Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dbuf.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dbuf.c (revision 337671) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dbuf.c (revision 337672) @@ -1,3905 +1,3915 @@ /* * 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 2011 Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2012, 2018 by Delphix. All rights reserved. * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. * Copyright (c) 2013, Joyent, Inc. All rights reserved. * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. * Copyright (c) 2014 Integros [integros.com] */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); #ifndef __lint extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func_sync, dmu_buf_evict_func_t *evict_func_async, dmu_buf_t **clear_on_evict_dbufp); #endif /* ! __lint */ /* * Global data structures and functions for the dbuf cache. */ static kmem_cache_t *dbuf_kmem_cache; static taskq_t *dbu_evict_taskq; static kthread_t *dbuf_cache_evict_thread; static kmutex_t dbuf_evict_lock; static kcondvar_t dbuf_evict_cv; static boolean_t dbuf_evict_thread_exit; /* * There are two dbuf caches; each dbuf can only be in one of them at a time. * * 1. Cache of metadata dbufs, to help make read-heavy administrative commands * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs * that represent the metadata that describes filesystems/snapshots/ * bookmarks/properties/etc. We only evict from this cache when we export a * pool, to short-circuit as much I/O as possible for all administrative * commands that need the metadata. There is no eviction policy for this * cache, because we try to only include types in it which would occupy a * very small amount of space per object but create a large impact on the * performance of these commands. Instead, after it reaches a maximum size * (which should only happen on very small memory systems with a very large * number of filesystem objects), we stop taking new dbufs into the * metadata cache, instead putting them in the normal dbuf cache. * * 2. LRU cache of dbufs. The "dbuf cache" maintains a list of dbufs that * are not currently held but have been recently released. These dbufs * are not eligible for arc eviction until they are aged out of the cache. * Dbufs that are aged out of the cache will be immediately destroyed and * become eligible for arc eviction. * * Dbufs are added to these caches once the last hold is released. If a dbuf is * later accessed and still exists in the dbuf cache, then it will be removed * from the cache and later re-added to the head of the cache. * * If a given dbuf meets the requirements for the metadata cache, it will go * there, otherwise it will be considered for the generic LRU dbuf cache. The * caches and the refcounts tracking their sizes are stored in an array indexed * by those caches' matching enum values (from dbuf_cached_state_t). */ typedef struct dbuf_cache { multilist_t *cache; refcount_t size; } dbuf_cache_t; dbuf_cache_t dbuf_caches[DB_CACHE_MAX]; /* Size limits for the caches */ uint64_t dbuf_cache_max_bytes = 0; uint64_t dbuf_metadata_cache_max_bytes = 0; /* Set the default sizes of the caches to log2 fraction of arc size */ int dbuf_cache_shift = 5; int dbuf_metadata_cache_shift = 6; /* * For diagnostic purposes, this is incremented whenever we can't add * something to the metadata cache because it's full, and instead put * the data in the regular dbuf cache. */ uint64_t dbuf_metadata_cache_overflow; /* * The LRU dbuf cache uses a three-stage eviction policy: * - A low water marker designates when the dbuf eviction thread * should stop evicting from the dbuf cache. * - When we reach the maximum size (aka mid water mark), we * signal the eviction thread to run. * - The high water mark indicates when the eviction thread * is unable to keep up with the incoming load and eviction must * happen in the context of the calling thread. * * The dbuf cache: * (max size) * low water mid water hi water * +----------------------------------------+----------+----------+ * | | | | * | | | | * | | | | * | | | | * +----------------------------------------+----------+----------+ * stop signal evict * evicting eviction directly * thread * * The high and low water marks indicate the operating range for the eviction * thread. The low water mark is, by default, 90% of the total size of the * cache and the high water mark is at 110% (both of these percentages can be * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, * respectively). The eviction thread will try to ensure that the cache remains * within this range by waking up every second and checking if the cache is * above the low water mark. The thread can also be woken up by callers adding * elements into the cache if the cache is larger than the mid water (i.e max * cache size). Once the eviction thread is woken up and eviction is required, * it will continue evicting buffers until it's able to reduce the cache size * to the low water mark. If the cache size continues to grow and hits the high * water mark, then callers adding elments to the cache will begin to evict * directly from the cache until the cache is no longer above the high water * mark. */ /* * The percentage above and below the maximum cache size. */ uint_t dbuf_cache_hiwater_pct = 10; uint_t dbuf_cache_lowater_pct = 10; SYSCTL_DECL(_vfs_zfs); SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_cache_max_bytes, CTLFLAG_RWTUN, &dbuf_cache_max_bytes, 0, "dbuf cache size in bytes"); SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_cache_shift, CTLFLAG_RDTUN, &dbuf_cache_shift, 0, "dbuf cache size as log2 fraction of ARC"); SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_hiwater_pct, CTLFLAG_RWTUN, &dbuf_cache_hiwater_pct, 0, "max percents above the dbuf cache size"); SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_lowater_pct, CTLFLAG_RWTUN, &dbuf_cache_lowater_pct, 0, "max percents below the dbuf cache size"); /* ARGSUSED */ static int dbuf_cons(void *vdb, void *unused, int kmflag) { dmu_buf_impl_t *db = vdb; bzero(db, sizeof (dmu_buf_impl_t)); mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); multilist_link_init(&db->db_cache_link); refcount_create(&db->db_holds); return (0); } /* ARGSUSED */ static void dbuf_dest(void *vdb, void *unused) { dmu_buf_impl_t *db = vdb; mutex_destroy(&db->db_mtx); cv_destroy(&db->db_changed); ASSERT(!multilist_link_active(&db->db_cache_link)); refcount_destroy(&db->db_holds); } /* * dbuf hash table routines */ static dbuf_hash_table_t dbuf_hash_table; static uint64_t dbuf_hash_count; /* * We use Cityhash for this. It's fast, and has good hash properties without * requiring any large static buffers. */ static uint64_t dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) { return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid)); } #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ ((dbuf)->db.db_object == (obj) && \ (dbuf)->db_objset == (os) && \ (dbuf)->db_level == (level) && \ (dbuf)->db_blkid == (blkid)) dmu_buf_impl_t * dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) { dbuf_hash_table_t *h = &dbuf_hash_table; uint64_t hv = dbuf_hash(os, obj, level, blkid); uint64_t idx = hv & h->hash_table_mask; dmu_buf_impl_t *db; mutex_enter(DBUF_HASH_MUTEX(h, idx)); for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { if (DBUF_EQUAL(db, os, obj, level, blkid)) { mutex_enter(&db->db_mtx); if (db->db_state != DB_EVICTING) { mutex_exit(DBUF_HASH_MUTEX(h, idx)); return (db); } mutex_exit(&db->db_mtx); } } mutex_exit(DBUF_HASH_MUTEX(h, idx)); return (NULL); } static dmu_buf_impl_t * dbuf_find_bonus(objset_t *os, uint64_t object) { dnode_t *dn; dmu_buf_impl_t *db = NULL; if (dnode_hold(os, object, FTAG, &dn) == 0) { rw_enter(&dn->dn_struct_rwlock, RW_READER); if (dn->dn_bonus != NULL) { db = dn->dn_bonus; mutex_enter(&db->db_mtx); } rw_exit(&dn->dn_struct_rwlock); dnode_rele(dn, FTAG); } return (db); } /* * Insert an entry into the hash table. If there is already an element * equal to elem in the hash table, then the already existing element * will be returned and the new element will not be inserted. * Otherwise returns NULL. */ static dmu_buf_impl_t * dbuf_hash_insert(dmu_buf_impl_t *db) { dbuf_hash_table_t *h = &dbuf_hash_table; objset_t *os = db->db_objset; uint64_t obj = db->db.db_object; int level = db->db_level; uint64_t blkid = db->db_blkid; uint64_t hv = dbuf_hash(os, obj, level, blkid); uint64_t idx = hv & h->hash_table_mask; dmu_buf_impl_t *dbf; mutex_enter(DBUF_HASH_MUTEX(h, idx)); for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { mutex_enter(&dbf->db_mtx); if (dbf->db_state != DB_EVICTING) { mutex_exit(DBUF_HASH_MUTEX(h, idx)); return (dbf); } mutex_exit(&dbf->db_mtx); } } mutex_enter(&db->db_mtx); db->db_hash_next = h->hash_table[idx]; h->hash_table[idx] = db; mutex_exit(DBUF_HASH_MUTEX(h, idx)); atomic_inc_64(&dbuf_hash_count); return (NULL); } /* * Remove an entry from the hash table. It must be in the EVICTING state. */ static void dbuf_hash_remove(dmu_buf_impl_t *db) { dbuf_hash_table_t *h = &dbuf_hash_table; uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object, db->db_level, db->db_blkid); uint64_t idx = hv & h->hash_table_mask; dmu_buf_impl_t *dbf, **dbp; /* * We musn't hold db_mtx to maintain lock ordering: * DBUF_HASH_MUTEX > db_mtx. */ ASSERT(refcount_is_zero(&db->db_holds)); ASSERT(db->db_state == DB_EVICTING); ASSERT(!MUTEX_HELD(&db->db_mtx)); mutex_enter(DBUF_HASH_MUTEX(h, idx)); dbp = &h->hash_table[idx]; while ((dbf = *dbp) != db) { dbp = &dbf->db_hash_next; ASSERT(dbf != NULL); } *dbp = db->db_hash_next; db->db_hash_next = NULL; mutex_exit(DBUF_HASH_MUTEX(h, idx)); atomic_dec_64(&dbuf_hash_count); } typedef enum { DBVU_EVICTING, DBVU_NOT_EVICTING } dbvu_verify_type_t; static void dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) { #ifdef ZFS_DEBUG int64_t holds; if (db->db_user == NULL) return; /* Only data blocks support the attachment of user data. */ ASSERT(db->db_level == 0); /* Clients must resolve a dbuf before attaching user data. */ ASSERT(db->db.db_data != NULL); ASSERT3U(db->db_state, ==, DB_CACHED); holds = refcount_count(&db->db_holds); if (verify_type == DBVU_EVICTING) { /* * Immediate eviction occurs when holds == dirtycnt. * For normal eviction buffers, holds is zero on * eviction, except when dbuf_fix_old_data() calls * dbuf_clear_data(). However, the hold count can grow * during eviction even though db_mtx is held (see * dmu_bonus_hold() for an example), so we can only * test the generic invariant that holds >= dirtycnt. */ ASSERT3U(holds, >=, db->db_dirtycnt); } else { if (db->db_user_immediate_evict == TRUE) ASSERT3U(holds, >=, db->db_dirtycnt); else ASSERT3U(holds, >, 0); } #endif } static void dbuf_evict_user(dmu_buf_impl_t *db) { dmu_buf_user_t *dbu = db->db_user; ASSERT(MUTEX_HELD(&db->db_mtx)); if (dbu == NULL) return; dbuf_verify_user(db, DBVU_EVICTING); db->db_user = NULL; #ifdef ZFS_DEBUG if (dbu->dbu_clear_on_evict_dbufp != NULL) *dbu->dbu_clear_on_evict_dbufp = NULL; #endif /* * There are two eviction callbacks - one that we call synchronously * and one that we invoke via a taskq. The async one is useful for * avoiding lock order reversals and limiting stack depth. * * Note that if we have a sync callback but no async callback, * it's likely that the sync callback will free the structure * containing the dbu. In that case we need to take care to not * dereference dbu after calling the sync evict func. */ boolean_t has_async = (dbu->dbu_evict_func_async != NULL); if (dbu->dbu_evict_func_sync != NULL) dbu->dbu_evict_func_sync(dbu); if (has_async) { taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, dbu, 0, &dbu->dbu_tqent); } } boolean_t dbuf_is_metadata(dmu_buf_impl_t *db) { if (db->db_level > 0) { return (B_TRUE); } else { boolean_t is_metadata; DB_DNODE_ENTER(db); is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); DB_DNODE_EXIT(db); return (is_metadata); } } /* * This returns whether this dbuf should be stored in the metadata cache, which * is based on whether it's from one of the dnode types that store data related * to traversing dataset hierarchies. */ static boolean_t dbuf_include_in_metadata_cache(dmu_buf_impl_t *db) { DB_DNODE_ENTER(db); dmu_object_type_t type = DB_DNODE(db)->dn_type; DB_DNODE_EXIT(db); /* Check if this dbuf is one of the types we care about */ if (DMU_OT_IS_METADATA_CACHED(type)) { /* If we hit this, then we set something up wrong in dmu_ot */ ASSERT(DMU_OT_IS_METADATA(type)); /* * Sanity check for small-memory systems: don't allocate too * much memory for this purpose. */ if (refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size) > dbuf_metadata_cache_max_bytes) { dbuf_metadata_cache_overflow++; DTRACE_PROBE1(dbuf__metadata__cache__overflow, dmu_buf_impl_t *, db); return (B_FALSE); } return (B_TRUE); } return (B_FALSE); } /* * This function *must* return indices evenly distributed between all * sublists of the multilist. This is needed due to how the dbuf eviction * code is laid out; dbuf_evict_thread() assumes dbufs are evenly * distributed between all sublists and uses this assumption when * deciding which sublist to evict from and how much to evict from it. */ unsigned int dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) { dmu_buf_impl_t *db = obj; /* * The assumption here, is the hash value for a given * dmu_buf_impl_t will remain constant throughout it's lifetime * (i.e. it's objset, object, level and blkid fields don't change). * Thus, we don't need to store the dbuf's sublist index * on insertion, as this index can be recalculated on removal. * * Also, the low order bits of the hash value are thought to be * distributed evenly. Otherwise, in the case that the multilist * has a power of two number of sublists, each sublists' usage * would not be evenly distributed. */ return (dbuf_hash(db->db_objset, db->db.db_object, db->db_level, db->db_blkid) % multilist_get_num_sublists(ml)); } static inline boolean_t dbuf_cache_above_hiwater(void) { uint64_t dbuf_cache_hiwater_bytes = (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100; return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes); } static inline boolean_t dbuf_cache_above_lowater(void) { uint64_t dbuf_cache_lowater_bytes = (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100; return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > dbuf_cache_max_bytes - dbuf_cache_lowater_bytes); } /* * Evict the oldest eligible dbuf from the dbuf cache. */ static void dbuf_evict_one(void) { int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache); multilist_sublist_t *mls = multilist_sublist_lock( dbuf_caches[DB_DBUF_CACHE].cache, idx); ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); dmu_buf_impl_t *db = multilist_sublist_tail(mls); while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { db = multilist_sublist_prev(mls, db); } DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, multilist_sublist_t *, mls); if (db != NULL) { multilist_sublist_remove(mls, db); multilist_sublist_unlock(mls); (void) refcount_remove_many(&dbuf_caches[DB_DBUF_CACHE].size, db->db.db_size, db); ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE); db->db_caching_status = DB_NO_CACHE; dbuf_destroy(db); } else { multilist_sublist_unlock(mls); } } /* * The dbuf evict thread is responsible for aging out dbufs from the * cache. Once the cache has reached it's maximum size, dbufs are removed * and destroyed. The eviction thread will continue running until the size * of the dbuf cache is at or below the maximum size. Once the dbuf is aged * out of the cache it is destroyed and becomes eligible for arc eviction. */ /* ARGSUSED */ static void dbuf_evict_thread(void *unused __unused) { callb_cpr_t cpr; CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); mutex_enter(&dbuf_evict_lock); while (!dbuf_evict_thread_exit) { while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { CALLB_CPR_SAFE_BEGIN(&cpr); (void) cv_timedwait_hires(&dbuf_evict_cv, &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); } mutex_exit(&dbuf_evict_lock); /* * Keep evicting as long as we're above the low water mark * for the cache. We do this without holding the locks to * minimize lock contention. */ while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { dbuf_evict_one(); } mutex_enter(&dbuf_evict_lock); } dbuf_evict_thread_exit = B_FALSE; cv_broadcast(&dbuf_evict_cv); CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ thread_exit(); } /* * Wake up the dbuf eviction thread if the dbuf cache is at its max size. * If the dbuf cache is at its high water mark, then evict a dbuf from the * dbuf cache using the callers context. */ static void dbuf_evict_notify(void) { /* * We check if we should evict without holding the dbuf_evict_lock, * because it's OK to occasionally make the wrong decision here, * and grabbing the lock results in massive lock contention. */ if (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > dbuf_cache_max_bytes) { if (dbuf_cache_above_hiwater()) dbuf_evict_one(); cv_signal(&dbuf_evict_cv); } } void dbuf_init(void) { uint64_t hsize = 1ULL << 16; dbuf_hash_table_t *h = &dbuf_hash_table; int i; /* * The hash table is big enough to fill all of physical memory * with an average 4K block size. The table will take up * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). */ while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) hsize <<= 1; retry: h->hash_table_mask = hsize - 1; h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); if (h->hash_table == NULL) { /* XXX - we should really return an error instead of assert */ ASSERT(hsize > (1ULL << 10)); hsize >>= 1; goto retry; } dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", sizeof (dmu_buf_impl_t), 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); for (i = 0; i < DBUF_MUTEXES; i++) mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); dbuf_stats_init(h); /* * Setup the parameters for the dbuf caches. We set the sizes of the * dbuf cache and the metadata cache to 1/32nd and 1/16th (default) * of the size of the ARC, respectively. If the values are set in * /etc/system and they're not greater than the size of the ARC, then * we honor that value. */ if (dbuf_cache_max_bytes == 0 || dbuf_cache_max_bytes >= arc_max_bytes()) { dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift; } if (dbuf_metadata_cache_max_bytes == 0 || dbuf_metadata_cache_max_bytes >= arc_max_bytes()) { dbuf_metadata_cache_max_bytes = arc_max_bytes() >> dbuf_metadata_cache_shift; } /* * All entries are queued via taskq_dispatch_ent(), so min/maxalloc * configuration is not required. */ dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { dbuf_caches[dcs].cache = multilist_create(sizeof (dmu_buf_impl_t), offsetof(dmu_buf_impl_t, db_cache_link), dbuf_cache_multilist_index_func); refcount_create(&dbuf_caches[dcs].size); } dbuf_evict_thread_exit = B_FALSE; mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, NULL, 0, &p0, TS_RUN, minclsyspri); } void dbuf_fini(void) { dbuf_hash_table_t *h = &dbuf_hash_table; int i; dbuf_stats_destroy(); for (i = 0; i < DBUF_MUTEXES; i++) mutex_destroy(&h->hash_mutexes[i]); kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); kmem_cache_destroy(dbuf_kmem_cache); taskq_destroy(dbu_evict_taskq); mutex_enter(&dbuf_evict_lock); dbuf_evict_thread_exit = B_TRUE; while (dbuf_evict_thread_exit) { cv_signal(&dbuf_evict_cv); cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); } mutex_exit(&dbuf_evict_lock); mutex_destroy(&dbuf_evict_lock); cv_destroy(&dbuf_evict_cv); for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { refcount_destroy(&dbuf_caches[dcs].size); multilist_destroy(dbuf_caches[dcs].cache); } } /* * Other stuff. */ #ifdef ZFS_DEBUG static void dbuf_verify(dmu_buf_impl_t *db) { dnode_t *dn; dbuf_dirty_record_t *dr; ASSERT(MUTEX_HELD(&db->db_mtx)); if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) return; ASSERT(db->db_objset != NULL); DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (dn == NULL) { ASSERT(db->db_parent == NULL); ASSERT(db->db_blkptr == NULL); } else { ASSERT3U(db->db.db_object, ==, dn->dn_object); ASSERT3P(db->db_objset, ==, dn->dn_objset); ASSERT3U(db->db_level, <, dn->dn_nlevels); ASSERT(db->db_blkid == DMU_BONUS_BLKID || db->db_blkid == DMU_SPILL_BLKID || !avl_is_empty(&dn->dn_dbufs)); } if (db->db_blkid == DMU_BONUS_BLKID) { ASSERT(dn != NULL); ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); } else if (db->db_blkid == DMU_SPILL_BLKID) { ASSERT(dn != NULL); ASSERT0(db->db.db_offset); } else { ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); } for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) ASSERT(dr->dr_dbuf == db); for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) ASSERT(dr->dr_dbuf == db); /* * We can't assert that db_size matches dn_datablksz because it * can be momentarily different when another thread is doing * dnode_set_blksz(). */ if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { dr = db->db_data_pending; /* * It should only be modified in syncing context, so * make sure we only have one copy of the data. */ ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); } /* verify db->db_blkptr */ if (db->db_blkptr) { if (db->db_parent == dn->dn_dbuf) { /* db is pointed to by the dnode */ /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) ASSERT(db->db_parent == NULL); else ASSERT(db->db_parent != NULL); if (db->db_blkid != DMU_SPILL_BLKID) ASSERT3P(db->db_blkptr, ==, &dn->dn_phys->dn_blkptr[db->db_blkid]); } else { /* db is pointed to by an indirect block */ int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); ASSERT3U(db->db_parent->db.db_object, ==, db->db.db_object); /* * dnode_grow_indblksz() can make this fail if we don't * have the struct_rwlock. XXX indblksz no longer * grows. safe to do this now? */ if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { ASSERT3P(db->db_blkptr, ==, ((blkptr_t *)db->db_parent->db.db_data + db->db_blkid % epb)); } } } if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && (db->db_buf == NULL || db->db_buf->b_data) && db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && db->db_state != DB_FILL && !dn->dn_free_txg) { /* * If the blkptr isn't set but they have nonzero data, * it had better be dirty, otherwise we'll lose that * data when we evict this buffer. * * There is an exception to this rule for indirect blocks; in * this case, if the indirect block is a hole, we fill in a few * fields on each of the child blocks (importantly, birth time) * to prevent hole birth times from being lost when you * partially fill in a hole. */ if (db->db_dirtycnt == 0) { if (db->db_level == 0) { uint64_t *buf = db->db.db_data; int i; for (i = 0; i < db->db.db_size >> 3; i++) { ASSERT(buf[i] == 0); } } else { blkptr_t *bps = db->db.db_data; ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, db->db.db_size); /* * We want to verify that all the blkptrs in the * indirect block are holes, but we may have * automatically set up a few fields for them. * We iterate through each blkptr and verify * they only have those fields set. */ for (int i = 0; i < db->db.db_size / sizeof (blkptr_t); i++) { blkptr_t *bp = &bps[i]; ASSERT(ZIO_CHECKSUM_IS_ZERO( &bp->blk_cksum)); ASSERT( DVA_IS_EMPTY(&bp->blk_dva[0]) && DVA_IS_EMPTY(&bp->blk_dva[1]) && DVA_IS_EMPTY(&bp->blk_dva[2])); ASSERT0(bp->blk_fill); ASSERT0(bp->blk_pad[0]); ASSERT0(bp->blk_pad[1]); ASSERT(!BP_IS_EMBEDDED(bp)); ASSERT(BP_IS_HOLE(bp)); ASSERT0(bp->blk_phys_birth); } } } } DB_DNODE_EXIT(db); } #endif static void dbuf_clear_data(dmu_buf_impl_t *db) { ASSERT(MUTEX_HELD(&db->db_mtx)); dbuf_evict_user(db); ASSERT3P(db->db_buf, ==, NULL); db->db.db_data = NULL; if (db->db_state != DB_NOFILL) db->db_state = DB_UNCACHED; } static void dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) { ASSERT(MUTEX_HELD(&db->db_mtx)); ASSERT(buf != NULL); db->db_buf = buf; ASSERT(buf->b_data != NULL); db->db.db_data = buf->b_data; } /* * Loan out an arc_buf for read. Return the loaned arc_buf. */ arc_buf_t * dbuf_loan_arcbuf(dmu_buf_impl_t *db) { arc_buf_t *abuf; ASSERT(db->db_blkid != DMU_BONUS_BLKID); mutex_enter(&db->db_mtx); if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { int blksz = db->db.db_size; spa_t *spa = db->db_objset->os_spa; mutex_exit(&db->db_mtx); abuf = arc_loan_buf(spa, B_FALSE, blksz); bcopy(db->db.db_data, abuf->b_data, blksz); } else { abuf = db->db_buf; arc_loan_inuse_buf(abuf, db); db->db_buf = NULL; dbuf_clear_data(db); mutex_exit(&db->db_mtx); } return (abuf); } /* * Calculate which level n block references the data at the level 0 offset * provided. */ uint64_t dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) { if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { /* * The level n blkid is equal to the level 0 blkid divided by * the number of level 0s in a level n block. * * The level 0 blkid is offset >> datablkshift = * offset / 2^datablkshift. * * The number of level 0s in a level n is the number of block * pointers in an indirect block, raised to the power of level. * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). * * Thus, the level n blkid is: offset / * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) * = offset / 2^(datablkshift + level * * (indblkshift - SPA_BLKPTRSHIFT)) * = offset >> (datablkshift + level * * (indblkshift - SPA_BLKPTRSHIFT)) */ return (offset >> (dn->dn_datablkshift + level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); } else { ASSERT3U(offset, <, dn->dn_datablksz); return (0); } } static void dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, arc_buf_t *buf, void *vdb) { dmu_buf_impl_t *db = vdb; mutex_enter(&db->db_mtx); ASSERT3U(db->db_state, ==, DB_READ); /* * All reads are synchronous, so we must have a hold on the dbuf */ ASSERT(refcount_count(&db->db_holds) > 0); ASSERT(db->db_buf == NULL); ASSERT(db->db.db_data == NULL); if (buf == NULL) { /* i/o error */ ASSERT(zio == NULL || zio->io_error != 0); ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT3P(db->db_buf, ==, NULL); db->db_state = DB_UNCACHED; } else if (db->db_level == 0 && db->db_freed_in_flight) { /* freed in flight */ ASSERT(zio == NULL || zio->io_error == 0); if (buf == NULL) { buf = arc_alloc_buf(db->db_objset->os_spa, db, DBUF_GET_BUFC_TYPE(db), db->db.db_size); } arc_release(buf, db); bzero(buf->b_data, db->db.db_size); arc_buf_freeze(buf); db->db_freed_in_flight = FALSE; dbuf_set_data(db, buf); db->db_state = DB_CACHED; } else { /* success */ ASSERT(zio == NULL || zio->io_error == 0); dbuf_set_data(db, buf); db->db_state = DB_CACHED; } cv_broadcast(&db->db_changed); dbuf_rele_and_unlock(db, NULL, B_FALSE); } static void dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) { dnode_t *dn; zbookmark_phys_t zb; arc_flags_t aflags = ARC_FLAG_NOWAIT; DB_DNODE_ENTER(db); dn = DB_DNODE(db); ASSERT(!refcount_is_zero(&db->db_holds)); /* We need the struct_rwlock to prevent db_blkptr from changing. */ ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); ASSERT(MUTEX_HELD(&db->db_mtx)); ASSERT(db->db_state == DB_UNCACHED); ASSERT(db->db_buf == NULL); if (db->db_blkid == DMU_BONUS_BLKID) { /* * The bonus length stored in the dnode may be less than * the maximum available space in the bonus buffer. */ int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); ASSERT3U(bonuslen, <=, db->db.db_size); db->db.db_data = zio_buf_alloc(max_bonuslen); arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); if (bonuslen < max_bonuslen) bzero(db->db.db_data, max_bonuslen); if (bonuslen) bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); DB_DNODE_EXIT(db); db->db_state = DB_CACHED; mutex_exit(&db->db_mtx); return; } /* * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() * processes the delete record and clears the bp while we are waiting * for the dn_mtx (resulting in a "no" from block_freed). */ if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || BP_IS_HOLE(db->db_blkptr)))) { arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, db->db.db_size)); bzero(db->db.db_data, db->db.db_size); if (db->db_blkptr != NULL && db->db_level > 0 && BP_IS_HOLE(db->db_blkptr) && db->db_blkptr->blk_birth != 0) { blkptr_t *bps = db->db.db_data; for (int i = 0; i < ((1 << DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); i++) { blkptr_t *bp = &bps[i]; ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 1 << dn->dn_indblkshift); BP_SET_LSIZE(bp, BP_GET_LEVEL(db->db_blkptr) == 1 ? dn->dn_datablksz : BP_GET_LSIZE(db->db_blkptr)); BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); BP_SET_LEVEL(bp, BP_GET_LEVEL(db->db_blkptr) - 1); BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); } } DB_DNODE_EXIT(db); db->db_state = DB_CACHED; mutex_exit(&db->db_mtx); return; } DB_DNODE_EXIT(db); db->db_state = DB_READ; mutex_exit(&db->db_mtx); if (DBUF_IS_L2CACHEABLE(db)) aflags |= ARC_FLAG_L2CACHE; SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, db->db.db_object, db->db_level, db->db_blkid); dbuf_add_ref(db, NULL); (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, &aflags, &zb); } /* * This is our just-in-time copy function. It makes a copy of buffers that * have been modified in a previous transaction group before we access them in * the current active group. * * This function is used in three places: when we are dirtying a buffer for the * first time in a txg, when we are freeing a range in a dnode that includes * this buffer, and when we are accessing a buffer which was received compressed * and later referenced in a WRITE_BYREF record. * * Note that when we are called from dbuf_free_range() we do not put a hold on * the buffer, we just traverse the active dbuf list for the dnode. */ static void dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) { dbuf_dirty_record_t *dr = db->db_last_dirty; ASSERT(MUTEX_HELD(&db->db_mtx)); ASSERT(db->db.db_data != NULL); ASSERT(db->db_level == 0); ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); if (dr == NULL || (dr->dt.dl.dr_data != ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) return; /* * If the last dirty record for this dbuf has not yet synced * and its referencing the dbuf data, either: * reset the reference to point to a new copy, * or (if there a no active holders) * just null out the current db_data pointer. */ ASSERT(dr->dr_txg >= txg - 2); if (db->db_blkid == DMU_BONUS_BLKID) { /* Note that the data bufs here are zio_bufs */ dnode_t *dn = DB_DNODE(db); int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); dr->dt.dl.dr_data = zio_buf_alloc(bonuslen); arc_space_consume(bonuslen, ARC_SPACE_BONUS); bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen); } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { int size = arc_buf_size(db->db_buf); arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); spa_t *spa = db->db_objset->os_spa; enum zio_compress compress_type = arc_get_compression(db->db_buf); if (compress_type == ZIO_COMPRESS_OFF) { dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); } else { ASSERT3U(type, ==, ARC_BUFC_DATA); dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, size, arc_buf_lsize(db->db_buf), compress_type); } bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); } else { db->db_buf = NULL; dbuf_clear_data(db); } } int dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) { int err = 0; boolean_t prefetch; dnode_t *dn; /* * We don't have to hold the mutex to check db_state because it * can't be freed while we have a hold on the buffer. */ ASSERT(!refcount_is_zero(&db->db_holds)); if (db->db_state == DB_NOFILL) return (SET_ERROR(EIO)); DB_DNODE_ENTER(db); dn = DB_DNODE(db); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_enter(&dn->dn_struct_rwlock, RW_READER); prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && DBUF_IS_CACHEABLE(db); mutex_enter(&db->db_mtx); if (db->db_state == DB_CACHED) { /* * If the arc buf is compressed, we need to decompress it to * read the data. This could happen during the "zfs receive" of * a stream which is compressed and deduplicated. */ if (db->db_buf != NULL && arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) { dbuf_fix_old_data(db, spa_syncing_txg(dmu_objset_spa(db->db_objset))); err = arc_decompress(db->db_buf); dbuf_set_data(db, db->db_buf); } mutex_exit(&db->db_mtx); if (prefetch) dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_exit(&dn->dn_struct_rwlock); DB_DNODE_EXIT(db); } else if (db->db_state == DB_UNCACHED) { spa_t *spa = dn->dn_objset->os_spa; boolean_t need_wait = B_FALSE; if (zio == NULL && db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); need_wait = B_TRUE; } dbuf_read_impl(db, zio, flags); /* dbuf_read_impl has dropped db_mtx for us */ if (prefetch) dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_exit(&dn->dn_struct_rwlock); DB_DNODE_EXIT(db); if (need_wait) err = zio_wait(zio); } else { /* * Another reader came in while the dbuf was in flight * between UNCACHED and CACHED. Either a writer will finish * writing the buffer (sending the dbuf to CACHED) or the * first reader's request will reach the read_done callback * and send the dbuf to CACHED. Otherwise, a failure * occurred and the dbuf went to UNCACHED. */ mutex_exit(&db->db_mtx); if (prefetch) dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); if ((flags & DB_RF_HAVESTRUCT) == 0) rw_exit(&dn->dn_struct_rwlock); DB_DNODE_EXIT(db); /* Skip the wait per the caller's request. */ mutex_enter(&db->db_mtx); if ((flags & DB_RF_NEVERWAIT) == 0) { while (db->db_state == DB_READ || db->db_state == DB_FILL) { ASSERT(db->db_state == DB_READ || (flags & DB_RF_HAVESTRUCT) == 0); DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, db, zio_t *, zio); cv_wait(&db->db_changed, &db->db_mtx); } if (db->db_state == DB_UNCACHED) err = SET_ERROR(EIO); } mutex_exit(&db->db_mtx); } return (err); } static void dbuf_noread(dmu_buf_impl_t *db) { ASSERT(!refcount_is_zero(&db->db_holds)); ASSERT(db->db_blkid != DMU_BONUS_BLKID); 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) { arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); spa_t *spa = db->db_objset->os_spa; ASSERT(db->db_buf == NULL); ASSERT(db->db.db_data == NULL); dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); db->db_state = DB_FILL; } else if (db->db_state == DB_NOFILL) { dbuf_clear_data(db); } else { ASSERT3U(db->db_state, ==, DB_CACHED); } mutex_exit(&db->db_mtx); } void dbuf_unoverride(dbuf_dirty_record_t *dr) { dmu_buf_impl_t *db = dr->dr_dbuf; blkptr_t *bp = &dr->dt.dl.dr_overridden_by; uint64_t txg = dr->dr_txg; ASSERT(MUTEX_HELD(&db->db_mtx)); /* * This assert is valid because dmu_sync() expects to be called by * a zilog's get_data while holding a range lock. This call only * comes from dbuf_dirty() callers who must also hold a range lock. */ ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); ASSERT(db->db_level == 0); if (db->db_blkid == DMU_BONUS_BLKID || dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) return; ASSERT(db->db_data_pending != dr); /* free this block */ if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) zio_free(db->db_objset->os_spa, txg, bp); dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; dr->dt.dl.dr_nopwrite = B_FALSE; /* * Release the already-written buffer, so we leave it in * a consistent dirty state. Note that all callers are * modifying the buffer, so they will immediately do * another (redundant) arc_release(). Therefore, leave * the buf thawed to save the effort of freezing & * immediately re-thawing it. */ arc_release(dr->dt.dl.dr_data, db); } /* * Evict (if its unreferenced) or clear (if its referenced) any level-0 * data blocks in the free range, so that any future readers will find * empty blocks. */ void dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, dmu_tx_t *tx) { dmu_buf_impl_t db_search; dmu_buf_impl_t *db, *db_next; uint64_t txg = tx->tx_txg; avl_index_t where; if (end_blkid > dn->dn_maxblkid && !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) end_blkid = dn->dn_maxblkid; dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); db_search.db_level = 0; db_search.db_blkid = start_blkid; db_search.db_state = DB_SEARCH; mutex_enter(&dn->dn_dbufs_mtx); db = avl_find(&dn->dn_dbufs, &db_search, &where); ASSERT3P(db, ==, NULL); db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); for (; db != NULL; db = db_next) { db_next = AVL_NEXT(&dn->dn_dbufs, db); ASSERT(db->db_blkid != DMU_BONUS_BLKID); if (db->db_level != 0 || db->db_blkid > end_blkid) { break; } ASSERT3U(db->db_blkid, >=, start_blkid); /* found a level 0 buffer in the range */ mutex_enter(&db->db_mtx); if (dbuf_undirty(db, tx)) { /* mutex has been dropped and dbuf destroyed */ continue; } if (db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL || db->db_state == DB_EVICTING) { ASSERT(db->db.db_data == NULL); mutex_exit(&db->db_mtx); continue; } if (db->db_state == DB_READ || db->db_state == DB_FILL) { /* will be handled in dbuf_read_done or dbuf_rele */ db->db_freed_in_flight = TRUE; mutex_exit(&db->db_mtx); continue; } if (refcount_count(&db->db_holds) == 0) { ASSERT(db->db_buf); dbuf_destroy(db); continue; } /* The dbuf is referenced */ if (db->db_last_dirty != NULL) { dbuf_dirty_record_t *dr = db->db_last_dirty; if (dr->dr_txg == txg) { /* * This buffer is "in-use", re-adjust the file * size to reflect that this buffer may * contain new data when we sync. */ if (db->db_blkid != DMU_SPILL_BLKID && db->db_blkid > dn->dn_maxblkid) dn->dn_maxblkid = db->db_blkid; dbuf_unoverride(dr); } else { /* * This dbuf is not dirty in the open context. * Either uncache it (if its not referenced in * the open context) or reset its contents to * empty. */ dbuf_fix_old_data(db, txg); } } /* clear the contents if its cached */ if (db->db_state == DB_CACHED) { ASSERT(db->db.db_data != NULL); arc_release(db->db_buf, db); bzero(db->db.db_data, db->db.db_size); arc_buf_freeze(db->db_buf); } mutex_exit(&db->db_mtx); } mutex_exit(&dn->dn_dbufs_mtx); } void dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) { arc_buf_t *buf, *obuf; int osize = db->db.db_size; arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); dnode_t *dn; ASSERT(db->db_blkid != DMU_BONUS_BLKID); DB_DNODE_ENTER(db); dn = DB_DNODE(db); /* XXX does *this* func really need the lock? */ ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); /* * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held * is OK, because there can be no other references to the db * when we are changing its size, so no concurrent DB_FILL can * be happening. */ /* * XXX we should be doing a dbuf_read, checking the return * value and returning that up to our callers */ dmu_buf_will_dirty(&db->db, tx); /* create the data buffer for the new block */ buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); /* copy old block data to the new block */ obuf = db->db_buf; bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); /* zero the remainder */ if (size > osize) bzero((uint8_t *)buf->b_data + osize, size - osize); mutex_enter(&db->db_mtx); dbuf_set_data(db, buf); arc_buf_destroy(obuf, db); db->db.db_size = size; if (db->db_level == 0) { ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); db->db_last_dirty->dt.dl.dr_data = buf; } mutex_exit(&db->db_mtx); dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); DB_DNODE_EXIT(db); } void dbuf_release_bp(dmu_buf_impl_t *db) { objset_t *os = db->db_objset; ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); ASSERT(arc_released(os->os_phys_buf) || list_link_active(&os->os_dsl_dataset->ds_synced_link)); ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); (void) arc_release(db->db_buf, db); } /* * We already have a dirty record for this TXG, and we are being * dirtied again. */ static void dbuf_redirty(dbuf_dirty_record_t *dr) { dmu_buf_impl_t *db = dr->dr_dbuf; ASSERT(MUTEX_HELD(&db->db_mtx)); if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { /* * If this buffer has already been written out, * we now need to reset its state. */ dbuf_unoverride(dr); if (db->db.db_object != DMU_META_DNODE_OBJECT && db->db_state != DB_NOFILL) { /* Already released on initial dirty, so just thaw. */ ASSERT(arc_released(db->db_buf)); arc_buf_thaw(db->db_buf); } } } dbuf_dirty_record_t * dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) { dnode_t *dn; objset_t *os; dbuf_dirty_record_t **drp, *dr; int drop_struct_lock = FALSE; int txgoff = tx->tx_txg & TXG_MASK; ASSERT(tx->tx_txg != 0); ASSERT(!refcount_is_zero(&db->db_holds)); DMU_TX_DIRTY_BUF(tx, db); DB_DNODE_ENTER(db); dn = DB_DNODE(db); /* * Shouldn't dirty a regular buffer in syncing context. Private * objects may be dirtied in syncing context, but only if they * were already pre-dirtied in open context. */ #ifdef DEBUG if (dn->dn_objset->os_dsl_dataset != NULL) { rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); } ASSERT(!dmu_tx_is_syncing(tx) || BP_IS_HOLE(dn->dn_objset->os_rootbp) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_objset->os_dsl_dataset == NULL); if (dn->dn_objset->os_dsl_dataset != NULL) rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); #endif /* * We make this assert for private objects as well, but after we * check if we're already dirty. They are allowed to re-dirty * in syncing context. */ ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); mutex_enter(&db->db_mtx); /* * XXX make this true for indirects too? The problem is that * transactions created with dmu_tx_create_assigned() from * syncing context don't bother holding ahead. */ ASSERT(db->db_level != 0 || db->db_state == DB_CACHED || db->db_state == DB_FILL || db->db_state == DB_NOFILL); mutex_enter(&dn->dn_mtx); /* * Don't set dirtyctx to SYNC if we're just modifying this as we * initialize the objset. */ if (dn->dn_dirtyctx == DN_UNDIRTIED) { if (dn->dn_objset->os_dsl_dataset != NULL) { rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); } if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); ASSERT(dn->dn_dirtyctx_firstset == NULL); dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); } if (dn->dn_objset->os_dsl_dataset != NULL) { rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); } } mutex_exit(&dn->dn_mtx); if (db->db_blkid == DMU_SPILL_BLKID) dn->dn_have_spill = B_TRUE; /* * If this buffer is already dirty, we're done. */ drp = &db->db_last_dirty; ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || db->db.db_object == DMU_META_DNODE_OBJECT); while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) drp = &dr->dr_next; if (dr && dr->dr_txg == tx->tx_txg) { DB_DNODE_EXIT(db); dbuf_redirty(dr); mutex_exit(&db->db_mtx); return (dr); } /* * Only valid if not already dirty. */ ASSERT(dn->dn_object == 0 || dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); ASSERT3U(dn->dn_nlevels, >, db->db_level); /* * We should only be dirtying in syncing context if it's the * mos or we're initializing the os or it's a special object. * However, we are allowed to dirty in syncing context provided * we already dirtied it in open context. Hence we must make * this assertion only if we're not already dirty. */ os = dn->dn_objset; VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); #ifdef DEBUG if (dn->dn_objset->os_dsl_dataset != NULL) rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); if (dn->dn_objset->os_dsl_dataset != NULL) rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); #endif ASSERT(db->db.db_size != 0); dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); if (db->db_blkid != DMU_BONUS_BLKID) { dmu_objset_willuse_space(os, db->db.db_size, tx); } /* * If this buffer is dirty in an old transaction group we need * to make a copy of it so that the changes we make in this * transaction group won't leak out when we sync the older txg. */ dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); if (db->db_level == 0) { void *data_old = db->db_buf; if (db->db_state != DB_NOFILL) { if (db->db_blkid == DMU_BONUS_BLKID) { dbuf_fix_old_data(db, tx->tx_txg); data_old = db->db.db_data; } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { /* * Release the data buffer from the cache so * that we can modify it without impacting * possible other users of this cached data * block. Note that indirect blocks and * private objects are not released until the * syncing state (since they are only modified * then). */ arc_release(db->db_buf, db); dbuf_fix_old_data(db, tx->tx_txg); data_old = db->db_buf; } ASSERT(data_old != NULL); } dr->dt.dl.dr_data = data_old; } else { mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); list_create(&dr->dt.di.dr_children, sizeof (dbuf_dirty_record_t), offsetof(dbuf_dirty_record_t, dr_dirty_node)); } if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) dr->dr_accounted = db->db.db_size; dr->dr_dbuf = db; dr->dr_txg = tx->tx_txg; dr->dr_next = *drp; *drp = dr; /* * We could have been freed_in_flight between the dbuf_noread * and dbuf_dirty. We win, as though the dbuf_noread() had * happened after the free. */ if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && db->db_blkid != DMU_SPILL_BLKID) { mutex_enter(&dn->dn_mtx); if (dn->dn_free_ranges[txgoff] != NULL) { range_tree_clear(dn->dn_free_ranges[txgoff], db->db_blkid, 1); } mutex_exit(&dn->dn_mtx); db->db_freed_in_flight = FALSE; } /* * This buffer is now part of this txg */ dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); db->db_dirtycnt += 1; ASSERT3U(db->db_dirtycnt, <=, 3); mutex_exit(&db->db_mtx); if (db->db_blkid == DMU_BONUS_BLKID || db->db_blkid == DMU_SPILL_BLKID) { mutex_enter(&dn->dn_mtx); ASSERT(!list_link_active(&dr->dr_dirty_node)); list_insert_tail(&dn->dn_dirty_records[txgoff], dr); mutex_exit(&dn->dn_mtx); dnode_setdirty(dn, tx); DB_DNODE_EXIT(db); return (dr); } /* * The dn_struct_rwlock prevents db_blkptr from changing * due to a write from syncing context completing * while we are running, so we want to acquire it before * looking at db_blkptr. */ if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { rw_enter(&dn->dn_struct_rwlock, RW_READER); drop_struct_lock = TRUE; } /* * We need to hold the dn_struct_rwlock to make this assertion, * because it protects dn_phys / dn_next_nlevels from changing. */ ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || dn->dn_phys->dn_nlevels > db->db_level || dn->dn_next_nlevels[txgoff] > db->db_level || dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); /* * If we are overwriting a dedup BP, then unless it is snapshotted, * when we get to syncing context we will need to decrement its * refcount in the DDT. Prefetch the relevant DDT block so that * syncing context won't have to wait for the i/o. */ ddt_prefetch(os->os_spa, db->db_blkptr); if (db->db_level == 0) { dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); ASSERT(dn->dn_maxblkid >= db->db_blkid); } if (db->db_level+1 < dn->dn_nlevels) { dmu_buf_impl_t *parent = db->db_parent; dbuf_dirty_record_t *di; int parent_held = FALSE; if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; parent = dbuf_hold_level(dn, db->db_level+1, db->db_blkid >> epbs, FTAG); ASSERT(parent != NULL); parent_held = TRUE; } if (drop_struct_lock) rw_exit(&dn->dn_struct_rwlock); ASSERT3U(db->db_level+1, ==, parent->db_level); di = dbuf_dirty(parent, tx); if (parent_held) dbuf_rele(parent, FTAG); mutex_enter(&db->db_mtx); /* * Since we've dropped the mutex, it's possible that * dbuf_undirty() might have changed this out from under us. */ if (db->db_last_dirty == dr || dn->dn_object == DMU_META_DNODE_OBJECT) { mutex_enter(&di->dt.di.dr_mtx); ASSERT3U(di->dr_txg, ==, tx->tx_txg); ASSERT(!list_link_active(&dr->dr_dirty_node)); list_insert_tail(&di->dt.di.dr_children, dr); mutex_exit(&di->dt.di.dr_mtx); dr->dr_parent = di; } mutex_exit(&db->db_mtx); } else { ASSERT(db->db_level+1 == dn->dn_nlevels); ASSERT(db->db_blkid < dn->dn_nblkptr); ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); mutex_enter(&dn->dn_mtx); ASSERT(!list_link_active(&dr->dr_dirty_node)); list_insert_tail(&dn->dn_dirty_records[txgoff], dr); mutex_exit(&dn->dn_mtx); if (drop_struct_lock) rw_exit(&dn->dn_struct_rwlock); } dnode_setdirty(dn, tx); DB_DNODE_EXIT(db); return (dr); } /* * Undirty a buffer in the transaction group referenced by the given * transaction. Return whether this evicted the dbuf. */ static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) { dnode_t *dn; uint64_t txg = tx->tx_txg; dbuf_dirty_record_t *dr, **drp; ASSERT(txg != 0); /* * Due to our use of dn_nlevels below, this can only be called * in open context, unless we are operating on the MOS. * From syncing context, dn_nlevels may be different from the * dn_nlevels used when dbuf was dirtied. */ ASSERT(db->db_objset == dmu_objset_pool(db->db_objset)->dp_meta_objset || txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT0(db->db_level); ASSERT(MUTEX_HELD(&db->db_mtx)); /* * If this buffer is not dirty, we're done. */ for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) if (dr->dr_txg <= txg) break; if (dr == NULL || dr->dr_txg < txg) return (B_FALSE); ASSERT(dr->dr_txg == txg); ASSERT(dr->dr_dbuf == db); DB_DNODE_ENTER(db); dn = DB_DNODE(db); dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); ASSERT(db->db.db_size != 0); dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), dr->dr_accounted, txg); *drp = dr->dr_next; /* * Note that there are three places in dbuf_dirty() * where this dirty record may be put on a list. * Make sure to do a list_remove corresponding to * every one of those list_insert calls. */ if (dr->dr_parent) { mutex_enter(&dr->dr_parent->dt.di.dr_mtx); list_remove(&dr->dr_parent->dt.di.dr_children, dr); mutex_exit(&dr->dr_parent->dt.di.dr_mtx); } else if (db->db_blkid == DMU_SPILL_BLKID || db->db_level + 1 == dn->dn_nlevels) { ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); mutex_enter(&dn->dn_mtx); list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); mutex_exit(&dn->dn_mtx); } DB_DNODE_EXIT(db); if (db->db_state != DB_NOFILL) { dbuf_unoverride(dr); ASSERT(db->db_buf != NULL); ASSERT(dr->dt.dl.dr_data != NULL); if (dr->dt.dl.dr_data != db->db_buf) arc_buf_destroy(dr->dt.dl.dr_data, db); } kmem_free(dr, sizeof (dbuf_dirty_record_t)); ASSERT(db->db_dirtycnt > 0); db->db_dirtycnt -= 1; if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); dbuf_destroy(db); return (B_TRUE); } return (B_FALSE); } void dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; ASSERT(tx->tx_txg != 0); ASSERT(!refcount_is_zero(&db->db_holds)); /* * Quick check for dirtyness. For already dirty blocks, this * reduces runtime of this function by >90%, and overall performance * by 50% for some workloads (e.g. file deletion with indirect blocks * cached). */ mutex_enter(&db->db_mtx); dbuf_dirty_record_t *dr; for (dr = db->db_last_dirty; dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { /* * It's possible that it is already dirty but not cached, * because there are some calls to dbuf_dirty() that don't * go through dmu_buf_will_dirty(). */ if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { /* This dbuf is already dirty and cached. */ dbuf_redirty(dr); mutex_exit(&db->db_mtx); return; } } mutex_exit(&db->db_mtx); DB_DNODE_ENTER(db); if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) rf |= DB_RF_HAVESTRUCT; DB_DNODE_EXIT(db); (void) dbuf_read(db, NULL, rf); (void) dbuf_dirty(db, tx); } void dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; db->db_state = DB_NOFILL; dmu_buf_will_fill(db_fake, tx); } void dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT(tx->tx_txg != 0); ASSERT(db->db_level == 0); ASSERT(!refcount_is_zero(&db->db_holds)); ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx)); dbuf_noread(db); (void) dbuf_dirty(db, tx); } #pragma weak dmu_buf_fill_done = dbuf_fill_done /* ARGSUSED */ void dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) { mutex_enter(&db->db_mtx); DBUF_VERIFY(db); if (db->db_state == DB_FILL) { if (db->db_level == 0 && db->db_freed_in_flight) { ASSERT(db->db_blkid != DMU_BONUS_BLKID); /* we were freed while filling */ /* XXX dbuf_undirty? */ bzero(db->db.db_data, db->db.db_size); db->db_freed_in_flight = FALSE; } db->db_state = DB_CACHED; cv_broadcast(&db->db_changed); } mutex_exit(&db->db_mtx); } void dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, bp_embedded_type_t etype, enum zio_compress comp, int uncompressed_size, int compressed_size, int byteorder, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; struct dirty_leaf *dl; dmu_object_type_t type; if (etype == BP_EMBEDDED_TYPE_DATA) { ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), SPA_FEATURE_EMBEDDED_DATA)); } DB_DNODE_ENTER(db); type = DB_DNODE(db)->dn_type; DB_DNODE_EXIT(db); ASSERT0(db->db_level); ASSERT(db->db_blkid != DMU_BONUS_BLKID); dmu_buf_will_not_fill(dbuf, tx); ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); dl = &db->db_last_dirty->dt.dl; encode_embedded_bp_compressed(&dl->dr_overridden_by, data, comp, uncompressed_size, compressed_size); BPE_SET_ETYPE(&dl->dr_overridden_by, etype); BP_SET_TYPE(&dl->dr_overridden_by, type); BP_SET_LEVEL(&dl->dr_overridden_by, 0); BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); dl->dr_override_state = DR_OVERRIDDEN; dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; } /* * Directly assign a provided arc buf to a given dbuf if it's not referenced * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. */ void dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) { ASSERT(!refcount_is_zero(&db->db_holds)); ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT(db->db_level == 0); ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); ASSERT(buf != NULL); ASSERT(arc_buf_lsize(buf) == db->db.db_size); ASSERT(tx->tx_txg != 0); arc_return_buf(buf, db); ASSERT(arc_released(buf)); mutex_enter(&db->db_mtx); while (db->db_state == DB_READ || db->db_state == DB_FILL) cv_wait(&db->db_changed, &db->db_mtx); ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); if (db->db_state == DB_CACHED && refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { mutex_exit(&db->db_mtx); (void) dbuf_dirty(db, tx); bcopy(buf->b_data, db->db.db_data, db->db.db_size); arc_buf_destroy(buf, db); xuio_stat_wbuf_copied(); return; } xuio_stat_wbuf_nocopy(); if (db->db_state == DB_CACHED) { dbuf_dirty_record_t *dr = db->db_last_dirty; ASSERT(db->db_buf != NULL); if (dr != NULL && dr->dr_txg == tx->tx_txg) { ASSERT(dr->dt.dl.dr_data == db->db_buf); if (!arc_released(db->db_buf)) { ASSERT(dr->dt.dl.dr_override_state == DR_OVERRIDDEN); arc_release(db->db_buf, db); } dr->dt.dl.dr_data = buf; arc_buf_destroy(db->db_buf, db); } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { arc_release(db->db_buf, db); arc_buf_destroy(db->db_buf, db); } db->db_buf = NULL; } ASSERT(db->db_buf == NULL); dbuf_set_data(db, buf); db->db_state = DB_FILL; mutex_exit(&db->db_mtx); (void) dbuf_dirty(db, tx); dmu_buf_fill_done(&db->db, tx); } void dbuf_destroy(dmu_buf_impl_t *db) { dnode_t *dn; dmu_buf_impl_t *parent = db->db_parent; dmu_buf_impl_t *dndb; ASSERT(MUTEX_HELD(&db->db_mtx)); ASSERT(refcount_is_zero(&db->db_holds)); if (db->db_buf != NULL) { arc_buf_destroy(db->db_buf, db); db->db_buf = NULL; } if (db->db_blkid == DMU_BONUS_BLKID) { int slots = DB_DNODE(db)->dn_num_slots; int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); if (db->db.db_data != NULL) { zio_buf_free(db->db.db_data, bonuslen); arc_space_return(bonuslen, ARC_SPACE_BONUS); db->db_state = DB_UNCACHED; } } dbuf_clear_data(db); if (multilist_link_active(&db->db_cache_link)) { ASSERT(db->db_caching_status == DB_DBUF_CACHE || db->db_caching_status == DB_DBUF_METADATA_CACHE); multilist_remove(dbuf_caches[db->db_caching_status].cache, db); (void) refcount_remove_many( &dbuf_caches[db->db_caching_status].size, db->db.db_size, db); db->db_caching_status = DB_NO_CACHE; } ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); ASSERT(db->db_data_pending == NULL); db->db_state = DB_EVICTING; db->db_blkptr = NULL; /* * Now that db_state is DB_EVICTING, nobody else can find this via * the hash table. We can now drop db_mtx, which allows us to * acquire the dn_dbufs_mtx. */ mutex_exit(&db->db_mtx); DB_DNODE_ENTER(db); dn = DB_DNODE(db); dndb = dn->dn_dbuf; if (db->db_blkid != DMU_BONUS_BLKID) { boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); if (needlock) mutex_enter(&dn->dn_dbufs_mtx); avl_remove(&dn->dn_dbufs, db); atomic_dec_32(&dn->dn_dbufs_count); membar_producer(); DB_DNODE_EXIT(db); if (needlock) mutex_exit(&dn->dn_dbufs_mtx); /* * Decrementing the dbuf count means that the hold corresponding * to the removed dbuf is no longer discounted in dnode_move(), * so the dnode cannot be moved until after we release the hold. * The membar_producer() ensures visibility of the decremented * value in dnode_move(), since DB_DNODE_EXIT doesn't actually * release any lock. */ mutex_enter(&dn->dn_mtx); dnode_rele_and_unlock(dn, db, B_TRUE); db->db_dnode_handle = NULL; dbuf_hash_remove(db); } else { DB_DNODE_EXIT(db); } ASSERT(refcount_is_zero(&db->db_holds)); db->db_parent = NULL; ASSERT(db->db_buf == NULL); ASSERT(db->db.db_data == NULL); ASSERT(db->db_hash_next == NULL); ASSERT(db->db_blkptr == NULL); ASSERT(db->db_data_pending == NULL); ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE); ASSERT(!multilist_link_active(&db->db_cache_link)); kmem_cache_free(dbuf_kmem_cache, db); arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); /* * If this dbuf is referenced from an indirect dbuf, * decrement the ref count on the indirect dbuf. */ if (parent && parent != dndb) { mutex_enter(&parent->db_mtx); dbuf_rele_and_unlock(parent, db, B_TRUE); } } /* * Note: While bpp will always be updated if the function returns success, * parentp will not be updated if the dnode does not have dn_dbuf filled in; * this happens when the dnode is the meta-dnode, or a userused or groupused * object. */ static int dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, dmu_buf_impl_t **parentp, blkptr_t **bpp) { *parentp = NULL; *bpp = NULL; ASSERT(blkid != DMU_BONUS_BLKID); if (blkid == DMU_SPILL_BLKID) { mutex_enter(&dn->dn_mtx); if (dn->dn_have_spill && (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) *bpp = DN_SPILL_BLKPTR(dn->dn_phys); else *bpp = NULL; dbuf_add_ref(dn->dn_dbuf, NULL); *parentp = dn->dn_dbuf; mutex_exit(&dn->dn_mtx); return (0); } int nlevels = (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; ASSERT3U(level * epbs, <, 64); ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); /* * This assertion shouldn't trip as long as the max indirect block size * is less than 1M. The reason for this is that up to that point, * the number of levels required to address an entire object with blocks * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 * (i.e. we can address the entire object), objects will all use at most * N-1 levels and the assertion won't overflow. However, once epbs is * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be * enough to address an entire object, so objects will have 5 levels, * but then this assertion will overflow. * * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we * need to redo this logic to handle overflows. */ ASSERT(level >= nlevels || ((nlevels - level - 1) * epbs) + highbit64(dn->dn_phys->dn_nblkptr) <= 64); if (level >= nlevels || blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << ((nlevels - level - 1) * epbs)) || (fail_sparse && blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { /* the buffer has no parent yet */ return (SET_ERROR(ENOENT)); } else if (level < nlevels-1) { /* this block is referenced from an indirect block */ int err = dbuf_hold_impl(dn, level+1, blkid >> epbs, fail_sparse, FALSE, NULL, parentp); if (err) return (err); err = dbuf_read(*parentp, NULL, (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); if (err) { dbuf_rele(*parentp, NULL); *parentp = NULL; return (err); } *bpp = ((blkptr_t *)(*parentp)->db.db_data) + (blkid & ((1ULL << epbs) - 1)); if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) ASSERT(BP_IS_HOLE(*bpp)); return (0); } else { /* the block is referenced from the dnode */ ASSERT3U(level, ==, nlevels-1); ASSERT(dn->dn_phys->dn_nblkptr == 0 || blkid < dn->dn_phys->dn_nblkptr); if (dn->dn_dbuf) { dbuf_add_ref(dn->dn_dbuf, NULL); *parentp = dn->dn_dbuf; } *bpp = &dn->dn_phys->dn_blkptr[blkid]; return (0); } } static dmu_buf_impl_t * dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, dmu_buf_impl_t *parent, blkptr_t *blkptr) { objset_t *os = dn->dn_objset; dmu_buf_impl_t *db, *odb; ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); ASSERT(dn->dn_type != DMU_OT_NONE); db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); db->db_objset = os; db->db.db_object = dn->dn_object; db->db_level = level; db->db_blkid = blkid; db->db_last_dirty = NULL; db->db_dirtycnt = 0; db->db_dnode_handle = dn->dn_handle; db->db_parent = parent; db->db_blkptr = blkptr; db->db_user = NULL; db->db_user_immediate_evict = FALSE; db->db_freed_in_flight = FALSE; db->db_pending_evict = FALSE; if (blkid == DMU_BONUS_BLKID) { ASSERT3P(parent, ==, dn->dn_dbuf); db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - (dn->dn_nblkptr-1) * sizeof (blkptr_t); ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); db->db.db_offset = DMU_BONUS_BLKID; db->db_state = DB_UNCACHED; db->db_caching_status = DB_NO_CACHE; /* the bonus dbuf is not placed in the hash table */ arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); return (db); } else if (blkid == DMU_SPILL_BLKID) { db->db.db_size = (blkptr != NULL) ? BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; db->db.db_offset = 0; } else { int blocksize = db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; db->db.db_size = blocksize; db->db.db_offset = db->db_blkid * blocksize; } /* * Hold the dn_dbufs_mtx while we get the new dbuf * in the hash table *and* added to the dbufs list. * This prevents a possible deadlock with someone * trying to look up this dbuf before its added to the * dn_dbufs list. */ mutex_enter(&dn->dn_dbufs_mtx); db->db_state = DB_EVICTING; if ((odb = dbuf_hash_insert(db)) != NULL) { /* someone else inserted it first */ kmem_cache_free(dbuf_kmem_cache, db); mutex_exit(&dn->dn_dbufs_mtx); return (odb); } avl_add(&dn->dn_dbufs, db); db->db_state = DB_UNCACHED; db->db_caching_status = DB_NO_CACHE; mutex_exit(&dn->dn_dbufs_mtx); arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); if (parent && parent != dn->dn_dbuf) dbuf_add_ref(parent, db); ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || refcount_count(&dn->dn_holds) > 0); (void) refcount_add(&dn->dn_holds, db); atomic_inc_32(&dn->dn_dbufs_count); dprintf_dbuf(db, "db=%p\n", db); return (db); } typedef struct dbuf_prefetch_arg { spa_t *dpa_spa; /* The spa to issue the prefetch in. */ zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ int dpa_curlevel; /* The current level that we're reading */ dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ } dbuf_prefetch_arg_t; /* * Actually issue the prefetch read for the block given. */ static void dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) { if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) return; arc_flags_t aflags = dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); ASSERT(dpa->dpa_zio != NULL); (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &aflags, &dpa->dpa_zb); } /* * Called when an indirect block above our prefetch target is read in. This * will either read in the next indirect block down the tree or issue the actual * prefetch if the next block down is our target. */ static void dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *iobp, arc_buf_t *abuf, void *private) { dbuf_prefetch_arg_t *dpa = private; ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); ASSERT3S(dpa->dpa_curlevel, >, 0); if (abuf == NULL) { ASSERT(zio == NULL || zio->io_error != 0); kmem_free(dpa, sizeof (*dpa)); return; } ASSERT(zio == NULL || zio->io_error == 0); /* * The dpa_dnode is only valid if we are called with a NULL * zio. This indicates that the arc_read() returned without * first calling zio_read() to issue a physical read. Once * a physical read is made the dpa_dnode must be invalidated * as the locks guarding it may have been dropped. If the * dpa_dnode is still valid, then we want to add it to the dbuf * cache. To do so, we must hold the dbuf associated with the block * we just prefetched, read its contents so that we associate it * with an arc_buf_t, and then release it. */ if (zio != NULL) { ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); if (zio->io_flags & ZIO_FLAG_RAW) { ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); } else { ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); } ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); dpa->dpa_dnode = NULL; } else if (dpa->dpa_dnode != NULL) { uint64_t curblkid = dpa->dpa_zb.zb_blkid >> (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, dpa->dpa_curlevel, curblkid, FTAG); (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); dbuf_rele(db, FTAG); } if (abuf == NULL) { kmem_free(dpa, sizeof(*dpa)); return; } dpa->dpa_curlevel--; uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); blkptr_t *bp = ((blkptr_t *)abuf->b_data) + P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); if (BP_IS_HOLE(bp)) { kmem_free(dpa, sizeof (*dpa)); } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); dbuf_issue_final_prefetch(dpa, bp); kmem_free(dpa, sizeof (*dpa)); } else { arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; zbookmark_phys_t zb; /* flag if L2ARC eligible, l2arc_noprefetch then decides */ if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) iter_aflags |= ARC_FLAG_L2CACHE; ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &iter_aflags, &zb); } arc_buf_destroy(abuf, private); } /* * Issue prefetch reads for the given block on the given level. If the indirect * blocks above that block are not in memory, we will read them in * asynchronously. As a result, this call never blocks waiting for a read to * complete. */ void dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, arc_flags_t aflags) { blkptr_t bp; int epbs, nlevels, curlevel; uint64_t curblkid; ASSERT(blkid != DMU_BONUS_BLKID); ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); if (blkid > dn->dn_maxblkid) return; if (dnode_block_freed(dn, blkid)) return; /* * This dnode hasn't been written to disk yet, so there's nothing to * prefetch. */ nlevels = dn->dn_phys->dn_nlevels; if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) return; epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) return; dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); if (db != NULL) { mutex_exit(&db->db_mtx); /* * This dbuf already exists. It is either CACHED, or * (we assume) about to be read or filled. */ return; } /* * Find the closest ancestor (indirect block) of the target block * that is present in the cache. In this indirect block, we will * find the bp that is at curlevel, curblkid. */ curlevel = level; curblkid = blkid; while (curlevel < nlevels - 1) { int parent_level = curlevel + 1; uint64_t parent_blkid = curblkid >> epbs; dmu_buf_impl_t *db; if (dbuf_hold_impl(dn, parent_level, parent_blkid, FALSE, TRUE, FTAG, &db) == 0) { blkptr_t *bpp = db->db_buf->b_data; bp = bpp[P2PHASE(curblkid, 1 << epbs)]; dbuf_rele(db, FTAG); break; } curlevel = parent_level; curblkid = parent_blkid; } if (curlevel == nlevels - 1) { /* No cached indirect blocks found. */ ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); bp = dn->dn_phys->dn_blkptr[curblkid]; } if (BP_IS_HOLE(&bp)) return; ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, ZIO_FLAG_CANFAIL); dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, dn->dn_object, level, blkid); dpa->dpa_curlevel = curlevel; dpa->dpa_prio = prio; dpa->dpa_aflags = aflags; dpa->dpa_spa = dn->dn_objset->os_spa; dpa->dpa_dnode = dn; dpa->dpa_epbs = epbs; dpa->dpa_zio = pio; /* flag if L2ARC eligible, l2arc_noprefetch then decides */ if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) dpa->dpa_aflags |= ARC_FLAG_L2CACHE; /* * If we have the indirect just above us, no need to do the asynchronous * prefetch chain; we'll just run the last step ourselves. If we're at * a higher level, though, we want to issue the prefetches for all the * indirect blocks asynchronously, so we can go on with whatever we were * doing. */ if (curlevel == level) { ASSERT3U(curblkid, ==, blkid); dbuf_issue_final_prefetch(dpa, &bp); kmem_free(dpa, sizeof (*dpa)); } else { arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; zbookmark_phys_t zb; /* flag if L2ARC eligible, l2arc_noprefetch then decides */ if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) iter_aflags |= ARC_FLAG_L2CACHE; SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, dn->dn_object, curlevel, curblkid); (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, &bp, dbuf_prefetch_indirect_done, dpa, prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &iter_aflags, &zb); } /* * We use pio here instead of dpa_zio since it's possible that * dpa may have already been freed. */ zio_nowait(pio); } /* * Returns with db_holds incremented, and db_mtx not held. * Note: dn_struct_rwlock must be held. */ int dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse, boolean_t fail_uncached, void *tag, dmu_buf_impl_t **dbp) { dmu_buf_impl_t *db, *parent = NULL; ASSERT(blkid != DMU_BONUS_BLKID); ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); ASSERT3U(dn->dn_nlevels, >, level); *dbp = NULL; top: /* dbuf_find() returns with db_mtx held */ db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); if (db == NULL) { blkptr_t *bp = NULL; int err; if (fail_uncached) return (SET_ERROR(ENOENT)); ASSERT3P(parent, ==, NULL); err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); if (fail_sparse) { if (err == 0 && bp && BP_IS_HOLE(bp)) err = SET_ERROR(ENOENT); if (err) { if (parent) dbuf_rele(parent, NULL); return (err); } } if (err && err != ENOENT) return (err); db = dbuf_create(dn, level, blkid, parent, bp); } if (fail_uncached && db->db_state != DB_CACHED) { mutex_exit(&db->db_mtx); return (SET_ERROR(ENOENT)); } if (db->db_buf != NULL) { arc_buf_access(db->db_buf); ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); } ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); /* * If this buffer is currently syncing out, and we are are * still referencing it from db_data, we need to make a copy * of it in case we decide we want to dirty it again in this txg. */ if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && dn->dn_object != DMU_META_DNODE_OBJECT && db->db_state == DB_CACHED && db->db_data_pending) { dbuf_dirty_record_t *dr = db->db_data_pending; if (dr->dt.dl.dr_data == db->db_buf) { arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, type, db->db.db_size)); bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, db->db.db_size); } } if (multilist_link_active(&db->db_cache_link)) { ASSERT(refcount_is_zero(&db->db_holds)); ASSERT(db->db_caching_status == DB_DBUF_CACHE || db->db_caching_status == DB_DBUF_METADATA_CACHE); multilist_remove(dbuf_caches[db->db_caching_status].cache, db); (void) refcount_remove_many( &dbuf_caches[db->db_caching_status].size, db->db.db_size, db); db->db_caching_status = DB_NO_CACHE; } (void) refcount_add(&db->db_holds, tag); DBUF_VERIFY(db); mutex_exit(&db->db_mtx); /* NOTE: we can't rele the parent until after we drop the db_mtx */ if (parent) dbuf_rele(parent, NULL); ASSERT3P(DB_DNODE(db), ==, dn); ASSERT3U(db->db_blkid, ==, blkid); ASSERT3U(db->db_level, ==, level); *dbp = db; return (0); } dmu_buf_impl_t * dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) { return (dbuf_hold_level(dn, 0, blkid, tag)); } dmu_buf_impl_t * dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) { dmu_buf_impl_t *db; int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); return (err ? NULL : db); } void dbuf_create_bonus(dnode_t *dn) { ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); ASSERT(dn->dn_bonus == NULL); dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); } int dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dnode_t *dn; if (db->db_blkid != DMU_SPILL_BLKID) return (SET_ERROR(ENOTSUP)); if (blksz == 0) blksz = SPA_MINBLOCKSIZE; ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); DB_DNODE_ENTER(db); dn = DB_DNODE(db); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); dbuf_new_size(db, blksz, tx); rw_exit(&dn->dn_struct_rwlock); DB_DNODE_EXIT(db); return (0); } void dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) { dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); } #pragma weak dmu_buf_add_ref = dbuf_add_ref void dbuf_add_ref(dmu_buf_impl_t *db, void *tag) { int64_t holds = refcount_add(&db->db_holds, tag); ASSERT3S(holds, >, 1); } #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref boolean_t dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, void *tag) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dmu_buf_impl_t *found_db; boolean_t result = B_FALSE; if (db->db_blkid == DMU_BONUS_BLKID) found_db = dbuf_find_bonus(os, obj); else found_db = dbuf_find(os, obj, 0, blkid); if (found_db != NULL) { if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { (void) refcount_add(&db->db_holds, tag); result = B_TRUE; } mutex_exit(&db->db_mtx); } return (result); } /* * If you call dbuf_rele() you had better not be referencing the dnode handle * unless you have some other direct or indirect hold on the dnode. (An indirect * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the * dnode's parent dbuf evicting its dnode handles. */ void dbuf_rele(dmu_buf_impl_t *db, void *tag) { mutex_enter(&db->db_mtx); dbuf_rele_and_unlock(db, tag, B_FALSE); } void dmu_buf_rele(dmu_buf_t *db, void *tag) { dbuf_rele((dmu_buf_impl_t *)db, tag); } /* * dbuf_rele() for an already-locked dbuf. This is necessary to allow * db_dirtycnt and db_holds to be updated atomically. The 'evicting' * argument should be set if we are already in the dbuf-evicting code * path, in which case we don't want to recursively evict. This allows us to * avoid deeply nested stacks that would have a call flow similar to this: * * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() * ^ | * | | * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ * */ void dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting) { int64_t holds; ASSERT(MUTEX_HELD(&db->db_mtx)); DBUF_VERIFY(db); /* * Remove the reference to the dbuf before removing its hold on the * dnode so we can guarantee in dnode_move() that a referenced bonus * buffer has a corresponding dnode hold. */ holds = refcount_remove(&db->db_holds, tag); ASSERT(holds >= 0); /* * We can't freeze indirects if there is a possibility that they * may be modified in the current syncing context. */ if (db->db_buf != NULL && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { arc_buf_freeze(db->db_buf); } if (holds == db->db_dirtycnt && db->db_level == 0 && db->db_user_immediate_evict) dbuf_evict_user(db); if (holds == 0) { if (db->db_blkid == DMU_BONUS_BLKID) { dnode_t *dn; boolean_t evict_dbuf = db->db_pending_evict; /* * If the dnode moves here, we cannot cross this * barrier until the move completes. */ DB_DNODE_ENTER(db); dn = DB_DNODE(db); atomic_dec_32(&dn->dn_dbufs_count); /* * Decrementing the dbuf count means that the bonus * buffer's dnode hold is no longer discounted in * dnode_move(). The dnode cannot move until after * the dnode_rele() below. */ DB_DNODE_EXIT(db); /* * Do not reference db after its lock is dropped. * Another thread may evict it. */ mutex_exit(&db->db_mtx); if (evict_dbuf) dnode_evict_bonus(dn); dnode_rele(dn, db); } else if (db->db_buf == NULL) { /* * This is a special case: we never associated this * dbuf with any data allocated from the ARC. */ ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); dbuf_destroy(db); } else if (arc_released(db->db_buf)) { /* * This dbuf has anonymous data associated with it. */ dbuf_destroy(db); } else { boolean_t do_arc_evict = B_FALSE; blkptr_t bp; spa_t *spa = dmu_objset_spa(db->db_objset); if (!DBUF_IS_CACHEABLE(db) && db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr) && !BP_IS_EMBEDDED(db->db_blkptr)) { do_arc_evict = B_TRUE; bp = *db->db_blkptr; } if (!DBUF_IS_CACHEABLE(db) || db->db_pending_evict) { dbuf_destroy(db); } else if (!multilist_link_active(&db->db_cache_link)) { ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE); dbuf_cached_state_t dcs = dbuf_include_in_metadata_cache(db) ? DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE; db->db_caching_status = dcs; multilist_insert(dbuf_caches[dcs].cache, db); (void) refcount_add_many(&dbuf_caches[dcs].size, db->db.db_size, db); mutex_exit(&db->db_mtx); if (db->db_caching_status == DB_DBUF_CACHE && !evicting) { dbuf_evict_notify(); } } if (do_arc_evict) arc_freed(spa, &bp); } } else { mutex_exit(&db->db_mtx); } } #pragma weak dmu_buf_refcount = dbuf_refcount uint64_t dbuf_refcount(dmu_buf_impl_t *db) { return (refcount_count(&db->db_holds)); } void * dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, dmu_buf_user_t *new_user) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; mutex_enter(&db->db_mtx); dbuf_verify_user(db, DBVU_NOT_EVICTING); if (db->db_user == old_user) db->db_user = new_user; else old_user = db->db_user; dbuf_verify_user(db, DBVU_NOT_EVICTING); mutex_exit(&db->db_mtx); return (old_user); } void * dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) { return (dmu_buf_replace_user(db_fake, NULL, user)); } void * dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; db->db_user_immediate_evict = TRUE; return (dmu_buf_set_user(db_fake, user)); } void * dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) { return (dmu_buf_replace_user(db_fake, user, NULL)); } void * dmu_buf_get_user(dmu_buf_t *db_fake) { dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; dbuf_verify_user(db, DBVU_NOT_EVICTING); return (db->db_user); } void dmu_buf_user_evict_wait() { taskq_wait(dbu_evict_taskq); } blkptr_t * dmu_buf_get_blkptr(dmu_buf_t *db) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; return (dbi->db_blkptr); } objset_t * dmu_buf_get_objset(dmu_buf_t *db) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; return (dbi->db_objset); } dnode_t * dmu_buf_dnode_enter(dmu_buf_t *db) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; DB_DNODE_ENTER(dbi); return (DB_DNODE(dbi)); } void dmu_buf_dnode_exit(dmu_buf_t *db) { dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; DB_DNODE_EXIT(dbi); } static void dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) { /* ASSERT(dmu_tx_is_syncing(tx) */ ASSERT(MUTEX_HELD(&db->db_mtx)); if (db->db_blkptr != NULL) return; if (db->db_blkid == DMU_SPILL_BLKID) { db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); BP_ZERO(db->db_blkptr); return; } if (db->db_level == dn->dn_phys->dn_nlevels-1) { /* * This buffer was allocated at a time when there was * no available blkptrs from the dnode, or it was * inappropriate to hook it in (i.e., nlevels mis-match). */ ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); ASSERT(db->db_parent == NULL); db->db_parent = dn->dn_dbuf; db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; DBUF_VERIFY(db); } else { dmu_buf_impl_t *parent = db->db_parent; int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; ASSERT(dn->dn_phys->dn_nlevels > 1); if (parent == NULL) { mutex_exit(&db->db_mtx); rw_enter(&dn->dn_struct_rwlock, RW_READER); parent = dbuf_hold_level(dn, db->db_level + 1, db->db_blkid >> epbs, db); rw_exit(&dn->dn_struct_rwlock); mutex_enter(&db->db_mtx); db->db_parent = parent; } db->db_blkptr = (blkptr_t *)parent->db.db_data + (db->db_blkid & ((1ULL << epbs) - 1)); DBUF_VERIFY(db); } } -static void +/* + * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it + * is critical the we not allow the compiler to inline this function in to + * dbuf_sync_list() thereby drastically bloating the stack usage. + */ +noinline static void dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) { dmu_buf_impl_t *db = dr->dr_dbuf; dnode_t *dn; zio_t *zio; ASSERT(dmu_tx_is_syncing(tx)); dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); mutex_enter(&db->db_mtx); ASSERT(db->db_level > 0); DBUF_VERIFY(db); /* Read the block if it hasn't been read yet. */ if (db->db_buf == NULL) { mutex_exit(&db->db_mtx); (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); mutex_enter(&db->db_mtx); } ASSERT3U(db->db_state, ==, DB_CACHED); ASSERT(db->db_buf != NULL); DB_DNODE_ENTER(db); dn = DB_DNODE(db); /* Indirect block size must match what the dnode thinks it is. */ ASSERT3U(db->db.db_size, ==, 1<dn_phys->dn_indblkshift); dbuf_check_blkptr(dn, db); DB_DNODE_EXIT(db); /* Provide the pending dirty record to child dbufs */ db->db_data_pending = dr; mutex_exit(&db->db_mtx); dbuf_write(dr, db->db_buf, tx); zio = dr->dr_zio; mutex_enter(&dr->dt.di.dr_mtx); dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); ASSERT(list_head(&dr->dt.di.dr_children) == NULL); mutex_exit(&dr->dt.di.dr_mtx); zio_nowait(zio); } -static void +/* + * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is + * critical the we not allow the compiler to inline this function in to + * dbuf_sync_list() thereby drastically bloating the stack usage. + */ +noinline static void dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) { arc_buf_t **datap = &dr->dt.dl.dr_data; dmu_buf_impl_t *db = dr->dr_dbuf; dnode_t *dn; objset_t *os; uint64_t txg = tx->tx_txg; ASSERT(dmu_tx_is_syncing(tx)); dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); mutex_enter(&db->db_mtx); /* * To be synced, we must be dirtied. But we * might have been freed after the dirty. */ if (db->db_state == DB_UNCACHED) { /* This buffer has been freed since it was dirtied */ ASSERT(db->db.db_data == NULL); } else if (db->db_state == DB_FILL) { /* This buffer was freed and is now being re-filled */ ASSERT(db->db.db_data != dr->dt.dl.dr_data); } else { ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); } DBUF_VERIFY(db); DB_DNODE_ENTER(db); dn = DB_DNODE(db); if (db->db_blkid == DMU_SPILL_BLKID) { mutex_enter(&dn->dn_mtx); if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { /* * In the previous transaction group, the bonus buffer * was entirely used to store the attributes for the * dnode which overrode the dn_spill field. However, * when adding more attributes to the file a spill * block was required to hold the extra attributes. * * Make sure to clear the garbage left in the dn_spill * field from the previous attributes in the bonus * buffer. Otherwise, after writing out the spill * block to the new allocated dva, it will free * the old block pointed to by the invalid dn_spill. */ db->db_blkptr = NULL; } dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; mutex_exit(&dn->dn_mtx); } /* * If this is a bonus buffer, simply copy the bonus data into the * dnode. It will be written out when the dnode is synced (and it * will be synced, since it must have been dirty for dbuf_sync to * be called). */ if (db->db_blkid == DMU_BONUS_BLKID) { dbuf_dirty_record_t **drp; ASSERT(*datap != NULL); ASSERT0(db->db_level); ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); bcopy(*datap, DN_BONUS(dn->dn_phys), DN_MAX_BONUS_LEN(dn->dn_phys)); DB_DNODE_EXIT(db); if (*datap != db->db.db_data) { int slots = DB_DNODE(db)->dn_num_slots; int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); zio_buf_free(*datap, bonuslen); arc_space_return(bonuslen, ARC_SPACE_BONUS); } db->db_data_pending = NULL; drp = &db->db_last_dirty; while (*drp != dr) drp = &(*drp)->dr_next; ASSERT(dr->dr_next == NULL); ASSERT(dr->dr_dbuf == db); *drp = dr->dr_next; if (dr->dr_dbuf->db_level != 0) { mutex_destroy(&dr->dt.di.dr_mtx); list_destroy(&dr->dt.di.dr_children); } kmem_free(dr, sizeof (dbuf_dirty_record_t)); ASSERT(db->db_dirtycnt > 0); db->db_dirtycnt -= 1; dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE); return; } os = dn->dn_objset; /* * This function may have dropped the db_mtx lock allowing a dmu_sync * operation to sneak in. As a result, we need to ensure that we * don't check the dr_override_state until we have returned from * dbuf_check_blkptr. */ dbuf_check_blkptr(dn, db); /* * If this buffer is in the middle of an immediate write, * wait for the synchronous IO to complete. */ while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); cv_wait(&db->db_changed, &db->db_mtx); ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); } if (db->db_state != DB_NOFILL && dn->dn_object != DMU_META_DNODE_OBJECT && refcount_count(&db->db_holds) > 1 && dr->dt.dl.dr_override_state != DR_OVERRIDDEN && *datap == db->db_buf) { /* * If this buffer is currently "in use" (i.e., there * are active holds and db_data still references it), * then make a copy before we start the write so that * any modifications from the open txg will not leak * into this write. * * NOTE: this copy does not need to be made for * objects only modified in the syncing context (e.g. * DNONE_DNODE blocks). */ int psize = arc_buf_size(*datap); arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); enum zio_compress compress_type = arc_get_compression(*datap); if (compress_type == ZIO_COMPRESS_OFF) { *datap = arc_alloc_buf(os->os_spa, db, type, psize); } else { ASSERT3U(type, ==, ARC_BUFC_DATA); int lsize = arc_buf_lsize(*datap); *datap = arc_alloc_compressed_buf(os->os_spa, db, psize, lsize, compress_type); } bcopy(db->db.db_data, (*datap)->b_data, psize); } db->db_data_pending = dr; mutex_exit(&db->db_mtx); dbuf_write(dr, *datap, tx); ASSERT(!list_link_active(&dr->dr_dirty_node)); if (dn->dn_object == DMU_META_DNODE_OBJECT) { list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); DB_DNODE_EXIT(db); } else { /* * Although zio_nowait() does not "wait for an IO", it does * initiate the IO. If this is an empty write it seems plausible * that the IO could actually be completed before the nowait * returns. We need to DB_DNODE_EXIT() first in case * zio_nowait() invalidates the dbuf. */ DB_DNODE_EXIT(db); zio_nowait(dr->dr_zio); } } void dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) { dbuf_dirty_record_t *dr; while (dr = list_head(list)) { if (dr->dr_zio != NULL) { /* * If we find an already initialized zio then we * are processing the meta-dnode, and we have finished. * The dbufs for all dnodes are put back on the list * during processing, so that we can zio_wait() * these IOs after initiating all child IOs. */ ASSERT3U(dr->dr_dbuf->db.db_object, ==, DMU_META_DNODE_OBJECT); break; } if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { VERIFY3U(dr->dr_dbuf->db_level, ==, level); } list_remove(list, dr); if (dr->dr_dbuf->db_level > 0) dbuf_sync_indirect(dr, tx); else dbuf_sync_leaf(dr, tx); } } /* ARGSUSED */ static void dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) { dmu_buf_impl_t *db = vdb; dnode_t *dn; blkptr_t *bp = zio->io_bp; blkptr_t *bp_orig = &zio->io_bp_orig; spa_t *spa = zio->io_spa; int64_t delta; uint64_t fill = 0; int i; ASSERT3P(db->db_blkptr, !=, NULL); ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); DB_DNODE_ENTER(db); dn = DB_DNODE(db); delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); dnode_diduse_space(dn, delta - zio->io_prev_space_delta); zio->io_prev_space_delta = delta; if (bp->blk_birth != 0) { ASSERT((db->db_blkid != DMU_SPILL_BLKID && BP_GET_TYPE(bp) == dn->dn_type) || (db->db_blkid == DMU_SPILL_BLKID && BP_GET_TYPE(bp) == dn->dn_bonustype) || BP_IS_EMBEDDED(bp)); ASSERT(BP_GET_LEVEL(bp) == db->db_level); } mutex_enter(&db->db_mtx); #ifdef ZFS_DEBUG if (db->db_blkid == DMU_SPILL_BLKID) { ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); ASSERT(!(BP_IS_HOLE(bp)) && db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); } #endif if (db->db_level == 0) { mutex_enter(&dn->dn_mtx); if (db->db_blkid > dn->dn_phys->dn_maxblkid && db->db_blkid != DMU_SPILL_BLKID) dn->dn_phys->dn_maxblkid = db->db_blkid; mutex_exit(&dn->dn_mtx); if (dn->dn_type == DMU_OT_DNODE) { i = 0; while (i < db->db.db_size) { dnode_phys_t *dnp = db->db.db_data + i; i += DNODE_MIN_SIZE; if (dnp->dn_type != DMU_OT_NONE) { fill++; i += dnp->dn_extra_slots * DNODE_MIN_SIZE; } } } else { if (BP_IS_HOLE(bp)) { fill = 0; } else { fill = 1; } } } else { blkptr_t *ibp = db->db.db_data; ASSERT3U(db->db.db_size, ==, 1<dn_phys->dn_indblkshift); for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { if (BP_IS_HOLE(ibp)) continue; fill += BP_GET_FILL(ibp); } } DB_DNODE_EXIT(db); if (!BP_IS_EMBEDDED(bp)) bp->blk_fill = fill; mutex_exit(&db->db_mtx); rw_enter(&dn->dn_struct_rwlock, RW_WRITER); *db->db_blkptr = *bp; rw_exit(&dn->dn_struct_rwlock); } /* ARGSUSED */ /* * This function gets called just prior to running through the compression * stage of the zio pipeline. If we're an indirect block comprised of only * holes, then we want this indirect to be compressed away to a hole. In * order to do that we must zero out any information about the holes that * this indirect points to prior to before we try to compress it. */ static void dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) { dmu_buf_impl_t *db = vdb; dnode_t *dn; blkptr_t *bp; unsigned int epbs, i; ASSERT3U(db->db_level, >, 0); DB_DNODE_ENTER(db); dn = DB_DNODE(db); epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; ASSERT3U(epbs, <, 31); /* Determine if all our children are holes */ for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { if (!BP_IS_HOLE(bp)) break; } /* * If all the children are holes, then zero them all out so that * we may get compressed away. */ if (i == 1 << epbs) { /* * We only found holes. Grab the rwlock to prevent * anybody from reading the blocks we're about to * zero out. */ rw_enter(&dn->dn_struct_rwlock, RW_WRITER); bzero(db->db.db_data, db->db.db_size); rw_exit(&dn->dn_struct_rwlock); } DB_DNODE_EXIT(db); } /* * The SPA will call this callback several times for each zio - once * for every physical child i/o (zio->io_phys_children times). This * allows the DMU to monitor the progress of each logical i/o. For example, * there may be 2 copies of an indirect block, or many fragments of a RAID-Z * block. There may be a long delay before all copies/fragments are completed, * so this callback allows us to retire dirty space gradually, as the physical * i/os complete. */ /* ARGSUSED */ static void dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) { dmu_buf_impl_t *db = arg; objset_t *os = db->db_objset; dsl_pool_t *dp = dmu_objset_pool(os); dbuf_dirty_record_t *dr; int delta = 0; dr = db->db_data_pending; ASSERT3U(dr->dr_txg, ==, zio->io_txg); /* * The callback will be called io_phys_children times. Retire one * portion of our dirty space each time we are called. Any rounding * error will be cleaned up by dsl_pool_sync()'s call to * dsl_pool_undirty_space(). */ delta = dr->dr_accounted / zio->io_phys_children; dsl_pool_undirty_space(dp, delta, zio->io_txg); } /* ARGSUSED */ static void dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) { dmu_buf_impl_t *db = vdb; blkptr_t *bp_orig = &zio->io_bp_orig; blkptr_t *bp = db->db_blkptr; objset_t *os = db->db_objset; dmu_tx_t *tx = os->os_synctx; dbuf_dirty_record_t **drp, *dr; ASSERT0(zio->io_error); ASSERT(db->db_blkptr == bp); /* * For nopwrites and rewrites we ensure that the bp matches our * original and bypass all the accounting. */ if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { ASSERT(BP_EQUAL(bp, bp_orig)); } else { dsl_dataset_t *ds = os->os_dsl_dataset; (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); dsl_dataset_block_born(ds, bp, tx); } mutex_enter(&db->db_mtx); DBUF_VERIFY(db); drp = &db->db_last_dirty; while ((dr = *drp) != db->db_data_pending) drp = &dr->dr_next; ASSERT(!list_link_active(&dr->dr_dirty_node)); ASSERT(dr->dr_dbuf == db); ASSERT(dr->dr_next == NULL); *drp = dr->dr_next; #ifdef ZFS_DEBUG if (db->db_blkid == DMU_SPILL_BLKID) { dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); DB_DNODE_EXIT(db); } #endif if (db->db_level == 0) { ASSERT(db->db_blkid != DMU_BONUS_BLKID); ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); if (db->db_state != DB_NOFILL) { if (dr->dt.dl.dr_data != db->db_buf) arc_buf_destroy(dr->dt.dl.dr_data, db); } } else { dnode_t *dn; DB_DNODE_ENTER(db); dn = DB_DNODE(db); ASSERT(list_head(&dr->dt.di.dr_children) == NULL); ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); if (!BP_IS_HOLE(db->db_blkptr)) { int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; ASSERT3U(db->db_blkid, <=, dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, db->db.db_size); } DB_DNODE_EXIT(db); mutex_destroy(&dr->dt.di.dr_mtx); list_destroy(&dr->dt.di.dr_children); } kmem_free(dr, sizeof (dbuf_dirty_record_t)); cv_broadcast(&db->db_changed); ASSERT(db->db_dirtycnt > 0); db->db_dirtycnt -= 1; db->db_data_pending = NULL; dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE); } static void dbuf_write_nofill_ready(zio_t *zio) { dbuf_write_ready(zio, NULL, zio->io_private); } static void dbuf_write_nofill_done(zio_t *zio) { dbuf_write_done(zio, NULL, zio->io_private); } static void dbuf_write_override_ready(zio_t *zio) { dbuf_dirty_record_t *dr = zio->io_private; dmu_buf_impl_t *db = dr->dr_dbuf; dbuf_write_ready(zio, NULL, db); } static void dbuf_write_override_done(zio_t *zio) { dbuf_dirty_record_t *dr = zio->io_private; dmu_buf_impl_t *db = dr->dr_dbuf; blkptr_t *obp = &dr->dt.dl.dr_overridden_by; mutex_enter(&db->db_mtx); if (!BP_EQUAL(zio->io_bp, obp)) { if (!BP_IS_HOLE(obp)) dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); arc_release(dr->dt.dl.dr_data, db); } mutex_exit(&db->db_mtx); dbuf_write_done(zio, NULL, db); if (zio->io_abd != NULL) abd_put(zio->io_abd); } typedef struct dbuf_remap_impl_callback_arg { objset_t *drica_os; uint64_t drica_blk_birth; dmu_tx_t *drica_tx; } dbuf_remap_impl_callback_arg_t; static void dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size, void *arg) { dbuf_remap_impl_callback_arg_t *drica = arg; objset_t *os = drica->drica_os; spa_t *spa = dmu_objset_spa(os); dmu_tx_t *tx = drica->drica_tx; ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); if (os == spa_meta_objset(spa)) { spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx); } else { dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset, size, drica->drica_blk_birth, tx); } } static void dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx) { blkptr_t bp_copy = *bp; spa_t *spa = dmu_objset_spa(dn->dn_objset); dbuf_remap_impl_callback_arg_t drica; ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); drica.drica_os = dn->dn_objset; drica.drica_blk_birth = bp->blk_birth; drica.drica_tx = tx; if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback, &drica)) { /* * The struct_rwlock prevents dbuf_read_impl() from * dereferencing the BP while we are changing it. To * avoid lock contention, only grab it when we are actually * changing the BP. */ rw_enter(&dn->dn_struct_rwlock, RW_WRITER); *bp = bp_copy; rw_exit(&dn->dn_struct_rwlock); } } /* * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting * to remap a copy of every bp in the dbuf. */ boolean_t dbuf_can_remap(const dmu_buf_impl_t *db) { spa_t *spa = dmu_objset_spa(db->db_objset); blkptr_t *bp = db->db.db_data; boolean_t ret = B_FALSE; ASSERT3U(db->db_level, >, 0); ASSERT3S(db->db_state, ==, DB_CACHED); ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { blkptr_t bp_copy = bp[i]; if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { ret = B_TRUE; break; } } spa_config_exit(spa, SCL_VDEV, FTAG); return (ret); } boolean_t dnode_needs_remap(const dnode_t *dn) { spa_t *spa = dmu_objset_spa(dn->dn_objset); boolean_t ret = B_FALSE; if (dn->dn_phys->dn_nlevels == 0) { return (B_FALSE); } ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) { blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j]; if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { ret = B_TRUE; break; } } spa_config_exit(spa, SCL_VDEV, FTAG); return (ret); } /* * Remap any existing BP's to concrete vdevs, if possible. */ static void dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx) { spa_t *spa = dmu_objset_spa(db->db_objset); ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)) return; if (db->db_level > 0) { blkptr_t *bp = db->db.db_data; for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { dbuf_remap_impl(dn, &bp[i], tx); } } else if (db->db.db_object == DMU_META_DNODE_OBJECT) { dnode_phys_t *dnp = db->db.db_data; ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==, DMU_OT_DNODE); for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; i++) { for (int j = 0; j < dnp[i].dn_nblkptr; j++) { dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx); } } } } /* Issue I/O to commit a dirty buffer to disk. */ static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) { dmu_buf_impl_t *db = dr->dr_dbuf; dnode_t *dn; objset_t *os; dmu_buf_impl_t *parent = db->db_parent; uint64_t txg = tx->tx_txg; zbookmark_phys_t zb; zio_prop_t zp; zio_t *zio; int wp_flag = 0; ASSERT(dmu_tx_is_syncing(tx)); DB_DNODE_ENTER(db); dn = DB_DNODE(db); os = dn->dn_objset; if (db->db_state != DB_NOFILL) { if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { /* * Private object buffers are released here rather * than in dbuf_dirty() since they are only modified * in the syncing context and we don't want the * overhead of making multiple copies of the data. */ if (BP_IS_HOLE(db->db_blkptr)) { arc_buf_thaw(data); } else { dbuf_release_bp(db); } dbuf_remap(dn, db, tx); } } if (parent != dn->dn_dbuf) { /* Our parent is an indirect block. */ /* We have a dirty parent that has been scheduled for write. */ ASSERT(parent && parent->db_data_pending); /* Our parent's buffer is one level closer to the dnode. */ ASSERT(db->db_level == parent->db_level-1); /* * We're about to modify our parent's db_data by modifying * our block pointer, so the parent must be released. */ ASSERT(arc_released(parent->db_buf)); zio = parent->db_data_pending->dr_zio; } else { /* Our parent is the dnode itself. */ ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && db->db_blkid != DMU_SPILL_BLKID) || (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); if (db->db_blkid != DMU_SPILL_BLKID) ASSERT3P(db->db_blkptr, ==, &dn->dn_phys->dn_blkptr[db->db_blkid]); zio = dn->dn_zio; } ASSERT(db->db_level == 0 || data == db->db_buf); ASSERT3U(db->db_blkptr->blk_birth, <=, txg); ASSERT(zio); SET_BOOKMARK(&zb, os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : DMU_META_OBJSET, db->db.db_object, db->db_level, db->db_blkid); if (db->db_blkid == DMU_SPILL_BLKID) wp_flag = WP_SPILL; wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); DB_DNODE_EXIT(db); /* * We copy the blkptr now (rather than when we instantiate the dirty * record), because its value can change between open context and * syncing context. We do not need to hold dn_struct_rwlock to read * db_blkptr because we are in syncing context. */ dr->dr_bp_copy = *db->db_blkptr; if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { /* * The BP for this block has been provided by open context * (by dmu_sync() or dmu_buf_write_embedded()). */ abd_t *contents = (data != NULL) ? abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size, &zp, dbuf_write_override_ready, NULL, NULL, dbuf_write_override_done, dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); mutex_enter(&db->db_mtx); dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); mutex_exit(&db->db_mtx); } else if (db->db_state == DB_NOFILL) { ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, dbuf_write_nofill_ready, NULL, NULL, dbuf_write_nofill_done, db, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); } else { ASSERT(arc_released(data)); /* * For indirect blocks, we want to setup the children * ready callback so that we can properly handle an indirect * block that only contains holes. */ arc_write_done_func_t *children_ready_cb = NULL; if (db->db_level != 0) children_ready_cb = dbuf_write_children_ready; dr->dr_zio = arc_write(zio, os->os_spa, txg, &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), &zp, dbuf_write_ready, children_ready_cb, dbuf_write_physdone, dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); } } Index: head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_send.c =================================================================== --- head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_send.c (revision 337671) +++ head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_send.c (revision 337672) @@ -1,3456 +1,3455 @@ /* * 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 2011 Nexenta Systems, Inc. All rights reserved. * Copyright (c) 2011, 2015 by Delphix. All rights reserved. * Copyright (c) 2014, Joyent, Inc. All rights reserved. * Copyright (c) 2012, Martin Matuska . All rights reserved. * Copyright 2014 HybridCluster. All rights reserved. * Copyright 2016 RackTop Systems. * Copyright (c) 2014 Integros [integros.com] */ #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 #ifdef __FreeBSD__ #undef dump_write #define dump_write dmu_dump_write #endif /* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */ int zfs_send_corrupt_data = B_FALSE; int zfs_send_queue_length = 16 * 1024 * 1024; int zfs_recv_queue_length = 16 * 1024 * 1024; /* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */ int zfs_send_set_freerecords_bit = B_TRUE; #ifdef _KERNEL TUNABLE_INT("vfs.zfs.send_set_freerecords_bit", &zfs_send_set_freerecords_bit); #endif static char *dmu_recv_tag = "dmu_recv_tag"; const char *recv_clone_name = "%recv"; /* * Use this to override the recordsize calculation for fast zfs send estimates. */ uint64_t zfs_override_estimate_recordsize = 0; #define BP_SPAN(datablkszsec, indblkshift, level) \ (((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \ (level) * (indblkshift - SPA_BLKPTRSHIFT))) static void byteswap_record(dmu_replay_record_t *drr); struct send_thread_arg { bqueue_t q; dsl_dataset_t *ds; /* Dataset to traverse */ uint64_t fromtxg; /* Traverse from this txg */ int flags; /* flags to pass to traverse_dataset */ int error_code; boolean_t cancel; zbookmark_phys_t resume; }; struct send_block_record { boolean_t eos_marker; /* Marks the end of the stream */ blkptr_t bp; zbookmark_phys_t zb; uint8_t indblkshift; uint16_t datablkszsec; bqueue_node_t ln; }; static int dump_bytes(dmu_sendarg_t *dsp, void *buf, int len) { dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os); struct uio auio; struct iovec aiov; /* * The code does not rely on this (len being a multiple of 8). We keep * this assertion because of the corresponding assertion in * receive_read(). Keeping this assertion ensures that we do not * inadvertently break backwards compatibility (causing the assertion * in receive_read() to trigger on old software). * * Removing the assertions could be rolled into a new feature that uses * data that isn't 8-byte aligned; if the assertions were removed, a * feature flag would have to be added. */ ASSERT0(len % 8); aiov.iov_base = buf; aiov.iov_len = len; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_resid = len; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_WRITE; auio.uio_offset = (off_t)-1; auio.uio_td = dsp->dsa_td; #ifdef _KERNEL if (dsp->dsa_fp->f_type == DTYPE_VNODE) bwillwrite(); dsp->dsa_err = fo_write(dsp->dsa_fp, &auio, dsp->dsa_td->td_ucred, 0, dsp->dsa_td); #else fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__); dsp->dsa_err = EOPNOTSUPP; #endif mutex_enter(&ds->ds_sendstream_lock); *dsp->dsa_off += len; mutex_exit(&ds->ds_sendstream_lock); return (dsp->dsa_err); } /* * For all record types except BEGIN, fill in the checksum (overlaid in * drr_u.drr_checksum.drr_checksum). The checksum verifies everything * up to the start of the checksum itself. */ static int dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len) { ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); (void) fletcher_4_incremental_native(dsp->dsa_drr, offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), &dsp->dsa_zc); if (dsp->dsa_drr->drr_type == DRR_BEGIN) { dsp->dsa_sent_begin = B_TRUE; } else { ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u. drr_checksum.drr_checksum)); dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc; } if (dsp->dsa_drr->drr_type == DRR_END) { dsp->dsa_sent_end = B_TRUE; } (void) fletcher_4_incremental_native(&dsp->dsa_drr-> drr_u.drr_checksum.drr_checksum, sizeof (zio_cksum_t), &dsp->dsa_zc); if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0) return (SET_ERROR(EINTR)); if (payload_len != 0) { (void) fletcher_4_incremental_native(payload, payload_len, &dsp->dsa_zc); if (dump_bytes(dsp, payload, payload_len) != 0) return (SET_ERROR(EINTR)); } return (0); } /* * Fill in the drr_free struct, or perform aggregation if the previous record is * also a free record, and the two are adjacent. * * Note that we send free records even for a full send, because we want to be * able to receive a full send as a clone, which requires a list of all the free * and freeobject records that were generated on the source. */ static int dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, uint64_t length) { struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free); /* * When we receive a free record, dbuf_free_range() assumes * that the receiving system doesn't have any dbufs in the range * being freed. This is always true because there is a one-record * constraint: we only send one WRITE record for any given * object,offset. We know that the one-record constraint is * true because we always send data in increasing order by * object,offset. * * If the increasing-order constraint ever changes, we should find * another way to assert that the one-record constraint is still * satisfied. */ ASSERT(object > dsp->dsa_last_data_object || (object == dsp->dsa_last_data_object && offset > dsp->dsa_last_data_offset)); if (length != -1ULL && offset + length < offset) length = -1ULL; /* * If there is a pending op, but it's not PENDING_FREE, push it out, * since free block aggregation can only be done for blocks of the * same type (i.e., DRR_FREE records can only be aggregated with * other DRR_FREE records. DRR_FREEOBJECTS records can only be * aggregated with other DRR_FREEOBJECTS records. */ if (dsp->dsa_pending_op != PENDING_NONE && dsp->dsa_pending_op != PENDING_FREE) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } if (dsp->dsa_pending_op == PENDING_FREE) { /* * There should never be a PENDING_FREE if length is -1 * (because dump_dnode is the only place where this * function is called with a -1, and only after flushing * any pending record). */ ASSERT(length != -1ULL); /* * Check to see whether this free block can be aggregated * with pending one. */ if (drrf->drr_object == object && drrf->drr_offset + drrf->drr_length == offset) { drrf->drr_length += length; return (0); } else { /* not a continuation. Push out pending record */ if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } } /* create a FREE record and make it pending */ bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_FREE; drrf->drr_object = object; drrf->drr_offset = offset; drrf->drr_length = length; drrf->drr_toguid = dsp->dsa_toguid; if (length == -1ULL) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); } else { dsp->dsa_pending_op = PENDING_FREE; } return (0); } static int dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type, uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp, void *data) { uint64_t payload_size; struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write); /* * We send data in increasing object, offset order. * See comment in dump_free() for details. */ ASSERT(object > dsp->dsa_last_data_object || (object == dsp->dsa_last_data_object && offset > dsp->dsa_last_data_offset)); dsp->dsa_last_data_object = object; dsp->dsa_last_data_offset = offset + lsize - 1; /* * If there is any kind of pending aggregation (currently either * a grouping of free objects or free blocks), push it out to * the stream, since aggregation can't be done across operations * of different types. */ if (dsp->dsa_pending_op != PENDING_NONE) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } /* write a WRITE record */ bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_WRITE; drrw->drr_object = object; drrw->drr_type = type; drrw->drr_offset = offset; drrw->drr_toguid = dsp->dsa_toguid; drrw->drr_logical_size = lsize; /* only set the compression fields if the buf is compressed */ if (lsize != psize) { ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED); ASSERT(!BP_IS_EMBEDDED(bp)); ASSERT(!BP_SHOULD_BYTESWAP(bp)); ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp))); ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF); ASSERT3S(psize, >, 0); ASSERT3S(lsize, >=, psize); drrw->drr_compressiontype = BP_GET_COMPRESS(bp); drrw->drr_compressed_size = psize; payload_size = drrw->drr_compressed_size; } else { payload_size = drrw->drr_logical_size; } if (bp == NULL || BP_IS_EMBEDDED(bp)) { /* * There's no pre-computed checksum for partial-block * writes or embedded BP's, so (like * fletcher4-checkummed blocks) userland will have to * compute a dedup-capable checksum itself. */ drrw->drr_checksumtype = ZIO_CHECKSUM_OFF; } else { drrw->drr_checksumtype = BP_GET_CHECKSUM(bp); if (zio_checksum_table[drrw->drr_checksumtype].ci_flags & ZCHECKSUM_FLAG_DEDUP) drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP; DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp)); DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp)); DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp)); drrw->drr_key.ddk_cksum = bp->blk_cksum; } if (dump_record(dsp, data, payload_size) != 0) return (SET_ERROR(EINTR)); return (0); } static int dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset, int blksz, const blkptr_t *bp) { char buf[BPE_PAYLOAD_SIZE]; struct drr_write_embedded *drrw = &(dsp->dsa_drr->drr_u.drr_write_embedded); if (dsp->dsa_pending_op != PENDING_NONE) { if (dump_record(dsp, NULL, 0) != 0) return (EINTR); dsp->dsa_pending_op = PENDING_NONE; } ASSERT(BP_IS_EMBEDDED(bp)); bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED; drrw->drr_object = object; drrw->drr_offset = offset; drrw->drr_length = blksz; drrw->drr_toguid = dsp->dsa_toguid; drrw->drr_compression = BP_GET_COMPRESS(bp); drrw->drr_etype = BPE_GET_ETYPE(bp); drrw->drr_lsize = BPE_GET_LSIZE(bp); drrw->drr_psize = BPE_GET_PSIZE(bp); decode_embedded_bp_compressed(bp, buf); if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0) return (EINTR); return (0); } static int dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data) { struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill); if (dsp->dsa_pending_op != PENDING_NONE) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } /* write a SPILL record */ bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_SPILL; drrs->drr_object = object; drrs->drr_length = blksz; drrs->drr_toguid = dsp->dsa_toguid; if (dump_record(dsp, data, blksz) != 0) return (SET_ERROR(EINTR)); return (0); } static int dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs) { struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects); /* * If there is a pending op, but it's not PENDING_FREEOBJECTS, * push it out, since free block aggregation can only be done for * blocks of the same type (i.e., DRR_FREE records can only be * aggregated with other DRR_FREE records. DRR_FREEOBJECTS records * can only be aggregated with other DRR_FREEOBJECTS records. */ if (dsp->dsa_pending_op != PENDING_NONE && dsp->dsa_pending_op != PENDING_FREEOBJECTS) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) { /* * See whether this free object array can be aggregated * with pending one */ if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) { drrfo->drr_numobjs += numobjs; return (0); } else { /* can't be aggregated. Push out pending record */ if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } } /* write a FREEOBJECTS record */ bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_FREEOBJECTS; drrfo->drr_firstobj = firstobj; drrfo->drr_numobjs = numobjs; drrfo->drr_toguid = dsp->dsa_toguid; dsp->dsa_pending_op = PENDING_FREEOBJECTS; return (0); } static int dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp) { struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object); if (object < dsp->dsa_resume_object) { /* * Note: when resuming, we will visit all the dnodes in * the block of dnodes that we are resuming from. In * this case it's unnecessary to send the dnodes prior to * the one we are resuming from. We should be at most one * block's worth of dnodes behind the resume point. */ ASSERT3U(dsp->dsa_resume_object - object, <, 1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT)); return (0); } if (dnp == NULL || dnp->dn_type == DMU_OT_NONE) return (dump_freeobjects(dsp, object, 1)); if (dsp->dsa_pending_op != PENDING_NONE) { if (dump_record(dsp, NULL, 0) != 0) return (SET_ERROR(EINTR)); dsp->dsa_pending_op = PENDING_NONE; } /* write an OBJECT record */ bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t)); dsp->dsa_drr->drr_type = DRR_OBJECT; drro->drr_object = object; drro->drr_type = dnp->dn_type; drro->drr_bonustype = dnp->dn_bonustype; drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; drro->drr_bonuslen = dnp->dn_bonuslen; drro->drr_dn_slots = dnp->dn_extra_slots + 1; drro->drr_checksumtype = dnp->dn_checksum; drro->drr_compress = dnp->dn_compress; drro->drr_toguid = dsp->dsa_toguid; if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE) drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE; if (dump_record(dsp, DN_BONUS(dnp), P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) { return (SET_ERROR(EINTR)); } /* Free anything past the end of the file. */ if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) * (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0) return (SET_ERROR(EINTR)); if (dsp->dsa_err != 0) return (SET_ERROR(EINTR)); return (0); } static boolean_t backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp) { if (!BP_IS_EMBEDDED(bp)) return (B_FALSE); /* * Compression function must be legacy, or explicitly enabled. */ if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS && !(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4))) return (B_FALSE); /* * Embed type must be explicitly enabled. */ switch (BPE_GET_ETYPE(bp)) { case BP_EMBEDDED_TYPE_DATA: if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) return (B_TRUE); break; default: return (B_FALSE); } return (B_FALSE); } /* * This is the callback function to traverse_dataset that acts as the worker * thread for dmu_send_impl. */ /*ARGSUSED*/ static int send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg) { struct send_thread_arg *sta = arg; struct send_block_record *record; uint64_t record_size; int err = 0; ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || zb->zb_object >= sta->resume.zb_object); if (sta->cancel) return (SET_ERROR(EINTR)); if (bp == NULL) { ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL); return (0); } else if (zb->zb_level < 0) { return (0); } record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP); record->eos_marker = B_FALSE; record->bp = *bp; record->zb = *zb; record->indblkshift = dnp->dn_indblkshift; record->datablkszsec = dnp->dn_datablkszsec; record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT; bqueue_enqueue(&sta->q, record, record_size); return (err); } /* * This function kicks off the traverse_dataset. It also handles setting the * error code of the thread in case something goes wrong, and pushes the End of * Stream record when the traverse_dataset call has finished. If there is no * dataset to traverse, the thread immediately pushes End of Stream marker. */ static void send_traverse_thread(void *arg) { struct send_thread_arg *st_arg = arg; int err; struct send_block_record *data; if (st_arg->ds != NULL) { err = traverse_dataset_resume(st_arg->ds, st_arg->fromtxg, &st_arg->resume, st_arg->flags, send_cb, st_arg); if (err != EINTR) st_arg->error_code = err; } data = kmem_zalloc(sizeof (*data), KM_SLEEP); data->eos_marker = B_TRUE; bqueue_enqueue(&st_arg->q, data, 1); thread_exit(); } /* * This function actually handles figuring out what kind of record needs to be * dumped, reading the data (which has hopefully been prefetched), and calling * the appropriate helper function. */ static int do_dump(dmu_sendarg_t *dsa, struct send_block_record *data) { dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os); const blkptr_t *bp = &data->bp; const zbookmark_phys_t *zb = &data->zb; uint8_t indblkshift = data->indblkshift; uint16_t dblkszsec = data->datablkszsec; spa_t *spa = ds->ds_dir->dd_pool->dp_spa; dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE; int err = 0; ASSERT3U(zb->zb_level, >=, 0); ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT || zb->zb_object >= dsa->dsa_resume_object); if (zb->zb_object != DMU_META_DNODE_OBJECT && DMU_OBJECT_IS_SPECIAL(zb->zb_object)) { return (0); } else if (BP_IS_HOLE(bp) && zb->zb_object == DMU_META_DNODE_OBJECT) { uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT; err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT); } else if (BP_IS_HOLE(bp)) { uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level); uint64_t offset = zb->zb_blkid * span; err = dump_free(dsa, zb->zb_object, offset, span); } else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) { return (0); } else if (type == DMU_OT_DNODE) { int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; arc_flags_t aflags = ARC_FLAG_WAIT; arc_buf_t *abuf; ASSERT0(zb->zb_level); if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &aflags, zb) != 0) return (SET_ERROR(EIO)); dnode_phys_t *blk = abuf->b_data; uint64_t dnobj = zb->zb_blkid * epb; for (int i = 0; i < epb; i += blk[i].dn_extra_slots + 1) { err = dump_dnode(dsa, dnobj + i, blk + i); if (err != 0) break; } arc_buf_destroy(abuf, &abuf); } else if (type == DMU_OT_SA) { arc_flags_t aflags = ARC_FLAG_WAIT; arc_buf_t *abuf; int blksz = BP_GET_LSIZE(bp); if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &aflags, zb) != 0) return (SET_ERROR(EIO)); err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data); arc_buf_destroy(abuf, &abuf); } else if (backup_do_embed(dsa, bp)) { /* it's an embedded level-0 block of a regular object */ int blksz = dblkszsec << SPA_MINBLOCKSHIFT; ASSERT0(zb->zb_level); err = dump_write_embedded(dsa, zb->zb_object, zb->zb_blkid * blksz, blksz, bp); } else { /* it's a level-0 block of a regular object */ arc_flags_t aflags = ARC_FLAG_WAIT; arc_buf_t *abuf; int blksz = dblkszsec << SPA_MINBLOCKSHIFT; uint64_t offset; /* * If we have large blocks stored on disk but the send flags * don't allow us to send large blocks, we split the data from * the arc buf into chunks. */ boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE && !(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS); /* * We should only request compressed data from the ARC if all * the following are true: * - stream compression was requested * - we aren't splitting large blocks into smaller chunks * - the data won't need to be byteswapped before sending * - this isn't an embedded block * - this isn't metadata (if receiving on a different endian * system it can be byteswapped more easily) */ boolean_t request_compressed = (dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) && !split_large_blocks && !BP_SHOULD_BYTESWAP(bp) && !BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp)); ASSERT0(zb->zb_level); ASSERT(zb->zb_object > dsa->dsa_resume_object || (zb->zb_object == dsa->dsa_resume_object && zb->zb_blkid * blksz >= dsa->dsa_resume_offset)); ASSERT0(zb->zb_level); ASSERT(zb->zb_object > dsa->dsa_resume_object || (zb->zb_object == dsa->dsa_resume_object && zb->zb_blkid * blksz >= dsa->dsa_resume_offset)); ASSERT3U(blksz, ==, BP_GET_LSIZE(bp)); enum zio_flag zioflags = ZIO_FLAG_CANFAIL; if (request_compressed) zioflags |= ZIO_FLAG_RAW; if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf, ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) { if (zfs_send_corrupt_data) { /* Send a block filled with 0x"zfs badd bloc" */ abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA, blksz); uint64_t *ptr; for (ptr = abuf->b_data; (char *)ptr < (char *)abuf->b_data + blksz; ptr++) *ptr = 0x2f5baddb10cULL; } else { return (SET_ERROR(EIO)); } } offset = zb->zb_blkid * blksz; if (split_large_blocks) { ASSERT3U(arc_get_compression(abuf), ==, ZIO_COMPRESS_OFF); char *buf = abuf->b_data; while (blksz > 0 && err == 0) { int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE); err = dump_write(dsa, type, zb->zb_object, offset, n, n, NULL, buf); offset += n; buf += n; blksz -= n; } } else { err = dump_write(dsa, type, zb->zb_object, offset, blksz, arc_buf_size(abuf), bp, abuf->b_data); } arc_buf_destroy(abuf, &abuf); } ASSERT(err == 0 || err == EINTR); return (err); } /* * Pop the new data off the queue, and free the old data. */ static struct send_block_record * get_next_record(bqueue_t *bq, struct send_block_record *data) { struct send_block_record *tmp = bqueue_dequeue(bq); kmem_free(data, sizeof (*data)); return (tmp); } /* * Actually do the bulk of the work in a zfs send. * * Note: Releases dp using the specified tag. */ static int dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds, zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone, boolean_t embedok, boolean_t large_block_ok, boolean_t compressok, int outfd, uint64_t resumeobj, uint64_t resumeoff, #ifdef illumos vnode_t *vp, offset_t *off) #else struct file *fp, offset_t *off) #endif { objset_t *os; dmu_replay_record_t *drr; dmu_sendarg_t *dsp; int err; uint64_t fromtxg = 0; uint64_t featureflags = 0; struct send_thread_arg to_arg = { 0 }; err = dmu_objset_from_ds(to_ds, &os); if (err != 0) { dsl_pool_rele(dp, tag); return (err); } drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP); drr->drr_type = DRR_BEGIN; drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC; DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo, DMU_SUBSTREAM); #ifdef _KERNEL if (dmu_objset_type(os) == DMU_OST_ZFS) { uint64_t version; if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) { kmem_free(drr, sizeof (dmu_replay_record_t)); dsl_pool_rele(dp, tag); return (SET_ERROR(EINVAL)); } if (version >= ZPL_VERSION_SA) { featureflags |= DMU_BACKUP_FEATURE_SA_SPILL; } } #endif if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS]) featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS; if (to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE]) featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE; if (embedok && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) { featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA; if (spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) featureflags |= DMU_BACKUP_FEATURE_LZ4; } if (compressok) { featureflags |= DMU_BACKUP_FEATURE_COMPRESSED; } if ((featureflags & (DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) != 0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) { featureflags |= DMU_BACKUP_FEATURE_LZ4; } if (resumeobj != 0 || resumeoff != 0) { featureflags |= DMU_BACKUP_FEATURE_RESUMING; } DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo, featureflags); drr->drr_u.drr_begin.drr_creation_time = dsl_dataset_phys(to_ds)->ds_creation_time; drr->drr_u.drr_begin.drr_type = dmu_objset_type(os); if (is_clone) drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE; drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid; if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET) drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA; if (zfs_send_set_freerecords_bit) drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS; if (ancestor_zb != NULL) { drr->drr_u.drr_begin.drr_fromguid = ancestor_zb->zbm_guid; fromtxg = ancestor_zb->zbm_creation_txg; } dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname); if (!to_ds->ds_is_snapshot) { (void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--", sizeof (drr->drr_u.drr_begin.drr_toname)); } dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP); dsp->dsa_drr = drr; dsp->dsa_outfd = outfd; dsp->dsa_proc = curproc; dsp->dsa_td = curthread; dsp->dsa_fp = fp; dsp->dsa_os = os; dsp->dsa_off = off; dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid; dsp->dsa_pending_op = PENDING_NONE; dsp->dsa_featureflags = featureflags; dsp->dsa_resume_object = resumeobj; dsp->dsa_resume_offset = resumeoff; mutex_enter(&to_ds->ds_sendstream_lock); list_insert_head(&to_ds->ds_sendstreams, dsp); mutex_exit(&to_ds->ds_sendstream_lock); dsl_dataset_long_hold(to_ds, FTAG); dsl_pool_rele(dp, tag); void *payload = NULL; size_t payload_len = 0; if (resumeobj != 0 || resumeoff != 0) { dmu_object_info_t to_doi; err = dmu_object_info(os, resumeobj, &to_doi); if (err != 0) goto out; SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0, resumeoff / to_doi.doi_data_block_size); nvlist_t *nvl = fnvlist_alloc(); fnvlist_add_uint64(nvl, "resume_object", resumeobj); fnvlist_add_uint64(nvl, "resume_offset", resumeoff); payload = fnvlist_pack(nvl, &payload_len); drr->drr_payloadlen = payload_len; fnvlist_free(nvl); } err = dump_record(dsp, payload, payload_len); fnvlist_pack_free(payload, payload_len); if (err != 0) { err = dsp->dsa_err; goto out; } err = bqueue_init(&to_arg.q, zfs_send_queue_length, offsetof(struct send_block_record, ln)); to_arg.error_code = 0; to_arg.cancel = B_FALSE; to_arg.ds = to_ds; to_arg.fromtxg = fromtxg; to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH; (void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, &p0, TS_RUN, minclsyspri); struct send_block_record *to_data; to_data = bqueue_dequeue(&to_arg.q); while (!to_data->eos_marker && err == 0) { err = do_dump(dsp, to_data); to_data = get_next_record(&to_arg.q, to_data); if (issig(JUSTLOOKING) && issig(FORREAL)) err = EINTR; } if (err != 0) { to_arg.cancel = B_TRUE; while (!to_data->eos_marker) { to_data = get_next_record(&to_arg.q, to_data); } } kmem_free(to_data, sizeof (*to_data)); bqueue_destroy(&to_arg.q); if (err == 0 && to_arg.error_code != 0) err = to_arg.error_code; if (err != 0) goto out; if (dsp->dsa_pending_op != PENDING_NONE) if (dump_record(dsp, NULL, 0) != 0) err = SET_ERROR(EINTR); if (err != 0) { if (err == EINTR && dsp->dsa_err != 0) err = dsp->dsa_err; goto out; } bzero(drr, sizeof (dmu_replay_record_t)); drr->drr_type = DRR_END; drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc; drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid; if (dump_record(dsp, NULL, 0) != 0) err = dsp->dsa_err; out: mutex_enter(&to_ds->ds_sendstream_lock); list_remove(&to_ds->ds_sendstreams, dsp); mutex_exit(&to_ds->ds_sendstream_lock); VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end)); kmem_free(drr, sizeof (dmu_replay_record_t)); kmem_free(dsp, sizeof (dmu_sendarg_t)); dsl_dataset_long_rele(to_ds, FTAG); return (err); } int dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap, boolean_t embedok, boolean_t large_block_ok, boolean_t compressok, #ifdef illumos int outfd, vnode_t *vp, offset_t *off) #else int outfd, struct file *fp, offset_t *off) #endif { dsl_pool_t *dp; dsl_dataset_t *ds; dsl_dataset_t *fromds = NULL; int err; err = dsl_pool_hold(pool, FTAG, &dp); if (err != 0) return (err); err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds); if (err != 0) { dsl_pool_rele(dp, FTAG); return (err); } if (fromsnap != 0) { zfs_bookmark_phys_t zb; boolean_t is_clone; err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds); if (err != 0) { dsl_dataset_rele(ds, FTAG); dsl_pool_rele(dp, FTAG); return (err); } if (!dsl_dataset_is_before(ds, fromds, 0)) err = SET_ERROR(EXDEV); zb.zbm_creation_time = dsl_dataset_phys(fromds)->ds_creation_time; zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg; zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; is_clone = (fromds->ds_dir != ds->ds_dir); dsl_dataset_rele(fromds, FTAG); err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok, large_block_ok, compressok, outfd, 0, 0, fp, off); } else { err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok, large_block_ok, compressok, outfd, 0, 0, fp, off); } dsl_dataset_rele(ds, FTAG); return (err); } int dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok, boolean_t large_block_ok, boolean_t compressok, int outfd, uint64_t resumeobj, uint64_t resumeoff, #ifdef illumos vnode_t *vp, offset_t *off) #else struct file *fp, offset_t *off) #endif { dsl_pool_t *dp; dsl_dataset_t *ds; int err; boolean_t owned = B_FALSE; if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL) return (SET_ERROR(EINVAL)); err = dsl_pool_hold(tosnap, FTAG, &dp); if (err != 0) return (err); if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) { /* * We are sending a filesystem or volume. Ensure * that it doesn't change by owning the dataset. */ err = dsl_dataset_own(dp, tosnap, FTAG, &ds); owned = B_TRUE; } else { err = dsl_dataset_hold(dp, tosnap, FTAG, &ds); } if (err != 0) { dsl_pool_rele(dp, FTAG); return (err); } if (fromsnap != NULL) { zfs_bookmark_phys_t zb; boolean_t is_clone = B_FALSE; int fsnamelen = strchr(tosnap, '@') - tosnap; /* * If the fromsnap is in a different filesystem, then * mark the send stream as a clone. */ if (strncmp(tosnap, fromsnap, fsnamelen) != 0 || (fromsnap[fsnamelen] != '@' && fromsnap[fsnamelen] != '#')) { is_clone = B_TRUE; } if (strchr(fromsnap, '@')) { dsl_dataset_t *fromds; err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds); if (err == 0) { if (!dsl_dataset_is_before(ds, fromds, 0)) err = SET_ERROR(EXDEV); zb.zbm_creation_time = dsl_dataset_phys(fromds)->ds_creation_time; zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg; zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid; is_clone = (ds->ds_dir != fromds->ds_dir); dsl_dataset_rele(fromds, FTAG); } } else { err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb); } if (err != 0) { dsl_dataset_rele(ds, FTAG); dsl_pool_rele(dp, FTAG); return (err); } err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok, large_block_ok, compressok, outfd, resumeobj, resumeoff, fp, off); } else { err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok, large_block_ok, compressok, outfd, resumeobj, resumeoff, fp, off); } if (owned) dsl_dataset_disown(ds, FTAG); else dsl_dataset_rele(ds, FTAG); return (err); } static int dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed, uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep) { int err = 0; uint64_t size; /* * Assume that space (both on-disk and in-stream) is dominated by * data. We will adjust for indirect blocks and the copies property, * but ignore per-object space used (eg, dnodes and DRR_OBJECT records). */ uint64_t recordsize; uint64_t record_count; objset_t *os; VERIFY0(dmu_objset_from_ds(ds, &os)); /* Assume all (uncompressed) blocks are recordsize. */ if (zfs_override_estimate_recordsize != 0) { recordsize = zfs_override_estimate_recordsize; } else if (os->os_phys->os_type == DMU_OST_ZVOL) { err = dsl_prop_get_int_ds(ds, zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize); } else { err = dsl_prop_get_int_ds(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize); } if (err != 0) return (err); record_count = uncompressed / recordsize; /* * If we're estimating a send size for a compressed stream, use the * compressed data size to estimate the stream size. Otherwise, use the * uncompressed data size. */ size = stream_compressed ? compressed : uncompressed; /* * Subtract out approximate space used by indirect blocks. * Assume most space is used by data blocks (non-indirect, non-dnode). * Assume no ditto blocks or internal fragmentation. * * Therefore, space used by indirect blocks is sizeof(blkptr_t) per * block. */ size -= record_count * sizeof (blkptr_t); /* Add in the space for the record associated with each block. */ size += record_count * sizeof (dmu_replay_record_t); *sizep = size; return (0); } int dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds, boolean_t stream_compressed, uint64_t *sizep) { dsl_pool_t *dp = ds->ds_dir->dd_pool; int err; uint64_t uncomp, comp; ASSERT(dsl_pool_config_held(dp)); /* tosnap must be a snapshot */ if (!ds->ds_is_snapshot) return (SET_ERROR(EINVAL)); /* fromsnap, if provided, must be a snapshot */ if (fromds != NULL && !fromds->ds_is_snapshot) return (SET_ERROR(EINVAL)); /* * fromsnap must be an earlier snapshot from the same fs as tosnap, * or the origin's fs. */ if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0)) return (SET_ERROR(EXDEV)); /* Get compressed and uncompressed size estimates of changed data. */ if (fromds == NULL) { uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes; comp = dsl_dataset_phys(ds)->ds_compressed_bytes; } else { uint64_t used; err = dsl_dataset_space_written(fromds, ds, &used, &comp, &uncomp); if (err != 0) return (err); } err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp, stream_compressed, sizep); /* * Add the size of the BEGIN and END records to the estimate. */ *sizep += 2 * sizeof (dmu_replay_record_t); return (err); } struct calculate_send_arg { uint64_t uncompressed; uint64_t compressed; }; /* * Simple callback used to traverse the blocks of a snapshot and sum their * uncompressed and compressed sizes. */ /* ARGSUSED */ static int dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) { struct calculate_send_arg *space = arg; if (bp != NULL && !BP_IS_HOLE(bp)) { space->uncompressed += BP_GET_UCSIZE(bp); space->compressed += BP_GET_PSIZE(bp); } return (0); } /* * Given a desination snapshot and a TXG, calculate the approximate size of a * send stream sent from that TXG. from_txg may be zero, indicating that the * whole snapshot will be sent. */ int dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg, boolean_t stream_compressed, uint64_t *sizep) { dsl_pool_t *dp = ds->ds_dir->dd_pool; int err; struct calculate_send_arg size = { 0 }; ASSERT(dsl_pool_config_held(dp)); /* tosnap must be a snapshot */ if (!ds->ds_is_snapshot) return (SET_ERROR(EINVAL)); /* verify that from_txg is before the provided snapshot was taken */ if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) { return (SET_ERROR(EXDEV)); } /* * traverse the blocks of the snapshot with birth times after * from_txg, summing their uncompressed size */ err = traverse_dataset(ds, from_txg, TRAVERSE_POST, dmu_calculate_send_traversal, &size); if (err) return (err); err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed, size.compressed, stream_compressed, sizep); return (err); } typedef struct dmu_recv_begin_arg { const char *drba_origin; dmu_recv_cookie_t *drba_cookie; cred_t *drba_cred; uint64_t drba_snapobj; } dmu_recv_begin_arg_t; static int recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds, uint64_t fromguid) { uint64_t val; int error; dsl_pool_t *dp = ds->ds_dir->dd_pool; /* temporary clone name must not exist */ error = zap_lookup(dp->dp_meta_objset, dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name, 8, 1, &val); if (error != ENOENT) return (error == 0 ? EBUSY : error); /* new snapshot name must not exist */ error = zap_lookup(dp->dp_meta_objset, dsl_dataset_phys(ds)->ds_snapnames_zapobj, drba->drba_cookie->drc_tosnap, 8, 1, &val); if (error != ENOENT) return (error == 0 ? EEXIST : error); /* * Check snapshot limit before receiving. We'll recheck again at the * end, but might as well abort before receiving if we're already over * the limit. * * Note that we do not check the file system limit with * dsl_dir_fscount_check because the temporary %clones don't count * against that limit. */ error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred); if (error != 0) return (error); if (fromguid != 0) { dsl_dataset_t *snap; uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj; /* Find snapshot in this dir that matches fromguid. */ while (obj != 0) { error = dsl_dataset_hold_obj(dp, obj, FTAG, &snap); if (error != 0) return (SET_ERROR(ENODEV)); if (snap->ds_dir != ds->ds_dir) { dsl_dataset_rele(snap, FTAG); return (SET_ERROR(ENODEV)); } if (dsl_dataset_phys(snap)->ds_guid == fromguid) break; obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; dsl_dataset_rele(snap, FTAG); } if (obj == 0) return (SET_ERROR(ENODEV)); if (drba->drba_cookie->drc_force) { drba->drba_snapobj = obj; } else { /* * If we are not forcing, there must be no * changes since fromsnap. */ if (dsl_dataset_modified_since_snap(ds, snap)) { dsl_dataset_rele(snap, FTAG); return (SET_ERROR(ETXTBSY)); } drba->drba_snapobj = ds->ds_prev->ds_object; } dsl_dataset_rele(snap, FTAG); } else { /* if full, then must be forced */ if (!drba->drba_cookie->drc_force) return (SET_ERROR(EEXIST)); /* start from $ORIGIN@$ORIGIN, if supported */ drba->drba_snapobj = dp->dp_origin_snap != NULL ? dp->dp_origin_snap->ds_object : 0; } return (0); } static int dmu_recv_begin_check(void *arg, dmu_tx_t *tx) { dmu_recv_begin_arg_t *drba = arg; dsl_pool_t *dp = dmu_tx_pool(tx); struct drr_begin *drrb = drba->drba_cookie->drc_drrb; uint64_t fromguid = drrb->drr_fromguid; int flags = drrb->drr_flags; int error; uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); dsl_dataset_t *ds; const char *tofs = drba->drba_cookie->drc_tofs; /* already checked */ ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING)); if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == DMU_COMPOUNDSTREAM || drrb->drr_type >= DMU_OST_NUMTYPES || ((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL)) return (SET_ERROR(EINVAL)); /* Verify pool version supports SA if SA_SPILL feature set */ if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && spa_version(dp->dp_spa) < SPA_VERSION_SA) return (SET_ERROR(ENOTSUP)); if (drba->drba_cookie->drc_resumable && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET)) return (SET_ERROR(ENOTSUP)); /* * The receiving code doesn't know how to translate a WRITE_EMBEDDED * record to a plain WRITE record, so the pool must have the * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED * records. Same with WRITE_EMBEDDED records that use LZ4 compression. */ if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) return (SET_ERROR(ENOTSUP)); if ((featureflags & DMU_BACKUP_FEATURE_LZ4) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) return (SET_ERROR(ENOTSUP)); /* * The receiving code doesn't know how to translate large blocks * to smaller ones, so the pool must have the LARGE_BLOCKS * feature enabled if the stream has LARGE_BLOCKS. */ if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS)) return (SET_ERROR(ENOTSUP)); /* * The receiving code doesn't know how to translate large dnodes * to smaller ones, so the pool must have the LARGE_DNODE * feature enabled if the stream has LARGE_DNODE. */ if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE)) return (SET_ERROR(ENOTSUP)); error = dsl_dataset_hold(dp, tofs, FTAG, &ds); if (error == 0) { /* target fs already exists; recv into temp clone */ /* Can't recv a clone into an existing fs */ if (flags & DRR_FLAG_CLONE || drba->drba_origin) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } error = recv_begin_check_existing_impl(drba, ds, fromguid); dsl_dataset_rele(ds, FTAG); } else if (error == ENOENT) { /* target fs does not exist; must be a full backup or clone */ char buf[ZFS_MAX_DATASET_NAME_LEN]; /* * If it's a non-clone incremental, we are missing the * target fs, so fail the recv. */ if (fromguid != 0 && !(flags & DRR_FLAG_CLONE || drba->drba_origin)) return (SET_ERROR(ENOENT)); /* * If we're receiving a full send as a clone, and it doesn't * contain all the necessary free records and freeobject * records, reject it. */ if (fromguid == 0 && drba->drba_origin && !(flags & DRR_FLAG_FREERECORDS)) return (SET_ERROR(EINVAL)); /* Open the parent of tofs */ ASSERT3U(strlen(tofs), <, sizeof (buf)); (void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1); error = dsl_dataset_hold(dp, buf, FTAG, &ds); if (error != 0) return (error); /* * Check filesystem and snapshot limits before receiving. We'll * recheck snapshot limits again at the end (we create the * filesystems and increment those counts during begin_sync). */ error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred); if (error != 0) { dsl_dataset_rele(ds, FTAG); return (error); } error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred); if (error != 0) { dsl_dataset_rele(ds, FTAG); return (error); } if (drba->drba_origin != NULL) { dsl_dataset_t *origin; error = dsl_dataset_hold(dp, drba->drba_origin, FTAG, &origin); if (error != 0) { dsl_dataset_rele(ds, FTAG); return (error); } if (!origin->ds_is_snapshot) { dsl_dataset_rele(origin, FTAG); dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } if (dsl_dataset_phys(origin)->ds_guid != fromguid && fromguid != 0) { dsl_dataset_rele(origin, FTAG); dsl_dataset_rele(ds, FTAG); return (SET_ERROR(ENODEV)); } dsl_dataset_rele(origin, FTAG); } dsl_dataset_rele(ds, FTAG); error = 0; } return (error); } static void dmu_recv_begin_sync(void *arg, dmu_tx_t *tx) { dmu_recv_begin_arg_t *drba = arg; dsl_pool_t *dp = dmu_tx_pool(tx); objset_t *mos = dp->dp_meta_objset; struct drr_begin *drrb = drba->drba_cookie->drc_drrb; const char *tofs = drba->drba_cookie->drc_tofs; dsl_dataset_t *ds, *newds; uint64_t dsobj; int error; uint64_t crflags = 0; if (drrb->drr_flags & DRR_FLAG_CI_DATA) crflags |= DS_FLAG_CI_DATASET; error = dsl_dataset_hold(dp, tofs, FTAG, &ds); if (error == 0) { /* create temporary clone */ dsl_dataset_t *snap = NULL; if (drba->drba_snapobj != 0) { VERIFY0(dsl_dataset_hold_obj(dp, drba->drba_snapobj, FTAG, &snap)); } dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name, snap, crflags, drba->drba_cred, tx); if (drba->drba_snapobj != 0) dsl_dataset_rele(snap, FTAG); dsl_dataset_rele(ds, FTAG); } else { dsl_dir_t *dd; const char *tail; dsl_dataset_t *origin = NULL; VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail)); if (drba->drba_origin != NULL) { VERIFY0(dsl_dataset_hold(dp, drba->drba_origin, FTAG, &origin)); } /* Create new dataset. */ dsobj = dsl_dataset_create_sync(dd, strrchr(tofs, '/') + 1, origin, crflags, drba->drba_cred, tx); if (origin != NULL) dsl_dataset_rele(origin, FTAG); dsl_dir_rele(dd, FTAG); drba->drba_cookie->drc_newfs = B_TRUE; } VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds)); if (drba->drba_cookie->drc_resumable) { dsl_dataset_zapify(newds, tx); if (drrb->drr_fromguid != 0) { VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID, 8, 1, &drrb->drr_fromguid, tx)); } VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID, 8, 1, &drrb->drr_toguid, tx)); VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME, 1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx)); uint64_t one = 1; uint64_t zero = 0; VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT, 8, 1, &one, tx)); VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET, 8, 1, &zero, tx)); VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES, 8, 1, &zero, tx)); if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & DMU_BACKUP_FEATURE_LARGE_BLOCKS) { VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK, 8, 1, &one, tx)); } if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & DMU_BACKUP_FEATURE_EMBED_DATA) { VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK, 8, 1, &one, tx)); } if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) & DMU_BACKUP_FEATURE_COMPRESSED) { VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK, 8, 1, &one, tx)); } } dmu_buf_will_dirty(newds->ds_dbuf, tx); dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT; /* * If we actually created a non-clone, we need to create the * objset in our new dataset. */ rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG); if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) { (void) dmu_objset_create_impl(dp->dp_spa, newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx); } rrw_exit(&newds->ds_bp_rwlock, FTAG); drba->drba_cookie->drc_ds = newds; spa_history_log_internal_ds(newds, "receive", tx, ""); } static int dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx) { dmu_recv_begin_arg_t *drba = arg; dsl_pool_t *dp = dmu_tx_pool(tx); struct drr_begin *drrb = drba->drba_cookie->drc_drrb; int error; uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo); dsl_dataset_t *ds; const char *tofs = drba->drba_cookie->drc_tofs; /* already checked */ ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC); ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING); if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) == DMU_COMPOUNDSTREAM || drrb->drr_type >= DMU_OST_NUMTYPES) return (SET_ERROR(EINVAL)); /* Verify pool version supports SA if SA_SPILL feature set */ if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) && spa_version(dp->dp_spa) < SPA_VERSION_SA) return (SET_ERROR(ENOTSUP)); /* * The receiving code doesn't know how to translate a WRITE_EMBEDDED * record to a plain WRITE record, so the pool must have the * EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED * records. Same with WRITE_EMBEDDED records that use LZ4 compression. */ if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) return (SET_ERROR(ENOTSUP)); if ((featureflags & DMU_BACKUP_FEATURE_LZ4) && !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) return (SET_ERROR(ENOTSUP)); /* 6 extra bytes for /%recv */ char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; (void) snprintf(recvname, sizeof (recvname), "%s/%s", tofs, recv_clone_name); if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { /* %recv does not exist; continue in tofs */ error = dsl_dataset_hold(dp, tofs, FTAG, &ds); if (error != 0) return (error); } /* check that ds is marked inconsistent */ if (!DS_IS_INCONSISTENT(ds)) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } /* check that there is resuming data, and that the toguid matches */ if (!dsl_dataset_is_zapified(ds)) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } uint64_t val; error = zap_lookup(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val); if (error != 0 || drrb->drr_toguid != val) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } /* * Check if the receive is still running. If so, it will be owned. * Note that nothing else can own the dataset (e.g. after the receive * fails) because it will be marked inconsistent. */ if (dsl_dataset_has_owner(ds)) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EBUSY)); } /* There should not be any snapshots of this fs yet. */ if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } /* * Note: resume point will be checked when we process the first WRITE * record. */ /* check that the origin matches */ val = 0; (void) zap_lookup(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val); if (drrb->drr_fromguid != val) { dsl_dataset_rele(ds, FTAG); return (SET_ERROR(EINVAL)); } dsl_dataset_rele(ds, FTAG); return (0); } static void dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx) { dmu_recv_begin_arg_t *drba = arg; dsl_pool_t *dp = dmu_tx_pool(tx); const char *tofs = drba->drba_cookie->drc_tofs; dsl_dataset_t *ds; uint64_t dsobj; /* 6 extra bytes for /%recv */ char recvname[ZFS_MAX_DATASET_NAME_LEN + 6]; (void) snprintf(recvname, sizeof (recvname), "%s/%s", tofs, recv_clone_name); if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) { /* %recv does not exist; continue in tofs */ VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds)); drba->drba_cookie->drc_newfs = B_TRUE; } /* clear the inconsistent flag so that we can own it */ ASSERT(DS_IS_INCONSISTENT(ds)); dmu_buf_will_dirty(ds->ds_dbuf, tx); dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; dsobj = ds->ds_object; dsl_dataset_rele(ds, FTAG); VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds)); dmu_buf_will_dirty(ds->ds_dbuf, tx); dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT; rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG); ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds))); rrw_exit(&ds->ds_bp_rwlock, FTAG); drba->drba_cookie->drc_ds = ds; spa_history_log_internal_ds(ds, "resume receive", tx, ""); } /* * NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin() * succeeds; otherwise we will leak the holds on the datasets. */ int dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin, boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc) { dmu_recv_begin_arg_t drba = { 0 }; bzero(drc, sizeof (dmu_recv_cookie_t)); drc->drc_drr_begin = drr_begin; drc->drc_drrb = &drr_begin->drr_u.drr_begin; drc->drc_tosnap = tosnap; drc->drc_tofs = tofs; drc->drc_force = force; drc->drc_resumable = resumable; drc->drc_cred = CRED(); drc->drc_clone = (origin != NULL); if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) { drc->drc_byteswap = B_TRUE; (void) fletcher_4_incremental_byteswap(drr_begin, sizeof (dmu_replay_record_t), &drc->drc_cksum); byteswap_record(drr_begin); } else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) { (void) fletcher_4_incremental_native(drr_begin, sizeof (dmu_replay_record_t), &drc->drc_cksum); } else { return (SET_ERROR(EINVAL)); } drba.drba_origin = origin; drba.drba_cookie = drc; drba.drba_cred = CRED(); if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) & DMU_BACKUP_FEATURE_RESUMING) { return (dsl_sync_task(tofs, dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync, &drba, 5, ZFS_SPACE_CHECK_NORMAL)); } else { return (dsl_sync_task(tofs, dmu_recv_begin_check, dmu_recv_begin_sync, &drba, 5, ZFS_SPACE_CHECK_NORMAL)); } } struct receive_record_arg { dmu_replay_record_t header; void *payload; /* Pointer to a buffer containing the payload */ /* * If the record is a write, pointer to the arc_buf_t containing the * payload. */ arc_buf_t *write_buf; int payload_size; uint64_t bytes_read; /* bytes read from stream when record created */ boolean_t eos_marker; /* Marks the end of the stream */ bqueue_node_t node; }; struct receive_writer_arg { objset_t *os; boolean_t byteswap; bqueue_t q; /* * These three args are used to signal to the main thread that we're * done. */ kmutex_t mutex; kcondvar_t cv; boolean_t done; int err; /* A map from guid to dataset to help handle dedup'd streams. */ avl_tree_t *guid_to_ds_map; boolean_t resumable; uint64_t last_object; uint64_t last_offset; uint64_t max_object; /* highest object ID referenced in stream */ uint64_t bytes_read; /* bytes read when current record created */ }; struct objlist { list_t list; /* List of struct receive_objnode. */ /* * Last object looked up. Used to assert that objects are being looked * up in ascending order. */ uint64_t last_lookup; }; struct receive_objnode { list_node_t node; uint64_t object; }; struct receive_arg { objset_t *os; kthread_t *td; struct file *fp; uint64_t voff; /* The current offset in the stream */ uint64_t bytes_read; /* * A record that has had its payload read in, but hasn't yet been handed * off to the worker thread. */ struct receive_record_arg *rrd; /* A record that has had its header read in, but not its payload. */ struct receive_record_arg *next_rrd; zio_cksum_t cksum; zio_cksum_t prev_cksum; int err; boolean_t byteswap; /* Sorted list of objects not to issue prefetches for. */ struct objlist ignore_objlist; }; typedef struct guid_map_entry { uint64_t guid; dsl_dataset_t *gme_ds; avl_node_t avlnode; } guid_map_entry_t; static int guid_compare(const void *arg1, const void *arg2) { const guid_map_entry_t *gmep1 = (const guid_map_entry_t *)arg1; const guid_map_entry_t *gmep2 = (const guid_map_entry_t *)arg2; return (AVL_CMP(gmep1->guid, gmep2->guid)); } static void free_guid_map_onexit(void *arg) { avl_tree_t *ca = arg; void *cookie = NULL; guid_map_entry_t *gmep; while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) { dsl_dataset_long_rele(gmep->gme_ds, gmep); dsl_dataset_rele(gmep->gme_ds, gmep); kmem_free(gmep, sizeof (guid_map_entry_t)); } avl_destroy(ca); kmem_free(ca, sizeof (avl_tree_t)); } static int restore_bytes(struct receive_arg *ra, void *buf, int len, off_t off, ssize_t *resid) { struct uio auio; struct iovec aiov; int error; aiov.iov_base = buf; aiov.iov_len = len; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_resid = len; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_READ; auio.uio_offset = off; auio.uio_td = ra->td; #ifdef _KERNEL error = fo_read(ra->fp, &auio, ra->td->td_ucred, FOF_OFFSET, ra->td); #else fprintf(stderr, "%s: returning EOPNOTSUPP\n", __func__); error = EOPNOTSUPP; #endif *resid = auio.uio_resid; return (error); } static int receive_read(struct receive_arg *ra, int len, void *buf) { int done = 0; /* * The code doesn't rely on this (lengths being multiples of 8). See * comment in dump_bytes. */ ASSERT0(len % 8); while (done < len) { ssize_t resid; ra->err = restore_bytes(ra, buf + done, len - done, ra->voff, &resid); if (resid == len - done) { /* * Note: ECKSUM indicates that the receive * was interrupted and can potentially be resumed. */ ra->err = SET_ERROR(ECKSUM); } ra->voff += len - done - resid; done = len - resid; if (ra->err != 0) return (ra->err); } ra->bytes_read += len; ASSERT3U(done, ==, len); return (0); } -static void +noinline static void byteswap_record(dmu_replay_record_t *drr) { #define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X)) #define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X)) drr->drr_type = BSWAP_32(drr->drr_type); drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen); switch (drr->drr_type) { case DRR_BEGIN: DO64(drr_begin.drr_magic); DO64(drr_begin.drr_versioninfo); DO64(drr_begin.drr_creation_time); DO32(drr_begin.drr_type); DO32(drr_begin.drr_flags); DO64(drr_begin.drr_toguid); DO64(drr_begin.drr_fromguid); break; case DRR_OBJECT: DO64(drr_object.drr_object); DO32(drr_object.drr_type); DO32(drr_object.drr_bonustype); DO32(drr_object.drr_blksz); DO32(drr_object.drr_bonuslen); DO64(drr_object.drr_toguid); break; case DRR_FREEOBJECTS: DO64(drr_freeobjects.drr_firstobj); DO64(drr_freeobjects.drr_numobjs); DO64(drr_freeobjects.drr_toguid); break; case DRR_WRITE: DO64(drr_write.drr_object); DO32(drr_write.drr_type); DO64(drr_write.drr_offset); DO64(drr_write.drr_logical_size); DO64(drr_write.drr_toguid); ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum); DO64(drr_write.drr_key.ddk_prop); DO64(drr_write.drr_compressed_size); break; case DRR_WRITE_BYREF: DO64(drr_write_byref.drr_object); DO64(drr_write_byref.drr_offset); DO64(drr_write_byref.drr_length); DO64(drr_write_byref.drr_toguid); DO64(drr_write_byref.drr_refguid); DO64(drr_write_byref.drr_refobject); DO64(drr_write_byref.drr_refoffset); ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref. drr_key.ddk_cksum); DO64(drr_write_byref.drr_key.ddk_prop); break; case DRR_WRITE_EMBEDDED: DO64(drr_write_embedded.drr_object); DO64(drr_write_embedded.drr_offset); DO64(drr_write_embedded.drr_length); DO64(drr_write_embedded.drr_toguid); DO32(drr_write_embedded.drr_lsize); DO32(drr_write_embedded.drr_psize); break; case DRR_FREE: DO64(drr_free.drr_object); DO64(drr_free.drr_offset); DO64(drr_free.drr_length); DO64(drr_free.drr_toguid); break; case DRR_SPILL: DO64(drr_spill.drr_object); DO64(drr_spill.drr_length); DO64(drr_spill.drr_toguid); break; case DRR_END: DO64(drr_end.drr_toguid); ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum); break; } if (drr->drr_type != DRR_BEGIN) { ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum); } #undef DO64 #undef DO32 } static inline uint8_t deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size) { if (bonus_type == DMU_OT_SA) { return (1); } else { return (1 + ((DN_OLD_MAX_BONUSLEN - MIN(DN_OLD_MAX_BONUSLEN, bonus_size)) >> SPA_BLKPTRSHIFT)); } } static void save_resume_state(struct receive_writer_arg *rwa, uint64_t object, uint64_t offset, dmu_tx_t *tx) { int txgoff = dmu_tx_get_txg(tx) & TXG_MASK; if (!rwa->resumable) return; /* * We use ds_resume_bytes[] != 0 to indicate that we need to * update this on disk, so it must not be 0. */ ASSERT(rwa->bytes_read != 0); /* * We only resume from write records, which have a valid * (non-meta-dnode) object number. */ ASSERT(object != 0); /* * For resuming to work correctly, we must receive records in order, * sorted by object,offset. This is checked by the callers, but * assert it here for good measure. */ ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]); ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] || offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]); ASSERT3U(rwa->bytes_read, >=, rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]); rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object; rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset; rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read; } -static int +noinline static int receive_object(struct receive_writer_arg *rwa, struct drr_object *drro, void *data) { dmu_object_info_t doi; dmu_tx_t *tx; uint64_t object; int err; if (drro->drr_type == DMU_OT_NONE || !DMU_OT_IS_VALID(drro->drr_type) || !DMU_OT_IS_VALID(drro->drr_bonustype) || drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS || drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS || P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) || drro->drr_blksz < SPA_MINBLOCKSIZE || drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) || drro->drr_bonuslen > DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(rwa->os)))) { return (SET_ERROR(EINVAL)); } err = dmu_object_info(rwa->os, drro->drr_object, &doi); if (err != 0 && err != ENOENT) return (SET_ERROR(EINVAL)); object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT; if (drro->drr_object > rwa->max_object) rwa->max_object = drro->drr_object; /* * If we are losing blkptrs or changing the block size this must * be a new file instance. We must clear out the previous file * contents before we can change this type of metadata in the dnode. */ if (err == 0) { int nblkptr; nblkptr = deduce_nblkptr(drro->drr_bonustype, drro->drr_bonuslen); if (drro->drr_blksz != doi.doi_data_block_size || nblkptr < doi.doi_nblkptr) { err = dmu_free_long_range(rwa->os, drro->drr_object, 0, DMU_OBJECT_END); if (err != 0) return (SET_ERROR(EINVAL)); } } tx = dmu_tx_create(rwa->os); dmu_tx_hold_bonus(tx, object); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); return (err); } if (object == DMU_NEW_OBJECT) { /* currently free, want to be allocated */ err = dmu_object_claim_dnsize(rwa->os, drro->drr_object, drro->drr_type, drro->drr_blksz, drro->drr_bonustype, drro->drr_bonuslen, drro->drr_dn_slots << DNODE_SHIFT, tx); } else if (drro->drr_type != doi.doi_type || drro->drr_blksz != doi.doi_data_block_size || drro->drr_bonustype != doi.doi_bonus_type || drro->drr_bonuslen != doi.doi_bonus_size) { /* currently allocated, but with different properties */ err = dmu_object_reclaim(rwa->os, drro->drr_object, drro->drr_type, drro->drr_blksz, drro->drr_bonustype, drro->drr_bonuslen, tx); } if (err != 0) { dmu_tx_commit(tx); return (SET_ERROR(EINVAL)); } dmu_object_set_checksum(rwa->os, drro->drr_object, drro->drr_checksumtype, tx); dmu_object_set_compress(rwa->os, drro->drr_object, drro->drr_compress, tx); if (data != NULL) { dmu_buf_t *db; VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db)); dmu_buf_will_dirty(db, tx); ASSERT3U(db->db_size, >=, drro->drr_bonuslen); bcopy(data, db->db_data, drro->drr_bonuslen); if (rwa->byteswap) { dmu_object_byteswap_t byteswap = DMU_OT_BYTESWAP(drro->drr_bonustype); dmu_ot_byteswap[byteswap].ob_func(db->db_data, drro->drr_bonuslen); } dmu_buf_rele(db, FTAG); } dmu_tx_commit(tx); return (0); } /* ARGSUSED */ -static int +noinline static int receive_freeobjects(struct receive_writer_arg *rwa, struct drr_freeobjects *drrfo) { uint64_t obj; int next_err = 0; if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj) return (SET_ERROR(EINVAL)); for (obj = drrfo->drr_firstobj == 0 ? 1 : drrfo->drr_firstobj; obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0; next_err = dmu_object_next(rwa->os, &obj, FALSE, 0)) { dmu_object_info_t doi; int err; err = dmu_object_info(rwa->os, obj, &doi); if (err == ENOENT) { obj++; continue; } else if (err != 0) { return (err); } err = dmu_free_long_object(rwa->os, obj); if (err != 0) return (err); if (obj > rwa->max_object) rwa->max_object = obj; } if (next_err != ESRCH) return (next_err); return (0); } -static int +noinline static int receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw, arc_buf_t *abuf) { dmu_tx_t *tx; int err; if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset || !DMU_OT_IS_VALID(drrw->drr_type)) return (SET_ERROR(EINVAL)); /* * For resuming to work, records must be in increasing order * by (object, offset). */ if (drrw->drr_object < rwa->last_object || (drrw->drr_object == rwa->last_object && drrw->drr_offset < rwa->last_offset)) { return (SET_ERROR(EINVAL)); } rwa->last_object = drrw->drr_object; rwa->last_offset = drrw->drr_offset; if (rwa->last_object > rwa->max_object) rwa->max_object = rwa->last_object; if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0) return (SET_ERROR(EINVAL)); tx = dmu_tx_create(rwa->os); - dmu_tx_hold_write(tx, drrw->drr_object, drrw->drr_offset, drrw->drr_logical_size); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); return (err); } if (rwa->byteswap) { dmu_object_byteswap_t byteswap = DMU_OT_BYTESWAP(drrw->drr_type); dmu_ot_byteswap[byteswap].ob_func(abuf->b_data, DRR_WRITE_PAYLOAD_SIZE(drrw)); } /* use the bonus buf to look up the dnode in dmu_assign_arcbuf */ dmu_buf_t *bonus; if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0) return (SET_ERROR(EINVAL)); dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx); /* * Note: If the receive fails, we want the resume stream to start * with the same record that we last successfully received (as opposed * to the next record), so that we can verify that we are * resuming from the correct location. */ save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx); dmu_tx_commit(tx); dmu_buf_rele(bonus, FTAG); return (0); } /* * Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed * streams to refer to a copy of the data that is already on the * system because it came in earlier in the stream. This function * finds the earlier copy of the data, and uses that copy instead of * data from the stream to fulfill this write. */ static int receive_write_byref(struct receive_writer_arg *rwa, struct drr_write_byref *drrwbr) { dmu_tx_t *tx; int err; guid_map_entry_t gmesrch; guid_map_entry_t *gmep; avl_index_t where; objset_t *ref_os = NULL; dmu_buf_t *dbp; if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset) return (SET_ERROR(EINVAL)); /* * If the GUID of the referenced dataset is different from the * GUID of the target dataset, find the referenced dataset. */ if (drrwbr->drr_toguid != drrwbr->drr_refguid) { gmesrch.guid = drrwbr->drr_refguid; if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch, &where)) == NULL) { return (SET_ERROR(EINVAL)); } if (dmu_objset_from_ds(gmep->gme_ds, &ref_os)) return (SET_ERROR(EINVAL)); } else { ref_os = rwa->os; } if (drrwbr->drr_object > rwa->max_object) rwa->max_object = drrwbr->drr_object; err = dmu_buf_hold(ref_os, drrwbr->drr_refobject, drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH); if (err != 0) return (err); tx = dmu_tx_create(rwa->os); dmu_tx_hold_write(tx, drrwbr->drr_object, drrwbr->drr_offset, drrwbr->drr_length); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); return (err); } dmu_write(rwa->os, drrwbr->drr_object, drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx); dmu_buf_rele(dbp, FTAG); /* See comment in restore_write. */ save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx); dmu_tx_commit(tx); return (0); } static int receive_write_embedded(struct receive_writer_arg *rwa, struct drr_write_embedded *drrwe, void *data) { dmu_tx_t *tx; int err; if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset) return (EINVAL); if (drrwe->drr_psize > BPE_PAYLOAD_SIZE) return (EINVAL); if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES) return (EINVAL); if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS) return (EINVAL); if (drrwe->drr_object > rwa->max_object) rwa->max_object = drrwe->drr_object; tx = dmu_tx_create(rwa->os); dmu_tx_hold_write(tx, drrwe->drr_object, drrwe->drr_offset, drrwe->drr_length); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_tx_abort(tx); return (err); } dmu_write_embedded(rwa->os, drrwe->drr_object, drrwe->drr_offset, data, drrwe->drr_etype, drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize, rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx); /* See comment in restore_write. */ save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx); dmu_tx_commit(tx); return (0); } static int receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs, void *data) { dmu_tx_t *tx; dmu_buf_t *db, *db_spill; int err; if (drrs->drr_length < SPA_MINBLOCKSIZE || drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os))) return (SET_ERROR(EINVAL)); if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0) return (SET_ERROR(EINVAL)); if (drrs->drr_object > rwa->max_object) rwa->max_object = drrs->drr_object; VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db)); if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) { dmu_buf_rele(db, FTAG); return (err); } tx = dmu_tx_create(rwa->os); dmu_tx_hold_spill(tx, db->db_object); err = dmu_tx_assign(tx, TXG_WAIT); if (err != 0) { dmu_buf_rele(db, FTAG); dmu_buf_rele(db_spill, FTAG); dmu_tx_abort(tx); return (err); } dmu_buf_will_dirty(db_spill, tx); if (db_spill->db_size < drrs->drr_length) VERIFY(0 == dbuf_spill_set_blksz(db_spill, drrs->drr_length, tx)); bcopy(data, db_spill->db_data, drrs->drr_length); dmu_buf_rele(db, FTAG); dmu_buf_rele(db_spill, FTAG); dmu_tx_commit(tx); return (0); } /* ARGSUSED */ -static int +noinline static int receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf) { int err; if (drrf->drr_length != -1ULL && drrf->drr_offset + drrf->drr_length < drrf->drr_offset) return (SET_ERROR(EINVAL)); if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0) return (SET_ERROR(EINVAL)); if (drrf->drr_object > rwa->max_object) rwa->max_object = drrf->drr_object; err = dmu_free_long_range(rwa->os, drrf->drr_object, drrf->drr_offset, drrf->drr_length); return (err); } /* used to destroy the drc_ds on error */ static void dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc) { if (drc->drc_resumable) { /* wait for our resume state to be written to disk */ txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0); dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); } else { char name[ZFS_MAX_DATASET_NAME_LEN]; dsl_dataset_name(drc->drc_ds, name); dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); (void) dsl_destroy_head(name); } } static void receive_cksum(struct receive_arg *ra, int len, void *buf) { if (ra->byteswap) { (void) fletcher_4_incremental_byteswap(buf, len, &ra->cksum); } else { (void) fletcher_4_incremental_native(buf, len, &ra->cksum); } } /* * Read the payload into a buffer of size len, and update the current record's * payload field. * Allocate ra->next_rrd and read the next record's header into * ra->next_rrd->header. * Verify checksum of payload and next record. */ static int receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf) { int err; if (len != 0) { ASSERT3U(len, <=, SPA_MAXBLOCKSIZE); err = receive_read(ra, len, buf); if (err != 0) return (err); receive_cksum(ra, len, buf); /* note: rrd is NULL when reading the begin record's payload */ if (ra->rrd != NULL) { ra->rrd->payload = buf; ra->rrd->payload_size = len; ra->rrd->bytes_read = ra->bytes_read; } } ra->prev_cksum = ra->cksum; ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP); err = receive_read(ra, sizeof (ra->next_rrd->header), &ra->next_rrd->header); ra->next_rrd->bytes_read = ra->bytes_read; if (err != 0) { kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); ra->next_rrd = NULL; return (err); } if (ra->next_rrd->header.drr_type == DRR_BEGIN) { kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); ra->next_rrd = NULL; return (SET_ERROR(EINVAL)); } /* * Note: checksum is of everything up to but not including the * checksum itself. */ ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), ==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t)); receive_cksum(ra, offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum), &ra->next_rrd->header); zio_cksum_t cksum_orig = ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; zio_cksum_t *cksump = &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum; if (ra->byteswap) byteswap_record(&ra->next_rrd->header); if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) && !ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) { kmem_free(ra->next_rrd, sizeof (*ra->next_rrd)); ra->next_rrd = NULL; return (SET_ERROR(ECKSUM)); } receive_cksum(ra, sizeof (cksum_orig), &cksum_orig); return (0); } static void objlist_create(struct objlist *list) { list_create(&list->list, sizeof (struct receive_objnode), offsetof(struct receive_objnode, node)); list->last_lookup = 0; } static void objlist_destroy(struct objlist *list) { for (struct receive_objnode *n = list_remove_head(&list->list); n != NULL; n = list_remove_head(&list->list)) { kmem_free(n, sizeof (*n)); } list_destroy(&list->list); } /* * This function looks through the objlist to see if the specified object number * is contained in the objlist. In the process, it will remove all object * numbers in the list that are smaller than the specified object number. Thus, * any lookup of an object number smaller than a previously looked up object * number will always return false; therefore, all lookups should be done in * ascending order. */ static boolean_t objlist_exists(struct objlist *list, uint64_t object) { struct receive_objnode *node = list_head(&list->list); ASSERT3U(object, >=, list->last_lookup); list->last_lookup = object; while (node != NULL && node->object < object) { VERIFY3P(node, ==, list_remove_head(&list->list)); kmem_free(node, sizeof (*node)); node = list_head(&list->list); } return (node != NULL && node->object == object); } /* * The objlist is a list of object numbers stored in ascending order. However, * the insertion of new object numbers does not seek out the correct location to * store a new object number; instead, it appends it to the list for simplicity. * Thus, any users must take care to only insert new object numbers in ascending * order. */ static void objlist_insert(struct objlist *list, uint64_t object) { struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP); node->object = object; #ifdef ZFS_DEBUG struct receive_objnode *last_object = list_tail(&list->list); uint64_t last_objnum = (last_object != NULL ? last_object->object : 0); ASSERT3U(node->object, >, last_objnum); #endif list_insert_tail(&list->list, node); } /* * Issue the prefetch reads for any necessary indirect blocks. * * We use the object ignore list to tell us whether or not to issue prefetches * for a given object. We do this for both correctness (in case the blocksize * of an object has changed) and performance (if the object doesn't exist, don't * needlessly try to issue prefetches). We also trim the list as we go through * the stream to prevent it from growing to an unbounded size. * * The object numbers within will always be in sorted order, and any write * records we see will also be in sorted order, but they're not sorted with * respect to each other (i.e. we can get several object records before * receiving each object's write records). As a result, once we've reached a * given object number, we can safely remove any reference to lower object * numbers in the ignore list. In practice, we receive up to 32 object records * before receiving write records, so the list can have up to 32 nodes in it. */ /* ARGSUSED */ static void receive_read_prefetch(struct receive_arg *ra, uint64_t object, uint64_t offset, uint64_t length) { if (!objlist_exists(&ra->ignore_objlist, object)) { dmu_prefetch(ra->os, object, 1, offset, length, ZIO_PRIORITY_SYNC_READ); } } /* * Read records off the stream, issuing any necessary prefetches. */ static int receive_read_record(struct receive_arg *ra) { int err; switch (ra->rrd->header.drr_type) { case DRR_OBJECT: { struct drr_object *drro = &ra->rrd->header.drr_u.drr_object; uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8); void *buf = kmem_zalloc(size, KM_SLEEP); dmu_object_info_t doi; err = receive_read_payload_and_next_header(ra, size, buf); if (err != 0) { kmem_free(buf, size); return (err); } err = dmu_object_info(ra->os, drro->drr_object, &doi); /* * See receive_read_prefetch for an explanation why we're * storing this object in the ignore_obj_list. */ if (err == ENOENT || (err == 0 && doi.doi_data_block_size != drro->drr_blksz)) { objlist_insert(&ra->ignore_objlist, drro->drr_object); err = 0; } return (err); } case DRR_FREEOBJECTS: { err = receive_read_payload_and_next_header(ra, 0, NULL); return (err); } case DRR_WRITE: { struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write; arc_buf_t *abuf; boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type); if (DRR_WRITE_COMPRESSED(drrw)) { ASSERT3U(drrw->drr_compressed_size, >, 0); ASSERT3U(drrw->drr_logical_size, >=, drrw->drr_compressed_size); ASSERT(!is_meta); abuf = arc_loan_compressed_buf( dmu_objset_spa(ra->os), drrw->drr_compressed_size, drrw->drr_logical_size, drrw->drr_compressiontype); } else { abuf = arc_loan_buf(dmu_objset_spa(ra->os), is_meta, drrw->drr_logical_size); } err = receive_read_payload_and_next_header(ra, DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data); if (err != 0) { dmu_return_arcbuf(abuf); return (err); } ra->rrd->write_buf = abuf; receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset, drrw->drr_logical_size); return (err); } case DRR_WRITE_BYREF: { struct drr_write_byref *drrwb = &ra->rrd->header.drr_u.drr_write_byref; err = receive_read_payload_and_next_header(ra, 0, NULL); receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset, drrwb->drr_length); return (err); } case DRR_WRITE_EMBEDDED: { struct drr_write_embedded *drrwe = &ra->rrd->header.drr_u.drr_write_embedded; uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8); void *buf = kmem_zalloc(size, KM_SLEEP); err = receive_read_payload_and_next_header(ra, size, buf); if (err != 0) { kmem_free(buf, size); return (err); } receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset, drrwe->drr_length); return (err); } case DRR_FREE: { /* * It might be beneficial to prefetch indirect blocks here, but * we don't really have the data to decide for sure. */ err = receive_read_payload_and_next_header(ra, 0, NULL); return (err); } case DRR_END: { struct drr_end *drre = &ra->rrd->header.drr_u.drr_end; if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum)) return (SET_ERROR(ECKSUM)); return (0); } case DRR_SPILL: { struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill; void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP); err = receive_read_payload_and_next_header(ra, drrs->drr_length, buf); if (err != 0) kmem_free(buf, drrs->drr_length); return (err); } default: return (SET_ERROR(EINVAL)); } } /* * Commit the records to the pool. */ static int receive_process_record(struct receive_writer_arg *rwa, struct receive_record_arg *rrd) { int err; /* Processing in order, therefore bytes_read should be increasing. */ ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read); rwa->bytes_read = rrd->bytes_read; switch (rrd->header.drr_type) { case DRR_OBJECT: { struct drr_object *drro = &rrd->header.drr_u.drr_object; err = receive_object(rwa, drro, rrd->payload); kmem_free(rrd->payload, rrd->payload_size); rrd->payload = NULL; return (err); } case DRR_FREEOBJECTS: { struct drr_freeobjects *drrfo = &rrd->header.drr_u.drr_freeobjects; return (receive_freeobjects(rwa, drrfo)); } case DRR_WRITE: { struct drr_write *drrw = &rrd->header.drr_u.drr_write; err = receive_write(rwa, drrw, rrd->write_buf); /* if receive_write() is successful, it consumes the arc_buf */ if (err != 0) dmu_return_arcbuf(rrd->write_buf); rrd->write_buf = NULL; rrd->payload = NULL; return (err); } case DRR_WRITE_BYREF: { struct drr_write_byref *drrwbr = &rrd->header.drr_u.drr_write_byref; return (receive_write_byref(rwa, drrwbr)); } case DRR_WRITE_EMBEDDED: { struct drr_write_embedded *drrwe = &rrd->header.drr_u.drr_write_embedded; err = receive_write_embedded(rwa, drrwe, rrd->payload); kmem_free(rrd->payload, rrd->payload_size); rrd->payload = NULL; return (err); } case DRR_FREE: { struct drr_free *drrf = &rrd->header.drr_u.drr_free; return (receive_free(rwa, drrf)); } case DRR_SPILL: { struct drr_spill *drrs = &rrd->header.drr_u.drr_spill; err = receive_spill(rwa, drrs, rrd->payload); kmem_free(rrd->payload, rrd->payload_size); rrd->payload = NULL; return (err); } default: return (SET_ERROR(EINVAL)); } } /* * dmu_recv_stream's worker thread; pull records off the queue, and then call * receive_process_record When we're done, signal the main thread and exit. */ static void receive_writer_thread(void *arg) { struct receive_writer_arg *rwa = arg; struct receive_record_arg *rrd; for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker; rrd = bqueue_dequeue(&rwa->q)) { /* * If there's an error, the main thread will stop putting things * on the queue, but we need to clear everything in it before we * can exit. */ if (rwa->err == 0) { rwa->err = receive_process_record(rwa, rrd); } else if (rrd->write_buf != NULL) { dmu_return_arcbuf(rrd->write_buf); rrd->write_buf = NULL; rrd->payload = NULL; } else if (rrd->payload != NULL) { kmem_free(rrd->payload, rrd->payload_size); rrd->payload = NULL; } kmem_free(rrd, sizeof (*rrd)); } kmem_free(rrd, sizeof (*rrd)); mutex_enter(&rwa->mutex); rwa->done = B_TRUE; cv_signal(&rwa->cv); mutex_exit(&rwa->mutex); thread_exit(); } static int resume_check(struct receive_arg *ra, nvlist_t *begin_nvl) { uint64_t val; objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset; uint64_t dsobj = dmu_objset_id(ra->os); uint64_t resume_obj, resume_off; if (nvlist_lookup_uint64(begin_nvl, "resume_object", &resume_obj) != 0 || nvlist_lookup_uint64(begin_nvl, "resume_offset", &resume_off) != 0) { return (SET_ERROR(EINVAL)); } VERIFY0(zap_lookup(mos, dsobj, DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val)); if (resume_obj != val) return (SET_ERROR(EINVAL)); VERIFY0(zap_lookup(mos, dsobj, DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val)); if (resume_off != val) return (SET_ERROR(EINVAL)); return (0); } /* * Read in the stream's records, one by one, and apply them to the pool. There * are two threads involved; the thread that calls this function will spin up a * worker thread, read the records off the stream one by one, and issue * prefetches for any necessary indirect blocks. It will then push the records * onto an internal blocking queue. The worker thread will pull the records off * the queue, and actually write the data into the DMU. This way, the worker * thread doesn't have to wait for reads to complete, since everything it needs * (the indirect blocks) will be prefetched. * * NB: callers *must* call dmu_recv_end() if this succeeds. */ int dmu_recv_stream(dmu_recv_cookie_t *drc, struct file *fp, offset_t *voffp, int cleanup_fd, uint64_t *action_handlep) { int err = 0; struct receive_arg ra = { 0 }; struct receive_writer_arg rwa = { 0 }; int featureflags; nvlist_t *begin_nvl = NULL; ra.byteswap = drc->drc_byteswap; ra.cksum = drc->drc_cksum; ra.td = curthread; ra.fp = fp; ra.voff = *voffp; if (dsl_dataset_is_zapified(drc->drc_ds)) { (void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset, drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES, sizeof (ra.bytes_read), 1, &ra.bytes_read); } objlist_create(&ra.ignore_objlist); /* these were verified in dmu_recv_begin */ ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==, DMU_SUBSTREAM); ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES); /* * Open the objset we are modifying. */ VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra.os)); ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT); featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo); /* if this stream is dedup'ed, set up the avl tree for guid mapping */ if (featureflags & DMU_BACKUP_FEATURE_DEDUP) { minor_t minor; if (cleanup_fd == -1) { ra.err = SET_ERROR(EBADF); goto out; } ra.err = zfs_onexit_fd_hold(cleanup_fd, &minor); if (ra.err != 0) { cleanup_fd = -1; goto out; } if (*action_handlep == 0) { rwa.guid_to_ds_map = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP); avl_create(rwa.guid_to_ds_map, guid_compare, sizeof (guid_map_entry_t), offsetof(guid_map_entry_t, avlnode)); err = zfs_onexit_add_cb(minor, free_guid_map_onexit, rwa.guid_to_ds_map, action_handlep); if (ra.err != 0) goto out; } else { err = zfs_onexit_cb_data(minor, *action_handlep, (void **)&rwa.guid_to_ds_map); if (ra.err != 0) goto out; } drc->drc_guid_to_ds_map = rwa.guid_to_ds_map; } uint32_t payloadlen = drc->drc_drr_begin->drr_payloadlen; void *payload = NULL; if (payloadlen != 0) payload = kmem_alloc(payloadlen, KM_SLEEP); err = receive_read_payload_and_next_header(&ra, payloadlen, payload); if (err != 0) { if (payloadlen != 0) kmem_free(payload, payloadlen); goto out; } if (payloadlen != 0) { err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP); kmem_free(payload, payloadlen); if (err != 0) goto out; } if (featureflags & DMU_BACKUP_FEATURE_RESUMING) { err = resume_check(&ra, begin_nvl); if (err != 0) goto out; } (void) bqueue_init(&rwa.q, zfs_recv_queue_length, offsetof(struct receive_record_arg, node)); cv_init(&rwa.cv, NULL, CV_DEFAULT, NULL); mutex_init(&rwa.mutex, NULL, MUTEX_DEFAULT, NULL); rwa.os = ra.os; rwa.byteswap = drc->drc_byteswap; rwa.resumable = drc->drc_resumable; (void) thread_create(NULL, 0, receive_writer_thread, &rwa, 0, &p0, TS_RUN, minclsyspri); /* * We're reading rwa.err without locks, which is safe since we are the * only reader, and the worker thread is the only writer. It's ok if we * miss a write for an iteration or two of the loop, since the writer * thread will keep freeing records we send it until we send it an eos * marker. * * We can leave this loop in 3 ways: First, if rwa.err is * non-zero. In that case, the writer thread will free the rrd we just * pushed. Second, if we're interrupted; in that case, either it's the * first loop and ra.rrd was never allocated, or it's later, and ra.rrd * has been handed off to the writer thread who will free it. Finally, * if receive_read_record fails or we're at the end of the stream, then * we free ra.rrd and exit. */ while (rwa.err == 0) { if (issig(JUSTLOOKING) && issig(FORREAL)) { err = SET_ERROR(EINTR); break; } ASSERT3P(ra.rrd, ==, NULL); ra.rrd = ra.next_rrd; ra.next_rrd = NULL; /* Allocates and loads header into ra.next_rrd */ err = receive_read_record(&ra); if (ra.rrd->header.drr_type == DRR_END || err != 0) { kmem_free(ra.rrd, sizeof (*ra.rrd)); ra.rrd = NULL; break; } bqueue_enqueue(&rwa.q, ra.rrd, sizeof (struct receive_record_arg) + ra.rrd->payload_size); ra.rrd = NULL; } if (ra.next_rrd == NULL) ra.next_rrd = kmem_zalloc(sizeof (*ra.next_rrd), KM_SLEEP); ra.next_rrd->eos_marker = B_TRUE; bqueue_enqueue(&rwa.q, ra.next_rrd, 1); mutex_enter(&rwa.mutex); while (!rwa.done) { cv_wait(&rwa.cv, &rwa.mutex); } mutex_exit(&rwa.mutex); /* * If we are receiving a full stream as a clone, all object IDs which * are greater than the maximum ID referenced in the stream are * by definition unused and must be freed. Note that it's possible that * we've resumed this send and the first record we received was the END * record. In that case, max_object would be 0, but we shouldn't start * freeing all objects from there; instead we should start from the * resumeobj. */ if (drc->drc_clone && drc->drc_drrb->drr_fromguid == 0) { uint64_t obj; if (nvlist_lookup_uint64(begin_nvl, "resume_object", &obj) != 0) obj = 0; if (rwa.max_object > obj) obj = rwa.max_object; obj++; int free_err = 0; int next_err = 0; while (next_err == 0) { free_err = dmu_free_long_object(rwa.os, obj); if (free_err != 0 && free_err != ENOENT) break; next_err = dmu_object_next(rwa.os, &obj, FALSE, 0); } if (err == 0) { if (free_err != 0 && free_err != ENOENT) err = free_err; else if (next_err != ESRCH) err = next_err; } } cv_destroy(&rwa.cv); mutex_destroy(&rwa.mutex); bqueue_destroy(&rwa.q); if (err == 0) err = rwa.err; out: nvlist_free(begin_nvl); if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1)) zfs_onexit_fd_rele(cleanup_fd); if (err != 0) { /* * Clean up references. If receive is not resumable, * destroy what we created, so we don't leave it in * the inconsistent state. */ dmu_recv_cleanup_ds(drc); } *voffp = ra.voff; objlist_destroy(&ra.ignore_objlist); return (err); } static int dmu_recv_end_check(void *arg, dmu_tx_t *tx) { dmu_recv_cookie_t *drc = arg; dsl_pool_t *dp = dmu_tx_pool(tx); int error; ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag); if (!drc->drc_newfs) { dsl_dataset_t *origin_head; error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head); if (error != 0) return (error); if (drc->drc_force) { /* * We will destroy any snapshots in tofs (i.e. before * origin_head) that are after the origin (which is * the snap before drc_ds, because drc_ds can not * have any snaps of its own). */ uint64_t obj; obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; while (obj != dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { dsl_dataset_t *snap; error = dsl_dataset_hold_obj(dp, obj, FTAG, &snap); if (error != 0) break; if (snap->ds_dir != origin_head->ds_dir) error = SET_ERROR(EINVAL); if (error == 0) { error = dsl_destroy_snapshot_check_impl( snap, B_FALSE); } obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; dsl_dataset_rele(snap, FTAG); if (error != 0) break; } if (error != 0) { dsl_dataset_rele(origin_head, FTAG); return (error); } } error = dsl_dataset_clone_swap_check_impl(drc->drc_ds, origin_head, drc->drc_force, drc->drc_owner, tx); if (error != 0) { dsl_dataset_rele(origin_head, FTAG); return (error); } error = dsl_dataset_snapshot_check_impl(origin_head, drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); dsl_dataset_rele(origin_head, FTAG); if (error != 0) return (error); error = dsl_destroy_head_check_impl(drc->drc_ds, 1); } else { error = dsl_dataset_snapshot_check_impl(drc->drc_ds, drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred); } return (error); } static void dmu_recv_end_sync(void *arg, dmu_tx_t *tx) { dmu_recv_cookie_t *drc = arg; dsl_pool_t *dp = dmu_tx_pool(tx); spa_history_log_internal_ds(drc->drc_ds, "finish receiving", tx, "snap=%s", drc->drc_tosnap); if (!drc->drc_newfs) { dsl_dataset_t *origin_head; VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head)); if (drc->drc_force) { /* * Destroy any snapshots of drc_tofs (origin_head) * after the origin (the snap before drc_ds). */ uint64_t obj; obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; while (obj != dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) { dsl_dataset_t *snap; VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &snap)); ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir); obj = dsl_dataset_phys(snap)->ds_prev_snap_obj; dsl_destroy_snapshot_sync_impl(snap, B_FALSE, tx); dsl_dataset_rele(snap, FTAG); } } VERIFY3P(drc->drc_ds->ds_prev, ==, origin_head->ds_prev); dsl_dataset_clone_swap_sync_impl(drc->drc_ds, origin_head, tx); dsl_dataset_snapshot_sync_impl(origin_head, drc->drc_tosnap, tx); /* set snapshot's creation time and guid */ dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx); dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time = drc->drc_drrb->drr_creation_time; dsl_dataset_phys(origin_head->ds_prev)->ds_guid = drc->drc_drrb->drr_toguid; dsl_dataset_phys(origin_head->ds_prev)->ds_flags &= ~DS_FLAG_INCONSISTENT; dmu_buf_will_dirty(origin_head->ds_dbuf, tx); dsl_dataset_phys(origin_head)->ds_flags &= ~DS_FLAG_INCONSISTENT; drc->drc_newsnapobj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj; dsl_dataset_rele(origin_head, FTAG); dsl_destroy_head_sync_impl(drc->drc_ds, tx); if (drc->drc_owner != NULL) VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner); } else { dsl_dataset_t *ds = drc->drc_ds; dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx); /* set snapshot's creation time and guid */ dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx); dsl_dataset_phys(ds->ds_prev)->ds_creation_time = drc->drc_drrb->drr_creation_time; dsl_dataset_phys(ds->ds_prev)->ds_guid = drc->drc_drrb->drr_toguid; dsl_dataset_phys(ds->ds_prev)->ds_flags &= ~DS_FLAG_INCONSISTENT; dmu_buf_will_dirty(ds->ds_dbuf, tx); dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT; if (dsl_dataset_has_resume_receive_state(ds)) { (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_FROMGUID, tx); (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_OBJECT, tx); (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_OFFSET, tx); (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_BYTES, tx); (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_TOGUID, tx); (void) zap_remove(dp->dp_meta_objset, ds->ds_object, DS_FIELD_RESUME_TONAME, tx); } drc->drc_newsnapobj = dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj; } /* * Release the hold from dmu_recv_begin. This must be done before * we return to open context, so that when we free the dataset's dnode, * we can evict its bonus buffer. */ dsl_dataset_disown(drc->drc_ds, dmu_recv_tag); drc->drc_ds = NULL; } static int add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj) { dsl_pool_t *dp; dsl_dataset_t *snapds; guid_map_entry_t *gmep; int err; ASSERT(guid_map != NULL); err = dsl_pool_hold(name, FTAG, &dp); if (err != 0) return (err); gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP); err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds); if (err == 0) { gmep->guid = dsl_dataset_phys(snapds)->ds_guid; gmep->gme_ds = snapds; avl_add(guid_map, gmep); dsl_dataset_long_hold(snapds, gmep); } else kmem_free(gmep, sizeof (*gmep)); dsl_pool_rele(dp, FTAG); return (err); } static int dmu_recv_end_modified_blocks = 3; static int dmu_recv_existing_end(dmu_recv_cookie_t *drc) { #ifdef _KERNEL /* * We will be destroying the ds; make sure its origin is unmounted if * necessary. */ char name[ZFS_MAX_DATASET_NAME_LEN]; dsl_dataset_name(drc->drc_ds, name); zfs_destroy_unmount_origin(name); #endif return (dsl_sync_task(drc->drc_tofs, dmu_recv_end_check, dmu_recv_end_sync, drc, dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL)); } static int dmu_recv_new_end(dmu_recv_cookie_t *drc) { return (dsl_sync_task(drc->drc_tofs, dmu_recv_end_check, dmu_recv_end_sync, drc, dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL)); } int dmu_recv_end(dmu_recv_cookie_t *drc, void *owner) { int error; drc->drc_owner = owner; if (drc->drc_newfs) error = dmu_recv_new_end(drc); else error = dmu_recv_existing_end(drc); if (error != 0) { dmu_recv_cleanup_ds(drc); } else if (drc->drc_guid_to_ds_map != NULL) { (void) add_ds_to_guidmap(drc->drc_tofs, drc->drc_guid_to_ds_map, drc->drc_newsnapobj); } return (error); } /* * Return TRUE if this objset is currently being received into. */ boolean_t dmu_objset_is_receiving(objset_t *os) { return (os->os_dsl_dataset != NULL && os->os_dsl_dataset->ds_owner == dmu_recv_tag); }