count. The containing object's * reference, if one exists, is counted using the VPRC_OBJREF bit in the * ref_count field. Additionally, the VPRC_BLOCKED bit is used to * atomically check for wirings and prevent new wirings via * pmap_extract_and_hold(). When a page belongs to an object, it may be * wired only when the object is locked, or the page is busy, or by * pmap_extract_and_hold(). As a result, if the object is locked and the * page is not busy (or is exclusively busied by the current thread), and * the page is unmapped, its wire count will not increase. The ref_count * field is updated using atomic operations in most cases, except when it * is known that no other references to the page exist, such as in the page * allocator. A page may be present in the page queues, or even actively * scanned by the page daemon, without an explicitly counted referenced. * The page daemon must therefore handle the possibility of a concurrent * free of the page. * * The queue field is the index of the page queue containing the page, * or PQ_NONE if the page is not enqueued. The queue lock of a page is * the page queue lock corresponding to the page queue index, or the * page lock (P) for the page if it is not enqueued. To modify the * queue field, the queue lock for the old value of the field must be * held. There is one exception to this rule: the page daemon may * transition the queue field from PQ_INACTIVE to PQ_NONE immediately * prior to freeing a page during an inactive queue scan. At that * point the page has already been physically dequeued and no other * references to that vm_page structure exist. * * To avoid contention on page queue locks, page queue operations * (enqueue, dequeue, requeue) are batched using per-CPU queues. A * deferred operation is requested by inserting an entry into a batch * queue; the entry is simply a pointer to the page, and the request * type is encoded in the page's aflags field using the values in * PGA_QUEUE_STATE_MASK. The type-stability of struct vm_pages is * crucial to this scheme since the processing of entries in a given * batch queue may be deferred indefinitely. In particular, a page may * be freed before its pending batch queue entries have been processed. * The page lock (P) must be held to schedule a batched queue * operation, and the page queue lock must be held in order to process * batch queue entries for the page queue. There is one exception to * this rule: the thread freeing a page may schedule a dequeue without * holding the page lock. In this scenario the only other thread which * may hold a reference to the page is the page daemon, which is * careful to avoid modifying the page's queue state once the dequeue * has been requested by setting PGA_DEQUEUE. */ #if PAGE_SIZE == 4096 #define VM_PAGE_BITS_ALL 0xffu typedef uint8_t vm_page_bits_t; #elif PAGE_SIZE == 8192 #define VM_PAGE_BITS_ALL 0xffffu typedef uint16_t vm_page_bits_t; #elif PAGE_SIZE == 16384 #define VM_PAGE_BITS_ALL 0xffffffffu typedef uint32_t vm_page_bits_t; #elif PAGE_SIZE == 32768 #define VM_PAGE_BITS_ALL 0xfffffffffffffffflu typedef uint64_t vm_page_bits_t; #endif typedef union vm_page_astate { struct { uint16_t flags; uint8_t queue; uint8_t act_count; }; uint32_t _bits; } vm_page_astate_t; struct vm_page { union { TAILQ_ENTRY(vm_page) q; /* page queue or free list (Q) */ struct { SLIST_ENTRY(vm_page) ss; /* private slists */ } s; struct { u_long p; u_long v; } memguard; struct { void *slab; void *zone; } uma; } plinks; TAILQ_ENTRY(vm_page) listq; /* pages in same object (O) */ vm_object_t object; /* which object am I in (O) */ vm_pindex_t pindex; /* offset into object (O,P) */ vm_paddr_t phys_addr; /* physical address of page (C) */ struct md_page md; /* machine dependent stuff */ u_int ref_count; /* page references (A) */ volatile u_int busy_lock; /* busy owners lock */ union vm_page_astate a; /* state accessed atomically */ uint8_t order; /* index of the buddy queue (F) */ uint8_t pool; /* vm_phys freepool index (F) */ uint8_t flags; /* page PG_* flags (P) */ uint8_t oflags; /* page VPO_* flags (O) */ int8_t psind; /* pagesizes[] index (O) */ int8_t segind; /* vm_phys segment index (C) */ /* NOTE that these must support one bit per DEV_BSIZE in a page */ /* so, on normal X86 kernels, they must be at least 8 bits wide */ vm_page_bits_t valid; /* valid DEV_BSIZE chunk map (O,B) */ vm_page_bits_t dirty; /* dirty DEV_BSIZE chunk map (M,B) */ }; /* * Special bits used in the ref_count field. * * ref_count is normally used to count wirings that prevent the page from being * reclaimed, but also supports several special types of references that do not * prevent reclamation. Accesses to the ref_count field must be atomic unless * the page is unallocated. * * VPRC_OBJREF is the reference held by the containing object. It can set or * cleared only when the corresponding object's write lock is held. * * VPRC_BLOCKED is used to atomically block wirings via pmap lookups while * attempting to tear down all mappings of a given page. The page lock and * object write lock must both be held in order to set or clear this bit. */ #define VPRC_BLOCKED 0x40000000u /* mappings are being removed */ #define VPRC_OBJREF 0x80000000u /* object reference, cleared with (O) */ #define VPRC_WIRE_COUNT(c) ((c) & ~(VPRC_BLOCKED | VPRC_OBJREF)) #define VPRC_WIRE_COUNT_MAX (~(VPRC_BLOCKED | VPRC_OBJREF)) /* * Page flags stored in oflags: * * Access to these page flags is synchronized by the lock on the object * containing the page (O). * * Note: VPO_UNMANAGED (used by OBJT_DEVICE, OBJT_PHYS and OBJT_SG) * indicates that the page is not under PV management but * otherwise should be treated as a normal page. Pages not * under PV management cannot be paged out via the * object/vm_page_t because there is no knowledge of their pte * mappings, and such pages are also not on any PQ queue. * */ #define VPO_KMEM_EXEC 0x01 /* kmem mapping allows execution */ #define VPO_SWAPSLEEP 0x02 /* waiting for swap to finish */ #define VPO_UNMANAGED 0x04 /* no PV management for page */ #define VPO_SWAPINPROG 0x08 /* swap I/O in progress on page */ /* * Busy page implementation details. * The algorithm is taken mostly by rwlock(9) and sx(9) locks implementation, * even if the support for owner identity is removed because of size * constraints. Checks on lock recursion are then not possible, while the * lock assertions effectiveness is someway reduced. */ #define VPB_BIT_SHARED 0x01 #define VPB_BIT_EXCLUSIVE 0x02 #define VPB_BIT_WAITERS 0x04 #define VPB_BIT_FLAGMASK \ (VPB_BIT_SHARED | VPB_BIT_EXCLUSIVE | VPB_BIT_WAITERS) #define VPB_SHARERS_SHIFT 3 #define VPB_SHARERS(x) \ (((x) & ~VPB_BIT_FLAGMASK) >> VPB_SHARERS_SHIFT) #define VPB_SHARERS_WORD(x) ((x) << VPB_SHARERS_SHIFT | VPB_BIT_SHARED) #define VPB_ONE_SHARER (1 << VPB_SHARERS_SHIFT) #define VPB_SINGLE_EXCLUSIVE VPB_BIT_EXCLUSIVE #ifdef INVARIANTS #define VPB_CURTHREAD_EXCLUSIVE \ (VPB_BIT_EXCLUSIVE | ((u_int)(uintptr_t)curthread & ~VPB_BIT_FLAGMASK)) #else #define VPB_CURTHREAD_EXCLUSIVE VPB_SINGLE_EXCLUSIVE #endif #define VPB_UNBUSIED VPB_SHARERS_WORD(0) #define PQ_NONE 255 #define PQ_INACTIVE 0 #define PQ_ACTIVE 1 #define PQ_LAUNDRY 2 #define PQ_UNSWAPPABLE 3 #define PQ_COUNT 4 #ifndef VM_PAGE_HAVE_PGLIST TAILQ_HEAD(pglist, vm_page); #define VM_PAGE_HAVE_PGLIST #endif SLIST_HEAD(spglist, vm_page); #ifdef _KERNEL extern vm_page_t bogus_page; #endif /* _KERNEL */ extern struct mtx_padalign pa_lock[]; #if defined(__arm__) #define PDRSHIFT PDR_SHIFT #elif !defined(PDRSHIFT) #define PDRSHIFT 21 #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define PA_LOCKPTR(pa) ((struct mtx *)(&pa_lock[pa_index(pa) % PA_LOCK_COUNT])) #define PA_LOCKOBJPTR(pa) ((struct lock_object *)PA_LOCKPTR((pa))) #define PA_LOCK(pa) mtx_lock(PA_LOCKPTR(pa)) #define PA_TRYLOCK(pa) mtx_trylock(PA_LOCKPTR(pa)) #define PA_UNLOCK(pa) mtx_unlock(PA_LOCKPTR(pa)) #define PA_UNLOCK_COND(pa) \ do { \ if ((pa) != 0) { \ PA_UNLOCK((pa)); \ (pa) = 0; \ } \ } while (0) #define PA_LOCK_ASSERT(pa, a) mtx_assert(PA_LOCKPTR(pa), (a)) #if defined(KLD_MODULE) && !defined(KLD_TIED) #define vm_page_lock(m) vm_page_lock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_unlock(m) vm_page_unlock_KBI((m), LOCK_FILE, LOCK_LINE) #define vm_page_trylock(m) vm_page_trylock_KBI((m), LOCK_FILE, LOCK_LINE) #else /* !KLD_MODULE */ #define vm_page_lockptr(m) (PA_LOCKPTR(VM_PAGE_TO_PHYS((m)))) #define vm_page_lock(m) mtx_lock(vm_page_lockptr((m))) #define vm_page_unlock(m) mtx_unlock(vm_page_lockptr((m))) #define vm_page_trylock(m) mtx_trylock(vm_page_lockptr((m))) #endif #if defined(INVARIANTS) #define vm_page_assert_locked(m) \ vm_page_assert_locked_KBI((m), __FILE__, __LINE__) #define vm_page_lock_assert(m, a) \ vm_page_lock_assert_KBI((m), (a), __FILE__, __LINE__) #else #define vm_page_assert_locked(m) #define vm_page_lock_assert(m, a) #endif /* * The vm_page's aflags are updated using atomic operations. To set or clear * these flags, the functions vm_page_aflag_set() and vm_page_aflag_clear() * must be used. Neither these flags nor these functions are part of the KBI. * * PGA_REFERENCED may be cleared only if the page is locked. It is set by * both the MI and MD VM layers. However, kernel loadable modules should not * directly set this flag. They should call vm_page_reference() instead. * * PGA_WRITEABLE is set exclusively on managed pages by pmap_enter(). * When it does so, the object must be locked, or the page must be * exclusive busied. The MI VM layer must never access this flag * directly. Instead, it should call pmap_page_is_write_mapped(). * * PGA_EXECUTABLE may be set by pmap routines, and indicates that a page has * at least one executable mapping. It is not consumed by the MI VM layer. * * PGA_NOSYNC must be set and cleared with the page busy lock held. * * PGA_ENQUEUED is set and cleared when a page is inserted into or removed * from a page queue, respectively. It determines whether the plinks.q field * of the page is valid. To set or clear this flag, the queue lock for the * page must be held: the page queue lock corresponding to the page's "queue" * field if its value is not PQ_NONE, and the page lock otherwise. * * PGA_DEQUEUE is set when the page is scheduled to be dequeued from a page * queue, and cleared when the dequeue request is processed. A page may * have PGA_DEQUEUE set and PGA_ENQUEUED cleared, for instance if a dequeue * is requested after the page is scheduled to be enqueued but before it is * actually inserted into the page queue. For allocated pages, the page lock * must be held to set this flag, but it may be set by vm_page_free_prep() * without the page lock held. The page queue lock must be held to clear the * PGA_DEQUEUE flag. * * PGA_REQUEUE is set when the page is scheduled to be enqueued or requeued * in its page queue. The page lock must be held to set this flag, and the * queue lock for the page must be held to clear it. * * PGA_REQUEUE_HEAD is a special flag for enqueuing pages near the head of * the inactive queue, thus bypassing LRU. The page lock must be held to * set this flag, and the queue lock for the page must be held to clear it. * * PGA_SWAP_FREE is used to defer freeing swap space to the pageout daemon * when the context that dirties the page does not have the object write lock * held. */ #define PGA_WRITEABLE 0x0001 /* page may be mapped writeable */ #define PGA_REFERENCED 0x0002 /* page has been referenced */ #define PGA_EXECUTABLE 0x0004 /* page may be mapped executable */ #define PGA_ENQUEUED 0x0008 /* page is enqueued in a page queue */ #define PGA_DEQUEUE 0x0010 /* page is due to be dequeued */ #define PGA_REQUEUE 0x0020 /* page is due to be requeued */ #define PGA_REQUEUE_HEAD 0x0040 /* page requeue should bypass LRU */ #define PGA_NOSYNC 0x0080 /* do not collect for syncer */ #define PGA_SWAP_FREE 0x0100 /* page with swap space was dirtied */ #define PGA_SWAP_SPACE 0x0200 /* page has allocated swap space */ #define PGA_QUEUE_OP_MASK (PGA_DEQUEUE | PGA_REQUEUE | PGA_REQUEUE_HEAD) #define PGA_QUEUE_STATE_MASK (PGA_ENQUEUED | PGA_QUEUE_OP_MASK) /* * Page flags. If changed at any other time than page allocation or * freeing, the modification must be protected by the vm_page lock. * * The PG_PCPU_CACHE flag is set at allocation time if the page was * allocated from a per-CPU cache. It is cleared the next time that the * page is allocated from the physical memory allocator. */ #define PG_PCPU_CACHE 0x01 /* was allocated from per-CPU caches */ #define PG_FICTITIOUS 0x02 /* physical page doesn't exist */ #define PG_ZERO 0x04 /* page is zeroed */ #define PG_MARKER 0x08 /* special queue marker page */ #define PG_NODUMP 0x10 /* don't include this page in a dump */ /* * Misc constants. */ #define ACT_DECLINE 1 #define ACT_ADVANCE 3 #define ACT_INIT 5 #define ACT_MAX 64 #ifdef _KERNEL #include #include /* * Each pageable resident page falls into one of five lists: * * free * Available for allocation now. * * inactive * Low activity, candidates for reclamation. * This list is approximately LRU ordered. * * laundry * This is the list of pages that should be * paged out next. * * unswappable * Dirty anonymous pages that cannot be paged * out because no swap device is configured. * * active * Pages that are "active", i.e., they have been * recently referenced. * */ extern vm_page_t vm_page_array; /* First resident page in table */ extern long vm_page_array_size; /* number of vm_page_t's */ extern long first_page; /* first physical page number */ #define VM_PAGE_TO_PHYS(entry) ((entry)->phys_addr) /* * PHYS_TO_VM_PAGE() returns the vm_page_t object that represents a memory * page to which the given physical address belongs. The correct vm_page_t * object is returned for addresses that are not page-aligned. */ vm_page_t PHYS_TO_VM_PAGE(vm_paddr_t pa); /* * Page allocation parameters for vm_page for the functions * vm_page_alloc(), vm_page_grab(), vm_page_alloc_contig() and * vm_page_alloc_freelist(). Some functions support only a subset * of the flags, and ignore others, see the flags legend. * * The meaning of VM_ALLOC_ZERO differs slightly between the vm_page_alloc*() * and the vm_page_grab*() functions. See these functions for details. * * Bits 0 - 1 define class. * Bits 2 - 15 dedicated for flags. * Legend: * (a) - vm_page_alloc() supports the flag. * (c) - vm_page_alloc_contig() supports the flag. * (f) - vm_page_alloc_freelist() supports the flag. * (g) - vm_page_grab() supports the flag. * (p) - vm_page_grab_pages() supports the flag. * Bits above 15 define the count of additional pages that the caller * intends to allocate. */ #define VM_ALLOC_NORMAL 0 #define VM_ALLOC_INTERRUPT 1 #define VM_ALLOC_SYSTEM 2 #define VM_ALLOC_CLASS_MASK 3 #define VM_ALLOC_WAITOK 0x0008 /* (acf) Sleep and retry */ #define VM_ALLOC_WAITFAIL 0x0010 /* (acf) Sleep and return error */ #define VM_ALLOC_WIRED 0x0020 /* (acfgp) Allocate a wired page */ #define VM_ALLOC_ZERO 0x0040 /* (acfgp) Allocate a prezeroed page */ #define VM_ALLOC_NOOBJ 0x0100 /* (acg) No associated object */ #define VM_ALLOC_NOBUSY 0x0200 /* (acgp) Do not excl busy the page */ #define VM_ALLOC_NOCREAT 0x0400 /* (gp) Don't create a page */ #define VM_ALLOC_IGN_SBUSY 0x1000 /* (gp) Ignore shared busy flag */ #define VM_ALLOC_NODUMP 0x2000 /* (ag) don't include in dump */ #define VM_ALLOC_SBUSY 0x4000 /* (acgp) Shared busy the page */ #define VM_ALLOC_NOWAIT 0x8000 /* (acfgp) Do not sleep */ #define VM_ALLOC_COUNT_SHIFT 16 #define VM_ALLOC_COUNT(count) ((count) << VM_ALLOC_COUNT_SHIFT) #ifdef M_NOWAIT static inline int malloc2vm_flags(int malloc_flags) { int pflags; KASSERT((malloc_flags & M_USE_RESERVE) == 0 || (malloc_flags & M_NOWAIT) != 0, ("M_USE_RESERVE requires M_NOWAIT")); pflags = (malloc_flags & M_USE_RESERVE) != 0 ? VM_ALLOC_INTERRUPT : VM_ALLOC_SYSTEM; if ((malloc_flags & M_ZERO) != 0) pflags |= VM_ALLOC_ZERO; if ((malloc_flags & M_NODUMP) != 0) pflags |= VM_ALLOC_NODUMP; if ((malloc_flags & M_NOWAIT)) pflags |= VM_ALLOC_NOWAIT; if ((malloc_flags & M_WAITOK)) pflags |= VM_ALLOC_WAITOK; return (pflags); } #endif /* * Predicates supported by vm_page_ps_test(): * * PS_ALL_DIRTY is true only if the entire (super)page is dirty. * However, it can be spuriously false when the (super)page has become * dirty in the pmap but that information has not been propagated to the * machine-independent layer. */ #define PS_ALL_DIRTY 0x1 #define PS_ALL_VALID 0x2 #define PS_NONE_BUSY 0x4 bool vm_page_busy_acquire(vm_page_t m, int allocflags); void vm_page_busy_downgrade(vm_page_t m); int vm_page_busy_tryupgrade(vm_page_t m); void vm_page_busy_sleep(vm_page_t m, const char *msg, bool nonshared); void vm_page_free(vm_page_t m); void vm_page_free_zero(vm_page_t m); void vm_page_activate (vm_page_t); void vm_page_advise(vm_page_t m, int advice); vm_page_t vm_page_alloc(vm_object_t, vm_pindex_t, int); vm_page_t vm_page_alloc_domain(vm_object_t, vm_pindex_t, int, int); vm_page_t vm_page_alloc_after(vm_object_t, vm_pindex_t, int, vm_page_t); vm_page_t vm_page_alloc_domain_after(vm_object_t, vm_pindex_t, int, int, vm_page_t); vm_page_t vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_contig_domain(vm_object_t object, vm_pindex_t pindex, int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary, vm_memattr_t memattr); vm_page_t vm_page_alloc_freelist(int, int); vm_page_t vm_page_alloc_freelist_domain(int, int, int); void vm_page_bits_set(vm_page_t m, vm_page_bits_t *bits, vm_page_bits_t set); bool vm_page_blacklist_add(vm_paddr_t pa, bool verbose); void vm_page_change_lock(vm_page_t m, struct mtx **mtx); vm_page_t vm_page_grab (vm_object_t, vm_pindex_t, int); int vm_page_grab_pages(vm_object_t object, vm_pindex_t pindex, int allocflags, vm_page_t *ma, int count); int vm_page_grab_valid(vm_page_t *mp, vm_object_t object, vm_pindex_t pindex, int allocflags); void vm_page_deactivate(vm_page_t); void vm_page_deactivate_noreuse(vm_page_t); void vm_page_dequeue(vm_page_t m); void vm_page_dequeue_deferred(vm_page_t m); vm_page_t vm_page_find_least(vm_object_t, vm_pindex_t); vm_page_t vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); int vm_page_insert (vm_page_t, vm_object_t, vm_pindex_t); void vm_page_invalid(vm_page_t m); void vm_page_launder(vm_page_t m); vm_page_t vm_page_lookup (vm_object_t, vm_pindex_t); vm_page_t vm_page_next(vm_page_t m); void vm_page_pqbatch_drain(void); void vm_page_pqbatch_submit(vm_page_t m, uint8_t queue); bool vm_page_pqstate_commit(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new); vm_page_t vm_page_prev(vm_page_t m); bool vm_page_ps_test(vm_page_t m, int flags, vm_page_t skip_m); void vm_page_putfake(vm_page_t m); void vm_page_readahead_finish(vm_page_t m); bool vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); bool vm_page_reclaim_contig_domain(int domain, int req, u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary); void vm_page_reference(vm_page_t m); #define VPR_TRYFREE 0x01 #define VPR_NOREUSE 0x02 void vm_page_release(vm_page_t m, int flags); void vm_page_release_locked(vm_page_t m, int flags); bool vm_page_remove(vm_page_t); bool vm_page_remove_xbusy(vm_page_t); int vm_page_rename(vm_page_t, vm_object_t, vm_pindex_t); void vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex, vm_page_t mold); void vm_page_requeue(vm_page_t m); int vm_page_sbusied(vm_page_t m); vm_page_t vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end, u_long alignment, vm_paddr_t boundary, int options); vm_page_bits_t vm_page_set_dirty(vm_page_t m); void vm_page_set_valid_range(vm_page_t m, int base, int size); int vm_page_sleep_if_busy(vm_page_t m, const char *msg); int vm_page_sleep_if_xbusy(vm_page_t m, const char *msg); vm_offset_t vm_page_startup(vm_offset_t vaddr); void vm_page_sunbusy(vm_page_t m); void vm_page_swapqueue(vm_page_t m, uint8_t oldq, uint8_t newq); bool vm_page_try_remove_all(vm_page_t m); bool vm_page_try_remove_write(vm_page_t m); int vm_page_trysbusy(vm_page_t m); int vm_page_tryxbusy(vm_page_t m); void vm_page_unhold_pages(vm_page_t *ma, int count); void vm_page_unswappable(vm_page_t m); void vm_page_unwire(vm_page_t m, uint8_t queue); bool vm_page_unwire_noq(vm_page_t m); void vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr); void vm_page_wire(vm_page_t); bool vm_page_wire_mapped(vm_page_t m); void vm_page_xunbusy_hard(vm_page_t m); void vm_page_xunbusy_hard_unchecked(vm_page_t m); void vm_page_set_validclean (vm_page_t, int, int); void vm_page_clear_dirty(vm_page_t, int, int); void vm_page_set_invalid(vm_page_t, int, int); void vm_page_valid(vm_page_t m); int vm_page_is_valid(vm_page_t, int, int); void vm_page_test_dirty(vm_page_t); vm_page_bits_t vm_page_bits(int base, int size); void vm_page_zero_invalid(vm_page_t m, boolean_t setvalid); void vm_page_free_pages_toq(struct spglist *free, bool update_wire_count); void vm_page_dirty_KBI(vm_page_t m); void vm_page_lock_KBI(vm_page_t m, const char *file, int line); void vm_page_unlock_KBI(vm_page_t m, const char *file, int line); int vm_page_trylock_KBI(vm_page_t m, const char *file, int line); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line); void vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line); #endif #define vm_page_assert_busied(m) \ KASSERT(vm_page_busied(m), \ ("vm_page_assert_busied: page %p not busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_sbusied(m) \ KASSERT(vm_page_sbusied(m), \ ("vm_page_assert_sbusied: page %p not shared busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_unbusied(m) \ KASSERT(!vm_page_busied(m), \ ("vm_page_assert_unbusied: page %p busy @ %s:%d", \ (m), __FILE__, __LINE__)) #define vm_page_assert_xbusied_unchecked(m) do { \ KASSERT(vm_page_xbusied(m), \ ("vm_page_assert_xbusied: page %p not exclusive busy @ %s:%d", \ (m), __FILE__, __LINE__)); \ } while (0) #define vm_page_assert_xbusied(m) do { \ vm_page_assert_xbusied_unchecked(m); \ KASSERT((m->busy_lock & ~VPB_BIT_WAITERS) == \ VPB_CURTHREAD_EXCLUSIVE, \ ("vm_page_assert_xbusied: page %p busy_lock %#x not owned" \ " by me @ %s:%d", \ (m), (m)->busy_lock, __FILE__, __LINE__)); \ } while (0) #define vm_page_busied(m) \ ((m)->busy_lock != VPB_UNBUSIED) #define vm_page_sbusy(m) do { \ if (!vm_page_trysbusy(m)) \ panic("%s: page %p failed shared busying", __func__, \ (m)); \ } while (0) #define vm_page_xbusied(m) \ (((m)->busy_lock & VPB_SINGLE_EXCLUSIVE) != 0) #define vm_page_xbusy(m) do { \ if (!vm_page_tryxbusy(m)) \ panic("%s: page %p failed exclusive busying", __func__, \ (m)); \ } while (0) /* Note: page m's lock must not be owned by the caller. */ #define vm_page_xunbusy(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard(m); \ } while (0) #define vm_page_xunbusy_unchecked(m) do { \ if (!atomic_cmpset_rel_int(&(m)->busy_lock, \ VPB_CURTHREAD_EXCLUSIVE, VPB_UNBUSIED)) \ vm_page_xunbusy_hard_unchecked(m); \ } while (0) #ifdef INVARIANTS void vm_page_object_busy_assert(vm_page_t m); #define VM_PAGE_OBJECT_BUSY_ASSERT(m) vm_page_object_busy_assert(m) void vm_page_assert_pga_writeable(vm_page_t m, uint16_t bits); #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) \ vm_page_assert_pga_writeable(m, bits) #else #define VM_PAGE_OBJECT_BUSY_ASSERT(m) (void)0 #define VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits) (void)0 #endif #if BYTE_ORDER == BIG_ENDIAN #define VM_PAGE_AFLAG_SHIFT 16 #else #define VM_PAGE_AFLAG_SHIFT 0 #endif /* * Load a snapshot of a page's 32-bit atomic state. */ static inline vm_page_astate_t vm_page_astate_load(vm_page_t m) { vm_page_astate_t a; a._bits = atomic_load_32(&m->a._bits); return (a); } /* * Atomically compare and set a page's atomic state. */ static inline bool vm_page_astate_fcmpset(vm_page_t m, vm_page_astate_t *old, vm_page_astate_t new) { KASSERT(new.queue == PQ_INACTIVE || (new.flags & PGA_REQUEUE_HEAD) == 0, ("%s: invalid head requeue request for page %p", __func__, m)); KASSERT((new.flags & PGA_ENQUEUED) == 0 || new.queue != PQ_NONE, ("%s: setting PGA_ENQUEUED with PQ_NONE in page %p", __func__, m)); KASSERT(new._bits != old->_bits, ("%s: bits are unchanged", __func__)); return (atomic_fcmpset_32(&m->a._bits, &old->_bits, new._bits) != 0); } /* * Clear the given bits in the specified page. */ static inline void vm_page_aflag_clear(vm_page_t m, uint16_t bits) { uint32_t *addr, val; /* - * The PGA_REFERENCED flag can only be cleared if the page is locked. - */ - if ((bits & PGA_REFERENCED) != 0) - vm_page_assert_locked(m); - - /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_clear_32(addr, val); } /* * Set the given bits in the specified page. */ static inline void vm_page_aflag_set(vm_page_t m, uint16_t bits) { uint32_t *addr, val; VM_PAGE_ASSERT_PGA_WRITEABLE(m, bits); /* * Access the whole 32-bit word containing the aflags field with an * atomic update. Parallel non-atomic updates to the other fields * within this word are handled properly by the atomic update. */ addr = (void *)&m->a; val = bits << VM_PAGE_AFLAG_SHIFT; atomic_set_32(addr, val); } /* * Atomically update the queue state of the page. The operation fails if * any of the queue flags in "fflags" are set or if the "queue" field of * the page does not match the expected value; if the operation is * successful, the flags in "nflags" are set and all other queue state * flags are cleared. */ static inline bool vm_page_pqstate_cmpset(vm_page_t m, uint32_t oldq, uint32_t newq, uint32_t fflags, uint32_t nflags) { vm_page_astate_t new, old; old = vm_page_astate_load(m); do { if ((old.flags & fflags) != 0 || old.queue != oldq) return (false); new = old; new.flags = (new.flags & ~PGA_QUEUE_OP_MASK) | nflags; new.queue = newq; } while (!vm_page_astate_fcmpset(m, &old, new)); return (true); } /* * vm_page_dirty: * * Set all bits in the page's dirty field. * * The object containing the specified page must be locked if the * call is made from the machine-independent layer. * * See vm_page_clear_dirty_mask(). */ static __inline void vm_page_dirty(vm_page_t m) { /* Use vm_page_dirty_KBI() under INVARIANTS to save memory. */ #if (defined(KLD_MODULE) && !defined(KLD_TIED)) || defined(INVARIANTS) vm_page_dirty_KBI(m); #else m->dirty = VM_PAGE_BITS_ALL; #endif } /* * vm_page_undirty: * * Set page to not be dirty. Note: does not clear pmap modify bits */ static __inline void vm_page_undirty(vm_page_t m) { VM_PAGE_OBJECT_BUSY_ASSERT(m); m->dirty = 0; } static inline uint8_t _vm_page_queue(vm_page_astate_t as) { if ((as.flags & PGA_DEQUEUE) != 0) return (PQ_NONE); return (as.queue); } /* * vm_page_queue: * * Return the index of the queue containing m. */ static inline uint8_t vm_page_queue(vm_page_t m) { - - vm_page_assert_locked(m); return (_vm_page_queue(vm_page_astate_load(m))); } static inline bool vm_page_active(vm_page_t m) { return (vm_page_queue(m) == PQ_ACTIVE); } static inline bool vm_page_inactive(vm_page_t m) { return (vm_page_queue(m) == PQ_INACTIVE); } static inline bool vm_page_in_laundry(vm_page_t m) { uint8_t queue; queue = vm_page_queue(m); return (queue == PQ_LAUNDRY || queue == PQ_UNSWAPPABLE); } /* * vm_page_drop: * * Release a reference to a page and return the old reference count. */ static inline u_int vm_page_drop(vm_page_t m, u_int val) { u_int old; /* * Synchronize with vm_page_free_prep(): ensure that all updates to the * page structure are visible before it is freed. */ atomic_thread_fence_rel(); old = atomic_fetchadd_int(&m->ref_count, -val); KASSERT(old != VPRC_BLOCKED, ("vm_page_drop: page %p has an invalid refcount value", m)); return (old); } /* * vm_page_wired: * * Perform a racy check to determine whether a reference prevents the page * from being reclaimable. If the page's object is locked, and the page is * unmapped and unbusied or exclusively busied by the current thread, no * new wirings may be created. */ static inline bool vm_page_wired(vm_page_t m) { return (VPRC_WIRE_COUNT(m->ref_count) > 0); } static inline bool vm_page_all_valid(vm_page_t m) { return (m->valid == VM_PAGE_BITS_ALL); } static inline bool vm_page_none_valid(vm_page_t m) { return (m->valid == 0); } #endif /* _KERNEL */ #endif /* !_VM_PAGE_ */ Index: head/sys/vm/vm_pageout.c =================================================================== --- head/sys/vm/vm_pageout.c (revision 356156) +++ head/sys/vm/vm_pageout.c (revision 356157) @@ -1,2285 +1,2238 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ /* * The proverbial page-out daemon. */ #include __FBSDID("$FreeBSD$"); #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * System initialization */ /* the kernel process "vm_pageout"*/ static void vm_pageout(void); static void vm_pageout_init(void); static int vm_pageout_clean(vm_page_t m, int *numpagedout); static int vm_pageout_cluster(vm_page_t m); static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage); SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init, NULL); struct proc *pageproc; static struct kproc_desc page_kp = { "pagedaemon", vm_pageout, &pageproc }; SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &page_kp); SDT_PROVIDER_DEFINE(vm); SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan); /* Pagedaemon activity rates, in subdivisions of one second. */ #define VM_LAUNDER_RATE 10 #define VM_INACT_SCAN_RATE 10 static int vm_pageout_oom_seq = 12; static int vm_pageout_update_period; static int disable_swap_pageouts; static int lowmem_period = 10; static int swapdev_enabled; static int vm_panic_on_oom = 0; SYSCTL_INT(_vm, OID_AUTO, panic_on_oom, CTLFLAG_RWTUN, &vm_panic_on_oom, 0, "panic on out of memory instead of killing the largest process"); SYSCTL_INT(_vm, OID_AUTO, pageout_update_period, CTLFLAG_RWTUN, &vm_pageout_update_period, 0, "Maximum active LRU update period"); SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0, "Low memory callback period"); SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); static int pageout_lock_miss; SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq, CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0, "back-to-back calls to oom detector to start OOM"); static int act_scan_laundry_weight = 3; SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN, &act_scan_laundry_weight, 0, "weight given to clean vs. dirty pages in active queue scans"); static u_int vm_background_launder_rate = 4096; SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN, &vm_background_launder_rate, 0, "background laundering rate, in kilobytes per second"); static u_int vm_background_launder_max = 20 * 1024; SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN, &vm_background_launder_max, 0, "background laundering cap, in kilobytes"); int vm_pageout_page_count = 32; u_long vm_page_max_user_wired; SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW, &vm_page_max_user_wired, 0, "system-wide limit to user-wired page count"); static u_int isqrt(u_int num); static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall); static void vm_pageout_laundry_worker(void *arg); struct scan_state { struct vm_batchqueue bq; struct vm_pagequeue *pq; vm_page_t marker; int maxscan; int scanned; }; static void vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq, vm_page_t marker, vm_page_t after, int maxscan) { vm_pagequeue_assert_locked(pq); KASSERT((marker->a.flags & PGA_ENQUEUED) == 0, ("marker %p already enqueued", marker)); if (after == NULL) TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q); else TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q); vm_page_aflag_set(marker, PGA_ENQUEUED); vm_batchqueue_init(&ss->bq); ss->pq = pq; ss->marker = marker; ss->maxscan = maxscan; ss->scanned = 0; vm_pagequeue_unlock(pq); } static void vm_pageout_end_scan(struct scan_state *ss) { struct vm_pagequeue *pq; pq = ss->pq; vm_pagequeue_assert_locked(pq); KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q); vm_page_aflag_clear(ss->marker, PGA_ENQUEUED); pq->pq_pdpages += ss->scanned; } /* * Add a small number of queued pages to a batch queue for later processing * without the corresponding queue lock held. The caller must have enqueued a * marker page at the desired start point for the scan. Pages will be * physically dequeued if the caller so requests. Otherwise, the returned * batch may contain marker pages, and it is up to the caller to handle them. * * When processing the batch queue, vm_page_queue() must be used to * determine whether the page has been logically dequeued by another thread. * Once this check is performed, the page lock guarantees that the page will * not be disassociated from the queue. */ static __always_inline void vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue) { struct vm_pagequeue *pq; vm_page_t m, marker, n; marker = ss->marker; pq = ss->pq; KASSERT((marker->a.flags & PGA_ENQUEUED) != 0, ("marker %p not enqueued", ss->marker)); vm_pagequeue_lock(pq); for (m = TAILQ_NEXT(marker, plinks.q); m != NULL && ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE; m = n, ss->scanned++) { n = TAILQ_NEXT(m, plinks.q); if ((m->flags & PG_MARKER) == 0) { KASSERT((m->a.flags & PGA_ENQUEUED) != 0, ("page %p not enqueued", m)); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("Fictitious page %p cannot be in page queue", m)); KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("Unmanaged page %p cannot be in page queue", m)); } else if (dequeue) continue; (void)vm_batchqueue_insert(&ss->bq, m); if (dequeue) { TAILQ_REMOVE(&pq->pq_pl, m, plinks.q); vm_page_aflag_clear(m, PGA_ENQUEUED); } } TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q); if (__predict_true(m != NULL)) TAILQ_INSERT_BEFORE(m, marker, plinks.q); else TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q); if (dequeue) vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt); vm_pagequeue_unlock(pq); } /* * Return the next page to be scanned, or NULL if the scan is complete. */ static __always_inline vm_page_t vm_pageout_next(struct scan_state *ss, const bool dequeue) { if (ss->bq.bq_cnt == 0) vm_pageout_collect_batch(ss, dequeue); return (vm_batchqueue_pop(&ss->bq)); } /* * Determine whether processing of a page should be deferred and ensure that any * outstanding queue operations are processed. */ static __always_inline bool vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued) { vm_page_astate_t as; as = vm_page_astate_load(m); if (__predict_false(as.queue != queue || ((as.flags & PGA_ENQUEUED) != 0) != enqueued)) return (true); if ((as.flags & PGA_QUEUE_OP_MASK) != 0) { vm_page_pqbatch_submit(m, queue); return (true); } return (false); } /* * Scan for pages at adjacent offsets within the given page's object that are * eligible for laundering, form a cluster of these pages and the given page, * and launder that cluster. */ static int vm_pageout_cluster(vm_page_t m) { vm_object_t object; vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps; vm_pindex_t pindex; int ib, is, page_base, pageout_count; object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); pindex = m->pindex; vm_page_assert_xbusied(m); mc[vm_pageout_page_count] = pb = ps = m; pageout_count = 1; page_base = vm_pageout_page_count; ib = 1; is = 1; /* * We can cluster only if the page is not clean, busy, or held, and * the page is in the laundry queue. * * During heavy mmap/modification loads the pageout * daemon can really fragment the underlying file * due to flushing pages out of order and not trying to * align the clusters (which leaves sporadic out-of-order * holes). To solve this problem we do the reverse scan * first and attempt to align our cluster, then do a * forward scan if room remains. */ more: while (ib != 0 && pageout_count < vm_pageout_page_count) { if (ib > pindex) { ib = 0; break; } if ((p = vm_page_prev(pb)) == NULL || vm_page_tryxbusy(p) == 0) { ib = 0; break; } if (vm_page_wired(p)) { ib = 0; vm_page_xunbusy(p); break; } vm_page_test_dirty(p); if (p->dirty == 0) { ib = 0; vm_page_xunbusy(p); break; } - vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { - vm_page_unlock(p); vm_page_xunbusy(p); ib = 0; break; } - vm_page_unlock(p); mc[--page_base] = pb = p; ++pageout_count; ++ib; /* * We are at an alignment boundary. Stop here, and switch * directions. Do not clear ib. */ if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) break; } while (pageout_count < vm_pageout_page_count && pindex + is < object->size) { if ((p = vm_page_next(ps)) == NULL || vm_page_tryxbusy(p) == 0) break; if (vm_page_wired(p)) { vm_page_xunbusy(p); break; } vm_page_test_dirty(p); if (p->dirty == 0) { vm_page_xunbusy(p); break; } - vm_page_lock(p); if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) { - vm_page_unlock(p); vm_page_xunbusy(p); break; } - vm_page_unlock(p); mc[page_base + pageout_count] = ps = p; ++pageout_count; ++is; } /* * If we exhausted our forward scan, continue with the reverse scan * when possible, even past an alignment boundary. This catches * boundary conditions. */ if (ib != 0 && pageout_count < vm_pageout_page_count) goto more; return (vm_pageout_flush(&mc[page_base], pageout_count, VM_PAGER_PUT_NOREUSE, 0, NULL, NULL)); } /* * vm_pageout_flush() - launder the given pages * * The given pages are laundered. Note that we setup for the start of * I/O ( i.e. busy the page ), mark it read-only, and bump the object * reference count all in here rather then in the parent. If we want * the parent to do more sophisticated things we may have to change * the ordering. * * Returned runlen is the count of pages between mreq and first * page after mreq with status VM_PAGER_AGAIN. * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL * for any page in runlen set. */ int vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen, boolean_t *eio) { vm_object_t object = mc[0]->object; int pageout_status[count]; int numpagedout = 0; int i, runlen; VM_OBJECT_ASSERT_WLOCKED(object); /* * Initiate I/O. Mark the pages shared busy and verify that they're * valid and read-only. * * We do not have to fixup the clean/dirty bits here... we can * allow the pager to do it after the I/O completes. * * NOTE! mc[i]->dirty may be partial or fragmented due to an * edge case with file fragments. */ for (i = 0; i < count; i++) { KASSERT(vm_page_all_valid(mc[i]), ("vm_pageout_flush: partially invalid page %p index %d/%d", mc[i], i, count)); KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0, ("vm_pageout_flush: writeable page %p", mc[i])); vm_page_busy_downgrade(mc[i]); } vm_object_pip_add(object, count); vm_pager_put_pages(object, mc, count, flags, pageout_status); runlen = count - mreq; if (eio != NULL) *eio = FALSE; for (i = 0; i < count; i++) { vm_page_t mt = mc[i]; KASSERT(pageout_status[i] == VM_PAGER_PEND || !pmap_page_is_write_mapped(mt), ("vm_pageout_flush: page %p is not write protected", mt)); switch (pageout_status[i]) { case VM_PAGER_OK: - vm_page_lock(mt); + /* + * The page may have moved since laundering started, in + * which case it should be left alone. + */ if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); - vm_page_unlock(mt); /* FALLTHROUGH */ case VM_PAGER_PEND: numpagedout++; break; case VM_PAGER_BAD: /* * The page is outside the object's range. We pretend * that the page out worked and clean the page, so the * changes will be lost if the page is reclaimed by * the page daemon. */ vm_page_undirty(mt); - vm_page_lock(mt); if (vm_page_in_laundry(mt)) vm_page_deactivate_noreuse(mt); - vm_page_unlock(mt); break; case VM_PAGER_ERROR: case VM_PAGER_FAIL: /* * If the page couldn't be paged out to swap because the * pager wasn't able to find space, place the page in * the PQ_UNSWAPPABLE holding queue. This is an * optimization that prevents the page daemon from * wasting CPU cycles on pages that cannot be reclaimed * becase no swap device is configured. * * Otherwise, reactivate the page so that it doesn't * clog the laundry and inactive queues. (We will try * paging it out again later.) */ - vm_page_lock(mt); if (object->type == OBJT_SWAP && pageout_status[i] == VM_PAGER_FAIL) { vm_page_unswappable(mt); numpagedout++; } else vm_page_activate(mt); - vm_page_unlock(mt); if (eio != NULL && i >= mreq && i - mreq < runlen) *eio = TRUE; break; case VM_PAGER_AGAIN: if (i >= mreq && i - mreq < runlen) runlen = i - mreq; break; } /* * If the operation is still going, leave the page busy to * block all other accesses. Also, leave the paging in * progress indicator set so that we don't attempt an object * collapse. */ if (pageout_status[i] != VM_PAGER_PEND) { vm_object_pip_wakeup(object); vm_page_sunbusy(mt); } } if (prunlen != NULL) *prunlen = runlen; return (numpagedout); } static void vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused) { atomic_store_rel_int(&swapdev_enabled, 1); } static void vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused) { if (swap_pager_nswapdev() == 1) atomic_store_rel_int(&swapdev_enabled, 0); } /* * Attempt to acquire all of the necessary locks to launder a page and * then call through the clustering layer to PUTPAGES. Wait a short * time for a vnode lock. * * Requires the page and object lock on entry, releases both before return. * Returns 0 on success and an errno otherwise. */ static int vm_pageout_clean(vm_page_t m, int *numpagedout) { struct vnode *vp; struct mount *mp; vm_object_t object; vm_pindex_t pindex; int error, lockmode; - vm_page_assert_locked(m); object = m->object; VM_OBJECT_ASSERT_WLOCKED(object); error = 0; vp = NULL; mp = NULL; /* * The object is already known NOT to be dead. It * is possible for the vget() to block the whole * pageout daemon, but the new low-memory handling * code should prevent it. * * We can't wait forever for the vnode lock, we might * deadlock due to a vn_read() getting stuck in * vm_wait while holding this vnode. We skip the * vnode if we can't get it in a reasonable amount * of time. */ if (object->type == OBJT_VNODE) { - vm_page_unlock(m); vm_page_xunbusy(m); vp = object->handle; if (vp->v_type == VREG && vn_start_write(vp, &mp, V_NOWAIT) != 0) { mp = NULL; error = EDEADLK; goto unlock_all; } KASSERT(mp != NULL, ("vp %p with NULL v_mount", vp)); vm_object_reference_locked(object); pindex = m->pindex; VM_OBJECT_WUNLOCK(object); lockmode = MNT_SHARED_WRITES(vp->v_mount) ? LK_SHARED : LK_EXCLUSIVE; if (vget(vp, lockmode | LK_TIMELOCK, curthread)) { vp = NULL; error = EDEADLK; goto unlock_mp; } VM_OBJECT_WLOCK(object); /* * Ensure that the object and vnode were not disassociated * while locks were dropped. */ if (vp->v_object != object) { error = ENOENT; goto unlock_all; } - vm_page_lock(m); /* - * While the object and page were unlocked, the page - * may have been: + * While the object was unlocked, the page may have been: * (1) moved to a different queue, * (2) reallocated to a different object, * (3) reallocated to a different offset, or * (4) cleaned. */ if (!vm_page_in_laundry(m) || m->object != object || m->pindex != pindex || m->dirty == 0) { - vm_page_unlock(m); error = ENXIO; goto unlock_all; } /* - * The page may have been busied while the object and page - * locks were released. + * The page may have been busied while the object lock was + * released. */ if (vm_page_tryxbusy(m) == 0) { - vm_page_unlock(m); error = EBUSY; goto unlock_all; } } /* * Remove all writeable mappings, failing if the page is wired. */ if (!vm_page_try_remove_write(m)) { vm_page_xunbusy(m); - vm_page_unlock(m); error = EBUSY; goto unlock_all; } - vm_page_unlock(m); /* * If a page is dirty, then it is either being washed * (but not yet cleaned) or it is still in the * laundry. If it is still in the laundry, then we * start the cleaning operation. */ if ((*numpagedout = vm_pageout_cluster(m)) == 0) error = EIO; unlock_all: VM_OBJECT_WUNLOCK(object); unlock_mp: - vm_page_lock_assert(m, MA_NOTOWNED); if (mp != NULL) { if (vp != NULL) vput(vp); vm_object_deallocate(object); vn_finished_write(mp); } return (error); } /* * Attempt to launder the specified number of pages. * * Returns the number of pages successfully laundered. */ static int vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall) { struct scan_state ss; struct vm_pagequeue *pq; - struct mtx *mtx; vm_object_t object; vm_page_t m, marker; vm_page_astate_t new, old; int act_delta, error, numpagedout, queue, refs, starting_target; int vnodes_skipped; bool pageout_ok; - mtx = NULL; object = NULL; starting_target = launder; vnodes_skipped = 0; /* * Scan the laundry queues for pages eligible to be laundered. We stop * once the target number of dirty pages have been laundered, or once * we've reached the end of the queue. A single iteration of this loop * may cause more than one page to be laundered because of clustering. * * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no * swap devices are configured. */ if (atomic_load_acq_int(&swapdev_enabled)) queue = PQ_UNSWAPPABLE; else queue = PQ_LAUNDRY; scan: marker = &vmd->vmd_markers[queue]; pq = &vmd->vmd_pagequeues[queue]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false((m->flags & PG_MARKER) != 0)) continue; - vm_page_change_lock(m, &mtx); - -recheck: /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, queue, true)) continue; - if (object != m->object) { + /* + * Lock the page's object. + */ + if (object == NULL || object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); - - /* - * A page's object pointer may be set to NULL before - * the object lock is acquired. - */ object = (vm_object_t)atomic_load_ptr(&m->object); - if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { - mtx_unlock(mtx); - /* Depends on type-stability. */ - VM_OBJECT_WLOCK(object); - mtx_lock(mtx); - goto recheck; + if (__predict_false(object == NULL)) + /* The page is being freed by another thread. */ + continue; + + /* Depends on type-stability. */ + VM_OBJECT_WLOCK(object); + if (__predict_false(m->object != object)) { + VM_OBJECT_WUNLOCK(object); + object = NULL; + continue; } } - if (__predict_false(m->object == NULL)) - /* - * The page has been removed from its object. - */ - continue; - KASSERT(m->object == object, ("page %p does not belong to %p", - m, object)); if (vm_page_tryxbusy(m) == 0) continue; /* * Check for wirings now that we hold the object lock and have - * verified that the page is unbusied. If the page is mapped, - * it may still be wired by pmap lookups. The call to + * exclusively busied the page. If the page is mapped, it may + * still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the - * wire count is guaranteed not to increase. + * wire count is guaranteed not to increase after this check. */ - if (__predict_false(vm_page_wired(m))) { - vm_page_dequeue_deferred(m); + if (__predict_false(vm_page_wired(m))) goto skip_page; - } /* * Invalid pages can be easily freed. They cannot be * mapped; vm_page_free() asserts this. */ if (vm_page_none_valid(m)) goto free_page; refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; for (old = vm_page_astate_load(m);;) { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) goto skip_page; new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta == 0) { ; } else if (object->ref_count != 0) { /* * Increase the activation count if the page was * referenced while in the laundry queue. This * makes it less likely that the page will be * returned prematurely to the laundry queue. */ new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; new.flags |= PGA_REQUEUE; new.queue = PQ_ACTIVE; if (!vm_page_pqstate_commit(m, &old, new)) continue; /* * If this was a background laundering, count * activated pages towards our target. The * purpose of background laundering is to ensure * that pages are eventually cycled through the * laundry queue, and an activation is a valid * way out. */ if (!in_shortfall) launder--; VM_CNT_INC(v_reactivated); goto skip_page; } else if ((object->flags & OBJ_DEAD) == 0) { new.flags |= PGA_REQUEUE; if (!vm_page_pqstate_commit(m, &old, new)) continue; goto skip_page; } break; } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); - if (m->dirty == 0 && !vm_page_try_remove_all(m)) { - vm_page_dequeue_deferred(m); + if (m->dirty == 0 && !vm_page_try_remove_all(m)) goto skip_page; - } } /* * Clean pages are freed, and dirty pages are paged out unless * they belong to a dead object. Requeueing dirty pages from * dead objects is pointless, as they are being paged out and * freed by the thread that destroyed the object. */ if (m->dirty == 0) { free_page: + /* + * Now we are guaranteed that no other threads are + * manipulating the page, check for a last-second + * reference. + */ + if (vm_pageout_defer(m, queue, true)) + goto skip_page; vm_page_free(m); VM_CNT_INC(v_dfree); } else if ((object->flags & OBJ_DEAD) == 0) { if (object->type != OBJT_SWAP && object->type != OBJT_DEFAULT) pageout_ok = true; else if (disable_swap_pageouts) pageout_ok = false; else pageout_ok = true; if (!pageout_ok) { vm_page_launder(m); goto skip_page; } /* * Form a cluster with adjacent, dirty pages from the * same object, and page out that entire cluster. * * The adjacent, dirty pages must also be in the * laundry. However, their mappings are not checked * for new references. Consequently, a recently * referenced page may be paged out. However, that * page will not be prematurely reclaimed. After page * out, the page will be placed in the inactive queue, * where any new references will be detected and the * page reactivated. */ error = vm_pageout_clean(m, &numpagedout); if (error == 0) { launder -= numpagedout; ss.scanned += numpagedout; } else if (error == EDEADLK) { pageout_lock_miss++; vnodes_skipped++; } - mtx = NULL; object = NULL; } else { skip_page: vm_page_xunbusy(m); } } - if (mtx != NULL) { - mtx_unlock(mtx); - mtx = NULL; - } if (object != NULL) { VM_OBJECT_WUNLOCK(object); object = NULL; } vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); if (launder > 0 && queue == PQ_UNSWAPPABLE) { queue = PQ_LAUNDRY; goto scan; } /* * Wakeup the sync daemon if we skipped a vnode in a writeable object * and we didn't launder enough pages. */ if (vnodes_skipped > 0 && launder > 0) (void)speedup_syncer(); return (starting_target - launder); } /* * Compute the integer square root. */ static u_int isqrt(u_int num) { u_int bit, root, tmp; bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0; root = 0; while (bit != 0) { tmp = root + bit; root >>= 1; if (num >= tmp) { num -= tmp; root += bit; } bit >>= 2; } return (root); } /* * Perform the work of the laundry thread: periodically wake up and determine * whether any pages need to be laundered. If so, determine the number of pages * that need to be laundered, and launder them. */ static void vm_pageout_laundry_worker(void *arg) { struct vm_domain *vmd; struct vm_pagequeue *pq; uint64_t nclean, ndirty, nfreed; int domain, last_target, launder, shortfall, shortfall_cycle, target; bool in_shortfall; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; KASSERT(vmd->vmd_segs != 0, ("domain without segments")); shortfall = 0; in_shortfall = false; shortfall_cycle = 0; last_target = target = 0; nfreed = 0; /* * Calls to these handlers are serialized by the swap syscall lock. */ (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd, EVENTHANDLER_PRI_ANY); (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd, EVENTHANDLER_PRI_ANY); /* * The pageout laundry worker is never done, so loop forever. */ for (;;) { KASSERT(target >= 0, ("negative target %d", target)); KASSERT(shortfall_cycle >= 0, ("negative cycle %d", shortfall_cycle)); launder = 0; /* * First determine whether we need to launder pages to meet a * shortage of free pages. */ if (shortfall > 0) { in_shortfall = true; shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE; target = shortfall; } else if (!in_shortfall) goto trybackground; else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) { /* * We recently entered shortfall and began laundering * pages. If we have completed that laundering run * (and we are no longer in shortfall) or we have met * our laundry target through other activity, then we * can stop laundering pages. */ in_shortfall = false; target = 0; goto trybackground; } launder = target / shortfall_cycle--; goto dolaundry; /* * There's no immediate need to launder any pages; see if we * meet the conditions to perform background laundering: * * 1. The ratio of dirty to clean inactive pages exceeds the * background laundering threshold, or * 2. we haven't yet reached the target of the current * background laundering run. * * The background laundering threshold is not a constant. * Instead, it is a slowly growing function of the number of * clean pages freed by the page daemon since the last * background laundering. Thus, as the ratio of dirty to * clean inactive pages grows, the amount of memory pressure * required to trigger laundering decreases. We ensure * that the threshold is non-zero after an inactive queue * scan, even if that scan failed to free a single clean page. */ trybackground: nclean = vmd->vmd_free_count + vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt; ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt; if (target == 0 && ndirty * isqrt(howmany(nfreed + 1, vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) { target = vmd->vmd_background_launder_target; } /* * We have a non-zero background laundering target. If we've * laundered up to our maximum without observing a page daemon * request, just stop. This is a safety belt that ensures we * don't launder an excessive amount if memory pressure is low * and the ratio of dirty to clean pages is large. Otherwise, * proceed at the background laundering rate. */ if (target > 0) { if (nfreed > 0) { nfreed = 0; last_target = target; } else if (last_target - target >= vm_background_launder_max * PAGE_SIZE / 1024) { target = 0; } launder = vm_background_launder_rate * PAGE_SIZE / 1024; launder /= VM_LAUNDER_RATE; if (launder > target) launder = target; } dolaundry: if (launder > 0) { /* * Because of I/O clustering, the number of laundered * pages could exceed "target" by the maximum size of * a cluster minus one. */ target -= min(vm_pageout_launder(vmd, launder, in_shortfall), target); pause("laundp", hz / VM_LAUNDER_RATE); } /* * If we're not currently laundering pages and the page daemon * hasn't posted a new request, sleep until the page daemon * kicks us. */ vm_pagequeue_lock(pq); if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE) (void)mtx_sleep(&vmd->vmd_laundry_request, vm_pagequeue_lockptr(pq), PVM, "launds", 0); /* * If the pagedaemon has indicated that it's in shortfall, start * a shortfall laundering unless we're already in the middle of * one. This may preempt a background laundering. */ if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL && (!in_shortfall || shortfall_cycle == 0)) { shortfall = vm_laundry_target(vmd) + vmd->vmd_pageout_deficit; target = 0; } else shortfall = 0; if (target == 0) vmd->vmd_laundry_request = VM_LAUNDRY_IDLE; nfreed += vmd->vmd_clean_pages_freed; vmd->vmd_clean_pages_freed = 0; vm_pagequeue_unlock(pq); } } /* * Compute the number of pages we want to try to move from the * active queue to either the inactive or laundry queue. * * When scanning active pages during a shortage, we make clean pages * count more heavily towards the page shortage than dirty pages. * This is because dirty pages must be laundered before they can be * reused and thus have less utility when attempting to quickly * alleviate a free page shortage. However, this weighting also * causes the scan to deactivate dirty pages more aggressively, * improving the effectiveness of clustering. */ static int vm_pageout_active_target(struct vm_domain *vmd) { int shortage; shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) - (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt + vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight); shortage *= act_scan_laundry_weight; return (shortage); } /* * Scan the active queue. If there is no shortage of inactive pages, scan a * small portion of the queue in order to maintain quasi-LRU. */ static void vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage) { struct scan_state ss; - struct mtx *mtx; vm_object_t object; vm_page_t m, marker; struct vm_pagequeue *pq; vm_page_astate_t old, new; long min_scan; int act_delta, max_scan, ps_delta, refs, scan_tick; uint8_t nqueue; marker = &vmd->vmd_markers[PQ_ACTIVE]; pq = &vmd->vmd_pagequeues[PQ_ACTIVE]; vm_pagequeue_lock(pq); /* * If we're just idle polling attempt to visit every * active page within 'update_period' seconds. */ scan_tick = ticks; if (vm_pageout_update_period != 0) { min_scan = pq->pq_cnt; min_scan *= scan_tick - vmd->vmd_last_active_scan; min_scan /= hz * vm_pageout_update_period; } else min_scan = 0; if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0)) vmd->vmd_last_active_scan = scan_tick; /* * Scan the active queue for pages that can be deactivated. Update * the per-page activity counter and use it to identify deactivation * candidates. Held pages may be deactivated. * * To avoid requeuing each page that remains in the active queue, we * implement the CLOCK algorithm. To keep the implementation of the * enqueue operation consistent for all page queues, we use two hands, * represented by marker pages. Scans begin at the first hand, which * precedes the second hand in the queue. When the two hands meet, * they are moved back to the head and tail of the queue, respectively, * and scanning resumes. */ max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan; - mtx = NULL; act_scan: vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan); while ((m = vm_pageout_next(&ss, false)) != NULL) { if (__predict_false(m == &vmd->vmd_clock[1])) { vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q); max_scan -= ss.scanned; vm_pageout_end_scan(&ss); goto act_scan; } if (__predict_false((m->flags & PG_MARKER) != 0)) continue; - vm_page_change_lock(m, &mtx); - /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, PQ_ACTIVE, true)) continue; /* * A page's object pointer may be set to NULL before * the object lock is acquired. */ object = (vm_object_t)atomic_load_ptr(&m->object); if (__predict_false(object == NULL)) /* * The page has been removed from its object. */ continue; /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0 && VM_OBJECT_TRYWLOCK(object)) { if (m->object == object) vm_pager_page_unswapped(m); VM_OBJECT_WUNLOCK(object); } /* * Check to see "how much" the page has been used. * * Test PGA_REFERENCED after calling pmap_ts_referenced() so * that a reference from a concurrently destroyed mapping is * observed here and now. * * Perform an unsynchronized object ref count check. While * the page lock ensures that the page is not reallocated to * another object, in particular, one with unmanaged mappings * that cannot support pmap_ts_referenced(), two races are, * nonetheless, possible: * 1) The count was transitioning to zero, but we saw a non- * zero value. pmap_ts_referenced() will return zero * because the page is not mapped. * 2) The count was transitioning to one, but we saw zero. * This race delays the detection of a new reference. At * worst, we will deactivate and reactivate the page. */ refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; old = vm_page_astate_load(m); do { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) break; /* * Advance or decay the act_count based on recent usage. */ new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta != 0) { new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; } else { new.act_count -= min(new.act_count, ACT_DECLINE); } if (new.act_count > 0) { /* * Adjust the activation count and keep the page * in the active queue. The count might be left * unchanged if it is saturated. The page may * have been moved to a different queue since we * started the scan, in which case we move it * back. */ ps_delta = 0; if (old.queue != PQ_ACTIVE) { old.queue = PQ_ACTIVE; old.flags |= PGA_REQUEUE; } } else { /* * When not short for inactive pages, let dirty * pages go through the inactive queue before * moving to the laundry queue. This gives them * some extra time to be reactivated, * potentially avoiding an expensive pageout. * However, during a page shortage, the inactive * queue is necessarily small, and so dirty * pages would only spend a trivial amount of * time in the inactive queue. Therefore, we * might as well place them directly in the * laundry queue to reduce queuing overhead. * * Calling vm_page_test_dirty() here would * require acquisition of the object's write * lock. However, during a page shortage, * directing dirty pages into the laundry queue * is only an optimization and not a * requirement. Therefore, we simply rely on * the opportunistic updates to the page's dirty * field by the pmap. */ if (page_shortage <= 0) { nqueue = PQ_INACTIVE; ps_delta = 0; } else if (m->dirty == 0) { nqueue = PQ_INACTIVE; ps_delta = act_scan_laundry_weight; } else { nqueue = PQ_LAUNDRY; ps_delta = 1; } new.flags |= PGA_REQUEUE; new.queue = nqueue; } } while (!vm_page_pqstate_commit(m, &old, new)); page_shortage -= ps_delta; } - if (mtx != NULL) { - mtx_unlock(mtx); - mtx = NULL; - } vm_pagequeue_lock(pq); TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q); TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); } static int vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker, vm_page_t m) { vm_page_astate_t as; vm_pagequeue_assert_locked(pq); as = vm_page_astate_load(m); if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0) return (0); vm_page_aflag_set(m, PGA_ENQUEUED); TAILQ_INSERT_BEFORE(marker, m, plinks.q); return (1); } /* * Re-add stuck pages to the inactive queue. We will examine them again * during the next scan. If the queue state of a page has changed since * it was physically removed from the page queue in * vm_pageout_collect_batch(), don't do anything with that page. */ static void vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq, vm_page_t m) { struct vm_pagequeue *pq; vm_page_t marker; int delta; delta = 0; marker = ss->marker; pq = ss->pq; if (m != NULL) { if (vm_batchqueue_insert(bq, m)) return; vm_pagequeue_lock(pq); delta += vm_pageout_reinsert_inactive_page(pq, marker, m); } else vm_pagequeue_lock(pq); while ((m = vm_batchqueue_pop(bq)) != NULL) delta += vm_pageout_reinsert_inactive_page(pq, marker, m); vm_pagequeue_cnt_add(pq, delta); vm_pagequeue_unlock(pq); vm_batchqueue_init(bq); } /* * Attempt to reclaim the requested number of pages from the inactive queue. * Returns true if the shortage was addressed. */ static int vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage) { struct scan_state ss; struct vm_batchqueue rq; - struct mtx *mtx; vm_page_t m, marker; struct vm_pagequeue *pq; vm_object_t object; vm_page_astate_t old, new; int act_delta, addl_page_shortage, deficit, page_shortage, refs; int starting_page_shortage; /* * The addl_page_shortage is an estimate of the number of temporarily * stuck pages in the inactive queue. In other words, the * number of pages from the inactive count that should be * discounted in setting the target for the active queue scan. */ addl_page_shortage = 0; /* * vmd_pageout_deficit counts the number of pages requested in * allocations that failed because of a free page shortage. We assume * that the allocations will be reattempted and thus include the deficit * in our scan target. */ deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit); starting_page_shortage = page_shortage = shortage + deficit; - mtx = NULL; object = NULL; vm_batchqueue_init(&rq); /* * Start scanning the inactive queue for pages that we can free. The * scan will stop when we reach the target or we have scanned the * entire queue. (Note that m->a.act_count is not used to make * decisions for the inactive queue, only for the active queue.) */ marker = &vmd->vmd_markers[PQ_INACTIVE]; pq = &vmd->vmd_pagequeues[PQ_INACTIVE]; vm_pagequeue_lock(pq); vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt); while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) { KASSERT((m->flags & PG_MARKER) == 0, ("marker page %p was dequeued", m)); - vm_page_change_lock(m, &mtx); - -recheck: /* * Don't touch a page that was removed from the queue after the * page queue lock was released. Otherwise, ensure that any * pending queue operations, such as dequeues for wired pages, * are handled. */ if (vm_pageout_defer(m, PQ_INACTIVE, false)) continue; - if (object != m->object) { + /* + * Lock the page's object. + */ + if (object == NULL || object != m->object) { if (object != NULL) VM_OBJECT_WUNLOCK(object); - - /* - * A page's object pointer may be set to NULL before - * the object lock is acquired. - */ object = (vm_object_t)atomic_load_ptr(&m->object); - if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) { - mtx_unlock(mtx); - /* Depends on type-stability. */ - VM_OBJECT_WLOCK(object); - mtx_lock(mtx); - goto recheck; + if (__predict_false(object == NULL)) + /* The page is being freed by another thread. */ + continue; + + /* Depends on type-stability. */ + VM_OBJECT_WLOCK(object); + if (__predict_false(m->object != object)) { + VM_OBJECT_WUNLOCK(object); + object = NULL; + goto reinsert; } } - if (__predict_false(m->object == NULL)) - /* - * The page has been removed from its object. - */ - continue; - KASSERT(m->object == object, ("page %p does not belong to %p", - m, object)); if (vm_page_tryxbusy(m) == 0) { /* * Don't mess with busy pages. Leave them at * the front of the queue. Most likely, they * are being paged out and will leave the * queue shortly after the scan finishes. So, * they ought to be discounted from the * inactive count. */ addl_page_shortage++; goto reinsert; } /* Deferred free of swap space. */ if ((m->a.flags & PGA_SWAP_FREE) != 0) vm_pager_page_unswapped(m); /* - * Re-check for wirings now that we hold the object lock and - * have verified that the page is unbusied. If the page is - * mapped, it may still be wired by pmap lookups. The call to + * Check for wirings now that we hold the object lock and have + * exclusively busied the page. If the page is mapped, it may + * still be wired by pmap lookups. The call to * vm_page_try_remove_all() below atomically checks for such * wirings and removes mappings. If the page is unmapped, the - * wire count is guaranteed not to increase. + * wire count is guaranteed not to increase after this check. */ - if (__predict_false(vm_page_wired(m))) { - vm_page_dequeue_deferred(m); + if (__predict_false(vm_page_wired(m))) goto skip_page; - } /* * Invalid pages can be easily freed. They cannot be * mapped, vm_page_free() asserts this. */ if (vm_page_none_valid(m)) goto free_page; refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0; for (old = vm_page_astate_load(m);;) { /* * Check to see if the page has been removed from the * queue since the first such check. Leave it alone if * so, discarding any references collected by * pmap_ts_referenced(). */ if (__predict_false(_vm_page_queue(old) == PQ_NONE)) goto skip_page; new = old; act_delta = refs; if ((old.flags & PGA_REFERENCED) != 0) { new.flags &= ~PGA_REFERENCED; act_delta++; } if (act_delta == 0) { ; } else if (object->ref_count != 0) { /* * Increase the activation count if the * page was referenced while in the * inactive queue. This makes it less * likely that the page will be returned * prematurely to the inactive queue. */ new.act_count += ACT_ADVANCE + act_delta; if (new.act_count > ACT_MAX) new.act_count = ACT_MAX; new.flags |= PGA_REQUEUE; new.queue = PQ_ACTIVE; if (!vm_page_pqstate_commit(m, &old, new)) continue; VM_CNT_INC(v_reactivated); goto skip_page; } else if ((object->flags & OBJ_DEAD) == 0) { new.queue = PQ_INACTIVE; new.flags |= PGA_REQUEUE; if (!vm_page_pqstate_commit(m, &old, new)) continue; goto skip_page; } break; } /* * If the page appears to be clean at the machine-independent * layer, then remove all of its mappings from the pmap in * anticipation of freeing it. If, however, any of the page's * mappings allow write access, then the page may still be * modified until the last of those mappings are removed. */ if (object->ref_count != 0) { vm_page_test_dirty(m); - if (m->dirty == 0 && !vm_page_try_remove_all(m)) { - vm_page_dequeue_deferred(m); + if (m->dirty == 0 && !vm_page_try_remove_all(m)) goto skip_page; - } } /* * Clean pages can be freed, but dirty pages must be sent back * to the laundry, unless they belong to a dead object. * Requeueing dirty pages from dead objects is pointless, as * they are being paged out and freed by the thread that * destroyed the object. */ if (m->dirty == 0) { free_page: /* - * Because we dequeued the page and have already - * checked for concurrent dequeue and enqueue - * requests, we can safely disassociate the page - * from the inactive queue. + * Now we are guaranteed that no other threads are + * manipulating the page, check for a last-second + * reference that would save it from doom. */ - KASSERT((m->a.flags & PGA_QUEUE_STATE_MASK) == 0, - ("page %p has queue state", m)); + if (vm_pageout_defer(m, PQ_INACTIVE, false)) + goto skip_page; + + /* + * Because we dequeued the page and have already checked + * for pending dequeue and enqueue requests, we can + * safely disassociate the page from the inactive queue + * without holding the queue lock. + */ m->a.queue = PQ_NONE; vm_page_free(m); page_shortage--; continue; } if ((object->flags & OBJ_DEAD) == 0) vm_page_launder(m); skip_page: vm_page_xunbusy(m); continue; reinsert: vm_pageout_reinsert_inactive(&ss, &rq, m); } - if (mtx != NULL) - mtx_unlock(mtx); if (object != NULL) VM_OBJECT_WUNLOCK(object); vm_pageout_reinsert_inactive(&ss, &rq, NULL); vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL); vm_pagequeue_lock(pq); vm_pageout_end_scan(&ss); vm_pagequeue_unlock(pq); VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage); /* * Wake up the laundry thread so that it can perform any needed * laundering. If we didn't meet our target, we're in shortfall and * need to launder more aggressively. If PQ_LAUNDRY is empty and no * swap devices are configured, the laundry thread has no work to do, so * don't bother waking it up. * * The laundry thread uses the number of inactive queue scans elapsed * since the last laundering to determine whether to launder again, so * keep count. */ if (starting_page_shortage > 0) { pq = &vmd->vmd_pagequeues[PQ_LAUNDRY]; vm_pagequeue_lock(pq); if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE && (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) { if (page_shortage > 0) { vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL; VM_CNT_INC(v_pdshortfalls); } else if (vmd->vmd_laundry_request != VM_LAUNDRY_SHORTFALL) vmd->vmd_laundry_request = VM_LAUNDRY_BACKGROUND; wakeup(&vmd->vmd_laundry_request); } vmd->vmd_clean_pages_freed += starting_page_shortage - page_shortage; vm_pagequeue_unlock(pq); } /* * Wakeup the swapout daemon if we didn't free the targeted number of * pages. */ if (page_shortage > 0) vm_swapout_run(); /* * If the inactive queue scan fails repeatedly to meet its * target, kill the largest process. */ vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage); /* * Reclaim pages by swapping out idle processes, if configured to do so. */ vm_swapout_run_idle(); /* * See the description of addl_page_shortage above. */ *addl_shortage = addl_page_shortage + deficit; return (page_shortage <= 0); } static int vm_pageout_oom_vote; /* * The pagedaemon threads randlomly select one to perform the * OOM. Trying to kill processes before all pagedaemons * failed to reach free target is premature. */ static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage, int starting_page_shortage) { int old_vote; if (starting_page_shortage <= 0 || starting_page_shortage != page_shortage) vmd->vmd_oom_seq = 0; else vmd->vmd_oom_seq++; if (vmd->vmd_oom_seq < vm_pageout_oom_seq) { if (vmd->vmd_oom) { vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } return; } /* * Do not follow the call sequence until OOM condition is * cleared. */ vmd->vmd_oom_seq = 0; if (vmd->vmd_oom) return; vmd->vmd_oom = TRUE; old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1); if (old_vote != vm_ndomains - 1) return; /* * The current pagedaemon thread is the last in the quorum to * start OOM. Initiate the selection and signaling of the * victim. */ vm_pageout_oom(VM_OOM_MEM); /* * After one round of OOM terror, recall our vote. On the * next pass, current pagedaemon would vote again if the low * memory condition is still there, due to vmd_oom being * false. */ vmd->vmd_oom = FALSE; atomic_subtract_int(&vm_pageout_oom_vote, 1); } /* * The OOM killer is the page daemon's action of last resort when * memory allocation requests have been stalled for a prolonged period * of time because it cannot reclaim memory. This function computes * the approximate number of physical pages that could be reclaimed if * the specified address space is destroyed. * * Private, anonymous memory owned by the address space is the * principal resource that we expect to recover after an OOM kill. * Since the physical pages mapped by the address space's COW entries * are typically shared pages, they are unlikely to be released and so * they are not counted. * * To get to the point where the page daemon runs the OOM killer, its * efforts to write-back vnode-backed pages may have stalled. This * could be caused by a memory allocation deadlock in the write path * that might be resolved by an OOM kill. Therefore, physical pages * belonging to vnode-backed objects are counted, because they might * be freed without being written out first if the address space holds * the last reference to an unlinked vnode. * * Similarly, physical pages belonging to OBJT_PHYS objects are * counted because the address space might hold the last reference to * the object. */ static long vm_pageout_oom_pagecount(struct vmspace *vmspace) { vm_map_t map; vm_map_entry_t entry; vm_object_t obj; long res; map = &vmspace->vm_map; KASSERT(!map->system_map, ("system map")); sx_assert(&map->lock, SA_LOCKED); res = 0; VM_MAP_ENTRY_FOREACH(entry, map) { if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) continue; obj = entry->object.vm_object; if (obj == NULL) continue; if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 && obj->ref_count != 1) continue; switch (obj->type) { case OBJT_DEFAULT: case OBJT_SWAP: case OBJT_PHYS: case OBJT_VNODE: res += obj->resident_page_count; break; } } return (res); } static int vm_oom_ratelim_last; static int vm_oom_pf_secs = 10; SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0, ""); static struct mtx vm_oom_ratelim_mtx; void vm_pageout_oom(int shortage) { struct proc *p, *bigproc; vm_offset_t size, bigsize; struct thread *td; struct vmspace *vm; int now; bool breakout; /* * For OOM requests originating from vm_fault(), there is a high * chance that a single large process faults simultaneously in * several threads. Also, on an active system running many * processes of middle-size, like buildworld, all of them * could fault almost simultaneously as well. * * To avoid killing too many processes, rate-limit OOMs * initiated by vm_fault() time-outs on the waits for free * pages. */ mtx_lock(&vm_oom_ratelim_mtx); now = ticks; if (shortage == VM_OOM_MEM_PF && (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) { mtx_unlock(&vm_oom_ratelim_mtx); return; } vm_oom_ratelim_last = now; mtx_unlock(&vm_oom_ratelim_mtx); /* * We keep the process bigproc locked once we find it to keep anyone * from messing with it; however, there is a possibility of * deadlock if process B is bigproc and one of its child processes * attempts to propagate a signal to B while we are waiting for A's * lock while walking this list. To avoid this, we don't block on * the process lock but just skip a process if it is already locked. */ bigproc = NULL; bigsize = 0; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); /* * If this is a system, protected or killed process, skip it. */ if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 || p->p_pid == 1 || P_KILLED(p) || (p->p_pid < 48 && swap_pager_avail != 0)) { PROC_UNLOCK(p); continue; } /* * If the process is in a non-running type state, * don't touch it. Check all the threads individually. */ breakout = false; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td) && !TD_IS_SWAPPED(td)) { thread_unlock(td); breakout = true; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get the process size */ vm = vmspace_acquire_ref(p); if (vm == NULL) { PROC_UNLOCK(p); continue; } _PHOLD_LITE(p); PROC_UNLOCK(p); sx_sunlock(&allproc_lock); if (!vm_map_trylock_read(&vm->vm_map)) { vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); continue; } size = vmspace_swap_count(vm); if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF) size += vm_pageout_oom_pagecount(vm); vm_map_unlock_read(&vm->vm_map); vmspace_free(vm); sx_slock(&allproc_lock); /* * If this process is bigger than the biggest one, * remember it. */ if (size > bigsize) { if (bigproc != NULL) PRELE(bigproc); bigproc = p; bigsize = size; } else { PRELE(p); } } sx_sunlock(&allproc_lock); if (bigproc != NULL) { if (vm_panic_on_oom != 0) panic("out of swap space"); PROC_LOCK(bigproc); killproc(bigproc, "out of swap space"); sched_nice(bigproc, PRIO_MIN); _PRELE(bigproc); PROC_UNLOCK(bigproc); } } /* * Signal a free page shortage to subsystems that have registered an event * handler. Reclaim memory from UMA in the event of a severe shortage. * Return true if the free page count should be re-evaluated. */ static bool vm_pageout_lowmem(void) { static int lowmem_ticks = 0; int last; bool ret; ret = false; last = atomic_load_int(&lowmem_ticks); while ((u_int)(ticks - last) / hz >= lowmem_period) { if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0) continue; /* * Decrease registered cache sizes. */ SDT_PROBE0(vm, , , vm__lowmem_scan); EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES); /* * We do this explicitly after the caches have been * drained above. */ uma_reclaim(UMA_RECLAIM_TRIM); ret = true; } /* * Kick off an asynchronous reclaim of cached memory if one of the * page daemons is failing to keep up with demand. Use the "severe" * threshold instead of "min" to ensure that we do not blow away the * caches if a subset of the NUMA domains are depleted by kernel memory * allocations; the domainset iterators automatically skip domains * below the "min" threshold on the first pass. * * UMA reclaim worker has its own rate-limiting mechanism, so don't * worry about kicking it too often. */ if (vm_page_count_severe()) uma_reclaim_wakeup(); return (ret); } static void vm_pageout_worker(void *arg) { struct vm_domain *vmd; u_int ofree; int addl_shortage, domain, shortage; bool target_met; domain = (uintptr_t)arg; vmd = VM_DOMAIN(domain); shortage = 0; target_met = true; /* * XXXKIB It could be useful to bind pageout daemon threads to * the cores belonging to the domain, from which vm_page_array * is allocated. */ KASSERT(vmd->vmd_segs != 0, ("domain without segments")); vmd->vmd_last_active_scan = ticks; /* * The pageout daemon worker is never done, so loop forever. */ while (TRUE) { vm_domain_pageout_lock(vmd); /* * We need to clear wanted before we check the limits. This * prevents races with wakers who will check wanted after they * reach the limit. */ atomic_store_int(&vmd->vmd_pageout_wanted, 0); /* * Might the page daemon need to run again? */ if (vm_paging_needed(vmd, vmd->vmd_free_count)) { /* * Yes. If the scan failed to produce enough free * pages, sleep uninterruptibly for some time in the * hope that the laundry thread will clean some pages. */ vm_domain_pageout_unlock(vmd); if (!target_met) pause("pwait", hz / VM_INACT_SCAN_RATE); } else { /* * No, sleep until the next wakeup or until pages * need to have their reference stats updated. */ if (mtx_sleep(&vmd->vmd_pageout_wanted, vm_domain_pageout_lockptr(vmd), PDROP | PVM, "psleep", hz / VM_INACT_SCAN_RATE) == 0) VM_CNT_INC(v_pdwakeups); } /* Prevent spurious wakeups by ensuring that wanted is set. */ atomic_store_int(&vmd->vmd_pageout_wanted, 1); /* * Use the controller to calculate how many pages to free in * this interval, and scan the inactive queue. If the lowmem * handlers appear to have freed up some pages, subtract the * difference from the inactive queue scan target. */ shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count); if (shortage > 0) { ofree = vmd->vmd_free_count; if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree) shortage -= min(vmd->vmd_free_count - ofree, (u_int)shortage); target_met = vm_pageout_scan_inactive(vmd, shortage, &addl_shortage); } else addl_shortage = 0; /* * Scan the active queue. A positive value for shortage * indicates that we must aggressively deactivate pages to avoid * a shortfall. */ shortage = vm_pageout_active_target(vmd) + addl_shortage; vm_pageout_scan_active(vmd, shortage); } } /* * Initialize basic pageout daemon settings. See the comment above the * definition of vm_domain for some explanation of how these thresholds are * used. */ static void vm_pageout_init_domain(int domain) { struct vm_domain *vmd; struct sysctl_oid *oid; vmd = VM_DOMAIN(domain); vmd->vmd_interrupt_free_min = 2; /* * v_free_reserved needs to include enough for the largest * swap pager structures plus enough for any pv_entry structs * when paging. */ vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE + vmd->vmd_interrupt_free_min; vmd->vmd_free_reserved = vm_pageout_page_count + vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768; vmd->vmd_free_min = vmd->vmd_page_count / 200; vmd->vmd_free_severe = vmd->vmd_free_min / 2; vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved; vmd->vmd_free_min += vmd->vmd_free_reserved; vmd->vmd_free_severe += vmd->vmd_free_reserved; vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2; if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3) vmd->vmd_inactive_target = vmd->vmd_free_count / 3; /* * Set the default wakeup threshold to be 10% below the paging * target. This keeps the steady state out of shortfall. */ vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9; /* * Target amount of memory to move out of the laundry queue during a * background laundering. This is proportional to the amount of system * memory. */ vmd->vmd_background_launder_target = (vmd->vmd_free_target - vmd->vmd_free_min) / 10; /* Initialize the pageout daemon pid controller. */ pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE, vmd->vmd_free_target, PIDCTRL_BOUND, PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO, "pidctrl", CTLFLAG_RD, NULL, ""); pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid)); } static void vm_pageout_init(void) { u_int freecount; int i; /* * Initialize some paging parameters. */ if (vm_cnt.v_page_count < 2000) vm_pageout_page_count = 8; freecount = 0; for (i = 0; i < vm_ndomains; i++) { struct vm_domain *vmd; vm_pageout_init_domain(i); vmd = VM_DOMAIN(i); vm_cnt.v_free_reserved += vmd->vmd_free_reserved; vm_cnt.v_free_target += vmd->vmd_free_target; vm_cnt.v_free_min += vmd->vmd_free_min; vm_cnt.v_inactive_target += vmd->vmd_inactive_target; vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min; vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min; vm_cnt.v_free_severe += vmd->vmd_free_severe; freecount += vmd->vmd_free_count; } /* * Set interval in seconds for active scan. We want to visit each * page at least once every ten minutes. This is to prevent worst * case paging behaviors with stale active LRU. */ if (vm_pageout_update_period == 0) vm_pageout_update_period = 600; if (vm_page_max_user_wired == 0) vm_page_max_user_wired = freecount / 3; } /* * vm_pageout is the high level pageout daemon. */ static void vm_pageout(void) { struct proc *p; struct thread *td; int error, first, i; p = curproc; td = curthread; mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF); swap_pager_swap_init(); for (first = -1, i = 0; i < vm_ndomains; i++) { if (VM_DOMAIN_EMPTY(i)) { if (bootverbose) printf("domain %d empty; skipping pageout\n", i); continue; } if (first == -1) first = i; else { error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i); if (error != 0) panic("starting pageout for domain %d: %d\n", i, error); } error = kthread_add(vm_pageout_laundry_worker, (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i); if (error != 0) panic("starting laundry for domain %d: %d", i, error); } error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma"); if (error != 0) panic("starting uma_reclaim helper, error %d\n", error); snprintf(td->td_name, sizeof(td->td_name), "dom%d", first); vm_pageout_worker((void *)(uintptr_t)first); } /* * Perform an advisory wakeup of the page daemon. */ void pagedaemon_wakeup(int domain) { struct vm_domain *vmd; vmd = VM_DOMAIN(domain); vm_domain_pageout_assert_unlocked(vmd); if (curproc == pageproc) return; if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) { vm_domain_pageout_lock(vmd); atomic_store_int(&vmd->vmd_pageout_wanted, 1); wakeup(&vmd->vmd_pageout_wanted); vm_domain_pageout_unlock(vmd); } } Index: head/sys/vm/vm_swapout.c =================================================================== --- head/sys/vm/vm_swapout.c (revision 356156) +++ head/sys/vm/vm_swapout.c (revision 356157) @@ -1,949 +1,947 @@ /*- * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU) * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005 Yahoo! Technologies Norway AS * All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Authors: Avadis Tevanian, Jr., Michael Wayne Young * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ #include __FBSDID("$FreeBSD$"); #include "opt_kstack_pages.h" #include "opt_kstack_max_pages.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* the kernel process "vm_daemon" */ static void vm_daemon(void); static struct proc *vmproc; static struct kproc_desc vm_kp = { "vmdaemon", vm_daemon, &vmproc }; SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp); static int vm_swap_enabled = 1; static int vm_swap_idle_enabled = 0; SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); /* * Swap_idle_threshold1 is the guaranteed swapped in time for a process */ static int swap_idle_threshold1 = 2; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold1, CTLFLAG_RW, &swap_idle_threshold1, 0, "Guaranteed swapped in time for a process"); /* * Swap_idle_threshold2 is the time that a process can be idle before * it will be swapped out, if idle swapping is enabled. */ static int swap_idle_threshold2 = 10; SYSCTL_INT(_vm, OID_AUTO, swap_idle_threshold2, CTLFLAG_RW, &swap_idle_threshold2, 0, "Time before a process will be swapped out"); static int vm_pageout_req_swapout; /* XXX */ static int vm_daemon_needed; static struct mtx vm_daemon_mtx; /* Allow for use by vm_pageout before vm_daemon is initialized. */ MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF); static int swapped_cnt; static int swap_inprogress; /* Pending swap-ins done outside swapper. */ static int last_swapin; static void swapclear(struct proc *); static int swapout(struct proc *); static void vm_swapout_map_deactivate_pages(vm_map_t, long); static void vm_swapout_object_deactivate(pmap_t, vm_object_t, long); static void swapout_procs(int action); static void vm_req_vmdaemon(int req); static void vm_thread_swapout(struct thread *td); static void vm_swapout_object_deactivate_page(pmap_t pmap, vm_page_t m, bool unmap) { /* * Ignore unreclaimable wired pages. Repeat the check after busying * since a busy holder may wire the page. */ if (vm_page_wired(m) || !vm_page_tryxbusy(m)) return; if (vm_page_wired(m) || !pmap_page_exists_quick(pmap, m)) { vm_page_xunbusy(m); return; } if (!pmap_is_referenced(m)) { - vm_page_lock(m); if (!vm_page_active(m)) (void)vm_page_try_remove_all(m); else if (unmap && vm_page_try_remove_all(m)) vm_page_deactivate(m); - vm_page_unlock(m); } vm_page_xunbusy(m); } /* * vm_swapout_object_deactivate * * Deactivate enough pages to satisfy the inactive target * requirements. * * The object and map must be locked. */ static void vm_swapout_object_deactivate(pmap_t pmap, vm_object_t first_object, long desired) { vm_object_t backing_object, object; vm_page_t m; bool unmap; VM_OBJECT_ASSERT_LOCKED(first_object); if ((first_object->flags & OBJ_FICTITIOUS) != 0) return; for (object = first_object;; object = backing_object) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; VM_OBJECT_ASSERT_LOCKED(object); if ((object->flags & OBJ_UNMANAGED) != 0 || REFCOUNT_COUNT(object->paging_in_progress) > 0) goto unlock_return; unmap = true; if (object->shadow_count > 1) unmap = false; /* * Scan the object's entire memory queue. */ TAILQ_FOREACH(m, &object->memq, listq) { if (pmap_resident_count(pmap) <= desired) goto unlock_return; if (should_yield()) goto unlock_return; vm_swapout_object_deactivate_page(pmap, m, unmap); } if ((backing_object = object->backing_object) == NULL) goto unlock_return; VM_OBJECT_RLOCK(backing_object); if (object != first_object) VM_OBJECT_RUNLOCK(object); } unlock_return: if (object != first_object) VM_OBJECT_RUNLOCK(object); } /* * deactivate some number of pages in a map, try to do it fairly, but * that is really hard to do. */ static void vm_swapout_map_deactivate_pages(vm_map_t map, long desired) { vm_map_entry_t tmpe; vm_object_t obj, bigobj; int nothingwired; if (!vm_map_trylock_read(map)) return; bigobj = NULL; nothingwired = TRUE; /* * first, search out the biggest object, and try to free pages from * that. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) { if (obj->shadow_count <= 1 && (bigobj == NULL || bigobj->resident_page_count < obj->resident_page_count)) { if (bigobj != NULL) VM_OBJECT_RUNLOCK(bigobj); bigobj = obj; } else VM_OBJECT_RUNLOCK(obj); } } if (tmpe->wired_count > 0) nothingwired = FALSE; } if (bigobj != NULL) { vm_swapout_object_deactivate(map->pmap, bigobj, desired); VM_OBJECT_RUNLOCK(bigobj); } /* * Next, hunt around for other pages to deactivate. We actually * do this search sort of wrong -- .text first is not the best idea. */ VM_MAP_ENTRY_FOREACH(tmpe, map) { if (pmap_resident_count(vm_map_pmap(map)) <= desired) break; if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { obj = tmpe->object.vm_object; if (obj != NULL) { VM_OBJECT_RLOCK(obj); vm_swapout_object_deactivate(map->pmap, obj, desired); VM_OBJECT_RUNLOCK(obj); } } } /* * Remove all mappings if a process is swapped out, this will free page * table pages. */ if (desired == 0 && nothingwired) { pmap_remove(vm_map_pmap(map), vm_map_min(map), vm_map_max(map)); } vm_map_unlock_read(map); } /* * Swap out requests */ #define VM_SWAP_NORMAL 1 #define VM_SWAP_IDLE 2 void vm_swapout_run(void) { if (vm_swap_enabled) vm_req_vmdaemon(VM_SWAP_NORMAL); } /* * Idle process swapout -- run once per second when pagedaemons are * reclaiming pages. */ void vm_swapout_run_idle(void) { static long lsec; if (!vm_swap_idle_enabled || time_second == lsec) return; vm_req_vmdaemon(VM_SWAP_IDLE); lsec = time_second; } static void vm_req_vmdaemon(int req) { static int lastrun = 0; mtx_lock(&vm_daemon_mtx); vm_pageout_req_swapout |= req; if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { wakeup(&vm_daemon_needed); lastrun = ticks; } mtx_unlock(&vm_daemon_mtx); } static void vm_daemon(void) { struct rlimit rsslim; struct proc *p; struct thread *td; struct vmspace *vm; int breakout, swapout_flags, tryagain, attempts; #ifdef RACCT uint64_t rsize, ravailable; #endif while (TRUE) { mtx_lock(&vm_daemon_mtx); msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", #ifdef RACCT racct_enable ? hz : 0 #else 0 #endif ); swapout_flags = vm_pageout_req_swapout; vm_pageout_req_swapout = 0; mtx_unlock(&vm_daemon_mtx); if (swapout_flags != 0) { /* * Drain the per-CPU page queue batches as a deadlock * avoidance measure. */ if ((swapout_flags & VM_SWAP_NORMAL) != 0) vm_page_pqbatch_drain(); swapout_procs(swapout_flags); } /* * scan the processes for exceeding their rlimits or if * process is swapped out -- deactivate pages */ tryagain = 0; attempts = 0; again: attempts++; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { vm_pindex_t limit, size; /* * if this is a system process or if we have already * looked at this process, skip it. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) { PROC_UNLOCK(p); continue; } /* * if the process is in a non-running type state, * don't touch it. */ breakout = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!TD_ON_RUNQ(td) && !TD_IS_RUNNING(td) && !TD_IS_SLEEPING(td) && !TD_IS_SUSPENDED(td)) { thread_unlock(td); breakout = 1; break; } thread_unlock(td); } if (breakout) { PROC_UNLOCK(p); continue; } /* * get a limit */ lim_rlimit_proc(p, RLIMIT_RSS, &rsslim); limit = OFF_TO_IDX( qmin(rsslim.rlim_cur, rsslim.rlim_max)); /* * let processes that are swapped out really be * swapped out set the limit to nothing (will force a * swap-out.) */ if ((p->p_flag & P_INMEM) == 0) limit = 0; /* XXX */ vm = vmspace_acquire_ref(p); _PHOLD_LITE(p); PROC_UNLOCK(p); if (vm == NULL) { PRELE(p); continue; } sx_sunlock(&allproc_lock); size = vmspace_resident_count(vm); if (size >= limit) { vm_swapout_map_deactivate_pages( &vm->vm_map, limit); size = vmspace_resident_count(vm); } #ifdef RACCT if (racct_enable) { rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); ravailable = racct_get_available(p, RACCT_RSS); PROC_UNLOCK(p); if (rsize > ravailable) { /* * Don't be overly aggressive; this * might be an innocent process, * and the limit could've been exceeded * by some memory hog. Don't try * to deactivate more than 1/4th * of process' resident set size. */ if (attempts <= 8) { if (ravailable < rsize - (rsize / 4)) { ravailable = rsize - (rsize / 4); } } vm_swapout_map_deactivate_pages( &vm->vm_map, OFF_TO_IDX(ravailable)); /* Update RSS usage after paging out. */ size = vmspace_resident_count(vm); rsize = IDX_TO_OFF(size); PROC_LOCK(p); if (p->p_state == PRS_NORMAL) racct_set(p, RACCT_RSS, rsize); PROC_UNLOCK(p); if (rsize > ravailable) tryagain = 1; } } #endif vmspace_free(vm); sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); if (tryagain != 0 && attempts <= 10) { maybe_yield(); goto again; } } } /* * Allow a thread's kernel stack to be paged out. */ static void vm_thread_swapout(struct thread *td) { vm_object_t ksobj; vm_page_t m; int i, pages; cpu_thread_swapout(td); pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; pmap_qremove(td->td_kstack, pages); VM_OBJECT_WLOCK(ksobj); for (i = 0; i < pages; i++) { m = vm_page_lookup(ksobj, i); if (m == NULL) panic("vm_thread_swapout: kstack already missing?"); vm_page_dirty(m); vm_page_unwire(m, PQ_LAUNDRY); } VM_OBJECT_WUNLOCK(ksobj); } /* * Bring the kernel stack for a specified thread back in. */ static void vm_thread_swapin(struct thread *td, int oom_alloc) { vm_object_t ksobj; vm_page_t ma[KSTACK_MAX_PAGES]; int a, count, i, j, pages, rv; pages = td->td_kstack_pages; ksobj = td->td_kstack_obj; VM_OBJECT_WLOCK(ksobj); (void)vm_page_grab_pages(ksobj, 0, oom_alloc | VM_ALLOC_WIRED, ma, pages); for (i = 0; i < pages;) { vm_page_assert_xbusied(ma[i]); if (vm_page_all_valid(ma[i])) { vm_page_xunbusy(ma[i]); i++; continue; } vm_object_pip_add(ksobj, 1); for (j = i + 1; j < pages; j++) if (vm_page_all_valid(ma[j])) break; rv = vm_pager_has_page(ksobj, ma[i]->pindex, NULL, &a); KASSERT(rv == 1, ("%s: missing page %p", __func__, ma[i])); count = min(a + 1, j - i); rv = vm_pager_get_pages(ksobj, ma + i, count, NULL, NULL); KASSERT(rv == VM_PAGER_OK, ("%s: cannot get kstack for proc %d", __func__, td->td_proc->p_pid)); vm_object_pip_wakeup(ksobj); for (j = i; j < i + count; j++) vm_page_xunbusy(ma[j]); i += count; } VM_OBJECT_WUNLOCK(ksobj); pmap_qenter(td->td_kstack, ma, pages); cpu_thread_swapin(td); } void faultin(struct proc *p) { struct thread *td; int oom_alloc; PROC_LOCK_ASSERT(p, MA_OWNED); /* * If another process is swapping in this process, * just wait until it finishes. */ if (p->p_flag & P_SWAPPINGIN) { while (p->p_flag & P_SWAPPINGIN) msleep(&p->p_flag, &p->p_mtx, PVM, "faultin", 0); return; } if ((p->p_flag & P_INMEM) == 0) { oom_alloc = (p->p_flag & P_WKILLED) != 0 ? VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; /* * Don't let another thread swap process p out while we are * busy swapping it in. */ ++p->p_lock; p->p_flag |= P_SWAPPINGIN; PROC_UNLOCK(p); sx_xlock(&allproc_lock); MPASS(swapped_cnt > 0); swapped_cnt--; if (curthread != &thread0) swap_inprogress++; sx_xunlock(&allproc_lock); /* * We hold no lock here because the list of threads * can not change while all threads in the process are * swapped out. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapin(td, oom_alloc); if (curthread != &thread0) { sx_xlock(&allproc_lock); MPASS(swap_inprogress > 0); swap_inprogress--; last_swapin = ticks; sx_xunlock(&allproc_lock); } PROC_LOCK(p); swapclear(p); p->p_swtick = ticks; /* Allow other threads to swap p out now. */ wakeup(&p->p_flag); --p->p_lock; } } /* * This swapin algorithm attempts to swap-in processes only if there * is enough space for them. Of course, if a process waits for a long * time, it will be swapped in anyway. */ static struct proc * swapper_selector(bool wkilled_only) { struct proc *p, *res; struct thread *td; int ppri, pri, slptime, swtime; sx_assert(&allproc_lock, SA_SLOCKED); if (swapped_cnt == 0) return (NULL); res = NULL; ppri = INT_MIN; FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW || (p->p_flag & (P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != 0) { PROC_UNLOCK(p); continue; } if (p->p_state == PRS_NORMAL && (p->p_flag & P_WKILLED) != 0) { /* * A swapped-out process might have mapped a * large portion of the system's pages as * anonymous memory. There is no other way to * release the memory other than to kill the * process, for which we need to swap it in. */ return (p); } if (wkilled_only) { PROC_UNLOCK(p); continue; } swtime = (ticks - p->p_swtick) / hz; FOREACH_THREAD_IN_PROC(p, td) { /* * An otherwise runnable thread of a process * swapped out has only the TDI_SWAPPED bit set. */ thread_lock(td); if (td->td_inhibitors == TDI_SWAPPED) { slptime = (ticks - td->td_slptick) / hz; pri = swtime + slptime; if ((td->td_flags & TDF_SWAPINREQ) == 0) pri -= p->p_nice * 8; /* * if this thread is higher priority * and there is enough space, then select * this process instead of the previous * selection. */ if (pri > ppri) { res = p; ppri = pri; } } thread_unlock(td); } PROC_UNLOCK(p); } if (res != NULL) PROC_LOCK(res); return (res); } #define SWAPIN_INTERVAL (MAXSLP * hz / 2) /* * Limit swapper to swap in one non-WKILLED process in MAXSLP/2 * interval, assuming that there is: * - at least one domain that is not suffering from a shortage of free memory; * - no parallel swap-ins; * - no other swap-ins in the current SWAPIN_INTERVAL. */ static bool swapper_wkilled_only(void) { return (vm_page_count_min_set(&all_domains) || swap_inprogress > 0 || (u_int)(ticks - last_swapin) < SWAPIN_INTERVAL); } void swapper(void) { struct proc *p; for (;;) { sx_slock(&allproc_lock); p = swapper_selector(swapper_wkilled_only()); sx_sunlock(&allproc_lock); if (p == NULL) { tsleep(&proc0, PVM, "swapin", SWAPIN_INTERVAL); } else { PROC_LOCK_ASSERT(p, MA_OWNED); /* * Another process may be bringing or may have * already brought this process in while we * traverse all threads. Or, this process may * have exited or even being swapped out * again. */ if (p->p_state == PRS_NORMAL && (p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == 0) { faultin(p); } PROC_UNLOCK(p); } } } /* * First, if any processes have been sleeping or stopped for at least * "swap_idle_threshold1" seconds, they are swapped out. If, however, * no such processes exist, then the longest-sleeping or stopped * process is swapped out. Finally, and only as a last resort, if * there are no sleeping or stopped processes, the longest-resident * process is swapped out. */ static void swapout_procs(int action) { struct proc *p; struct thread *td; int slptime; bool didswap, doswap; MPASS((action & (VM_SWAP_NORMAL | VM_SWAP_IDLE)) != 0); didswap = false; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { /* * Filter out not yet fully constructed processes. Do * not swap out held processes. Avoid processes which * are system, exiting, execing, traced, already swapped * out or are in the process of being swapped in or out. */ PROC_LOCK(p); if (p->p_state != PRS_NORMAL || p->p_lock != 0 || (p->p_flag & (P_SYSTEM | P_WEXIT | P_INEXEC | P_STOPPED_SINGLE | P_TRACED | P_SWAPPINGOUT | P_SWAPPINGIN | P_INMEM)) != P_INMEM) { PROC_UNLOCK(p); continue; } /* * Further consideration of this process for swap out * requires iterating over its threads. We release * allproc_lock here so that process creation and * destruction are not blocked while we iterate. * * To later reacquire allproc_lock and resume * iteration over the allproc list, we will first have * to release the lock on the process. We place a * hold on the process so that it remains in the * allproc list while it is unlocked. */ _PHOLD_LITE(p); sx_sunlock(&allproc_lock); /* * Do not swapout a realtime process. * Guarantee swap_idle_threshold1 time in memory. * If the system is under memory stress, or if we are * swapping idle processes >= swap_idle_threshold2, * then swap the process out. */ doswap = true; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); slptime = (ticks - td->td_slptick) / hz; if (PRI_IS_REALTIME(td->td_pri_class) || slptime < swap_idle_threshold1 || !thread_safetoswapout(td) || ((action & VM_SWAP_NORMAL) == 0 && slptime < swap_idle_threshold2)) doswap = false; thread_unlock(td); if (!doswap) break; } if (doswap && swapout(p) == 0) didswap = true; PROC_UNLOCK(p); if (didswap) { sx_xlock(&allproc_lock); swapped_cnt++; sx_downgrade(&allproc_lock); } else sx_slock(&allproc_lock); PRELE(p); } sx_sunlock(&allproc_lock); /* * If we swapped something out, and another process needed memory, * then wakeup the sched process. */ if (didswap) wakeup(&proc0); } static void swapclear(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); td->td_flags |= TDF_INMEM; td->td_flags &= ~TDF_SWAPINREQ; TD_CLR_SWAPPED(td); if (TD_CAN_RUN(td)) { if (setrunnable(td, 0)) { #ifdef INVARIANTS /* * XXX: We just cleared TDI_SWAPPED * above and set TDF_INMEM, so this * should never happen. */ panic("not waking up swapper"); #endif } } else thread_unlock(td); } p->p_flag &= ~(P_SWAPPINGIN | P_SWAPPINGOUT); p->p_flag |= P_INMEM; } static int swapout(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); /* * The states of this process and its threads may have changed * by now. Assuming that there is only one pageout daemon thread, * this process should still be in memory. */ KASSERT((p->p_flag & (P_INMEM | P_SWAPPINGOUT | P_SWAPPINGIN)) == P_INMEM, ("swapout: lost a swapout race?")); /* * Remember the resident count. */ p->p_vmspace->vm_swrss = vmspace_resident_count(p->p_vmspace); /* * Check and mark all threads before we proceed. */ p->p_flag &= ~P_INMEM; p->p_flag |= P_SWAPPINGOUT; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (!thread_safetoswapout(td)) { thread_unlock(td); swapclear(p); return (EBUSY); } td->td_flags &= ~TDF_INMEM; TD_SET_SWAPPED(td); thread_unlock(td); } td = FIRST_THREAD_IN_PROC(p); ++td->td_ru.ru_nswap; PROC_UNLOCK(p); /* * This list is stable because all threads are now prevented from * running. The list is only modified in the context of a running * thread in this process. */ FOREACH_THREAD_IN_PROC(p, td) vm_thread_swapout(td); PROC_LOCK(p); p->p_flag &= ~P_SWAPPINGOUT; p->p_swtick = ticks; return (0); }