Index: head/sys/kern/vfs_cache.c =================================================================== --- head/sys/kern/vfs_cache.c (revision 367534) +++ head/sys/kern/vfs_cache.c (revision 367535) @@ -1,4761 +1,4763 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993, 1995 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * Poul-Henning Kamp of the FreeBSD Project. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #ifdef DDB #include #endif #include static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Name cache"); SDT_PROVIDER_DECLARE(vfs); SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *", "char *"); SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *", "const char *"); SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *", "struct namecache *", "int", "int"); SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *"); SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *"); SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int", "struct vnode *", "char *"); SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative, "struct vnode *", "char *"); SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *", "char *"); SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *", "struct componentname *"); SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *", "struct componentname *"); SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int"); SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *"); SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *", "char *"); SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *", "char *"); SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool"); SDT_PROBE_DECLARE(vfs, namei, lookup, entry); SDT_PROBE_DECLARE(vfs, namei, lookup, return); /* * This structure describes the elements in the cache of recent * names looked up by namei. */ struct negstate { u_char neg_flag; u_char neg_hit; }; _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *), "the state must fit in a union with a pointer without growing it"); struct namecache { LIST_ENTRY(namecache) nc_src; /* source vnode list */ TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */ struct vnode *nc_dvp; /* vnode of parent of name */ union { struct vnode *nu_vp; /* vnode the name refers to */ struct negstate nu_neg;/* negative entry state */ } n_un; u_char nc_flag; /* flag bits */ u_char nc_nlen; /* length of name */ char nc_name[0]; /* segment name + nul */ }; /* * struct namecache_ts repeats struct namecache layout up to the * nc_nlen member. * struct namecache_ts is used in place of struct namecache when time(s) need * to be stored. The nc_dotdottime field is used when a cache entry is mapping * both a non-dotdot directory name plus dotdot for the directory's * parent. * * See below for alignment requirement. */ struct namecache_ts { struct timespec nc_time; /* timespec provided by fs */ struct timespec nc_dotdottime; /* dotdot timespec provided by fs */ int nc_ticks; /* ticks value when entry was added */ int nc_pad; struct namecache nc_nc; }; TAILQ_HEAD(cache_freebatch, namecache); /* * At least mips n32 performs 64-bit accesses to timespec as found * in namecache_ts and requires them to be aligned. Since others * may be in the same spot suffer a little bit and enforce the * alignment for everyone. Note this is a nop for 64-bit platforms. */ #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t) /* * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the * 4.4 BSD codebase. Later on struct namecache was tweaked to become * smaller and the value was bumped to retain the total size, but it * was never re-evaluated for suitability. A simple test counting * lengths during package building shows that the value of 45 covers * about 86% of all added entries, reaching 99% at 65. * * Regardless of the above, use of dedicated zones instead of malloc may be * inducing additional waste. This may be hard to address as said zones are * tied to VFS SMR. Even if retaining them, the current split should be * re-evaluated. */ #ifdef __LP64__ #define CACHE_PATH_CUTOFF 45 #define CACHE_LARGE_PAD 6 #else #define CACHE_PATH_CUTOFF 41 #define CACHE_LARGE_PAD 2 #endif #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1) #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE) #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD) #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE) _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); #define nc_vp n_un.nu_vp #define nc_neg n_un.nu_neg /* * Flags in namecache.nc_flag */ #define NCF_WHITE 0x01 #define NCF_ISDOTDOT 0x02 #define NCF_TS 0x04 #define NCF_DTS 0x08 #define NCF_DVDROP 0x10 #define NCF_NEGATIVE 0x20 #define NCF_INVALID 0x40 #define NCF_WIP 0x80 /* * Flags in negstate.neg_flag */ #define NEG_HOT 0x01 static bool cache_neg_evict_cond(u_long lnumcache); /* * Mark an entry as invalid. * * This is called before it starts getting deconstructed. */ static void cache_ncp_invalidate(struct namecache *ncp) { KASSERT((ncp->nc_flag & NCF_INVALID) == 0, ("%s: entry %p already invalid", __func__, ncp)); atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID); atomic_thread_fence_rel(); } /* * Check whether the entry can be safely used. * * All places which elide locks are supposed to call this after they are * done with reading from an entry. */ #define cache_ncp_canuse(ncp) ({ \ struct namecache *_ncp = (ncp); \ u_char _nc_flag; \ \ atomic_thread_fence_acq(); \ _nc_flag = atomic_load_char(&_ncp->nc_flag); \ __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \ }) /* * Name caching works as follows: * * Names found by directory scans are retained in a cache * for future reference. It is managed LRU, so frequently * used names will hang around. Cache is indexed by hash value * obtained from (dvp, name) where dvp refers to the directory * containing name. * * If it is a "negative" entry, (i.e. for a name that is known NOT to * exist) the vnode pointer will be NULL. * * Upon reaching the last segment of a path, if the reference * is for DELETE, or NOCACHE is set (rewrite), and the * name is located in the cache, it will be dropped. * * These locks are used (in the order in which they can be taken): * NAME TYPE ROLE * vnodelock mtx vnode lists and v_cache_dd field protection * bucketlock mtx for access to given set of hash buckets * neglist mtx negative entry LRU management * * It is legal to take multiple vnodelock and bucketlock locks. The locking * order is lower address first. Both are recursive. * * "." lookups are lockless. * * ".." and vnode -> name lookups require vnodelock. * * name -> vnode lookup requires the relevant bucketlock to be held for reading. * * Insertions and removals of entries require involved vnodes and bucketlocks * to be locked to provide safe operation against other threads modifying the * cache. * * Some lookups result in removal of the found entry (e.g. getting rid of a * negative entry with the intent to create a positive one), which poses a * problem when multiple threads reach the state. Similarly, two different * threads can purge two different vnodes and try to remove the same name. * * If the already held vnode lock is lower than the second required lock, we * can just take the other lock. However, in the opposite case, this could * deadlock. As such, this is resolved by trylocking and if that fails unlocking * the first node, locking everything in order and revalidating the state. */ VFS_SMR_DECLARE; static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Name cache parameters"); static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */ SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0, "Total namecache capacity"); u_int ncsizefactor = 2; SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0, "Size factor for namecache"); static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */ SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0, "Ratio of negative namecache entries"); /* * Negative entry % of namecache capacity above which automatic eviction is allowed. * * Check cache_neg_evict_cond for details. */ static u_int ncnegminpct = 3; static u_int __read_mostly neg_min; /* the above recomputed against ncsize */ SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0, "Negative entry count above which automatic eviction is allowed"); /* * Structures associated with name caching. */ #define NCHHASH(hash) \ (&nchashtbl[(hash) & nchash]) static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */ static u_long __read_mostly nchash; /* size of hash table */ SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, "Size of namecache hash table"); static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */ static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */ struct nchstats nchstats; /* cache effectiveness statistics */ static bool __read_frequently cache_fast_revlookup = true; SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW, &cache_fast_revlookup, 0, ""); static u_int __exclusive_cache_line neg_cycle; #define ncneghash 3 #define numneglists (ncneghash + 1) struct neglist { struct mtx nl_evict_lock; struct mtx nl_lock __aligned(CACHE_LINE_SIZE); TAILQ_HEAD(, namecache) nl_list; TAILQ_HEAD(, namecache) nl_hotlist; u_long nl_hotnum; } __aligned(CACHE_LINE_SIZE); static struct neglist neglists[numneglists]; static inline struct neglist * NCP2NEGLIST(struct namecache *ncp) { return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]); } static inline struct negstate * NCP2NEGSTATE(struct namecache *ncp) { MPASS(ncp->nc_flag & NCF_NEGATIVE); return (&ncp->nc_neg); } #define numbucketlocks (ncbuckethash + 1) static u_int __read_mostly ncbuckethash; static struct mtx_padalign __read_mostly *bucketlocks; #define HASH2BUCKETLOCK(hash) \ ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)])) #define numvnodelocks (ncvnodehash + 1) static u_int __read_mostly ncvnodehash; static struct mtx __read_mostly *vnodelocks; static inline struct mtx * VP2VNODELOCK(struct vnode *vp) { return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]); } static void cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp) { struct namecache_ts *ncp_ts; KASSERT((ncp->nc_flag & NCF_TS) != 0 || (tsp == NULL && ticksp == NULL), ("No NCF_TS")); if (tsp == NULL) return; ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); *tsp = ncp_ts->nc_time; *ticksp = ncp_ts->nc_ticks; } #ifdef DEBUG_CACHE static int __read_mostly doingcache = 1; /* 1 => enable the cache */ SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0, "VFS namecache enabled"); #endif /* Export size information to userland */ SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, sizeof(struct namecache), "sizeof(struct namecache)"); /* * The new name cache statistics */ static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Name cache statistics"); #define STATNODE_ULONG(name, varname, descr) \ SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); #define STATNODE_COUNTER(name, varname, descr) \ static COUNTER_U64_DEFINE_EARLY(varname); \ SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \ descr); STATNODE_ULONG(neg, numneg, "Number of negative cache entries"); STATNODE_ULONG(count, numcache, "Number of cache entries"); STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held"); STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit"); STATNODE_COUNTER(dothits, dothits, "Number of '.' hits"); STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits"); STATNODE_COUNTER(miss, nummiss, "Number of cache misses"); STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache"); STATNODE_COUNTER(posszaps, numposzaps, "Number of cache hits (positive) we do not want to cache"); STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)"); STATNODE_COUNTER(negzaps, numnegzaps, "Number of cache hits (negative) we do not want to cache"); STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)"); /* These count for vn_getcwd(), too. */ STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls"); STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)"); STATNODE_COUNTER(fullpathfail2, numfullpathfail2, "Number of fullpath search errors (VOP_VPTOCNP failures)"); STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)"); STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls"); /* * Debug or developer statistics. */ static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Name cache debugging"); #define DEBUGNODE_ULONG(name, varname, descr) \ SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); #define DEBUGNODE_COUNTER(name, varname, descr) \ static COUNTER_U64_DEFINE_EARLY(varname); \ SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \ descr); DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success, "Number of successful removals after relocking"); static long zap_bucket_fail; DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, ""); static long zap_bucket_fail2; DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, ""); static long cache_lock_vnodes_cel_3_failures; DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures, "Number of times 3-way vnode locking failed"); static void cache_zap_locked(struct namecache *ncp); static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf, size_t *buflen); static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *buflen, size_t addend); static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *buflen); static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *len, size_t addend); static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); static inline void cache_assert_vlp_locked(struct mtx *vlp) { if (vlp != NULL) mtx_assert(vlp, MA_OWNED); } static inline void cache_assert_vnode_locked(struct vnode *vp) { struct mtx *vlp; vlp = VP2VNODELOCK(vp); cache_assert_vlp_locked(vlp); } /* * Directory vnodes with entries are held for two reasons: * 1. make them less of a target for reclamation in vnlru * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided * * Note this preferably would not be done and it's a hold over from. It will be * feasible to eliminate altogether if all filesystems start supporting * lockless lookup. */ static void cache_hold_vnode(struct vnode *vp) { cache_assert_vnode_locked(vp); VNPASS(LIST_EMPTY(&vp->v_cache_src), vp); vhold(vp); counter_u64_add(numcachehv, 1); } static void cache_drop_vnode(struct vnode *vp) { /* * Called after all locks are dropped, meaning we can't assert * on the state of v_cache_src. */ vdrop(vp); counter_u64_add(numcachehv, -1); } /* * UMA zones. */ static uma_zone_t __read_mostly cache_zone_small; static uma_zone_t __read_mostly cache_zone_small_ts; static uma_zone_t __read_mostly cache_zone_large; static uma_zone_t __read_mostly cache_zone_large_ts; static struct namecache * cache_alloc_uma(int len, bool ts) { struct namecache_ts *ncp_ts; struct namecache *ncp; if (__predict_false(ts)) { if (len <= CACHE_PATH_CUTOFF) ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK); else ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK); ncp = &ncp_ts->nc_nc; } else { if (len <= CACHE_PATH_CUTOFF) ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK); else ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK); } return (ncp); } static void cache_free_uma(struct namecache *ncp) { struct namecache_ts *ncp_ts; if (__predict_false(ncp->nc_flag & NCF_TS)) { ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) uma_zfree_smr(cache_zone_small_ts, ncp_ts); else uma_zfree_smr(cache_zone_large_ts, ncp_ts); } else { if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) uma_zfree_smr(cache_zone_small, ncp); else uma_zfree_smr(cache_zone_large, ncp); } } static struct namecache * cache_alloc(int len, bool ts) { u_long lnumcache; /* * Avoid blowout in namecache entries. * * Bugs: * 1. filesystems may end up trying to add an already existing entry * (for example this can happen after a cache miss during concurrent * lookup), in which case we will call cache_neg_evict despite not * adding anything. * 2. the routine may fail to free anything and no provisions are made * to make it try harder (see the inside for failure modes) * 3. it only ever looks at negative entries. */ lnumcache = atomic_fetchadd_long(&numcache, 1) + 1; if (cache_neg_evict_cond(lnumcache)) { lnumcache = atomic_load_long(&numcache); } if (__predict_false(lnumcache >= ncsize)) { atomic_subtract_long(&numcache, 1); counter_u64_add(numdrops, 1); return (NULL); } return (cache_alloc_uma(len, ts)); } static void cache_free(struct namecache *ncp) { MPASS(ncp != NULL); if ((ncp->nc_flag & NCF_DVDROP) != 0) { cache_drop_vnode(ncp->nc_dvp); } cache_free_uma(ncp); atomic_subtract_long(&numcache, 1); } static void cache_free_batch(struct cache_freebatch *batch) { struct namecache *ncp, *nnp; int i; i = 0; if (TAILQ_EMPTY(batch)) goto out; TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) { if ((ncp->nc_flag & NCF_DVDROP) != 0) { cache_drop_vnode(ncp->nc_dvp); } cache_free_uma(ncp); i++; } atomic_subtract_long(&numcache, i); out: SDT_PROBE1(vfs, namecache, purge, batch, i); } /* * TODO: With the value stored we can do better than computing the hash based * on the address. The choice of FNV should also be revisited. */ static void cache_prehash(struct vnode *vp) { vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT); } static uint32_t cache_get_hash(char *name, u_char len, struct vnode *dvp) { return (fnv_32_buf(name, len, dvp->v_nchash)); } static inline struct nchashhead * NCP2BUCKET(struct namecache *ncp) { uint32_t hash; hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); return (NCHHASH(hash)); } static inline struct mtx * NCP2BUCKETLOCK(struct namecache *ncp) { uint32_t hash; hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); return (HASH2BUCKETLOCK(hash)); } #ifdef INVARIANTS static void cache_assert_bucket_locked(struct namecache *ncp) { struct mtx *blp; blp = NCP2BUCKETLOCK(ncp); mtx_assert(blp, MA_OWNED); } static void cache_assert_bucket_unlocked(struct namecache *ncp) { struct mtx *blp; blp = NCP2BUCKETLOCK(ncp); mtx_assert(blp, MA_NOTOWNED); } #else #define cache_assert_bucket_locked(x) do { } while (0) #define cache_assert_bucket_unlocked(x) do { } while (0) #endif #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y)) static void _cache_sort_vnodes(void **p1, void **p2) { void *tmp; MPASS(*p1 != NULL || *p2 != NULL); if (*p1 > *p2) { tmp = *p2; *p2 = *p1; *p1 = tmp; } } static void cache_lock_all_buckets(void) { u_int i; for (i = 0; i < numbucketlocks; i++) mtx_lock(&bucketlocks[i]); } static void cache_unlock_all_buckets(void) { u_int i; for (i = 0; i < numbucketlocks; i++) mtx_unlock(&bucketlocks[i]); } static void cache_lock_all_vnodes(void) { u_int i; for (i = 0; i < numvnodelocks; i++) mtx_lock(&vnodelocks[i]); } static void cache_unlock_all_vnodes(void) { u_int i; for (i = 0; i < numvnodelocks; i++) mtx_unlock(&vnodelocks[i]); } static int cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2) { cache_sort_vnodes(&vlp1, &vlp2); if (vlp1 != NULL) { if (!mtx_trylock(vlp1)) return (EAGAIN); } if (!mtx_trylock(vlp2)) { if (vlp1 != NULL) mtx_unlock(vlp1); return (EAGAIN); } return (0); } static void cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2) { MPASS(vlp1 != NULL || vlp2 != NULL); MPASS(vlp1 <= vlp2); if (vlp1 != NULL) mtx_lock(vlp1); if (vlp2 != NULL) mtx_lock(vlp2); } static void cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2) { MPASS(vlp1 != NULL || vlp2 != NULL); if (vlp1 != NULL) mtx_unlock(vlp1); if (vlp2 != NULL) mtx_unlock(vlp2); } static int sysctl_nchstats(SYSCTL_HANDLER_ARGS) { struct nchstats snap; if (req->oldptr == NULL) return (SYSCTL_OUT(req, 0, sizeof(snap))); snap = nchstats; snap.ncs_goodhits = counter_u64_fetch(numposhits); snap.ncs_neghits = counter_u64_fetch(numneghits); snap.ncs_badhits = counter_u64_fetch(numposzaps) + counter_u64_fetch(numnegzaps); snap.ncs_miss = counter_u64_fetch(nummisszap) + counter_u64_fetch(nummiss); return (SYSCTL_OUT(req, &snap, sizeof(snap))); } SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU", "VFS cache effectiveness statistics"); static void cache_recalc_neg_min(u_int val) { neg_min = (ncsize * val) / 100; } static int sysctl_negminpct(SYSCTL_HANDLER_ARGS) { u_int val; int error; val = ncnegminpct; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val == ncnegminpct) return (0); if (val < 0 || val > 99) return (EINVAL); ncnegminpct = val; cache_recalc_neg_min(val); return (0); } SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct, "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed"); #ifdef DIAGNOSTIC /* * Grab an atomic snapshot of the name cache hash chain lengths */ static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "hash table stats"); static int sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS) { struct nchashhead *ncpp; struct namecache *ncp; int i, error, n_nchash, *cntbuf; retry: n_nchash = nchash + 1; /* nchash is max index, not count */ if (req->oldptr == NULL) return SYSCTL_OUT(req, 0, n_nchash * sizeof(int)); cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK); cache_lock_all_buckets(); if (n_nchash != nchash + 1) { cache_unlock_all_buckets(); free(cntbuf, M_TEMP); goto retry; } /* Scan hash tables counting entries */ for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++) CK_SLIST_FOREACH(ncp, ncpp, nc_hash) cntbuf[i]++; cache_unlock_all_buckets(); for (error = 0, i = 0; i < n_nchash; i++) if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0) break; free(cntbuf, M_TEMP); return (error); } SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD| CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int", "nchash chain lengths"); static int sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS) { int error; struct nchashhead *ncpp; struct namecache *ncp; int n_nchash; int count, maxlength, used, pct; if (!req->oldptr) return SYSCTL_OUT(req, 0, 4 * sizeof(int)); cache_lock_all_buckets(); n_nchash = nchash + 1; /* nchash is max index, not count */ used = 0; maxlength = 0; /* Scan hash tables for applicable entries */ for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) { count = 0; CK_SLIST_FOREACH(ncp, ncpp, nc_hash) { count++; } if (count) used++; if (maxlength < count) maxlength = count; } n_nchash = nchash + 1; cache_unlock_all_buckets(); pct = (used * 100) / (n_nchash / 100); error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash)); if (error) return (error); error = SYSCTL_OUT(req, &used, sizeof(used)); if (error) return (error); error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength)); if (error) return (error); error = SYSCTL_OUT(req, &pct, sizeof(pct)); if (error) return (error); return (0); } SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD| CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I", "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)"); #endif /* * Negative entries management * * Various workloads create plenty of negative entries and barely use them * afterwards. Moreover malicious users can keep performing bogus lookups * adding even more entries. For example "make tinderbox" as of writing this * comment ends up with 2.6M namecache entries in total, 1.2M of which are * negative. * * As such, a rather aggressive eviction method is needed. The currently * employed method is a placeholder. * * Entries are split over numneglists separate lists, each of which is further * split into hot and cold entries. Entries get promoted after getting a hit. * Eviction happens on addition of new entry. */ static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Name cache negative entry statistics"); SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0, "Number of negative cache entries"); static COUNTER_U64_DEFINE_EARLY(neg_created); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created, "Number of created negative entries"); static COUNTER_U64_DEFINE_EARLY(neg_evicted); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted, "Number of evicted negative entries"); static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD, &neg_evict_skipped_empty, "Number of times evicting failed due to lack of entries"); static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD, &neg_evict_skipped_missed, "Number of times evicting failed due to target entry disappearing"); static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD, &neg_evict_skipped_contended, "Number of times evicting failed due to contention"); SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits, "Number of cache hits (negative)"); static int sysctl_neg_hot(SYSCTL_HANDLER_ARGS) { int i, out; out = 0; for (i = 0; i < numneglists; i++) out += neglists[i].nl_hotnum; return (SYSCTL_OUT(req, &out, sizeof(out))); } SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I", "Number of hot negative entries"); static void cache_neg_init(struct namecache *ncp) { struct negstate *ns; ncp->nc_flag |= NCF_NEGATIVE; ns = NCP2NEGSTATE(ncp); ns->neg_flag = 0; ns->neg_hit = 0; counter_u64_add(neg_created, 1); } #define CACHE_NEG_PROMOTION_THRESH 2 static bool cache_neg_hit_prep(struct namecache *ncp) { struct negstate *ns; u_char n; ns = NCP2NEGSTATE(ncp); n = atomic_load_char(&ns->neg_hit); for (;;) { if (n >= CACHE_NEG_PROMOTION_THRESH) return (false); if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1)) break; } return (n + 1 == CACHE_NEG_PROMOTION_THRESH); } /* * Nothing to do here but it is provided for completeness as some * cache_neg_hit_prep callers may end up returning without even * trying to promote. */ #define cache_neg_hit_abort(ncp) do { } while (0) static void cache_neg_hit_finish(struct namecache *ncp) { SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name); counter_u64_add(numneghits, 1); } /* * Move a negative entry to the hot list. */ static void cache_neg_promote_locked(struct namecache *ncp) { struct neglist *nl; struct negstate *ns; ns = NCP2NEGSTATE(ncp); nl = NCP2NEGLIST(ncp); mtx_assert(&nl->nl_lock, MA_OWNED); if ((ns->neg_flag & NEG_HOT) == 0) { TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst); nl->nl_hotnum++; ns->neg_flag |= NEG_HOT; } } /* * Move a hot negative entry to the cold list. */ static void cache_neg_demote_locked(struct namecache *ncp) { struct neglist *nl; struct negstate *ns; ns = NCP2NEGSTATE(ncp); nl = NCP2NEGLIST(ncp); mtx_assert(&nl->nl_lock, MA_OWNED); MPASS(ns->neg_flag & NEG_HOT); TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); nl->nl_hotnum--; ns->neg_flag &= ~NEG_HOT; atomic_store_char(&ns->neg_hit, 0); } /* * Move a negative entry to the hot list if it matches the lookup. * * We have to take locks, but they may be contended and in the worst * case we may need to go off CPU. We don't want to spin within the * smr section and we can't block with it. Exiting the section means * the found entry could have been evicted. We are going to look it * up again. */ static bool cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp, struct namecache *oncp, uint32_t hash) { struct namecache *ncp; struct neglist *nl; u_char nc_flag; nl = NCP2NEGLIST(oncp); mtx_lock(&nl->nl_lock); /* * For hash iteration. */ vfs_smr_enter(); /* * Avoid all surprises by only succeeding if we got the same entry and * bailing completely otherwise. * XXX There are no provisions to keep the vnode around, meaning we may * end up promoting a negative entry for a *new* vnode and returning * ENOENT on its account. This is the error we want to return anyway * and promotion is harmless. * * In particular at this point there can be a new ncp which matches the * search but hashes to a different neglist. */ CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp == oncp) break; } /* * No match to begin with. */ if (__predict_false(ncp == NULL)) { goto out_abort; } /* * The newly found entry may be something different... */ if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) { goto out_abort; } /* * ... and not even negative. */ nc_flag = atomic_load_char(&ncp->nc_flag); if ((nc_flag & NCF_NEGATIVE) == 0) { goto out_abort; } if (!cache_ncp_canuse(ncp)) { goto out_abort; } cache_neg_promote_locked(ncp); cache_neg_hit_finish(ncp); vfs_smr_exit(); mtx_unlock(&nl->nl_lock); return (true); out_abort: vfs_smr_exit(); mtx_unlock(&nl->nl_lock); return (false); } static void cache_neg_promote(struct namecache *ncp) { struct neglist *nl; nl = NCP2NEGLIST(ncp); mtx_lock(&nl->nl_lock); cache_neg_promote_locked(ncp); mtx_unlock(&nl->nl_lock); } static void cache_neg_insert(struct namecache *ncp) { struct neglist *nl; MPASS(ncp->nc_flag & NCF_NEGATIVE); cache_assert_bucket_locked(ncp); nl = NCP2NEGLIST(ncp); mtx_lock(&nl->nl_lock); TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); mtx_unlock(&nl->nl_lock); atomic_add_long(&numneg, 1); } static void cache_neg_remove(struct namecache *ncp) { struct neglist *nl; struct negstate *ns; cache_assert_bucket_locked(ncp); nl = NCP2NEGLIST(ncp); ns = NCP2NEGSTATE(ncp); mtx_lock(&nl->nl_lock); if ((ns->neg_flag & NEG_HOT) != 0) { TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); nl->nl_hotnum--; } else { TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); } mtx_unlock(&nl->nl_lock); atomic_subtract_long(&numneg, 1); } static struct neglist * cache_neg_evict_select_list(void) { struct neglist *nl; u_int c; c = atomic_fetchadd_int(&neg_cycle, 1) + 1; nl = &neglists[c % numneglists]; if (!mtx_trylock(&nl->nl_evict_lock)) { counter_u64_add(neg_evict_skipped_contended, 1); return (NULL); } return (nl); } static struct namecache * cache_neg_evict_select_entry(struct neglist *nl) { struct namecache *ncp, *lncp; struct negstate *ns, *lns; int i; mtx_assert(&nl->nl_evict_lock, MA_OWNED); mtx_assert(&nl->nl_lock, MA_OWNED); ncp = TAILQ_FIRST(&nl->nl_list); if (ncp == NULL) return (NULL); lncp = ncp; lns = NCP2NEGSTATE(lncp); for (i = 1; i < 4; i++) { ncp = TAILQ_NEXT(ncp, nc_dst); if (ncp == NULL) break; ns = NCP2NEGSTATE(ncp); if (ns->neg_hit < lns->neg_hit) { lncp = ncp; lns = ns; } } return (lncp); } static bool cache_neg_evict(void) { struct namecache *ncp, *ncp2; struct neglist *nl; struct vnode *dvp; struct mtx *dvlp; struct mtx *blp; uint32_t hash; u_char nlen; bool evicted; nl = cache_neg_evict_select_list(); if (nl == NULL) { return (false); } mtx_lock(&nl->nl_lock); ncp = TAILQ_FIRST(&nl->nl_hotlist); if (ncp != NULL) { cache_neg_demote_locked(ncp); } ncp = cache_neg_evict_select_entry(nl); if (ncp == NULL) { counter_u64_add(neg_evict_skipped_empty, 1); mtx_unlock(&nl->nl_lock); mtx_unlock(&nl->nl_evict_lock); return (false); } nlen = ncp->nc_nlen; dvp = ncp->nc_dvp; hash = cache_get_hash(ncp->nc_name, nlen, dvp); dvlp = VP2VNODELOCK(dvp); blp = HASH2BUCKETLOCK(hash); mtx_unlock(&nl->nl_lock); mtx_unlock(&nl->nl_evict_lock); mtx_lock(dvlp); mtx_lock(blp); /* * Note that since all locks were dropped above, the entry may be * gone or reallocated to be something else. */ CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) { if (ncp2 == ncp && ncp2->nc_dvp == dvp && ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0) break; } if (ncp2 == NULL) { counter_u64_add(neg_evict_skipped_missed, 1); ncp = NULL; evicted = false; } else { MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp)); MPASS(blp == NCP2BUCKETLOCK(ncp)); SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp, ncp->nc_name); cache_zap_locked(ncp); counter_u64_add(neg_evicted, 1); evicted = true; } mtx_unlock(blp); mtx_unlock(dvlp); if (ncp != NULL) cache_free(ncp); return (evicted); } /* * Maybe evict a negative entry to create more room. * * The ncnegfactor parameter limits what fraction of the total count * can comprise of negative entries. However, if the cache is just * warming up this leads to excessive evictions. As such, ncnegminpct * (recomputed to neg_min) dictates whether the above should be * applied. * * Try evicting if the cache is close to full capacity regardless of * other considerations. */ static bool cache_neg_evict_cond(u_long lnumcache) { u_long lnumneg; if (ncsize - 1000 < lnumcache) goto out_evict; lnumneg = atomic_load_long(&numneg); if (lnumneg < neg_min) return (false); if (lnumneg * ncnegfactor < lnumcache) return (false); out_evict: return (cache_neg_evict()); } /* * cache_zap_locked(): * * Removes a namecache entry from cache, whether it contains an actual * pointer to a vnode or if it is just a negative cache entry. */ static void cache_zap_locked(struct namecache *ncp) { struct nchashhead *ncpp; if (!(ncp->nc_flag & NCF_NEGATIVE)) cache_assert_vnode_locked(ncp->nc_vp); cache_assert_vnode_locked(ncp->nc_dvp); cache_assert_bucket_locked(ncp); cache_ncp_invalidate(ncp); ncpp = NCP2BUCKET(ncp); CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash); if (!(ncp->nc_flag & NCF_NEGATIVE)) { SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp, ncp->nc_name, ncp->nc_vp); TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst); if (ncp == ncp->nc_vp->v_cache_dd) { vn_seqc_write_begin_unheld(ncp->nc_vp); ncp->nc_vp->v_cache_dd = NULL; vn_seqc_write_end(ncp->nc_vp); } } else { SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp, ncp->nc_name); cache_neg_remove(ncp); } if (ncp->nc_flag & NCF_ISDOTDOT) { if (ncp == ncp->nc_dvp->v_cache_dd) { vn_seqc_write_begin_unheld(ncp->nc_dvp); ncp->nc_dvp->v_cache_dd = NULL; vn_seqc_write_end(ncp->nc_dvp); } } else { LIST_REMOVE(ncp, nc_src); if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) { ncp->nc_flag |= NCF_DVDROP; } } } static void cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp) { struct mtx *blp; MPASS(ncp->nc_dvp == vp); MPASS(ncp->nc_flag & NCF_NEGATIVE); cache_assert_vnode_locked(vp); blp = NCP2BUCKETLOCK(ncp); mtx_lock(blp); cache_zap_locked(ncp); mtx_unlock(blp); } static bool cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp, struct mtx **vlpp) { struct mtx *pvlp, *vlp1, *vlp2, *to_unlock; struct mtx *blp; MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp); cache_assert_vnode_locked(vp); if (ncp->nc_flag & NCF_NEGATIVE) { if (*vlpp != NULL) { mtx_unlock(*vlpp); *vlpp = NULL; } cache_zap_negative_locked_vnode_kl(ncp, vp); return (true); } pvlp = VP2VNODELOCK(vp); blp = NCP2BUCKETLOCK(ncp); vlp1 = VP2VNODELOCK(ncp->nc_dvp); vlp2 = VP2VNODELOCK(ncp->nc_vp); if (*vlpp == vlp1 || *vlpp == vlp2) { to_unlock = *vlpp; *vlpp = NULL; } else { if (*vlpp != NULL) { mtx_unlock(*vlpp); *vlpp = NULL; } cache_sort_vnodes(&vlp1, &vlp2); if (vlp1 == pvlp) { mtx_lock(vlp2); to_unlock = vlp2; } else { if (!mtx_trylock(vlp1)) goto out_relock; to_unlock = vlp1; } } mtx_lock(blp); cache_zap_locked(ncp); mtx_unlock(blp); if (to_unlock != NULL) mtx_unlock(to_unlock); return (true); out_relock: mtx_unlock(vlp2); mtx_lock(vlp1); mtx_lock(vlp2); MPASS(*vlpp == NULL); *vlpp = vlp1; return (false); } /* * If trylocking failed we can get here. We know enough to take all needed locks * in the right order and re-lookup the entry. */ static int cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp, struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash, struct mtx *blp) { struct namecache *rncp; cache_assert_bucket_unlocked(ncp); cache_sort_vnodes(&dvlp, &vlp); cache_lock_vnodes(dvlp, vlp); mtx_lock(blp); CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) { if (rncp == ncp && rncp->nc_dvp == dvp && rncp->nc_nlen == cnp->cn_namelen && !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen)) break; } if (rncp != NULL) { cache_zap_locked(rncp); mtx_unlock(blp); cache_unlock_vnodes(dvlp, vlp); counter_u64_add(zap_bucket_relock_success, 1); return (0); } mtx_unlock(blp); cache_unlock_vnodes(dvlp, vlp); return (EAGAIN); } static int __noinline cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp, uint32_t hash, struct mtx *blp) { struct mtx *dvlp, *vlp; struct vnode *dvp; cache_assert_bucket_locked(ncp); dvlp = VP2VNODELOCK(ncp->nc_dvp); vlp = NULL; if (!(ncp->nc_flag & NCF_NEGATIVE)) vlp = VP2VNODELOCK(ncp->nc_vp); if (cache_trylock_vnodes(dvlp, vlp) == 0) { cache_zap_locked(ncp); mtx_unlock(blp); cache_unlock_vnodes(dvlp, vlp); return (0); } dvp = ncp->nc_dvp; mtx_unlock(blp); return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp)); } static __noinline int cache_remove_cnp(struct vnode *dvp, struct componentname *cnp) { struct namecache *ncp; struct mtx *blp; struct mtx *dvlp, *dvlp2; uint32_t hash; int error; if (cnp->cn_namelen == 2 && cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') { dvlp = VP2VNODELOCK(dvp); dvlp2 = NULL; mtx_lock(dvlp); retry_dotdot: ncp = dvp->v_cache_dd; if (ncp == NULL) { mtx_unlock(dvlp); if (dvlp2 != NULL) mtx_unlock(dvlp2); SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); return (0); } if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2)) goto retry_dotdot; MPASS(dvp->v_cache_dd == NULL); mtx_unlock(dvlp); if (dvlp2 != NULL) mtx_unlock(dvlp2); cache_free(ncp); } else { vn_seqc_write_begin(dvp); dvp->v_cache_dd = NULL; vn_seqc_write_end(dvp); mtx_unlock(dvlp); if (dvlp2 != NULL) mtx_unlock(dvlp2); } SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); return (1); } hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); blp = HASH2BUCKETLOCK(hash); retry: if (CK_SLIST_EMPTY(NCHHASH(hash))) goto out_no_entry; mtx_lock(blp); CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) break; } if (ncp == NULL) { mtx_unlock(blp); goto out_no_entry; } error = cache_zap_locked_bucket(ncp, cnp, hash, blp); if (__predict_false(error != 0)) { zap_bucket_fail++; goto retry; } counter_u64_add(numposzaps, 1); SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); cache_free(ncp); return (1); out_no_entry: counter_u64_add(nummisszap, 1); SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); return (0); } static int __noinline cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp) { int ltype; *vpp = dvp; counter_u64_add(dothits, 1); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp); if (tsp != NULL) timespecclear(tsp); if (ticksp != NULL) *ticksp = ticks; vrefact(*vpp); /* * When we lookup "." we still can be asked to lock it * differently... */ ltype = cnp->cn_lkflags & LK_TYPE_MASK; if (ltype != VOP_ISLOCKED(*vpp)) { if (ltype == LK_EXCLUSIVE) { vn_lock(*vpp, LK_UPGRADE | LK_RETRY); if (VN_IS_DOOMED((*vpp))) { /* forced unmount */ vrele(*vpp); *vpp = NULL; return (ENOENT); } } else vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY); } return (-1); } static int __noinline cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp) { struct namecache_ts *ncp_ts; struct namecache *ncp; struct mtx *dvlp; enum vgetstate vs; int error, ltype; bool whiteout; MPASS((cnp->cn_flags & ISDOTDOT) != 0); if ((cnp->cn_flags & MAKEENTRY) == 0) { cache_remove_cnp(dvp, cnp); return (0); } counter_u64_add(dotdothits, 1); retry: dvlp = VP2VNODELOCK(dvp); mtx_lock(dvlp); ncp = dvp->v_cache_dd; if (ncp == NULL) { SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL); mtx_unlock(dvlp); return (0); } if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { if (ncp->nc_flag & NCF_NEGATIVE) *vpp = NULL; else *vpp = ncp->nc_vp; } else *vpp = ncp->nc_dvp; if (*vpp == NULL) goto negative_success; SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp); cache_out_ts(ncp, tsp, ticksp); if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) == NCF_DTS && tsp != NULL) { ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); *tsp = ncp_ts->nc_dotdottime; } MPASS(dvp != *vpp); ltype = VOP_ISLOCKED(dvp); VOP_UNLOCK(dvp); vs = vget_prep(*vpp); mtx_unlock(dvlp); error = vget_finish(*vpp, cnp->cn_lkflags, vs); vn_lock(dvp, ltype | LK_RETRY); if (VN_IS_DOOMED(dvp)) { if (error == 0) vput(*vpp); *vpp = NULL; return (ENOENT); } if (error) { *vpp = NULL; goto retry; } return (-1); negative_success: if (__predict_false(cnp->cn_nameiop == CREATE)) { if (cnp->cn_flags & ISLASTCN) { counter_u64_add(numnegzaps, 1); cache_zap_negative_locked_vnode_kl(ncp, dvp); mtx_unlock(dvlp); cache_free(ncp); return (0); } } whiteout = (ncp->nc_flag & NCF_WHITE); cache_out_ts(ncp, tsp, ticksp); if (cache_neg_hit_prep(ncp)) cache_neg_promote(ncp); else cache_neg_hit_finish(ncp); mtx_unlock(dvlp); if (whiteout) cnp->cn_flags |= ISWHITEOUT; return (ENOENT); } /** * Lookup a name in the name cache * * # Arguments * * - dvp: Parent directory in which to search. * - vpp: Return argument. Will contain desired vnode on cache hit. * - cnp: Parameters of the name search. The most interesting bits of * the cn_flags field have the following meanings: * - MAKEENTRY: If clear, free an entry from the cache rather than look * it up. * - ISDOTDOT: Must be set if and only if cn_nameptr == ".." * - tsp: Return storage for cache timestamp. On a successful (positive * or negative) lookup, tsp will be filled with any timespec that * was stored when this cache entry was created. However, it will * be clear for "." entries. * - ticks: Return storage for alternate cache timestamp. On a successful * (positive or negative) lookup, it will contain the ticks value * that was current when the cache entry was created, unless cnp * was ".". * * Either both tsp and ticks have to be provided or neither of them. * * # Returns * * - -1: A positive cache hit. vpp will contain the desired vnode. * - ENOENT: A negative cache hit, or dvp was recycled out from under us due * to a forced unmount. vpp will not be modified. If the entry * is a whiteout, then the ISWHITEOUT flag will be set in * cnp->cn_flags. * - 0: A cache miss. vpp will not be modified. * * # Locking * * On a cache hit, vpp will be returned locked and ref'd. If we're looking up * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the * lock is not recursively acquired. */ static int __noinline cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp) { struct namecache *ncp; struct mtx *blp; uint32_t hash; enum vgetstate vs; int error; bool whiteout; MPASS((cnp->cn_flags & ISDOTDOT) == 0); MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0); retry: hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); blp = HASH2BUCKETLOCK(hash); mtx_lock(blp); CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) break; } if (__predict_false(ncp == NULL)) { mtx_unlock(blp); SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL); counter_u64_add(nummiss, 1); return (0); } if (ncp->nc_flag & NCF_NEGATIVE) goto negative_success; counter_u64_add(numposhits, 1); *vpp = ncp->nc_vp; SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); cache_out_ts(ncp, tsp, ticksp); MPASS(dvp != *vpp); vs = vget_prep(*vpp); mtx_unlock(blp); error = vget_finish(*vpp, cnp->cn_lkflags, vs); if (error) { *vpp = NULL; goto retry; } return (-1); negative_success: if (__predict_false(cnp->cn_nameiop == CREATE)) { if (cnp->cn_flags & ISLASTCN) { counter_u64_add(numnegzaps, 1); error = cache_zap_locked_bucket(ncp, cnp, hash, blp); if (__predict_false(error != 0)) { zap_bucket_fail2++; goto retry; } cache_free(ncp); return (0); } } whiteout = (ncp->nc_flag & NCF_WHITE); cache_out_ts(ncp, tsp, ticksp); if (cache_neg_hit_prep(ncp)) cache_neg_promote(ncp); else cache_neg_hit_finish(ncp); mtx_unlock(blp); if (whiteout) cnp->cn_flags |= ISWHITEOUT; return (ENOENT); } int cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp) { struct namecache *ncp; uint32_t hash; enum vgetstate vs; int error; bool whiteout, neg_promote; u_short nc_flag; MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL)); #ifdef DEBUG_CACHE if (__predict_false(!doingcache)) { cnp->cn_flags &= ~MAKEENTRY; return (0); } #endif if (__predict_false(cnp->cn_nameptr[0] == '.')) { if (cnp->cn_namelen == 1) return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp)); if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp)); } MPASS((cnp->cn_flags & ISDOTDOT) == 0); if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) { cache_remove_cnp(dvp, cnp); return (0); } hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); vfs_smr_enter(); CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) break; } if (__predict_false(ncp == NULL)) { vfs_smr_exit(); SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL); counter_u64_add(nummiss, 1); return (0); } nc_flag = atomic_load_char(&ncp->nc_flag); if (nc_flag & NCF_NEGATIVE) goto negative_success; counter_u64_add(numposhits, 1); *vpp = ncp->nc_vp; SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); cache_out_ts(ncp, tsp, ticksp); MPASS(dvp != *vpp); if (!cache_ncp_canuse(ncp)) { vfs_smr_exit(); *vpp = NULL; goto out_fallback; } vs = vget_prep_smr(*vpp); vfs_smr_exit(); if (__predict_false(vs == VGET_NONE)) { *vpp = NULL; goto out_fallback; } error = vget_finish(*vpp, cnp->cn_lkflags, vs); if (error) { *vpp = NULL; goto out_fallback; } return (-1); negative_success: if (__predict_false(cnp->cn_nameiop == CREATE)) { if (cnp->cn_flags & ISLASTCN) { vfs_smr_exit(); goto out_fallback; } } cache_out_ts(ncp, tsp, ticksp); whiteout = (ncp->nc_flag & NCF_WHITE); neg_promote = cache_neg_hit_prep(ncp); if (!cache_ncp_canuse(ncp)) { cache_neg_hit_abort(ncp); vfs_smr_exit(); goto out_fallback; } if (neg_promote) { vfs_smr_exit(); if (!cache_neg_promote_cond(dvp, cnp, ncp, hash)) goto out_fallback; } else { cache_neg_hit_finish(ncp); vfs_smr_exit(); } if (whiteout) cnp->cn_flags |= ISWHITEOUT; return (ENOENT); out_fallback: return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp)); } struct celockstate { struct mtx *vlp[3]; struct mtx *blp[2]; }; CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3)); CTASSERT((nitems(((struct celockstate *)0)->blp) == 2)); static inline void cache_celockstate_init(struct celockstate *cel) { bzero(cel, sizeof(*cel)); } static void cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp, struct vnode *dvp) { struct mtx *vlp1, *vlp2; MPASS(cel->vlp[0] == NULL); MPASS(cel->vlp[1] == NULL); MPASS(cel->vlp[2] == NULL); MPASS(vp != NULL || dvp != NULL); vlp1 = VP2VNODELOCK(vp); vlp2 = VP2VNODELOCK(dvp); cache_sort_vnodes(&vlp1, &vlp2); if (vlp1 != NULL) { mtx_lock(vlp1); cel->vlp[0] = vlp1; } mtx_lock(vlp2); cel->vlp[1] = vlp2; } static void cache_unlock_vnodes_cel(struct celockstate *cel) { MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL); if (cel->vlp[0] != NULL) mtx_unlock(cel->vlp[0]); if (cel->vlp[1] != NULL) mtx_unlock(cel->vlp[1]); if (cel->vlp[2] != NULL) mtx_unlock(cel->vlp[2]); } static bool cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp) { struct mtx *vlp; bool ret; cache_assert_vlp_locked(cel->vlp[0]); cache_assert_vlp_locked(cel->vlp[1]); MPASS(cel->vlp[2] == NULL); MPASS(vp != NULL); vlp = VP2VNODELOCK(vp); ret = true; if (vlp >= cel->vlp[1]) { mtx_lock(vlp); } else { if (mtx_trylock(vlp)) goto out; cache_lock_vnodes_cel_3_failures++; cache_unlock_vnodes_cel(cel); if (vlp < cel->vlp[0]) { mtx_lock(vlp); mtx_lock(cel->vlp[0]); mtx_lock(cel->vlp[1]); } else { if (cel->vlp[0] != NULL) mtx_lock(cel->vlp[0]); mtx_lock(vlp); mtx_lock(cel->vlp[1]); } ret = false; } out: cel->vlp[2] = vlp; return (ret); } static void cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1, struct mtx *blp2) { MPASS(cel->blp[0] == NULL); MPASS(cel->blp[1] == NULL); cache_sort_vnodes(&blp1, &blp2); if (blp1 != NULL) { mtx_lock(blp1); cel->blp[0] = blp1; } mtx_lock(blp2); cel->blp[1] = blp2; } static void cache_unlock_buckets_cel(struct celockstate *cel) { if (cel->blp[0] != NULL) mtx_unlock(cel->blp[0]); mtx_unlock(cel->blp[1]); } /* * Lock part of the cache affected by the insertion. * * This means vnodelocks for dvp, vp and the relevant bucketlock. * However, insertion can result in removal of an old entry. In this * case we have an additional vnode and bucketlock pair to lock. * * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while * preserving the locking order (smaller address first). */ static void cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, uint32_t hash) { struct namecache *ncp; struct mtx *blps[2]; blps[0] = HASH2BUCKETLOCK(hash); for (;;) { blps[1] = NULL; cache_lock_vnodes_cel(cel, dvp, vp); if (vp == NULL || vp->v_type != VDIR) break; ncp = vp->v_cache_dd; if (ncp == NULL) break; if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) break; MPASS(ncp->nc_dvp == vp); blps[1] = NCP2BUCKETLOCK(ncp); if (ncp->nc_flag & NCF_NEGATIVE) break; if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) break; /* * All vnodes got re-locked. Re-validate the state and if * nothing changed we are done. Otherwise restart. */ if (ncp == vp->v_cache_dd && (ncp->nc_flag & NCF_ISDOTDOT) != 0 && blps[1] == NCP2BUCKETLOCK(ncp) && VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) break; cache_unlock_vnodes_cel(cel); cel->vlp[0] = NULL; cel->vlp[1] = NULL; cel->vlp[2] = NULL; } cache_lock_buckets_cel(cel, blps[0], blps[1]); } static void cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, uint32_t hash) { struct namecache *ncp; struct mtx *blps[2]; blps[0] = HASH2BUCKETLOCK(hash); for (;;) { blps[1] = NULL; cache_lock_vnodes_cel(cel, dvp, vp); ncp = dvp->v_cache_dd; if (ncp == NULL) break; if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) break; MPASS(ncp->nc_dvp == dvp); blps[1] = NCP2BUCKETLOCK(ncp); if (ncp->nc_flag & NCF_NEGATIVE) break; if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) break; if (ncp == dvp->v_cache_dd && (ncp->nc_flag & NCF_ISDOTDOT) != 0 && blps[1] == NCP2BUCKETLOCK(ncp) && VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) break; cache_unlock_vnodes_cel(cel); cel->vlp[0] = NULL; cel->vlp[1] = NULL; cel->vlp[2] = NULL; } cache_lock_buckets_cel(cel, blps[0], blps[1]); } static void cache_enter_unlock(struct celockstate *cel) { cache_unlock_buckets_cel(cel); cache_unlock_vnodes_cel(cel); } static void __noinline cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { struct celockstate cel; struct namecache *ncp; uint32_t hash; int len; if (dvp->v_cache_dd == NULL) return; len = cnp->cn_namelen; cache_celockstate_init(&cel); hash = cache_get_hash(cnp->cn_nameptr, len, dvp); cache_enter_lock_dd(&cel, dvp, vp, hash); vn_seqc_write_begin(dvp); ncp = dvp->v_cache_dd; if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) { KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent")); cache_zap_locked(ncp); } else { ncp = NULL; } dvp->v_cache_dd = NULL; vn_seqc_write_end(dvp); cache_enter_unlock(&cel); if (ncp != NULL) cache_free(ncp); } /* * Add an entry to the cache. */ void cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, struct timespec *tsp, struct timespec *dtsp) { struct celockstate cel; struct namecache *ncp, *n2, *ndd; struct namecache_ts *ncp_ts; struct nchashhead *ncpp; uint32_t hash; int flag; int len; VNPASS(dvp != vp, dvp); VNPASS(!VN_IS_DOOMED(dvp), dvp); VNPASS(dvp->v_type != VNON, dvp); if (vp != NULL) { VNPASS(!VN_IS_DOOMED(vp), vp); VNPASS(vp->v_type != VNON, vp); } #ifdef DEBUG_CACHE if (__predict_false(!doingcache)) return; #endif flag = 0; if (__predict_false(cnp->cn_nameptr[0] == '.')) { if (cnp->cn_namelen == 1) return; if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { cache_enter_dotdot_prep(dvp, vp, cnp); flag = NCF_ISDOTDOT; } } ncp = cache_alloc(cnp->cn_namelen, tsp != NULL); if (ncp == NULL) return; cache_celockstate_init(&cel); ndd = NULL; ncp_ts = NULL; /* * Calculate the hash key and setup as much of the new * namecache entry as possible before acquiring the lock. */ ncp->nc_flag = flag | NCF_WIP; ncp->nc_vp = vp; if (vp == NULL) cache_neg_init(ncp); ncp->nc_dvp = dvp; if (tsp != NULL) { ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); ncp_ts->nc_time = *tsp; ncp_ts->nc_ticks = ticks; ncp_ts->nc_nc.nc_flag |= NCF_TS; if (dtsp != NULL) { ncp_ts->nc_dotdottime = *dtsp; ncp_ts->nc_nc.nc_flag |= NCF_DTS; } } len = ncp->nc_nlen = cnp->cn_namelen; hash = cache_get_hash(cnp->cn_nameptr, len, dvp); memcpy(ncp->nc_name, cnp->cn_nameptr, len); ncp->nc_name[len] = '\0'; cache_enter_lock(&cel, dvp, vp, hash); /* * See if this vnode or negative entry is already in the cache * with this name. This can happen with concurrent lookups of * the same path name. */ ncpp = NCHHASH(hash); CK_SLIST_FOREACH(n2, ncpp, nc_hash) { if (n2->nc_dvp == dvp && n2->nc_nlen == cnp->cn_namelen && !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) { MPASS(cache_ncp_canuse(n2)); if ((n2->nc_flag & NCF_NEGATIVE) != 0) KASSERT(vp == NULL, ("%s: found entry pointing to a different vnode (%p != %p)", __func__, NULL, vp)); else KASSERT(n2->nc_vp == vp, ("%s: found entry pointing to a different vnode (%p != %p)", __func__, n2->nc_vp, vp)); /* * Entries are supposed to be immutable unless in the * process of getting destroyed. Accommodating for * changing timestamps is possible but not worth it. * This should be harmless in terms of correctness, in * the worst case resulting in an earlier expiration. * Alternatively, the found entry can be replaced * altogether. */ MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS))); #if 0 if (tsp != NULL) { KASSERT((n2->nc_flag & NCF_TS) != 0, ("no NCF_TS")); n2_ts = __containerof(n2, struct namecache_ts, nc_nc); n2_ts->nc_time = ncp_ts->nc_time; n2_ts->nc_ticks = ncp_ts->nc_ticks; if (dtsp != NULL) { n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime; n2_ts->nc_nc.nc_flag |= NCF_DTS; } } #endif SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name, vp); goto out_unlock_free; } } if (flag == NCF_ISDOTDOT) { /* * See if we are trying to add .. entry, but some other lookup * has populated v_cache_dd pointer already. */ if (dvp->v_cache_dd != NULL) goto out_unlock_free; KASSERT(vp == NULL || vp->v_type == VDIR, ("wrong vnode type %p", vp)); vn_seqc_write_begin(dvp); dvp->v_cache_dd = ncp; vn_seqc_write_end(dvp); } if (vp != NULL) { if (flag != NCF_ISDOTDOT) { /* * For this case, the cache entry maps both the * directory name in it and the name ".." for the * directory's parent. */ vn_seqc_write_begin(vp); if ((ndd = vp->v_cache_dd) != NULL) { if ((ndd->nc_flag & NCF_ISDOTDOT) != 0) cache_zap_locked(ndd); else ndd = NULL; } vp->v_cache_dd = ncp; vn_seqc_write_end(vp); } else if (vp->v_type != VDIR) { if (vp->v_cache_dd != NULL) { vn_seqc_write_begin(vp); vp->v_cache_dd = NULL; vn_seqc_write_end(vp); } } } if (flag != NCF_ISDOTDOT) { if (LIST_EMPTY(&dvp->v_cache_src)) { cache_hold_vnode(dvp); } LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src); } /* * If the entry is "negative", we place it into the * "negative" cache queue, otherwise, we place it into the * destination vnode's cache entries queue. */ if (vp != NULL) { TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name, vp); } else { if (cnp->cn_flags & ISWHITEOUT) ncp->nc_flag |= NCF_WHITE; cache_neg_insert(ncp); SDT_PROBE2(vfs, namecache, enter_negative, done, dvp, ncp->nc_name); } /* * Insert the new namecache entry into the appropriate chain * within the cache entries table. */ CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash); atomic_thread_fence_rel(); /* * Mark the entry as fully constructed. * It is immutable past this point until its removal. */ atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP); cache_enter_unlock(&cel); if (ndd != NULL) cache_free(ndd); return; out_unlock_free: cache_enter_unlock(&cel); cache_free(ncp); return; } static u_int cache_roundup_2(u_int val) { u_int res; for (res = 1; res <= val; res <<= 1) continue; return (res); } static struct nchashhead * nchinittbl(u_long elements, u_long *hashmask) { struct nchashhead *hashtbl; u_long hashsize, i; hashsize = cache_roundup_2(elements) / 2; hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK); for (i = 0; i < hashsize; i++) CK_SLIST_INIT(&hashtbl[i]); *hashmask = hashsize - 1; return (hashtbl); } static void ncfreetbl(struct nchashhead *hashtbl) { free(hashtbl, M_VFSCACHE); } /* * Name cache initialization, from vfs_init() when we are booting */ static void nchinit(void *dummy __unused) { u_int i; cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE, NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE, NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE, NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE, NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); VFS_SMR_ZONE_SET(cache_zone_small); VFS_SMR_ZONE_SET(cache_zone_small_ts); VFS_SMR_ZONE_SET(cache_zone_large); VFS_SMR_ZONE_SET(cache_zone_large_ts); ncsize = desiredvnodes * ncsizefactor; cache_recalc_neg_min(ncnegminpct); nchashtbl = nchinittbl(desiredvnodes * 2, &nchash); ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1; if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */ ncbuckethash = 7; if (ncbuckethash > nchash) ncbuckethash = nchash; bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE, M_WAITOK | M_ZERO); for (i = 0; i < numbucketlocks; i++) mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE); ncvnodehash = ncbuckethash; vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE, M_WAITOK | M_ZERO); for (i = 0; i < numvnodelocks; i++) mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE); for (i = 0; i < numneglists; i++) { mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF); mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF); TAILQ_INIT(&neglists[i].nl_list); TAILQ_INIT(&neglists[i].nl_hotlist); } } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL); void cache_vnode_init(struct vnode *vp) { LIST_INIT(&vp->v_cache_src); TAILQ_INIT(&vp->v_cache_dst); vp->v_cache_dd = NULL; cache_prehash(vp); } void cache_changesize(u_long newmaxvnodes) { struct nchashhead *new_nchashtbl, *old_nchashtbl; u_long new_nchash, old_nchash; struct namecache *ncp; uint32_t hash; u_long newncsize; int i; newncsize = newmaxvnodes * ncsizefactor; newmaxvnodes = cache_roundup_2(newmaxvnodes * 2); if (newmaxvnodes < numbucketlocks) newmaxvnodes = numbucketlocks; new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash); /* If same hash table size, nothing to do */ if (nchash == new_nchash) { ncfreetbl(new_nchashtbl); return; } /* * Move everything from the old hash table to the new table. * None of the namecache entries in the table can be removed * because to do so, they have to be removed from the hash table. */ cache_lock_all_vnodes(); cache_lock_all_buckets(); old_nchashtbl = nchashtbl; old_nchash = nchash; nchashtbl = new_nchashtbl; nchash = new_nchash; for (i = 0; i <= old_nchash; i++) { while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) { hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash); CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash); } } ncsize = newncsize; cache_recalc_neg_min(ncnegminpct); cache_unlock_all_buckets(); cache_unlock_all_vnodes(); ncfreetbl(old_nchashtbl); } /* * Invalidate all entries from and to a particular vnode. */ static void cache_purge_impl(struct vnode *vp) { struct cache_freebatch batch; struct namecache *ncp; struct mtx *vlp, *vlp2; TAILQ_INIT(&batch); vlp = VP2VNODELOCK(vp); vlp2 = NULL; mtx_lock(vlp); retry: while (!LIST_EMPTY(&vp->v_cache_src)) { ncp = LIST_FIRST(&vp->v_cache_src); if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) goto retry; TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); } while (!TAILQ_EMPTY(&vp->v_cache_dst)) { ncp = TAILQ_FIRST(&vp->v_cache_dst); if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) goto retry; TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); } ncp = vp->v_cache_dd; if (ncp != NULL) { KASSERT(ncp->nc_flag & NCF_ISDOTDOT, ("lost dotdot link")); if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) goto retry; TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); } KASSERT(vp->v_cache_dd == NULL, ("incomplete purge")); mtx_unlock(vlp); if (vlp2 != NULL) mtx_unlock(vlp2); cache_free_batch(&batch); } /* * Opportunistic check to see if there is anything to do. */ static bool cache_has_entries(struct vnode *vp) { if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) && vp->v_cache_dd == NULL) return (false); return (true); } void cache_purge(struct vnode *vp) { SDT_PROBE1(vfs, namecache, purge, done, vp); if (!cache_has_entries(vp)) return; cache_purge_impl(vp); } /* * Only to be used by vgone. */ void cache_purge_vgone(struct vnode *vp) { struct mtx *vlp; VNPASS(VN_IS_DOOMED(vp), vp); if (cache_has_entries(vp)) { cache_purge_impl(vp); return; } /* * Serialize against a potential thread doing cache_purge. */ vlp = VP2VNODELOCK(vp); mtx_wait_unlocked(vlp); if (cache_has_entries(vp)) { cache_purge_impl(vp); return; } return; } /* * Invalidate all negative entries for a particular directory vnode. */ void cache_purge_negative(struct vnode *vp) { struct cache_freebatch batch; struct namecache *ncp, *nnp; struct mtx *vlp; SDT_PROBE1(vfs, namecache, purge_negative, done, vp); if (LIST_EMPTY(&vp->v_cache_src)) return; TAILQ_INIT(&batch); vlp = VP2VNODELOCK(vp); mtx_lock(vlp); LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) { if (!(ncp->nc_flag & NCF_NEGATIVE)) continue; cache_zap_negative_locked_vnode_kl(ncp, vp); TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); } mtx_unlock(vlp); cache_free_batch(&batch); } /* * Entry points for modifying VOP operations. */ void cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp, struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp) { ASSERT_VOP_IN_SEQC(fdvp); ASSERT_VOP_IN_SEQC(fvp); ASSERT_VOP_IN_SEQC(tdvp); if (tvp != NULL) ASSERT_VOP_IN_SEQC(tvp); cache_purge(fvp); if (tvp != NULL) { cache_purge(tvp); KASSERT(!cache_remove_cnp(tdvp, tcnp), ("%s: lingering negative entry", __func__)); } else { cache_remove_cnp(tdvp, tcnp); } } void cache_vop_rmdir(struct vnode *dvp, struct vnode *vp) { ASSERT_VOP_IN_SEQC(dvp); ASSERT_VOP_IN_SEQC(vp); cache_purge(vp); } #ifdef INVARIANTS /* * Validate that if an entry exists it matches. */ void cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { struct namecache *ncp; struct mtx *blp; uint32_t hash; hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); if (CK_SLIST_EMPTY(NCHHASH(hash))) return; blp = HASH2BUCKETLOCK(hash); mtx_lock(blp); CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) { if (ncp->nc_vp != vp) panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p vp %p\n", __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp, ncp->nc_vp); } } mtx_unlock(blp); } #endif /* * Flush all entries referencing a particular filesystem. */ void cache_purgevfs(struct mount *mp) { struct vnode *vp, *mvp; SDT_PROBE1(vfs, namecache, purgevfs, done, mp); /* * Somewhat wasteful iteration over all vnodes. Would be better to * support filtering and avoid the interlock to begin with. */ MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { if (!cache_has_entries(vp)) { VI_UNLOCK(vp); continue; } vholdl(vp); VI_UNLOCK(vp); cache_purge(vp); vdrop(vp); } } /* * Perform canonical checks and cache lookup and pass on to filesystem * through the vop_cachedlookup only if needed. */ int vfs_cache_lookup(struct vop_lookup_args *ap) { struct vnode *dvp; int error; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; int flags = cnp->cn_flags; *vpp = NULL; dvp = ap->a_dvp; if (dvp->v_type != VDIR) return (ENOTDIR); if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) return (EROFS); error = vn_dir_check_exec(dvp, cnp); if (error != 0) return (error); error = cache_lookup(dvp, vpp, cnp, NULL, NULL); if (error == 0) return (VOP_CACHEDLOOKUP(dvp, vpp, cnp)); if (error == -1) return (0); return (error); } /* Implementation of the getcwd syscall. */ int sys___getcwd(struct thread *td, struct __getcwd_args *uap) { char *buf, *retbuf; size_t buflen; int error; buflen = uap->buflen; if (__predict_false(buflen < 2)) return (EINVAL); if (buflen > MAXPATHLEN) buflen = MAXPATHLEN; buf = uma_zalloc(namei_zone, M_WAITOK); error = vn_getcwd(buf, &retbuf, &buflen); if (error == 0) error = copyout(retbuf, uap->buf, buflen); uma_zfree(namei_zone, buf); return (error); } int vn_getcwd(char *buf, char **retbuf, size_t *buflen) { struct pwd *pwd; int error; vfs_smr_enter(); pwd = pwd_get_smr(); error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, buflen, 0); VFS_SMR_ASSERT_NOT_ENTERED(); if (error < 0) { pwd = pwd_hold(curthread); error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, buflen); pwd_drop(pwd); } #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI) && error == 0) ktrnamei(*retbuf); #endif return (error); } static int kern___realpathat(struct thread *td, int fd, const char *path, char *buf, size_t size, int flags, enum uio_seg pathseg) { struct nameidata nd; char *retbuf, *freebuf; int error; if (flags != 0) return (EINVAL); NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1, pathseg, path, fd, &cap_fstat_rights, td); if ((error = namei(&nd)) != 0) return (error); error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size); if (error == 0) { error = copyout(retbuf, buf, size); free(freebuf, M_TEMP); } NDFREE(&nd, 0); return (error); } int sys___realpathat(struct thread *td, struct __realpathat_args *uap) { return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size, uap->flags, UIO_USERSPACE)); } /* * Retrieve the full filesystem path that correspond to a vnode from the name * cache (if available) */ int vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf) { struct pwd *pwd; char *buf; size_t buflen; int error; if (__predict_false(vp == NULL)) return (EINVAL); buflen = MAXPATHLEN; buf = malloc(buflen, M_TEMP, M_WAITOK); vfs_smr_enter(); pwd = pwd_get_smr(); error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0); VFS_SMR_ASSERT_NOT_ENTERED(); if (error < 0) { pwd = pwd_hold(curthread); error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen); pwd_drop(pwd); } if (error == 0) *freebuf = buf; else free(buf, M_TEMP); return (error); } /* * This function is similar to vn_fullpath, but it attempts to lookup the * pathname relative to the global root mount point. This is required for the * auditing sub-system, as audited pathnames must be absolute, relative to the * global root mount point. */ int vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf) { char *buf; size_t buflen; int error; if (__predict_false(vp == NULL)) return (EINVAL); buflen = MAXPATHLEN; buf = malloc(buflen, M_TEMP, M_WAITOK); vfs_smr_enter(); error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0); VFS_SMR_ASSERT_NOT_ENTERED(); if (error < 0) { error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen); } if (error == 0) *freebuf = buf; else free(buf, M_TEMP); return (error); } static struct namecache * vn_dd_from_dst(struct vnode *vp) { struct namecache *ncp; cache_assert_vnode_locked(vp); TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) { if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) return (ncp); } return (NULL); } int vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen) { struct vnode *dvp; struct namecache *ncp; struct mtx *vlp; int error; vlp = VP2VNODELOCK(*vp); mtx_lock(vlp); ncp = (*vp)->v_cache_dd; if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) { KASSERT(ncp == vn_dd_from_dst(*vp), ("%s: mismatch for dd entry (%p != %p)", __func__, ncp, vn_dd_from_dst(*vp))); } else { ncp = vn_dd_from_dst(*vp); } if (ncp != NULL) { if (*buflen < ncp->nc_nlen) { mtx_unlock(vlp); vrele(*vp); counter_u64_add(numfullpathfail4, 1); error = ENOMEM; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } *buflen -= ncp->nc_nlen; memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp, ncp->nc_name, vp); dvp = *vp; *vp = ncp->nc_dvp; vref(*vp); mtx_unlock(vlp); vrele(dvp); return (0); } SDT_PROBE1(vfs, namecache, fullpath, miss, vp); mtx_unlock(vlp); vn_lock(*vp, LK_SHARED | LK_RETRY); error = VOP_VPTOCNP(*vp, &dvp, buf, buflen); vput(*vp); if (error) { counter_u64_add(numfullpathfail2, 1); SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } *vp = dvp; if (VN_IS_DOOMED(dvp)) { /* forced unmount */ vrele(dvp); error = ENOENT; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } /* * *vp has its use count incremented still. */ return (0); } /* * Resolve a directory to a pathname. * * The name of the directory can always be found in the namecache or fetched * from the filesystem. There is also guaranteed to be only one parent, meaning * we can just follow vnodes up until we find the root. * * The vnode must be referenced. */ static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *len, size_t addend) { #ifdef KDTRACE_HOOKS struct vnode *startvp = vp; #endif struct vnode *vp1; size_t buflen; int error; bool slash_prefixed; VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp); VNPASS(vp->v_usecount > 0, vp); buflen = *len; slash_prefixed = true; if (addend == 0) { MPASS(*len >= 2); buflen--; buf[buflen] = '\0'; slash_prefixed = false; } error = 0; SDT_PROBE1(vfs, namecache, fullpath, entry, vp); counter_u64_add(numfullpathcalls, 1); while (vp != rdir && vp != rootvnode) { /* * The vp vnode must be already fully constructed, * since it is either found in namecache or obtained * from VOP_VPTOCNP(). We may test for VV_ROOT safely * without obtaining the vnode lock. */ if ((vp->v_vflag & VV_ROOT) != 0) { vn_lock(vp, LK_RETRY | LK_SHARED); /* * With the vnode locked, check for races with * unmount, forced or not. Note that we * already verified that vp is not equal to * the root vnode, which means that * mnt_vnodecovered can be NULL only for the * case of unmount. */ if (VN_IS_DOOMED(vp) || (vp1 = vp->v_mount->mnt_vnodecovered) == NULL || vp1->v_mountedhere != vp->v_mount) { vput(vp); error = ENOENT; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); break; } vref(vp1); vput(vp); vp = vp1; continue; } if (vp->v_type != VDIR) { vrele(vp); counter_u64_add(numfullpathfail1, 1); error = ENOTDIR; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); break; } error = vn_vptocnp(&vp, buf, &buflen); if (error) break; if (buflen == 0) { vrele(vp); error = ENOMEM; SDT_PROBE3(vfs, namecache, fullpath, return, error, startvp, NULL); break; } buf[--buflen] = '/'; slash_prefixed = true; } if (error) return (error); if (!slash_prefixed) { if (buflen == 0) { vrele(vp); counter_u64_add(numfullpathfail4, 1); SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM, startvp, NULL); return (ENOMEM); } buf[--buflen] = '/'; } counter_u64_add(numfullpathfound, 1); vrele(vp); *retbuf = buf + buflen; SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf); *len -= buflen; *len += addend; return (0); } /* * Resolve an arbitrary vnode to a pathname. * * Note 2 caveats: * - hardlinks are not tracked, thus if the vnode is not a directory this can * resolve to a different path than the one used to find it * - namecache is not mandatory, meaning names are not guaranteed to be added * (in which case resolving fails) */ static void __inline cache_rev_failed_impl(int *reason, int line) { *reason = line; } #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__) static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *buflen, size_t addend) { #ifdef KDTRACE_HOOKS struct vnode *startvp = vp; #endif struct vnode *tvp; struct mount *mp; struct namecache *ncp; size_t orig_buflen; int reason; int error; #ifdef KDTRACE_HOOKS int i; #endif seqc_t vp_seqc, tvp_seqc; u_char nc_flag; VFS_SMR_ASSERT_ENTERED(); if (!cache_fast_revlookup) { vfs_smr_exit(); return (-1); } orig_buflen = *buflen; if (addend == 0) { MPASS(*buflen >= 2); *buflen -= 1; buf[*buflen] = '\0'; } if (vp == rdir || vp == rootvnode) { if (addend == 0) { *buflen -= 1; buf[*buflen] = '/'; } goto out_ok; } #ifdef KDTRACE_HOOKS i = 0; #endif error = -1; ncp = NULL; /* for sdt probe down below */ vp_seqc = vn_seqc_read_any(vp); if (seqc_in_modify(vp_seqc)) { cache_rev_failed(&reason); goto out_abort; } for (;;) { #ifdef KDTRACE_HOOKS i++; #endif if ((vp->v_vflag & VV_ROOT) != 0) { mp = atomic_load_ptr(&vp->v_mount); if (mp == NULL) { cache_rev_failed(&reason); goto out_abort; } tvp = atomic_load_ptr(&mp->mnt_vnodecovered); tvp_seqc = vn_seqc_read_any(tvp); if (seqc_in_modify(tvp_seqc)) { cache_rev_failed(&reason); goto out_abort; } if (!vn_seqc_consistent(vp, vp_seqc)) { cache_rev_failed(&reason); goto out_abort; } vp = tvp; vp_seqc = tvp_seqc; continue; } ncp = atomic_load_ptr(&vp->v_cache_dd); if (ncp == NULL) { cache_rev_failed(&reason); goto out_abort; } nc_flag = atomic_load_char(&ncp->nc_flag); if ((nc_flag & NCF_ISDOTDOT) != 0) { cache_rev_failed(&reason); goto out_abort; } if (!cache_ncp_canuse(ncp)) { cache_rev_failed(&reason); goto out_abort; } if (ncp->nc_nlen >= *buflen) { cache_rev_failed(&reason); error = ENOMEM; goto out_abort; } *buflen -= ncp->nc_nlen; memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); *buflen -= 1; buf[*buflen] = '/'; tvp = ncp->nc_dvp; tvp_seqc = vn_seqc_read_any(tvp); if (seqc_in_modify(tvp_seqc)) { cache_rev_failed(&reason); goto out_abort; } if (!vn_seqc_consistent(vp, vp_seqc)) { cache_rev_failed(&reason); goto out_abort; } vp = tvp; vp_seqc = tvp_seqc; if (vp == rdir || vp == rootvnode) break; } out_ok: vfs_smr_exit(); *retbuf = buf + *buflen; *buflen = orig_buflen - *buflen + addend; SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf); return (0); out_abort: *buflen = orig_buflen; SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i); vfs_smr_exit(); return (error); } static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, size_t *buflen) { size_t orig_buflen, addend; int error; if (*buflen < 2) return (EINVAL); orig_buflen = *buflen; vref(vp); addend = 0; if (vp->v_type != VDIR) { *buflen -= 1; buf[*buflen] = '\0'; error = vn_vptocnp(&vp, buf, buflen); if (error) return (error); if (*buflen == 0) { vrele(vp); return (ENOMEM); } *buflen -= 1; buf[*buflen] = '/'; addend = orig_buflen - *buflen; } return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend)); } /* * Resolve an arbitrary vnode to a pathname (taking care of hardlinks). * * Since the namecache does not track hardlinks, the caller is expected to first * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei. * * Then we have 2 cases: * - if the found vnode is a directory, the path can be constructed just by * following names up the chain * - otherwise we populate the buffer with the saved name and start resolving * from the parent */ static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf, size_t *buflen) { char *buf, *tmpbuf; struct pwd *pwd; struct componentname *cnp; struct vnode *vp; size_t addend; int error; enum vtype type; if (*buflen < 2) return (EINVAL); if (*buflen > MAXPATHLEN) *buflen = MAXPATHLEN; buf = malloc(*buflen, M_TEMP, M_WAITOK); addend = 0; vp = ndp->ni_vp; /* * Check for VBAD to work around the vp_crossmp bug in lookup(). * * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be * set to mount point's root vnode while ni_dvp will be vp_crossmp. * If the type is VDIR (like in this very case) we can skip looking * at ni_dvp in the first place. However, since vnodes get passed here * unlocked the target may transition to doomed state (type == VBAD) * before we get to evaluate the condition. If this happens, we will * populate part of the buffer and descend to vn_fullpath_dir with * vp == vp_crossmp. Prevent the problem by checking for VBAD. * * This should be atomic_load(&vp->v_type) but it is illegal to take * an address of a bit field, even if said field is sized to char. * Work around the problem by reading the value into a full-sized enum * and then re-reading it with atomic_load which will still prevent * the compiler from re-reading down the road. */ type = vp->v_type; type = atomic_load_int(&type); if (type == VBAD) { error = ENOENT; goto out_bad; } if (type != VDIR) { cnp = &ndp->ni_cnd; addend = cnp->cn_namelen + 2; if (*buflen < addend) { error = ENOMEM; goto out_bad; } *buflen -= addend; tmpbuf = buf + *buflen; tmpbuf[0] = '/'; memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen); tmpbuf[addend - 1] = '\0'; vp = ndp->ni_dvp; } vfs_smr_enter(); pwd = pwd_get_smr(); error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen, addend); VFS_SMR_ASSERT_NOT_ENTERED(); if (error < 0) { pwd = pwd_hold(curthread); vref(vp); error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen, addend); pwd_drop(pwd); if (error != 0) goto out_bad; } *freebuf = buf; return (0); out_bad: free(buf, M_TEMP); return (error); } struct vnode * vn_dir_dd_ino(struct vnode *vp) { struct namecache *ncp; struct vnode *ddvp; struct mtx *vlp; enum vgetstate vs; ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino"); vlp = VP2VNODELOCK(vp); mtx_lock(vlp); TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) { if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) continue; ddvp = ncp->nc_dvp; vs = vget_prep(ddvp); mtx_unlock(vlp); if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs)) return (NULL); return (ddvp); } mtx_unlock(vlp); return (NULL); } int vn_commname(struct vnode *vp, char *buf, u_int buflen) { struct namecache *ncp; struct mtx *vlp; int l; vlp = VP2VNODELOCK(vp); mtx_lock(vlp); TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) break; if (ncp == NULL) { mtx_unlock(vlp); return (ENOENT); } l = min(ncp->nc_nlen, buflen - 1); memcpy(buf, ncp->nc_name, l); mtx_unlock(vlp); buf[l] = '\0'; return (0); } /* * This function updates path string to vnode's full global path * and checks the size of the new path string against the pathlen argument. * * Requires a locked, referenced vnode. * Vnode is re-locked on success or ENODEV, otherwise unlocked. * * If vp is a directory, the call to vn_fullpath_global() always succeeds * because it falls back to the ".." lookup if the namecache lookup fails. */ int vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, u_int pathlen) { struct nameidata nd; struct vnode *vp1; char *rpath, *fbuf; int error; ASSERT_VOP_ELOCKED(vp, __func__); /* Construct global filesystem path from vp. */ VOP_UNLOCK(vp); error = vn_fullpath_global(vp, &rpath, &fbuf); if (error != 0) { vrele(vp); return (error); } if (strlen(rpath) >= pathlen) { vrele(vp); error = ENAMETOOLONG; goto out; } /* * Re-lookup the vnode by path to detect a possible rename. * As a side effect, the vnode is relocked. * If vnode was renamed, return ENOENT. */ NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path, td); error = namei(&nd); if (error != 0) { vrele(vp); goto out; } NDFREE(&nd, NDF_ONLY_PNBUF); vp1 = nd.ni_vp; vrele(vp); if (vp1 == vp) strcpy(path, rpath); else { vput(vp1); error = ENOENT; } out: free(fbuf, M_TEMP); return (error); } #ifdef DDB static void db_print_vpath(struct vnode *vp) { while (vp != NULL) { db_printf("%p: ", vp); if (vp == rootvnode) { db_printf("/"); vp = NULL; } else { if (vp->v_vflag & VV_ROOT) { db_printf(""); vp = vp->v_mount->mnt_vnodecovered; } else { struct namecache *ncp; char *ncn; int i; ncp = TAILQ_FIRST(&vp->v_cache_dst); if (ncp != NULL) { ncn = ncp->nc_name; for (i = 0; i < ncp->nc_nlen; i++) db_printf("%c", *ncn++); vp = ncp->nc_dvp; } else { vp = NULL; } } } db_printf("\n"); } return; } DB_SHOW_COMMAND(vpath, db_show_vpath) { struct vnode *vp; if (!have_addr) { db_printf("usage: show vpath \n"); return; } vp = (struct vnode *)addr; db_print_vpath(vp); } #endif static bool __read_frequently cache_fast_lookup = true; SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_lookup, CTLFLAG_RW, &cache_fast_lookup, 0, ""); #define CACHE_FPL_FAILED -2020 static void cache_fpl_cleanup_cnp(struct componentname *cnp) { uma_zfree(namei_zone, cnp->cn_pnbuf); #ifdef DIAGNOSTIC cnp->cn_pnbuf = NULL; cnp->cn_nameptr = NULL; #endif } static void cache_fpl_handle_root(struct nameidata *ndp, struct vnode **dpp) { struct componentname *cnp; cnp = &ndp->ni_cnd; while (*(cnp->cn_nameptr) == '/') { cnp->cn_nameptr++; ndp->ni_pathlen--; } *dpp = ndp->ni_rootdir; } /* * Components of nameidata (or objects it can point to) which may * need restoring in case fast path lookup fails. */ struct nameidata_saved { long cn_namelen; char *cn_nameptr; size_t ni_pathlen; int cn_flags; }; struct cache_fpl { struct nameidata *ndp; struct componentname *cnp; struct pwd *pwd; struct vnode *dvp; struct vnode *tvp; seqc_t dvp_seqc; seqc_t tvp_seqc; struct nameidata_saved snd; int line; enum cache_fpl_status status:8; bool in_smr; bool fsearch; }; static void cache_fpl_checkpoint(struct cache_fpl *fpl, struct nameidata_saved *snd) { snd->cn_flags = fpl->ndp->ni_cnd.cn_flags; snd->cn_namelen = fpl->ndp->ni_cnd.cn_namelen; snd->cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr; snd->ni_pathlen = fpl->ndp->ni_pathlen; } static void cache_fpl_restore_partial(struct cache_fpl *fpl, struct nameidata_saved *snd) { fpl->ndp->ni_cnd.cn_flags = snd->cn_flags; fpl->ndp->ni_cnd.cn_namelen = snd->cn_namelen; fpl->ndp->ni_cnd.cn_nameptr = snd->cn_nameptr; fpl->ndp->ni_pathlen = snd->ni_pathlen; } static void cache_fpl_restore_abort(struct cache_fpl *fpl, struct nameidata_saved *snd) { cache_fpl_restore_partial(fpl, snd); /* * It is 0 on entry by API contract. */ fpl->ndp->ni_resflags = 0; } #ifdef INVARIANTS #define cache_fpl_smr_assert_entered(fpl) ({ \ struct cache_fpl *_fpl = (fpl); \ MPASS(_fpl->in_smr == true); \ VFS_SMR_ASSERT_ENTERED(); \ }) #define cache_fpl_smr_assert_not_entered(fpl) ({ \ struct cache_fpl *_fpl = (fpl); \ MPASS(_fpl->in_smr == false); \ VFS_SMR_ASSERT_NOT_ENTERED(); \ }) #else #define cache_fpl_smr_assert_entered(fpl) do { } while (0) #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0) #endif #define cache_fpl_smr_enter_initial(fpl) ({ \ struct cache_fpl *_fpl = (fpl); \ vfs_smr_enter(); \ _fpl->in_smr = true; \ }) #define cache_fpl_smr_enter(fpl) ({ \ struct cache_fpl *_fpl = (fpl); \ MPASS(_fpl->in_smr == false); \ vfs_smr_enter(); \ _fpl->in_smr = true; \ }) #define cache_fpl_smr_exit(fpl) ({ \ struct cache_fpl *_fpl = (fpl); \ MPASS(_fpl->in_smr == true); \ vfs_smr_exit(); \ _fpl->in_smr = false; \ }) static int cache_fpl_aborted_impl(struct cache_fpl *fpl, int line) { if (fpl->status != CACHE_FPL_STATUS_UNSET) { KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL, ("%s: converting to abort from %d at %d, set at %d\n", __func__, fpl->status, line, fpl->line)); } fpl->status = CACHE_FPL_STATUS_ABORTED; fpl->line = line; return (CACHE_FPL_FAILED); } #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__) static int cache_fpl_partial_impl(struct cache_fpl *fpl, int line) { KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, ("%s: setting to partial at %d, but already set to %d at %d\n", __func__, line, fpl->status, fpl->line)); cache_fpl_smr_assert_entered(fpl); fpl->status = CACHE_FPL_STATUS_PARTIAL; fpl->line = line; return (CACHE_FPL_FAILED); } #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__) static int cache_fpl_handled_impl(struct cache_fpl *fpl, int error, int line) { KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, ("%s: setting to handled at %d, but already set to %d at %d\n", __func__, line, fpl->status, fpl->line)); cache_fpl_smr_assert_not_entered(fpl); MPASS(error != CACHE_FPL_FAILED); fpl->status = CACHE_FPL_STATUS_HANDLED; fpl->line = line; return (error); } #define cache_fpl_handled(x, e) cache_fpl_handled_impl((x), (e), __LINE__) #define CACHE_FPL_SUPPORTED_CN_FLAGS \ (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \ FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | ISOPEN | \ NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK) #define CACHE_FPL_INTERNAL_CN_FLAGS \ (ISDOTDOT | MAKEENTRY | ISLASTCN) _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0, "supported and internal flags overlap"); static bool cache_fpl_islastcn(struct nameidata *ndp) { return (*ndp->ni_next == 0); } static bool cache_fpl_isdotdot(struct componentname *cnp) { if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.') return (true); return (false); } static bool cache_can_fplookup(struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; struct thread *td; ndp = fpl->ndp; cnp = fpl->cnp; td = cnp->cn_thread; if (!cache_fast_lookup) { cache_fpl_aborted(fpl); return (false); } #ifdef MAC if (mac_vnode_check_lookup_enabled()) { cache_fpl_aborted(fpl); return (false); } #endif if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) { cache_fpl_aborted(fpl); return (false); } if (IN_CAPABILITY_MODE(td)) { cache_fpl_aborted(fpl); return (false); } if (AUDITING_TD(td)) { cache_fpl_aborted(fpl); return (false); } if (ndp->ni_startdir != NULL) { cache_fpl_aborted(fpl); return (false); } return (true); } static int cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp) { struct nameidata *ndp; int error; bool fsearch; ndp = fpl->ndp; error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch); if (__predict_false(error != 0)) { cache_fpl_smr_exit(fpl); return (cache_fpl_aborted(fpl)); } fpl->fsearch = fsearch; return (0); } static bool cache_fplookup_vnode_supported(struct vnode *vp) { return (vp->v_type != VLNK); } static int __noinline cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp, uint32_t hash) { struct componentname *cnp; struct vnode *dvp; cnp = fpl->cnp; dvp = fpl->dvp; cache_fpl_smr_exit(fpl); if (cache_neg_promote_cond(dvp, cnp, oncp, hash)) return (cache_fpl_handled(fpl, ENOENT)); else return (cache_fpl_aborted(fpl)); } /* * The target vnode is not supported, prepare for the slow path to take over. */ static int __noinline cache_fplookup_partial_setup(struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; enum vgetstate dvs; struct vnode *dvp; struct pwd *pwd; seqc_t dvp_seqc; ndp = fpl->ndp; cnp = fpl->cnp; pwd = fpl->pwd; dvp = fpl->dvp; dvp_seqc = fpl->dvp_seqc; if (!pwd_hold_smr(pwd)) { cache_fpl_smr_exit(fpl); return (cache_fpl_aborted(fpl)); } dvs = vget_prep_smr(dvp); cache_fpl_smr_exit(fpl); if (__predict_false(dvs == VGET_NONE)) { pwd_drop(pwd); return (cache_fpl_aborted(fpl)); } vget_finish_ref(dvp, dvs); if (!vn_seqc_consistent(dvp, dvp_seqc)) { vrele(dvp); pwd_drop(pwd); return (cache_fpl_aborted(fpl)); } cache_fpl_restore_partial(fpl, &fpl->snd); ndp->ni_startdir = dvp; cnp->cn_flags |= MAKEENTRY; if (cache_fpl_islastcn(ndp)) cnp->cn_flags |= ISLASTCN; if (cache_fpl_isdotdot(cnp)) cnp->cn_flags |= ISDOTDOT; return (0); } static int cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs) { struct componentname *cnp; struct vnode *tvp; seqc_t tvp_seqc; int error, lkflags; cnp = fpl->cnp; tvp = fpl->tvp; tvp_seqc = fpl->tvp_seqc; if ((cnp->cn_flags & LOCKLEAF) != 0) { lkflags = LK_SHARED; if ((cnp->cn_flags & LOCKSHARED) == 0) lkflags = LK_EXCLUSIVE; error = vget_finish(tvp, lkflags, tvs); if (__predict_false(error != 0)) { return (cache_fpl_aborted(fpl)); } } else { vget_finish_ref(tvp, tvs); } if (!vn_seqc_consistent(tvp, tvp_seqc)) { if ((cnp->cn_flags & LOCKLEAF) != 0) vput(tvp); else vrele(tvp); return (cache_fpl_aborted(fpl)); } return (cache_fpl_handled(fpl, 0)); } /* * They want to possibly modify the state of the namecache. * * Don't try to match the API contract, just leave. * TODO: this leaves scalability on the table */ static int cache_fplookup_final_modifying(struct cache_fpl *fpl) { struct componentname *cnp; cnp = fpl->cnp; MPASS(cnp->cn_nameiop != LOOKUP); return (cache_fpl_partial(fpl)); } static int __noinline cache_fplookup_final_withparent(struct cache_fpl *fpl) { struct componentname *cnp; enum vgetstate dvs, tvs; struct vnode *dvp, *tvp; seqc_t dvp_seqc; int error; cnp = fpl->cnp; dvp = fpl->dvp; dvp_seqc = fpl->dvp_seqc; tvp = fpl->tvp; MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0); /* * This is less efficient than it can be for simplicity. */ dvs = vget_prep_smr(dvp); if (__predict_false(dvs == VGET_NONE)) { return (cache_fpl_aborted(fpl)); } tvs = vget_prep_smr(tvp); if (__predict_false(tvs == VGET_NONE)) { cache_fpl_smr_exit(fpl); vget_abort(dvp, dvs); return (cache_fpl_aborted(fpl)); } cache_fpl_smr_exit(fpl); if ((cnp->cn_flags & LOCKPARENT) != 0) { error = vget_finish(dvp, LK_EXCLUSIVE, dvs); if (__predict_false(error != 0)) { vget_abort(tvp, tvs); return (cache_fpl_aborted(fpl)); } } else { vget_finish_ref(dvp, dvs); } if (!vn_seqc_consistent(dvp, dvp_seqc)) { vget_abort(tvp, tvs); if ((cnp->cn_flags & LOCKPARENT) != 0) vput(dvp); else vrele(dvp); return (cache_fpl_aborted(fpl)); } error = cache_fplookup_final_child(fpl, tvs); if (__predict_false(error != 0)) { MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED); if ((cnp->cn_flags & LOCKPARENT) != 0) vput(dvp); else vrele(dvp); return (error); } MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED); return (0); } static int cache_fplookup_final(struct cache_fpl *fpl) { struct componentname *cnp; enum vgetstate tvs; struct vnode *dvp, *tvp; seqc_t dvp_seqc; cnp = fpl->cnp; dvp = fpl->dvp; dvp_seqc = fpl->dvp_seqc; tvp = fpl->tvp; VNPASS(cache_fplookup_vnode_supported(dvp), dvp); if (cnp->cn_nameiop != LOOKUP) { return (cache_fplookup_final_modifying(fpl)); } if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) return (cache_fplookup_final_withparent(fpl)); tvs = vget_prep_smr(tvp); if (__predict_false(tvs == VGET_NONE)) { return (cache_fpl_partial(fpl)); } if (!vn_seqc_consistent(dvp, dvp_seqc)) { cache_fpl_smr_exit(fpl); vget_abort(tvp, tvs); return (cache_fpl_aborted(fpl)); } cache_fpl_smr_exit(fpl); return (cache_fplookup_final_child(fpl, tvs)); } static int __noinline cache_fplookup_dot(struct cache_fpl *fpl) { struct vnode *dvp; dvp = fpl->dvp; fpl->tvp = dvp; fpl->tvp_seqc = vn_seqc_read_any(dvp); if (seqc_in_modify(fpl->tvp_seqc)) { return (cache_fpl_aborted(fpl)); } counter_u64_add(dothits, 1); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", dvp); return (0); } static int __noinline cache_fplookup_dotdot(struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; struct namecache *ncp; struct vnode *dvp; struct prison *pr; u_char nc_flag; ndp = fpl->ndp; cnp = fpl->cnp; dvp = fpl->dvp; /* * XXX this is racy the same way regular lookup is */ for (pr = cnp->cn_cred->cr_prison; pr != NULL; pr = pr->pr_parent) if (dvp == pr->pr_root) break; if (dvp == ndp->ni_rootdir || dvp == ndp->ni_topdir || dvp == rootvnode || pr != NULL) { fpl->tvp = dvp; fpl->tvp_seqc = vn_seqc_read_any(dvp); if (seqc_in_modify(fpl->tvp_seqc)) { return (cache_fpl_aborted(fpl)); } return (0); } if ((dvp->v_vflag & VV_ROOT) != 0) { /* * TODO * The opposite of climb mount is needed here. */ return (cache_fpl_aborted(fpl)); } ncp = atomic_load_ptr(&dvp->v_cache_dd); if (ncp == NULL) { return (cache_fpl_aborted(fpl)); } nc_flag = atomic_load_char(&ncp->nc_flag); if ((nc_flag & NCF_ISDOTDOT) != 0) { if ((nc_flag & NCF_NEGATIVE) != 0) return (cache_fpl_aborted(fpl)); fpl->tvp = ncp->nc_vp; } else { fpl->tvp = ncp->nc_dvp; } if (!cache_ncp_canuse(ncp)) { return (cache_fpl_aborted(fpl)); } fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp); if (seqc_in_modify(fpl->tvp_seqc)) { return (cache_fpl_partial(fpl)); } counter_u64_add(dotdothits, 1); return (0); } static int __noinline cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash) { u_char nc_flag; bool neg_promote; nc_flag = atomic_load_char(&ncp->nc_flag); MPASS((nc_flag & NCF_NEGATIVE) != 0); /* * If they want to create an entry we need to replace this one. */ if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) { /* * TODO * This should call something similar to * cache_fplookup_final_modifying. */ return (cache_fpl_partial(fpl)); } neg_promote = cache_neg_hit_prep(ncp); if (!cache_ncp_canuse(ncp)) { cache_neg_hit_abort(ncp); return (cache_fpl_partial(fpl)); } if (__predict_false((nc_flag & NCF_WHITE) != 0)) { cache_neg_hit_abort(ncp); return (cache_fpl_partial(fpl)); } if (neg_promote) { return (cache_fplookup_negative_promote(fpl, ncp, hash)); } cache_neg_hit_finish(ncp); cache_fpl_smr_exit(fpl); return (cache_fpl_handled(fpl, ENOENT)); } static int cache_fplookup_next(struct cache_fpl *fpl) { struct componentname *cnp; struct namecache *ncp; struct vnode *dvp, *tvp; u_char nc_flag; uint32_t hash; cnp = fpl->cnp; dvp = fpl->dvp; if (__predict_false(cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.')) { return (cache_fplookup_dot(fpl)); } hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) break; } /* * If there is no entry we have to punt to the slow path to perform * actual lookup. Should there be nothing with this name a negative * entry will be created. */ if (__predict_false(ncp == NULL)) { return (cache_fpl_partial(fpl)); } tvp = atomic_load_ptr(&ncp->nc_vp); nc_flag = atomic_load_char(&ncp->nc_flag); if ((nc_flag & NCF_NEGATIVE) != 0) { return (cache_fplookup_neg(fpl, ncp, hash)); } if (!cache_ncp_canuse(ncp)) { return (cache_fpl_partial(fpl)); } fpl->tvp = tvp; fpl->tvp_seqc = vn_seqc_read_any(tvp); if (seqc_in_modify(fpl->tvp_seqc)) { return (cache_fpl_partial(fpl)); } if (!cache_fplookup_vnode_supported(tvp)) { return (cache_fpl_partial(fpl)); } counter_u64_add(numposhits, 1); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp); return (0); } static bool cache_fplookup_mp_supported(struct mount *mp) { if (mp == NULL) return (false); if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0) return (false); return (true); } /* * Walk up the mount stack (if any). * * Correctness is provided in the following ways: * - all vnodes are protected from freeing with SMR * - struct mount objects are type stable making them always safe to access * - stability of the particular mount is provided by busying it * - relationship between the vnode which is mounted on and the mount is * verified with the vnode sequence counter after busying * - association between root vnode of the mount and the mount is protected * by busy * * From that point on we can read the sequence counter of the root vnode * and get the next mount on the stack (if any) using the same protection. * * By the end of successful walk we are guaranteed the reached state was * indeed present at least at some point which matches the regular lookup. */ static int __noinline cache_fplookup_climb_mount(struct cache_fpl *fpl) { struct mount *mp, *prev_mp; + struct mount_pcpu *mpcpu, *prev_mpcpu; struct vnode *vp; seqc_t vp_seqc; vp = fpl->tvp; vp_seqc = fpl->tvp_seqc; VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp); mp = atomic_load_ptr(&vp->v_mountedhere); if (mp == NULL) return (0); prev_mp = NULL; for (;;) { - if (!vfs_op_thread_enter_crit(mp)) { + if (!vfs_op_thread_enter_crit(mp, mpcpu)) { if (prev_mp != NULL) - vfs_op_thread_exit_crit(prev_mp); + vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); return (cache_fpl_partial(fpl)); } if (prev_mp != NULL) - vfs_op_thread_exit_crit(prev_mp); + vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); if (!vn_seqc_consistent(vp, vp_seqc)) { - vfs_op_thread_exit_crit(mp); + vfs_op_thread_exit_crit(mp, mpcpu); return (cache_fpl_partial(fpl)); } if (!cache_fplookup_mp_supported(mp)) { - vfs_op_thread_exit_crit(mp); + vfs_op_thread_exit_crit(mp, mpcpu); return (cache_fpl_partial(fpl)); } vp = atomic_load_ptr(&mp->mnt_rootvnode); if (vp == NULL || VN_IS_DOOMED(vp)) { - vfs_op_thread_exit_crit(mp); + vfs_op_thread_exit_crit(mp, mpcpu); return (cache_fpl_partial(fpl)); } vp_seqc = vn_seqc_read_any(vp); if (seqc_in_modify(vp_seqc)) { - vfs_op_thread_exit_crit(mp); + vfs_op_thread_exit_crit(mp, mpcpu); return (cache_fpl_partial(fpl)); } prev_mp = mp; + prev_mpcpu = mpcpu; mp = atomic_load_ptr(&vp->v_mountedhere); if (mp == NULL) break; } - vfs_op_thread_exit_crit(prev_mp); + vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); fpl->tvp = vp; fpl->tvp_seqc = vp_seqc; return (0); } static bool cache_fplookup_need_climb_mount(struct cache_fpl *fpl) { struct mount *mp; struct vnode *vp; vp = fpl->tvp; /* * Hack: while this is a union, the pointer tends to be NULL so save on * a branch. */ mp = atomic_load_ptr(&vp->v_mountedhere); if (mp == NULL) return (false); if (vp->v_type == VDIR) return (true); return (false); } /* * Parse the path. * * The code was originally copy-pasted from regular lookup and despite * clean ups leaves performance on the table. Any modifications here * must take into account that in case off fallback the resulting * nameidata state has to be compatible with the original. */ static int cache_fplookup_parse(struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; char *cp; ndp = fpl->ndp; cnp = fpl->cnp; /* * Search a new directory. * * The last component of the filename is left accessible via * cnp->cn_nameptr for callers that need the name. Callers needing * the name set the SAVENAME flag. When done, they assume * responsibility for freeing the pathname buffer. */ for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++) continue; cnp->cn_namelen = cp - cnp->cn_nameptr; if (__predict_false(cnp->cn_namelen > NAME_MAX)) { cache_fpl_smr_exit(fpl); return (cache_fpl_handled(fpl, ENAMETOOLONG)); } ndp->ni_pathlen -= cnp->cn_namelen; KASSERT(ndp->ni_pathlen <= PATH_MAX, ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen)); ndp->ni_next = cp; /* * Replace multiple slashes by a single slash and trailing slashes * by a null. This must be done before VOP_LOOKUP() because some * fs's don't know about trailing slashes. Remember if there were * trailing slashes to handle symlinks, existing non-directories * and non-existing files that won't be directories specially later. */ while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) { cp++; ndp->ni_pathlen--; if (*cp == '\0') { /* * TODO * Regular lookup performs the following: * *ndp->ni_next = '\0'; * cnp->cn_flags |= TRAILINGSLASH; * * Which is problematic since it modifies data read * from userspace. Then if fast path lookup was to * abort we would have to either restore it or convey * the flag. Since this is a corner case just ignore * it for simplicity. */ return (cache_fpl_partial(fpl)); } } ndp->ni_next = cp; /* * Check for degenerate name (e.g. / or "") * which is a way of talking about a directory, * e.g. like "/." or ".". * * TODO * Another corner case handled by the regular lookup */ if (__predict_false(cnp->cn_nameptr[0] == '\0')) { return (cache_fpl_partial(fpl)); } return (0); } static void cache_fplookup_parse_advance(struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; ndp = fpl->ndp; cnp = fpl->cnp; cnp->cn_nameptr = ndp->ni_next; while (*cnp->cn_nameptr == '/') { cnp->cn_nameptr++; ndp->ni_pathlen--; } } /* * See the API contract for VOP_FPLOOKUP_VEXEC. */ static int __noinline cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error) { struct vnode *dvp; seqc_t dvp_seqc; dvp = fpl->dvp; dvp_seqc = fpl->dvp_seqc; /* * Hack: they may be looking up foo/bar, where foo is a * regular file. In such a case we need to turn ENOTDIR, * but we may happen to get here with a different error. */ if (dvp->v_type != VDIR) { /* * The check here is predominantly to catch * EOPNOTSUPP from dead_vnodeops. If the vnode * gets doomed past this point it is going to * fail seqc verification. */ if (VN_IS_DOOMED(dvp)) { return (cache_fpl_aborted(fpl)); } error = ENOTDIR; } /* * Hack: handle O_SEARCH. * * Open Group Base Specifications Issue 7, 2018 edition states: * If the access mode of the open file description associated with the * file descriptor is not O_SEARCH, the function shall check whether * directory searches are permitted using the current permissions of * the directory underlying the file descriptor. If the access mode is * O_SEARCH, the function shall not perform the check. * * Regular lookup tests for the NOEXECCHECK flag for every path * component to decide whether to do the permission check. However, * since most lookups never have the flag (and when they do it is only * present for the first path component), lockless lookup only acts on * it if there is a permission problem. Here the flag is represented * with a boolean so that we don't have to clear it on the way out. * * For simplicity this always aborts. * TODO: check if this is the first lookup and ignore the permission * problem. Note the flag has to survive fallback (if it happens to be * performed). */ if (fpl->fsearch) { return (cache_fpl_aborted(fpl)); } switch (error) { case EAGAIN: if (!vn_seqc_consistent(dvp, dvp_seqc)) { error = cache_fpl_aborted(fpl); } else { cache_fpl_partial(fpl); } break; default: if (!vn_seqc_consistent(dvp, dvp_seqc)) { error = cache_fpl_aborted(fpl); } else { cache_fpl_smr_exit(fpl); cache_fpl_handled(fpl, error); } break; } return (error); } static int cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl) { struct nameidata *ndp; struct componentname *cnp; struct mount *mp; int error; error = CACHE_FPL_FAILED; ndp = fpl->ndp; cnp = fpl->cnp; cache_fpl_checkpoint(fpl, &fpl->snd); fpl->dvp = dvp; fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp); if (seqc_in_modify(fpl->dvp_seqc)) { cache_fpl_aborted(fpl); goto out; } mp = atomic_load_ptr(&fpl->dvp->v_mount); if (!cache_fplookup_mp_supported(mp)) { cache_fpl_aborted(fpl); goto out; } VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp); for (;;) { error = cache_fplookup_parse(fpl); if (__predict_false(error != 0)) { break; } VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp); error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred); if (__predict_false(error != 0)) { error = cache_fplookup_failed_vexec(fpl, error); break; } if (__predict_false(cache_fpl_isdotdot(cnp))) { error = cache_fplookup_dotdot(fpl); if (__predict_false(error != 0)) { break; } } else { error = cache_fplookup_next(fpl); if (__predict_false(error != 0)) { break; } VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp); if (cache_fplookup_need_climb_mount(fpl)) { error = cache_fplookup_climb_mount(fpl); if (__predict_false(error != 0)) { break; } } } VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp); if (cache_fpl_islastcn(ndp)) { error = cache_fplookup_final(fpl); break; } if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) { error = cache_fpl_aborted(fpl); break; } fpl->dvp = fpl->tvp; fpl->dvp_seqc = fpl->tvp_seqc; cache_fplookup_parse_advance(fpl); cache_fpl_checkpoint(fpl, &fpl->snd); } out: switch (fpl->status) { case CACHE_FPL_STATUS_UNSET: __assert_unreachable(); break; case CACHE_FPL_STATUS_PARTIAL: cache_fpl_smr_assert_entered(fpl); return (cache_fplookup_partial_setup(fpl)); case CACHE_FPL_STATUS_ABORTED: if (fpl->in_smr) cache_fpl_smr_exit(fpl); return (CACHE_FPL_FAILED); case CACHE_FPL_STATUS_HANDLED: MPASS(error != CACHE_FPL_FAILED); cache_fpl_smr_assert_not_entered(fpl); if (__predict_false(error != 0)) { ndp->ni_dvp = NULL; ndp->ni_vp = NULL; cache_fpl_cleanup_cnp(cnp); return (error); } ndp->ni_dvp = fpl->dvp; ndp->ni_vp = fpl->tvp; if (cnp->cn_flags & SAVENAME) cnp->cn_flags |= HASBUF; else cache_fpl_cleanup_cnp(cnp); return (error); } } /* * Fast path lookup protected with SMR and sequence counters. * * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one. * * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria * outlined below. * * Traditional vnode lookup conceptually looks like this: * * vn_lock(current); * for (;;) { * next = find(); * vn_lock(next); * vn_unlock(current); * current = next; * if (last) * break; * } * return (current); * * Each jump to the next vnode is safe memory-wise and atomic with respect to * any modifications thanks to holding respective locks. * * The same guarantee can be provided with a combination of safe memory * reclamation and sequence counters instead. If all operations which affect * the relationship between the current vnode and the one we are looking for * also modify the counter, we can verify whether all the conditions held as * we made the jump. This includes things like permissions, mount points etc. * Counter modification is provided by enclosing relevant places in * vn_seqc_write_begin()/end() calls. * * Thus this translates to: * * vfs_smr_enter(); * dvp_seqc = seqc_read_any(dvp); * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode * abort(); * for (;;) { * tvp = find(); * tvp_seqc = seqc_read_any(tvp); * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode * abort(); * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode * abort(); * dvp = tvp; // we know nothing of importance has changed * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration * if (last) * break; * } * vget(); // secure the vnode * if (!seqc_consistent(tvp, tvp_seqc) // final check * abort(); * // at this point we know nothing has changed for any parent<->child pair * // as they were crossed during the lookup, meaning we matched the guarantee * // of the locked variant * return (tvp); * * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows: * - they are called while within vfs_smr protection which they must never exit * - EAGAIN can be returned to denote checking could not be performed, it is * always valid to return it * - if the sequence counter has not changed the result must be valid * - if the sequence counter has changed both false positives and false negatives * are permitted (since the result will be rejected later) * - for simple cases of unix permission checks vaccess_vexec_smr can be used * * Caveats to watch out for: * - vnodes are passed unlocked and unreferenced with nothing stopping * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised * to use atomic_load_ptr to fetch it. * - the aforementioned object can also get freed, meaning absent other means it * should be protected with vfs_smr * - either safely checking permissions as they are modified or guaranteeing * their stability is left to the routine */ int cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status, struct pwd **pwdp) { struct cache_fpl fpl; struct pwd *pwd; struct vnode *dvp; struct componentname *cnp; struct nameidata_saved orig; int error; MPASS(ndp->ni_lcf == 0); fpl.status = CACHE_FPL_STATUS_UNSET; fpl.ndp = ndp; fpl.cnp = &ndp->ni_cnd; MPASS(curthread == fpl.cnp->cn_thread); if ((fpl.cnp->cn_flags & SAVESTART) != 0) MPASS(fpl.cnp->cn_nameiop != LOOKUP); if (!cache_can_fplookup(&fpl)) { SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); *status = fpl.status; return (EOPNOTSUPP); } cache_fpl_checkpoint(&fpl, &orig); cache_fpl_smr_enter_initial(&fpl); fpl.fsearch = false; pwd = pwd_get_smr(); fpl.pwd = pwd; ndp->ni_rootdir = pwd->pwd_rdir; ndp->ni_topdir = pwd->pwd_jdir; cnp = fpl.cnp; cnp->cn_nameptr = cnp->cn_pnbuf; if (cnp->cn_pnbuf[0] == '/') { cache_fpl_handle_root(ndp, &dvp); ndp->ni_resflags |= NIRES_ABS; } else { if (ndp->ni_dirfd == AT_FDCWD) { dvp = pwd->pwd_cdir; } else { error = cache_fplookup_dirfd(&fpl, &dvp); if (__predict_false(error != 0)) { goto out; } } } SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true); error = cache_fplookup_impl(dvp, &fpl); out: cache_fpl_smr_assert_not_entered(&fpl); SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); *status = fpl.status; switch (fpl.status) { case CACHE_FPL_STATUS_UNSET: __assert_unreachable(); break; case CACHE_FPL_STATUS_HANDLED: SDT_PROBE3(vfs, namei, lookup, return, error, (error == 0 ? ndp->ni_vp : NULL), true); break; case CACHE_FPL_STATUS_PARTIAL: *pwdp = fpl.pwd; /* * Status restored by cache_fplookup_partial_setup. */ break; case CACHE_FPL_STATUS_ABORTED: cache_fpl_restore_abort(&fpl, &orig); break; } return (error); } Index: head/sys/kern/vfs_default.c =================================================================== --- head/sys/kern/vfs_default.c (revision 367534) +++ head/sys/kern/vfs_default.c (revision 367535) @@ -1,1593 +1,1594 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed * to Berkeley by John Heidemann of the UCLA Ficus project. * * Source: * @(#)i405_init.c 2.10 92/04/27 UCLA Ficus project * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static int vop_nolookup(struct vop_lookup_args *); static int vop_norename(struct vop_rename_args *); static int vop_nostrategy(struct vop_strategy_args *); static int get_next_dirent(struct vnode *vp, struct dirent **dpp, char *dirbuf, int dirbuflen, off_t *off, char **cpos, int *len, int *eofflag, struct thread *td); static int dirent_exists(struct vnode *vp, const char *dirname, struct thread *td); #define DIRENT_MINSIZE (sizeof(struct dirent) - (MAXNAMLEN+1) + 4) static int vop_stdis_text(struct vop_is_text_args *ap); static int vop_stdunset_text(struct vop_unset_text_args *ap); static int vop_stdadd_writecount(struct vop_add_writecount_args *ap); static int vop_stdcopy_file_range(struct vop_copy_file_range_args *ap); static int vop_stdfdatasync(struct vop_fdatasync_args *ap); static int vop_stdgetpages_async(struct vop_getpages_async_args *ap); static int vop_stdread_pgcache(struct vop_read_pgcache_args *ap); static int vop_stdstat(struct vop_stat_args *ap); /* * This vnode table stores what we want to do if the filesystem doesn't * implement a particular VOP. * * If there is no specific entry here, we will return EOPNOTSUPP. * * Note that every filesystem has to implement either vop_access * or vop_accessx; failing to do so will result in immediate crash * due to stack overflow, as vop_stdaccess() calls vop_stdaccessx(), * which calls vop_stdaccess() etc. */ struct vop_vector default_vnodeops = { .vop_default = NULL, .vop_bypass = VOP_EOPNOTSUPP, .vop_access = vop_stdaccess, .vop_accessx = vop_stdaccessx, .vop_advise = vop_stdadvise, .vop_advlock = vop_stdadvlock, .vop_advlockasync = vop_stdadvlockasync, .vop_advlockpurge = vop_stdadvlockpurge, .vop_allocate = vop_stdallocate, .vop_bmap = vop_stdbmap, .vop_close = VOP_NULL, .vop_fsync = VOP_NULL, .vop_stat = vop_stdstat, .vop_fdatasync = vop_stdfdatasync, .vop_getpages = vop_stdgetpages, .vop_getpages_async = vop_stdgetpages_async, .vop_getwritemount = vop_stdgetwritemount, .vop_inactive = VOP_NULL, .vop_need_inactive = vop_stdneed_inactive, .vop_ioctl = vop_stdioctl, .vop_kqfilter = vop_stdkqfilter, .vop_islocked = vop_stdislocked, .vop_lock1 = vop_stdlock, .vop_lookup = vop_nolookup, .vop_open = VOP_NULL, .vop_pathconf = VOP_EINVAL, .vop_poll = vop_nopoll, .vop_putpages = vop_stdputpages, .vop_readlink = VOP_EINVAL, .vop_read_pgcache = vop_stdread_pgcache, .vop_rename = vop_norename, .vop_revoke = VOP_PANIC, .vop_strategy = vop_nostrategy, .vop_unlock = vop_stdunlock, .vop_vptocnp = vop_stdvptocnp, .vop_vptofh = vop_stdvptofh, .vop_unp_bind = vop_stdunp_bind, .vop_unp_connect = vop_stdunp_connect, .vop_unp_detach = vop_stdunp_detach, .vop_is_text = vop_stdis_text, .vop_set_text = vop_stdset_text, .vop_unset_text = vop_stdunset_text, .vop_add_writecount = vop_stdadd_writecount, .vop_copy_file_range = vop_stdcopy_file_range, }; VFS_VOP_VECTOR_REGISTER(default_vnodeops); /* * Series of placeholder functions for various error returns for * VOPs. */ int vop_eopnotsupp(struct vop_generic_args *ap) { /* printf("vop_notsupp[%s]\n", ap->a_desc->vdesc_name); */ return (EOPNOTSUPP); } int vop_ebadf(struct vop_generic_args *ap) { return (EBADF); } int vop_enotty(struct vop_generic_args *ap) { return (ENOTTY); } int vop_einval(struct vop_generic_args *ap) { return (EINVAL); } int vop_enoent(struct vop_generic_args *ap) { return (ENOENT); } int vop_eagain(struct vop_generic_args *ap) { return (EAGAIN); } int vop_null(struct vop_generic_args *ap) { return (0); } /* * Helper function to panic on some bad VOPs in some filesystems. */ int vop_panic(struct vop_generic_args *ap) { panic("filesystem goof: vop_panic[%s]", ap->a_desc->vdesc_name); } /* * vop_std and vop_no are default functions for use by * filesystems that need the "default reasonable" implementation for a * particular operation. * * The documentation for the operations they implement exists (if it exists) * in the VOP_(9) manpage (all uppercase). */ /* * Default vop for filesystems that do not support name lookup */ static int vop_nolookup(ap) struct vop_lookup_args /* { struct vnode *a_dvp; struct vnode **a_vpp; struct componentname *a_cnp; } */ *ap; { *ap->a_vpp = NULL; return (ENOTDIR); } /* * vop_norename: * * Handle unlock and reference counting for arguments of vop_rename * for filesystems that do not implement rename operation. */ static int vop_norename(struct vop_rename_args *ap) { vop_rename_fail(ap); return (EOPNOTSUPP); } /* * vop_nostrategy: * * Strategy routine for VFS devices that have none. * * BIO_ERROR and B_INVAL must be cleared prior to calling any strategy * routine. Typically this is done for a BIO_READ strategy call. * Typically B_INVAL is assumed to already be clear prior to a write * and should not be cleared manually unless you just made the buffer * invalid. BIO_ERROR should be cleared either way. */ static int vop_nostrategy (struct vop_strategy_args *ap) { printf("No strategy for buffer at %p\n", ap->a_bp); vn_printf(ap->a_vp, "vnode "); ap->a_bp->b_ioflags |= BIO_ERROR; ap->a_bp->b_error = EOPNOTSUPP; bufdone(ap->a_bp); return (EOPNOTSUPP); } static int get_next_dirent(struct vnode *vp, struct dirent **dpp, char *dirbuf, int dirbuflen, off_t *off, char **cpos, int *len, int *eofflag, struct thread *td) { int error, reclen; struct uio uio; struct iovec iov; struct dirent *dp; KASSERT(VOP_ISLOCKED(vp), ("vp %p is not locked", vp)); KASSERT(vp->v_type == VDIR, ("vp %p is not a directory", vp)); if (*len == 0) { iov.iov_base = dirbuf; iov.iov_len = dirbuflen; uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = *off; uio.uio_resid = dirbuflen; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = td; *eofflag = 0; #ifdef MAC error = mac_vnode_check_readdir(td->td_ucred, vp); if (error == 0) #endif error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag, NULL, NULL); if (error) return (error); *off = uio.uio_offset; *cpos = dirbuf; *len = (dirbuflen - uio.uio_resid); if (*len == 0) return (ENOENT); } dp = (struct dirent *)(*cpos); reclen = dp->d_reclen; *dpp = dp; /* check for malformed directory.. */ if (reclen < DIRENT_MINSIZE) return (EINVAL); *cpos += reclen; *len -= reclen; return (0); } /* * Check if a named file exists in a given directory vnode. */ static int dirent_exists(struct vnode *vp, const char *dirname, struct thread *td) { char *dirbuf, *cpos; int error, eofflag, dirbuflen, len, found; off_t off; struct dirent *dp; struct vattr va; KASSERT(VOP_ISLOCKED(vp), ("vp %p is not locked", vp)); KASSERT(vp->v_type == VDIR, ("vp %p is not a directory", vp)); found = 0; error = VOP_GETATTR(vp, &va, td->td_ucred); if (error) return (found); dirbuflen = DEV_BSIZE; if (dirbuflen < va.va_blocksize) dirbuflen = va.va_blocksize; dirbuf = (char *)malloc(dirbuflen, M_TEMP, M_WAITOK); off = 0; len = 0; do { error = get_next_dirent(vp, &dp, dirbuf, dirbuflen, &off, &cpos, &len, &eofflag, td); if (error) goto out; if (dp->d_type != DT_WHT && dp->d_fileno != 0 && strcmp(dp->d_name, dirname) == 0) { found = 1; goto out; } } while (len > 0 || !eofflag); out: free(dirbuf, M_TEMP); return (found); } int vop_stdaccess(struct vop_access_args *ap) { KASSERT((ap->a_accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, ("invalid bit in accmode")); return (VOP_ACCESSX(ap->a_vp, ap->a_accmode, ap->a_cred, ap->a_td)); } int vop_stdaccessx(struct vop_accessx_args *ap) { int error; accmode_t accmode = ap->a_accmode; error = vfs_unixify_accmode(&accmode); if (error != 0) return (error); if (accmode == 0) return (0); return (VOP_ACCESS(ap->a_vp, accmode, ap->a_cred, ap->a_td)); } /* * Advisory record locking support */ int vop_stdadvlock(struct vop_advlock_args *ap) { struct vnode *vp; struct vattr vattr; int error; vp = ap->a_vp; if (ap->a_fl->l_whence == SEEK_END) { /* * The NFSv4 server must avoid doing a vn_lock() here, since it * can deadlock the nfsd threads, due to a LOR. Fortunately * the NFSv4 server always uses SEEK_SET and this code is * only required for the SEEK_END case. */ vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, curthread->td_ucred); VOP_UNLOCK(vp); if (error) return (error); } else vattr.va_size = 0; return (lf_advlock(ap, &(vp->v_lockf), vattr.va_size)); } int vop_stdadvlockasync(struct vop_advlockasync_args *ap) { struct vnode *vp; struct vattr vattr; int error; vp = ap->a_vp; if (ap->a_fl->l_whence == SEEK_END) { /* The size argument is only needed for SEEK_END. */ vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, curthread->td_ucred); VOP_UNLOCK(vp); if (error) return (error); } else vattr.va_size = 0; return (lf_advlockasync(ap, &(vp->v_lockf), vattr.va_size)); } int vop_stdadvlockpurge(struct vop_advlockpurge_args *ap) { struct vnode *vp; vp = ap->a_vp; lf_purgelocks(vp, &vp->v_lockf); return (0); } /* * vop_stdpathconf: * * Standard implementation of POSIX pathconf, to get information about limits * for a filesystem. * Override per filesystem for the case where the filesystem has smaller * limits. */ int vop_stdpathconf(ap) struct vop_pathconf_args /* { struct vnode *a_vp; int a_name; int *a_retval; } */ *ap; { switch (ap->a_name) { case _PC_ASYNC_IO: *ap->a_retval = _POSIX_ASYNCHRONOUS_IO; return (0); case _PC_PATH_MAX: *ap->a_retval = PATH_MAX; return (0); case _PC_ACL_EXTENDED: case _PC_ACL_NFS4: case _PC_CAP_PRESENT: case _PC_INF_PRESENT: case _PC_MAC_PRESENT: *ap->a_retval = 0; return (0); default: return (EINVAL); } /* NOTREACHED */ } /* * Standard lock, unlock and islocked functions. */ int vop_stdlock(ap) struct vop_lock1_args /* { struct vnode *a_vp; int a_flags; char *file; int line; } */ *ap; { struct vnode *vp = ap->a_vp; struct mtx *ilk; ilk = VI_MTX(vp); return (lockmgr_lock_flags(vp->v_vnlock, ap->a_flags, &ilk->lock_object, ap->a_file, ap->a_line)); } /* See above. */ int vop_stdunlock(ap) struct vop_unlock_args /* { struct vnode *a_vp; } */ *ap; { struct vnode *vp = ap->a_vp; return (lockmgr_unlock(vp->v_vnlock)); } /* See above. */ int vop_stdislocked(ap) struct vop_islocked_args /* { struct vnode *a_vp; } */ *ap; { return (lockstatus(ap->a_vp->v_vnlock)); } /* * Variants of the above set. * * Differences are: * - shared locking disablement is not supported * - v_vnlock pointer is not honored */ int vop_lock(ap) struct vop_lock1_args /* { struct vnode *a_vp; int a_flags; char *file; int line; } */ *ap; { struct vnode *vp = ap->a_vp; int flags = ap->a_flags; struct mtx *ilk; MPASS(vp->v_vnlock == &vp->v_lock); if (__predict_false((flags & ~(LK_TYPE_MASK | LK_NODDLKTREAT | LK_RETRY)) != 0)) goto other; switch (flags & LK_TYPE_MASK) { case LK_SHARED: return (lockmgr_slock(&vp->v_lock, flags, ap->a_file, ap->a_line)); case LK_EXCLUSIVE: return (lockmgr_xlock(&vp->v_lock, flags, ap->a_file, ap->a_line)); } other: ilk = VI_MTX(vp); return (lockmgr_lock_flags(&vp->v_lock, flags, &ilk->lock_object, ap->a_file, ap->a_line)); } int vop_unlock(ap) struct vop_unlock_args /* { struct vnode *a_vp; } */ *ap; { struct vnode *vp = ap->a_vp; MPASS(vp->v_vnlock == &vp->v_lock); return (lockmgr_unlock(&vp->v_lock)); } int vop_islocked(ap) struct vop_islocked_args /* { struct vnode *a_vp; } */ *ap; { struct vnode *vp = ap->a_vp; MPASS(vp->v_vnlock == &vp->v_lock); return (lockstatus(&vp->v_lock)); } /* * Return true for select/poll. */ int vop_nopoll(ap) struct vop_poll_args /* { struct vnode *a_vp; int a_events; struct ucred *a_cred; struct thread *a_td; } */ *ap; { if (ap->a_events & ~POLLSTANDARD) return (POLLNVAL); return (ap->a_events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)); } /* * Implement poll for local filesystems that support it. */ int vop_stdpoll(ap) struct vop_poll_args /* { struct vnode *a_vp; int a_events; struct ucred *a_cred; struct thread *a_td; } */ *ap; { if (ap->a_events & ~POLLSTANDARD) return (vn_pollrecord(ap->a_vp, ap->a_td, ap->a_events)); return (ap->a_events & (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)); } /* * Return our mount point, as we will take charge of the writes. */ int vop_stdgetwritemount(ap) struct vop_getwritemount_args /* { struct vnode *a_vp; struct mount **a_mpp; } */ *ap; { struct mount *mp; + struct mount_pcpu *mpcpu; struct vnode *vp; /* * Note that having a reference does not prevent forced unmount from * setting ->v_mount to NULL after the lock gets released. This is of * no consequence for typical consumers (most notably vn_start_write) * since in this case the vnode is VIRF_DOOMED. Unmount might have * progressed far enough that its completion is only delayed by the * reference obtained here. The consumer only needs to concern itself * with releasing it. */ vp = ap->a_vp; mp = vp->v_mount; if (mp == NULL) { *(ap->a_mpp) = NULL; return (0); } - if (vfs_op_thread_enter(mp)) { + if (vfs_op_thread_enter(mp, mpcpu)) { if (mp == vp->v_mount) { - vfs_mp_count_add_pcpu(mp, ref, 1); - vfs_op_thread_exit(mp); + vfs_mp_count_add_pcpu(mpcpu, ref, 1); + vfs_op_thread_exit(mp, mpcpu); } else { - vfs_op_thread_exit(mp); + vfs_op_thread_exit(mp, mpcpu); mp = NULL; } } else { MNT_ILOCK(mp); if (mp == vp->v_mount) { MNT_REF(mp); MNT_IUNLOCK(mp); } else { MNT_IUNLOCK(mp); mp = NULL; } } *(ap->a_mpp) = mp; return (0); } /* * If the file system doesn't implement VOP_BMAP, then return sensible defaults: * - Return the vnode's bufobj instead of any underlying device's bufobj * - Calculate the physical block number as if there were equal size * consecutive blocks, but * - Report no contiguous runs of blocks. */ int vop_stdbmap(ap) struct vop_bmap_args /* { struct vnode *a_vp; daddr_t a_bn; struct bufobj **a_bop; daddr_t *a_bnp; int *a_runp; int *a_runb; } */ *ap; { if (ap->a_bop != NULL) *ap->a_bop = &ap->a_vp->v_bufobj; if (ap->a_bnp != NULL) *ap->a_bnp = ap->a_bn * btodb(ap->a_vp->v_mount->mnt_stat.f_iosize); if (ap->a_runp != NULL) *ap->a_runp = 0; if (ap->a_runb != NULL) *ap->a_runb = 0; return (0); } int vop_stdfsync(ap) struct vop_fsync_args /* { struct vnode *a_vp; int a_waitfor; struct thread *a_td; } */ *ap; { return (vn_fsync_buf(ap->a_vp, ap->a_waitfor)); } static int vop_stdfdatasync(struct vop_fdatasync_args *ap) { return (VOP_FSYNC(ap->a_vp, MNT_WAIT, ap->a_td)); } int vop_stdfdatasync_buf(struct vop_fdatasync_args *ap) { return (vn_fsync_buf(ap->a_vp, MNT_WAIT)); } /* XXX Needs good comment and more info in the manpage (VOP_GETPAGES(9)). */ int vop_stdgetpages(ap) struct vop_getpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int *a_rbehind; int *a_rahead; } */ *ap; { return vnode_pager_generic_getpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, NULL, NULL); } static int vop_stdgetpages_async(struct vop_getpages_async_args *ap) { int error; error = VOP_GETPAGES(ap->a_vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead); if (ap->a_iodone != NULL) ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error); return (error); } int vop_stdkqfilter(struct vop_kqfilter_args *ap) { return vfs_kqfilter(ap); } /* XXX Needs good comment and more info in the manpage (VOP_PUTPAGES(9)). */ int vop_stdputpages(ap) struct vop_putpages_args /* { struct vnode *a_vp; vm_page_t *a_m; int a_count; int a_sync; int *a_rtvals; } */ *ap; { return vnode_pager_generic_putpages(ap->a_vp, ap->a_m, ap->a_count, ap->a_sync, ap->a_rtvals); } int vop_stdvptofh(struct vop_vptofh_args *ap) { return (EOPNOTSUPP); } int vop_stdvptocnp(struct vop_vptocnp_args *ap) { struct vnode *vp = ap->a_vp; struct vnode **dvp = ap->a_vpp; struct ucred *cred; char *buf = ap->a_buf; size_t *buflen = ap->a_buflen; char *dirbuf, *cpos; int i, error, eofflag, dirbuflen, flags, locked, len, covered; off_t off; ino_t fileno; struct vattr va; struct nameidata nd; struct thread *td; struct dirent *dp; struct vnode *mvp; i = *buflen; error = 0; covered = 0; td = curthread; cred = td->td_ucred; if (vp->v_type != VDIR) return (ENOENT); error = VOP_GETATTR(vp, &va, cred); if (error) return (error); VREF(vp); locked = VOP_ISLOCKED(vp); VOP_UNLOCK(vp); NDINIT_ATVP(&nd, LOOKUP, FOLLOW | LOCKSHARED | LOCKLEAF, UIO_SYSSPACE, "..", vp, td); flags = FREAD; error = vn_open_cred(&nd, &flags, 0, VN_OPEN_NOAUDIT, cred, NULL); if (error) { vn_lock(vp, locked | LK_RETRY); return (error); } NDFREE(&nd, NDF_ONLY_PNBUF); mvp = *dvp = nd.ni_vp; if (vp->v_mount != (*dvp)->v_mount && ((*dvp)->v_vflag & VV_ROOT) && ((*dvp)->v_mount->mnt_flag & MNT_UNION)) { *dvp = (*dvp)->v_mount->mnt_vnodecovered; VREF(mvp); VOP_UNLOCK(mvp); vn_close(mvp, FREAD, cred, td); VREF(*dvp); vn_lock(*dvp, LK_SHARED | LK_RETRY); covered = 1; } fileno = va.va_fileid; dirbuflen = DEV_BSIZE; if (dirbuflen < va.va_blocksize) dirbuflen = va.va_blocksize; dirbuf = (char *)malloc(dirbuflen, M_TEMP, M_WAITOK); if ((*dvp)->v_type != VDIR) { error = ENOENT; goto out; } off = 0; len = 0; do { /* call VOP_READDIR of parent */ error = get_next_dirent(*dvp, &dp, dirbuf, dirbuflen, &off, &cpos, &len, &eofflag, td); if (error) goto out; if ((dp->d_type != DT_WHT) && (dp->d_fileno == fileno)) { if (covered) { VOP_UNLOCK(*dvp); vn_lock(mvp, LK_SHARED | LK_RETRY); if (dirent_exists(mvp, dp->d_name, td)) { error = ENOENT; VOP_UNLOCK(mvp); vn_lock(*dvp, LK_SHARED | LK_RETRY); goto out; } VOP_UNLOCK(mvp); vn_lock(*dvp, LK_SHARED | LK_RETRY); } i -= dp->d_namlen; if (i < 0) { error = ENOMEM; goto out; } if (dp->d_namlen == 1 && dp->d_name[0] == '.') { error = ENOENT; } else { bcopy(dp->d_name, buf + i, dp->d_namlen); error = 0; } goto out; } } while (len > 0 || !eofflag); error = ENOENT; out: free(dirbuf, M_TEMP); if (!error) { *buflen = i; vref(*dvp); } if (covered) { vput(*dvp); vrele(mvp); } else { VOP_UNLOCK(mvp); vn_close(mvp, FREAD, cred, td); } vn_lock(vp, locked | LK_RETRY); return (error); } int vop_stdallocate(struct vop_allocate_args *ap) { #ifdef __notyet__ struct statfs *sfs; off_t maxfilesize = 0; #endif struct iovec aiov; struct vattr vattr, *vap; struct uio auio; off_t fsize, len, cur, offset; uint8_t *buf; struct thread *td; struct vnode *vp; size_t iosize; int error; buf = NULL; error = 0; td = curthread; vap = &vattr; vp = ap->a_vp; len = *ap->a_len; offset = *ap->a_offset; error = VOP_GETATTR(vp, vap, td->td_ucred); if (error != 0) goto out; fsize = vap->va_size; iosize = vap->va_blocksize; if (iosize == 0) iosize = BLKDEV_IOSIZE; if (iosize > MAXPHYS) iosize = MAXPHYS; buf = malloc(iosize, M_TEMP, M_WAITOK); #ifdef __notyet__ /* * Check if the filesystem sets f_maxfilesize; if not use * VOP_SETATTR to perform the check. */ sfs = malloc(sizeof(struct statfs), M_STATFS, M_WAITOK); error = VFS_STATFS(vp->v_mount, sfs, td); if (error == 0) maxfilesize = sfs->f_maxfilesize; free(sfs, M_STATFS); if (error != 0) goto out; if (maxfilesize) { if (offset > maxfilesize || len > maxfilesize || offset + len > maxfilesize) { error = EFBIG; goto out; } } else #endif if (offset + len > vap->va_size) { /* * Test offset + len against the filesystem's maxfilesize. */ VATTR_NULL(vap); vap->va_size = offset + len; error = VOP_SETATTR(vp, vap, td->td_ucred); if (error != 0) goto out; VATTR_NULL(vap); vap->va_size = fsize; error = VOP_SETATTR(vp, vap, td->td_ucred); if (error != 0) goto out; } for (;;) { /* * Read and write back anything below the nominal file * size. There's currently no way outside the filesystem * to know whether this area is sparse or not. */ cur = iosize; if ((offset % iosize) != 0) cur -= (offset % iosize); if (cur > len) cur = len; if (offset < fsize) { aiov.iov_base = buf; aiov.iov_len = cur; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = offset; auio.uio_resid = cur; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_READ; auio.uio_td = td; error = VOP_READ(vp, &auio, 0, td->td_ucred); if (error != 0) break; if (auio.uio_resid > 0) { bzero(buf + cur - auio.uio_resid, auio.uio_resid); } } else { bzero(buf, cur); } aiov.iov_base = buf; aiov.iov_len = cur; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = offset; auio.uio_resid = cur; auio.uio_segflg = UIO_SYSSPACE; auio.uio_rw = UIO_WRITE; auio.uio_td = td; error = VOP_WRITE(vp, &auio, 0, td->td_ucred); if (error != 0) break; len -= cur; offset += cur; if (len == 0) break; if (should_yield()) break; } out: *ap->a_len = len; *ap->a_offset = offset; free(buf, M_TEMP); return (error); } int vop_stdadvise(struct vop_advise_args *ap) { struct vnode *vp; struct bufobj *bo; daddr_t startn, endn; off_t bstart, bend, start, end; int bsize, error; vp = ap->a_vp; switch (ap->a_advice) { case POSIX_FADV_WILLNEED: /* * Do nothing for now. Filesystems should provide a * custom method which starts an asynchronous read of * the requested region. */ error = 0; break; case POSIX_FADV_DONTNEED: error = 0; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); if (VN_IS_DOOMED(vp)) { VOP_UNLOCK(vp); break; } /* * Round to block boundaries (and later possibly further to * page boundaries). Applications cannot reasonably be aware * of the boundaries, and the rounding must be to expand at * both extremities to cover enough. It still doesn't cover * read-ahead. For partial blocks, this gives unnecessary * discarding of buffers but is efficient enough since the * pages usually remain in VMIO for some time. */ bsize = vp->v_bufobj.bo_bsize; bstart = rounddown(ap->a_start, bsize); bend = roundup(ap->a_end, bsize); /* * Deactivate pages in the specified range from the backing VM * object. Pages that are resident in the buffer cache will * remain wired until their corresponding buffers are released * below. */ if (vp->v_object != NULL) { start = trunc_page(bstart); end = round_page(bend); VM_OBJECT_RLOCK(vp->v_object); vm_object_page_noreuse(vp->v_object, OFF_TO_IDX(start), OFF_TO_IDX(end)); VM_OBJECT_RUNLOCK(vp->v_object); } bo = &vp->v_bufobj; BO_RLOCK(bo); startn = bstart / bsize; endn = bend / bsize; error = bnoreuselist(&bo->bo_clean, bo, startn, endn); if (error == 0) error = bnoreuselist(&bo->bo_dirty, bo, startn, endn); BO_RUNLOCK(bo); VOP_UNLOCK(vp); break; default: error = EINVAL; break; } return (error); } int vop_stdunp_bind(struct vop_unp_bind_args *ap) { ap->a_vp->v_unpcb = ap->a_unpcb; return (0); } int vop_stdunp_connect(struct vop_unp_connect_args *ap) { *ap->a_unpcb = ap->a_vp->v_unpcb; return (0); } int vop_stdunp_detach(struct vop_unp_detach_args *ap) { ap->a_vp->v_unpcb = NULL; return (0); } static int vop_stdis_text(struct vop_is_text_args *ap) { return (ap->a_vp->v_writecount < 0); } int vop_stdset_text(struct vop_set_text_args *ap) { struct vnode *vp; struct mount *mp; int error; vp = ap->a_vp; VI_LOCK(vp); if (vp->v_writecount > 0) { error = ETXTBSY; } else { /* * If requested by fs, keep a use reference to the * vnode until the last text reference is released. */ mp = vp->v_mount; if (mp != NULL && (mp->mnt_kern_flag & MNTK_TEXT_REFS) != 0 && vp->v_writecount == 0) { VNPASS((vp->v_iflag & VI_TEXT_REF) == 0, vp); vp->v_iflag |= VI_TEXT_REF; vrefl(vp); } vp->v_writecount--; error = 0; } VI_UNLOCK(vp); return (error); } static int vop_stdunset_text(struct vop_unset_text_args *ap) { struct vnode *vp; int error; bool last; vp = ap->a_vp; last = false; VI_LOCK(vp); if (vp->v_writecount < 0) { if ((vp->v_iflag & VI_TEXT_REF) != 0 && vp->v_writecount == -1) { last = true; vp->v_iflag &= ~VI_TEXT_REF; } vp->v_writecount++; error = 0; } else { error = EINVAL; } VI_UNLOCK(vp); if (last) vunref(vp); return (error); } static int vop_stdadd_writecount(struct vop_add_writecount_args *ap) { struct vnode *vp; struct mount *mp; int error; vp = ap->a_vp; VI_LOCK_FLAGS(vp, MTX_DUPOK); if (vp->v_writecount < 0) { error = ETXTBSY; } else { VNASSERT(vp->v_writecount + ap->a_inc >= 0, vp, ("neg writecount increment %d", ap->a_inc)); if (vp->v_writecount == 0) { mp = vp->v_mount; if (mp != NULL && (mp->mnt_kern_flag & MNTK_NOMSYNC) == 0) vlazy(vp); } vp->v_writecount += ap->a_inc; error = 0; } VI_UNLOCK(vp); return (error); } int vop_stdneed_inactive(struct vop_need_inactive_args *ap) { return (1); } int vop_stdioctl(struct vop_ioctl_args *ap) { struct vnode *vp; struct vattr va; off_t *offp; int error; switch (ap->a_command) { case FIOSEEKDATA: case FIOSEEKHOLE: vp = ap->a_vp; error = vn_lock(vp, LK_SHARED); if (error != 0) return (EBADF); if (vp->v_type == VREG) error = VOP_GETATTR(vp, &va, ap->a_cred); else error = ENOTTY; if (error == 0) { offp = ap->a_data; if (*offp < 0 || *offp >= va.va_size) error = ENXIO; else if (ap->a_command == FIOSEEKHOLE) *offp = va.va_size; } VOP_UNLOCK(vp); break; default: error = ENOTTY; break; } return (error); } /* * vfs default ops * used to fill the vfs function table to get reasonable default return values. */ int vfs_stdroot (mp, flags, vpp) struct mount *mp; int flags; struct vnode **vpp; { return (EOPNOTSUPP); } int vfs_stdstatfs (mp, sbp) struct mount *mp; struct statfs *sbp; { return (EOPNOTSUPP); } int vfs_stdquotactl (mp, cmds, uid, arg) struct mount *mp; int cmds; uid_t uid; void *arg; { return (EOPNOTSUPP); } int vfs_stdsync(mp, waitfor) struct mount *mp; int waitfor; { struct vnode *vp, *mvp; struct thread *td; int error, lockreq, allerror = 0; td = curthread; lockreq = LK_EXCLUSIVE | LK_INTERLOCK; if (waitfor != MNT_WAIT) lockreq |= LK_NOWAIT; /* * Force stale buffer cache information to be flushed. */ loop: MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { if (vp->v_bufobj.bo_dirty.bv_cnt == 0) { VI_UNLOCK(vp); continue; } if ((error = vget(vp, lockreq)) != 0) { if (error == ENOENT) { MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); goto loop; } continue; } error = VOP_FSYNC(vp, waitfor, td); if (error) allerror = error; vput(vp); } return (allerror); } int vfs_stdnosync (mp, waitfor) struct mount *mp; int waitfor; { return (0); } static int vop_stdcopy_file_range(struct vop_copy_file_range_args *ap) { int error; error = vn_generic_copy_file_range(ap->a_invp, ap->a_inoffp, ap->a_outvp, ap->a_outoffp, ap->a_lenp, ap->a_flags, ap->a_incred, ap->a_outcred, ap->a_fsizetd); return (error); } int vfs_stdvget (mp, ino, flags, vpp) struct mount *mp; ino_t ino; int flags; struct vnode **vpp; { return (EOPNOTSUPP); } int vfs_stdfhtovp (mp, fhp, flags, vpp) struct mount *mp; struct fid *fhp; int flags; struct vnode **vpp; { return (EOPNOTSUPP); } int vfs_stdinit (vfsp) struct vfsconf *vfsp; { return (0); } int vfs_stduninit (vfsp) struct vfsconf *vfsp; { return(0); } int vfs_stdextattrctl(mp, cmd, filename_vp, attrnamespace, attrname) struct mount *mp; int cmd; struct vnode *filename_vp; int attrnamespace; const char *attrname; { if (filename_vp != NULL) VOP_UNLOCK(filename_vp); return (EOPNOTSUPP); } int vfs_stdsysctl(mp, op, req) struct mount *mp; fsctlop_t op; struct sysctl_req *req; { return (EOPNOTSUPP); } static vop_bypass_t * bp_by_off(struct vop_vector *vop, struct vop_generic_args *a) { return (*(vop_bypass_t **)((char *)vop + a->a_desc->vdesc_vop_offset)); } int vop_sigdefer(struct vop_vector *vop, struct vop_generic_args *a) { vop_bypass_t *bp; int prev_stops, rc; bp = bp_by_off(vop, a); MPASS(bp != NULL); prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = bp(a); sigallowstop(prev_stops); return (rc); } static int vop_stdstat(struct vop_stat_args *a) { struct vattr vattr; struct vattr *vap; struct vnode *vp; struct stat *sb; int error; u_short mode; vp = a->a_vp; sb = a->a_sb; error = vop_stat_helper_pre(a); if (error != 0) return (error); vap = &vattr; /* * Initialize defaults for new and unusual fields, so that file * systems which don't support these fields don't need to know * about them. */ vap->va_birthtime.tv_sec = -1; vap->va_birthtime.tv_nsec = 0; vap->va_fsid = VNOVAL; vap->va_rdev = NODEV; error = VOP_GETATTR(vp, vap, a->a_active_cred); if (error) goto out; /* * Zero the spare stat fields */ bzero(sb, sizeof *sb); /* * Copy from vattr table */ if (vap->va_fsid != VNOVAL) sb->st_dev = vap->va_fsid; else sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; sb->st_ino = vap->va_fileid; mode = vap->va_mode; switch (vap->va_type) { case VREG: mode |= S_IFREG; break; case VDIR: mode |= S_IFDIR; break; case VBLK: mode |= S_IFBLK; break; case VCHR: mode |= S_IFCHR; break; case VLNK: mode |= S_IFLNK; break; case VSOCK: mode |= S_IFSOCK; break; case VFIFO: mode |= S_IFIFO; break; default: error = EBADF; goto out; } sb->st_mode = mode; sb->st_nlink = vap->va_nlink; sb->st_uid = vap->va_uid; sb->st_gid = vap->va_gid; sb->st_rdev = vap->va_rdev; if (vap->va_size > OFF_MAX) { error = EOVERFLOW; goto out; } sb->st_size = vap->va_size; sb->st_atim.tv_sec = vap->va_atime.tv_sec; sb->st_atim.tv_nsec = vap->va_atime.tv_nsec; sb->st_mtim.tv_sec = vap->va_mtime.tv_sec; sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec; sb->st_ctim.tv_sec = vap->va_ctime.tv_sec; sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec; sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec; sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec; /* * According to www.opengroup.org, the meaning of st_blksize is * "a filesystem-specific preferred I/O block size for this * object. In some filesystem types, this may vary from file * to file" * Use minimum/default of PAGE_SIZE (e.g. for VCHR). */ sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize); sb->st_flags = vap->va_flags; sb->st_blocks = vap->va_bytes / S_BLKSIZE; sb->st_gen = vap->va_gen; out: return (vop_stat_helper_post(a, error)); } static int vop_stdread_pgcache(struct vop_read_pgcache_args *ap __unused) { return (EJUSTRETURN); } Index: head/sys/kern/vfs_mount.c =================================================================== --- head/sys/kern/vfs_mount.c (revision 367534) +++ head/sys/kern/vfs_mount.c (revision 367535) @@ -1,2573 +1,2586 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1999-2004 Poul-Henning Kamp * Copyright (c) 1999 Michael Smith * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define VFS_MOUNTARG_SIZE_MAX (1024 * 64) static int vfs_domount(struct thread *td, const char *fstype, char *fspath, uint64_t fsflags, struct vfsoptlist **optlist); static void free_mntarg(struct mntarg *ma); static int usermount = 0; SYSCTL_INT(_vfs, OID_AUTO, usermount, CTLFLAG_RW, &usermount, 0, "Unprivileged users may mount and unmount file systems"); static bool default_autoro = false; SYSCTL_BOOL(_vfs, OID_AUTO, default_autoro, CTLFLAG_RW, &default_autoro, 0, "Retry failed r/w mount as r/o if no explicit ro/rw option is specified"); MALLOC_DEFINE(M_MOUNT, "mount", "vfs mount structure"); MALLOC_DEFINE(M_STATFS, "statfs", "statfs structure"); static uma_zone_t mount_zone; /* List of mounted filesystems. */ struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist); /* For any iteration/modification of mountlist */ struct mtx_padalign __exclusive_cache_line mountlist_mtx; MTX_SYSINIT(mountlist, &mountlist_mtx, "mountlist", MTX_DEF); EVENTHANDLER_LIST_DEFINE(vfs_mounted); EVENTHANDLER_LIST_DEFINE(vfs_unmounted); static void mount_devctl_event(const char *type, struct mount *mp, bool donew); /* * Global opts, taken by all filesystems */ static const char *global_opts[] = { "errmsg", "fstype", "fspath", "ro", "rw", "nosuid", "noexec", NULL }; static int mount_init(void *mem, int size, int flags) { struct mount *mp; mp = (struct mount *)mem; mtx_init(&mp->mnt_mtx, "struct mount mtx", NULL, MTX_DEF); mtx_init(&mp->mnt_listmtx, "struct mount vlist mtx", NULL, MTX_DEF); lockinit(&mp->mnt_explock, PVFS, "explock", 0, 0); - mp->mnt_thread_in_ops_pcpu = uma_zalloc_pcpu(pcpu_zone_4, - M_WAITOK | M_ZERO); - mp->mnt_ref_pcpu = uma_zalloc_pcpu(pcpu_zone_4, - M_WAITOK | M_ZERO); - mp->mnt_lockref_pcpu = uma_zalloc_pcpu(pcpu_zone_4, - M_WAITOK | M_ZERO); - mp->mnt_writeopcount_pcpu = uma_zalloc_pcpu(pcpu_zone_4, - M_WAITOK | M_ZERO); + mp->mnt_pcpu = uma_zalloc_pcpu(pcpu_zone_16, M_WAITOK | M_ZERO); mp->mnt_ref = 0; mp->mnt_vfs_ops = 1; mp->mnt_rootvnode = NULL; return (0); } static void mount_fini(void *mem, int size) { struct mount *mp; mp = (struct mount *)mem; - uma_zfree_pcpu(pcpu_zone_4, mp->mnt_writeopcount_pcpu); - uma_zfree_pcpu(pcpu_zone_4, mp->mnt_lockref_pcpu); - uma_zfree_pcpu(pcpu_zone_4, mp->mnt_ref_pcpu); - uma_zfree_pcpu(pcpu_zone_4, mp->mnt_thread_in_ops_pcpu); + uma_zfree_pcpu(pcpu_zone_16, mp->mnt_pcpu); lockdestroy(&mp->mnt_explock); mtx_destroy(&mp->mnt_listmtx); mtx_destroy(&mp->mnt_mtx); } static void vfs_mount_init(void *dummy __unused) { mount_zone = uma_zcreate("Mountpoints", sizeof(struct mount), NULL, NULL, mount_init, mount_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE); } SYSINIT(vfs_mount, SI_SUB_VFS, SI_ORDER_ANY, vfs_mount_init, NULL); /* * --------------------------------------------------------------------- * Functions for building and sanitizing the mount options */ /* Remove one mount option. */ static void vfs_freeopt(struct vfsoptlist *opts, struct vfsopt *opt) { TAILQ_REMOVE(opts, opt, link); free(opt->name, M_MOUNT); if (opt->value != NULL) free(opt->value, M_MOUNT); free(opt, M_MOUNT); } /* Release all resources related to the mount options. */ void vfs_freeopts(struct vfsoptlist *opts) { struct vfsopt *opt; while (!TAILQ_EMPTY(opts)) { opt = TAILQ_FIRST(opts); vfs_freeopt(opts, opt); } free(opts, M_MOUNT); } void vfs_deleteopt(struct vfsoptlist *opts, const char *name) { struct vfsopt *opt, *temp; if (opts == NULL) return; TAILQ_FOREACH_SAFE(opt, opts, link, temp) { if (strcmp(opt->name, name) == 0) vfs_freeopt(opts, opt); } } static int vfs_isopt_ro(const char *opt) { if (strcmp(opt, "ro") == 0 || strcmp(opt, "rdonly") == 0 || strcmp(opt, "norw") == 0) return (1); return (0); } static int vfs_isopt_rw(const char *opt) { if (strcmp(opt, "rw") == 0 || strcmp(opt, "noro") == 0) return (1); return (0); } /* * Check if options are equal (with or without the "no" prefix). */ static int vfs_equalopts(const char *opt1, const char *opt2) { char *p; /* "opt" vs. "opt" or "noopt" vs. "noopt" */ if (strcmp(opt1, opt2) == 0) return (1); /* "noopt" vs. "opt" */ if (strncmp(opt1, "no", 2) == 0 && strcmp(opt1 + 2, opt2) == 0) return (1); /* "opt" vs. "noopt" */ if (strncmp(opt2, "no", 2) == 0 && strcmp(opt1, opt2 + 2) == 0) return (1); while ((p = strchr(opt1, '.')) != NULL && !strncmp(opt1, opt2, ++p - opt1)) { opt2 += p - opt1; opt1 = p; /* "foo.noopt" vs. "foo.opt" */ if (strncmp(opt1, "no", 2) == 0 && strcmp(opt1 + 2, opt2) == 0) return (1); /* "foo.opt" vs. "foo.noopt" */ if (strncmp(opt2, "no", 2) == 0 && strcmp(opt1, opt2 + 2) == 0) return (1); } /* "ro" / "rdonly" / "norw" / "rw" / "noro" */ if ((vfs_isopt_ro(opt1) || vfs_isopt_rw(opt1)) && (vfs_isopt_ro(opt2) || vfs_isopt_rw(opt2))) return (1); return (0); } /* * If a mount option is specified several times, * (with or without the "no" prefix) only keep * the last occurrence of it. */ static void vfs_sanitizeopts(struct vfsoptlist *opts) { struct vfsopt *opt, *opt2, *tmp; TAILQ_FOREACH_REVERSE(opt, opts, vfsoptlist, link) { opt2 = TAILQ_PREV(opt, vfsoptlist, link); while (opt2 != NULL) { if (vfs_equalopts(opt->name, opt2->name)) { tmp = TAILQ_PREV(opt2, vfsoptlist, link); vfs_freeopt(opts, opt2); opt2 = tmp; } else { opt2 = TAILQ_PREV(opt2, vfsoptlist, link); } } } } /* * Build a linked list of mount options from a struct uio. */ int vfs_buildopts(struct uio *auio, struct vfsoptlist **options) { struct vfsoptlist *opts; struct vfsopt *opt; size_t memused, namelen, optlen; unsigned int i, iovcnt; int error; opts = malloc(sizeof(struct vfsoptlist), M_MOUNT, M_WAITOK); TAILQ_INIT(opts); memused = 0; iovcnt = auio->uio_iovcnt; for (i = 0; i < iovcnt; i += 2) { namelen = auio->uio_iov[i].iov_len; optlen = auio->uio_iov[i + 1].iov_len; memused += sizeof(struct vfsopt) + optlen + namelen; /* * Avoid consuming too much memory, and attempts to overflow * memused. */ if (memused > VFS_MOUNTARG_SIZE_MAX || optlen > VFS_MOUNTARG_SIZE_MAX || namelen > VFS_MOUNTARG_SIZE_MAX) { error = EINVAL; goto bad; } opt = malloc(sizeof(struct vfsopt), M_MOUNT, M_WAITOK); opt->name = malloc(namelen, M_MOUNT, M_WAITOK); opt->value = NULL; opt->len = 0; opt->pos = i / 2; opt->seen = 0; /* * Do this early, so jumps to "bad" will free the current * option. */ TAILQ_INSERT_TAIL(opts, opt, link); if (auio->uio_segflg == UIO_SYSSPACE) { bcopy(auio->uio_iov[i].iov_base, opt->name, namelen); } else { error = copyin(auio->uio_iov[i].iov_base, opt->name, namelen); if (error) goto bad; } /* Ensure names are null-terminated strings. */ if (namelen == 0 || opt->name[namelen - 1] != '\0') { error = EINVAL; goto bad; } if (optlen != 0) { opt->len = optlen; opt->value = malloc(optlen, M_MOUNT, M_WAITOK); if (auio->uio_segflg == UIO_SYSSPACE) { bcopy(auio->uio_iov[i + 1].iov_base, opt->value, optlen); } else { error = copyin(auio->uio_iov[i + 1].iov_base, opt->value, optlen); if (error) goto bad; } } } vfs_sanitizeopts(opts); *options = opts; return (0); bad: vfs_freeopts(opts); return (error); } /* * Merge the old mount options with the new ones passed * in the MNT_UPDATE case. * * XXX: This function will keep a "nofoo" option in the new * options. E.g, if the option's canonical name is "foo", * "nofoo" ends up in the mount point's active options. */ static void vfs_mergeopts(struct vfsoptlist *toopts, struct vfsoptlist *oldopts) { struct vfsopt *opt, *new; TAILQ_FOREACH(opt, oldopts, link) { new = malloc(sizeof(struct vfsopt), M_MOUNT, M_WAITOK); new->name = strdup(opt->name, M_MOUNT); if (opt->len != 0) { new->value = malloc(opt->len, M_MOUNT, M_WAITOK); bcopy(opt->value, new->value, opt->len); } else new->value = NULL; new->len = opt->len; new->seen = opt->seen; TAILQ_INSERT_HEAD(toopts, new, link); } vfs_sanitizeopts(toopts); } /* * Mount a filesystem. */ #ifndef _SYS_SYSPROTO_H_ struct nmount_args { struct iovec *iovp; unsigned int iovcnt; int flags; }; #endif int sys_nmount(struct thread *td, struct nmount_args *uap) { struct uio *auio; int error; u_int iovcnt; uint64_t flags; /* * Mount flags are now 64-bits. On 32-bit archtectures only * 32-bits are passed in, but from here on everything handles * 64-bit flags correctly. */ flags = uap->flags; AUDIT_ARG_FFLAGS(flags); CTR4(KTR_VFS, "%s: iovp %p with iovcnt %d and flags %d", __func__, uap->iovp, uap->iovcnt, flags); /* * Filter out MNT_ROOTFS. We do not want clients of nmount() in * userspace to set this flag, but we must filter it out if we want * MNT_UPDATE on the root file system to work. * MNT_ROOTFS should only be set by the kernel when mounting its * root file system. */ flags &= ~MNT_ROOTFS; iovcnt = uap->iovcnt; /* * Check that we have an even number of iovec's * and that we have at least two options. */ if ((iovcnt & 1) || (iovcnt < 4)) { CTR2(KTR_VFS, "%s: failed for invalid iovcnt %d", __func__, uap->iovcnt); return (EINVAL); } error = copyinuio(uap->iovp, iovcnt, &auio); if (error) { CTR2(KTR_VFS, "%s: failed for invalid uio op with %d errno", __func__, error); return (error); } error = vfs_donmount(td, flags, auio); free(auio, M_IOV); return (error); } /* * --------------------------------------------------------------------- * Various utility functions */ void vfs_ref(struct mount *mp) { + struct mount_pcpu *mpcpu; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); - if (vfs_op_thread_enter(mp)) { - vfs_mp_count_add_pcpu(mp, ref, 1); - vfs_op_thread_exit(mp); + if (vfs_op_thread_enter(mp, mpcpu)) { + vfs_mp_count_add_pcpu(mpcpu, ref, 1); + vfs_op_thread_exit(mp, mpcpu); return; } MNT_ILOCK(mp); MNT_REF(mp); MNT_IUNLOCK(mp); } void vfs_rel(struct mount *mp) { + struct mount_pcpu *mpcpu; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); - if (vfs_op_thread_enter(mp)) { - vfs_mp_count_sub_pcpu(mp, ref, 1); - vfs_op_thread_exit(mp); + if (vfs_op_thread_enter(mp, mpcpu)) { + vfs_mp_count_sub_pcpu(mpcpu, ref, 1); + vfs_op_thread_exit(mp, mpcpu); return; } MNT_ILOCK(mp); MNT_REL(mp); MNT_IUNLOCK(mp); } /* * Allocate and initialize the mount point struct. */ struct mount * vfs_mount_alloc(struct vnode *vp, struct vfsconf *vfsp, const char *fspath, struct ucred *cred) { struct mount *mp; mp = uma_zalloc(mount_zone, M_WAITOK); bzero(&mp->mnt_startzero, __rangeof(struct mount, mnt_startzero, mnt_endzero)); + mp->mnt_kern_flag = 0; + mp->mnt_flag = 0; + mp->mnt_rootvnode = NULL; + mp->mnt_vnodecovered = NULL; + mp->mnt_op = NULL; + mp->mnt_vfc = NULL; TAILQ_INIT(&mp->mnt_nvnodelist); mp->mnt_nvnodelistsize = 0; TAILQ_INIT(&mp->mnt_lazyvnodelist); mp->mnt_lazyvnodelistsize = 0; if (mp->mnt_ref != 0 || mp->mnt_lockref != 0 || mp->mnt_writeopcount != 0) panic("%s: non-zero counters on new mp %p\n", __func__, mp); if (mp->mnt_vfs_ops != 1) panic("%s: vfs_ops should be 1 but %d found\n", __func__, mp->mnt_vfs_ops); (void) vfs_busy(mp, MBF_NOWAIT); atomic_add_acq_int(&vfsp->vfc_refcount, 1); mp->mnt_op = vfsp->vfc_vfsops; mp->mnt_vfc = vfsp; mp->mnt_stat.f_type = vfsp->vfc_typenum; mp->mnt_gen++; strlcpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN); mp->mnt_vnodecovered = vp; mp->mnt_cred = crdup(cred); mp->mnt_stat.f_owner = cred->cr_uid; strlcpy(mp->mnt_stat.f_mntonname, fspath, MNAMELEN); mp->mnt_iosize_max = DFLTPHYS; #ifdef MAC mac_mount_init(mp); mac_mount_create(cred, mp); #endif arc4rand(&mp->mnt_hashseed, sizeof mp->mnt_hashseed, 0); TAILQ_INIT(&mp->mnt_uppers); return (mp); } /* * Destroy the mount struct previously allocated by vfs_mount_alloc(). */ void vfs_mount_destroy(struct mount *mp) { if (mp->mnt_vfs_ops == 0) panic("%s: entered with zero vfs_ops\n", __func__); vfs_assert_mount_counters(mp); MNT_ILOCK(mp); mp->mnt_kern_flag |= MNTK_REFEXPIRE; if (mp->mnt_kern_flag & MNTK_MWAIT) { mp->mnt_kern_flag &= ~MNTK_MWAIT; wakeup(mp); } while (mp->mnt_ref) msleep(mp, MNT_MTX(mp), PVFS, "mntref", 0); KASSERT(mp->mnt_ref == 0, ("%s: invalid refcount in the drain path @ %s:%d", __func__, __FILE__, __LINE__)); if (mp->mnt_writeopcount != 0) panic("vfs_mount_destroy: nonzero writeopcount"); if (mp->mnt_secondary_writes != 0) panic("vfs_mount_destroy: nonzero secondary_writes"); atomic_subtract_rel_int(&mp->mnt_vfc->vfc_refcount, 1); if (!TAILQ_EMPTY(&mp->mnt_nvnodelist)) { struct vnode *vp; TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) vn_printf(vp, "dangling vnode "); panic("unmount: dangling vnode"); } KASSERT(TAILQ_EMPTY(&mp->mnt_uppers), ("mnt_uppers")); if (mp->mnt_nvnodelistsize != 0) panic("vfs_mount_destroy: nonzero nvnodelistsize"); if (mp->mnt_lazyvnodelistsize != 0) panic("vfs_mount_destroy: nonzero lazyvnodelistsize"); if (mp->mnt_lockref != 0) panic("vfs_mount_destroy: nonzero lock refcount"); MNT_IUNLOCK(mp); if (mp->mnt_vfs_ops != 1) panic("%s: vfs_ops should be 1 but %d found\n", __func__, mp->mnt_vfs_ops); if (mp->mnt_rootvnode != NULL) panic("%s: mount point still has a root vnode %p\n", __func__, mp->mnt_rootvnode); if (mp->mnt_vnodecovered != NULL) vrele(mp->mnt_vnodecovered); #ifdef MAC mac_mount_destroy(mp); #endif if (mp->mnt_opt != NULL) vfs_freeopts(mp->mnt_opt); crfree(mp->mnt_cred); uma_zfree(mount_zone, mp); } static bool vfs_should_downgrade_to_ro_mount(uint64_t fsflags, int error) { /* This is an upgrade of an exisiting mount. */ if ((fsflags & MNT_UPDATE) != 0) return (false); /* This is already an R/O mount. */ if ((fsflags & MNT_RDONLY) != 0) return (false); switch (error) { case ENODEV: /* generic, geom, ... */ case EACCES: /* cam/scsi, ... */ case EROFS: /* md, mmcsd, ... */ /* * These errors can be returned by the storage layer to signal * that the media is read-only. No harm in the R/O mount * attempt if the error was returned for some other reason. */ return (true); default: return (false); } } int vfs_donmount(struct thread *td, uint64_t fsflags, struct uio *fsoptions) { struct vfsoptlist *optlist; struct vfsopt *opt, *tmp_opt; char *fstype, *fspath, *errmsg; int error, fstypelen, fspathlen, errmsg_len, errmsg_pos; bool autoro; errmsg = fspath = NULL; errmsg_len = fspathlen = 0; errmsg_pos = -1; autoro = default_autoro; error = vfs_buildopts(fsoptions, &optlist); if (error) return (error); if (vfs_getopt(optlist, "errmsg", (void **)&errmsg, &errmsg_len) == 0) errmsg_pos = vfs_getopt_pos(optlist, "errmsg"); /* * We need these two options before the others, * and they are mandatory for any filesystem. * Ensure they are NUL terminated as well. */ fstypelen = 0; error = vfs_getopt(optlist, "fstype", (void **)&fstype, &fstypelen); if (error || fstypelen <= 0 || fstype[fstypelen - 1] != '\0') { error = EINVAL; if (errmsg != NULL) strncpy(errmsg, "Invalid fstype", errmsg_len); goto bail; } fspathlen = 0; error = vfs_getopt(optlist, "fspath", (void **)&fspath, &fspathlen); if (error || fspathlen <= 0 || fspath[fspathlen - 1] != '\0') { error = EINVAL; if (errmsg != NULL) strncpy(errmsg, "Invalid fspath", errmsg_len); goto bail; } /* * We need to see if we have the "update" option * before we call vfs_domount(), since vfs_domount() has special * logic based on MNT_UPDATE. This is very important * when we want to update the root filesystem. */ TAILQ_FOREACH_SAFE(opt, optlist, link, tmp_opt) { int do_freeopt = 0; if (strcmp(opt->name, "update") == 0) { fsflags |= MNT_UPDATE; do_freeopt = 1; } else if (strcmp(opt->name, "async") == 0) fsflags |= MNT_ASYNC; else if (strcmp(opt->name, "force") == 0) { fsflags |= MNT_FORCE; do_freeopt = 1; } else if (strcmp(opt->name, "reload") == 0) { fsflags |= MNT_RELOAD; do_freeopt = 1; } else if (strcmp(opt->name, "multilabel") == 0) fsflags |= MNT_MULTILABEL; else if (strcmp(opt->name, "noasync") == 0) fsflags &= ~MNT_ASYNC; else if (strcmp(opt->name, "noatime") == 0) fsflags |= MNT_NOATIME; else if (strcmp(opt->name, "atime") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonoatime", M_MOUNT); } else if (strcmp(opt->name, "noclusterr") == 0) fsflags |= MNT_NOCLUSTERR; else if (strcmp(opt->name, "clusterr") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonoclusterr", M_MOUNT); } else if (strcmp(opt->name, "noclusterw") == 0) fsflags |= MNT_NOCLUSTERW; else if (strcmp(opt->name, "clusterw") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonoclusterw", M_MOUNT); } else if (strcmp(opt->name, "noexec") == 0) fsflags |= MNT_NOEXEC; else if (strcmp(opt->name, "exec") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonoexec", M_MOUNT); } else if (strcmp(opt->name, "nosuid") == 0) fsflags |= MNT_NOSUID; else if (strcmp(opt->name, "suid") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonosuid", M_MOUNT); } else if (strcmp(opt->name, "nosymfollow") == 0) fsflags |= MNT_NOSYMFOLLOW; else if (strcmp(opt->name, "symfollow") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("nonosymfollow", M_MOUNT); } else if (strcmp(opt->name, "noro") == 0) { fsflags &= ~MNT_RDONLY; autoro = false; } else if (strcmp(opt->name, "rw") == 0) { fsflags &= ~MNT_RDONLY; autoro = false; } else if (strcmp(opt->name, "ro") == 0) { fsflags |= MNT_RDONLY; autoro = false; } else if (strcmp(opt->name, "rdonly") == 0) { free(opt->name, M_MOUNT); opt->name = strdup("ro", M_MOUNT); fsflags |= MNT_RDONLY; autoro = false; } else if (strcmp(opt->name, "autoro") == 0) { do_freeopt = 1; autoro = true; } else if (strcmp(opt->name, "suiddir") == 0) fsflags |= MNT_SUIDDIR; else if (strcmp(opt->name, "sync") == 0) fsflags |= MNT_SYNCHRONOUS; else if (strcmp(opt->name, "union") == 0) fsflags |= MNT_UNION; else if (strcmp(opt->name, "automounted") == 0) { fsflags |= MNT_AUTOMOUNTED; do_freeopt = 1; } else if (strcmp(opt->name, "nocover") == 0) { fsflags |= MNT_NOCOVER; do_freeopt = 1; } else if (strcmp(opt->name, "cover") == 0) { fsflags &= ~MNT_NOCOVER; do_freeopt = 1; } else if (strcmp(opt->name, "emptydir") == 0) { fsflags |= MNT_EMPTYDIR; do_freeopt = 1; } else if (strcmp(opt->name, "noemptydir") == 0) { fsflags &= ~MNT_EMPTYDIR; do_freeopt = 1; } if (do_freeopt) vfs_freeopt(optlist, opt); } /* * Be ultra-paranoid about making sure the type and fspath * variables will fit in our mp buffers, including the * terminating NUL. */ if (fstypelen > MFSNAMELEN || fspathlen > MNAMELEN) { error = ENAMETOOLONG; goto bail; } error = vfs_domount(td, fstype, fspath, fsflags, &optlist); /* * See if we can mount in the read-only mode if the error code suggests * that it could be possible and the mount options allow for that. * Never try it if "[no]{ro|rw}" has been explicitly requested and not * overridden by "autoro". */ if (autoro && vfs_should_downgrade_to_ro_mount(fsflags, error)) { printf("%s: R/W mount failed, possibly R/O media," " trying R/O mount\n", __func__); fsflags |= MNT_RDONLY; error = vfs_domount(td, fstype, fspath, fsflags, &optlist); } bail: /* copyout the errmsg */ if (errmsg_pos != -1 && ((2 * errmsg_pos + 1) < fsoptions->uio_iovcnt) && errmsg_len > 0 && errmsg != NULL) { if (fsoptions->uio_segflg == UIO_SYSSPACE) { bcopy(errmsg, fsoptions->uio_iov[2 * errmsg_pos + 1].iov_base, fsoptions->uio_iov[2 * errmsg_pos + 1].iov_len); } else { copyout(errmsg, fsoptions->uio_iov[2 * errmsg_pos + 1].iov_base, fsoptions->uio_iov[2 * errmsg_pos + 1].iov_len); } } if (optlist != NULL) vfs_freeopts(optlist); return (error); } /* * Old mount API. */ #ifndef _SYS_SYSPROTO_H_ struct mount_args { char *type; char *path; int flags; caddr_t data; }; #endif /* ARGSUSED */ int sys_mount(struct thread *td, struct mount_args *uap) { char *fstype; struct vfsconf *vfsp = NULL; struct mntarg *ma = NULL; uint64_t flags; int error; /* * Mount flags are now 64-bits. On 32-bit architectures only * 32-bits are passed in, but from here on everything handles * 64-bit flags correctly. */ flags = uap->flags; AUDIT_ARG_FFLAGS(flags); /* * Filter out MNT_ROOTFS. We do not want clients of mount() in * userspace to set this flag, but we must filter it out if we want * MNT_UPDATE on the root file system to work. * MNT_ROOTFS should only be set by the kernel when mounting its * root file system. */ flags &= ~MNT_ROOTFS; fstype = malloc(MFSNAMELEN, M_TEMP, M_WAITOK); error = copyinstr(uap->type, fstype, MFSNAMELEN, NULL); if (error) { free(fstype, M_TEMP); return (error); } AUDIT_ARG_TEXT(fstype); vfsp = vfs_byname_kld(fstype, td, &error); free(fstype, M_TEMP); if (vfsp == NULL) return (ENOENT); if (((vfsp->vfc_flags & VFCF_SBDRY) != 0 && vfsp->vfc_vfsops_sd->vfs_cmount == NULL) || ((vfsp->vfc_flags & VFCF_SBDRY) == 0 && vfsp->vfc_vfsops->vfs_cmount == NULL)) return (EOPNOTSUPP); ma = mount_argsu(ma, "fstype", uap->type, MFSNAMELEN); ma = mount_argsu(ma, "fspath", uap->path, MNAMELEN); ma = mount_argb(ma, flags & MNT_RDONLY, "noro"); ma = mount_argb(ma, !(flags & MNT_NOSUID), "nosuid"); ma = mount_argb(ma, !(flags & MNT_NOEXEC), "noexec"); if ((vfsp->vfc_flags & VFCF_SBDRY) != 0) return (vfsp->vfc_vfsops_sd->vfs_cmount(ma, uap->data, flags)); return (vfsp->vfc_vfsops->vfs_cmount(ma, uap->data, flags)); } /* * vfs_domount_first(): first file system mount (not update) */ static int vfs_domount_first( struct thread *td, /* Calling thread. */ struct vfsconf *vfsp, /* File system type. */ char *fspath, /* Mount path. */ struct vnode *vp, /* Vnode to be covered. */ uint64_t fsflags, /* Flags common to all filesystems. */ struct vfsoptlist **optlist /* Options local to the filesystem. */ ) { struct vattr va; struct mount *mp; struct vnode *newdp, *rootvp; int error, error1; ASSERT_VOP_ELOCKED(vp, __func__); KASSERT((fsflags & MNT_UPDATE) == 0, ("MNT_UPDATE shouldn't be here")); if ((fsflags & MNT_EMPTYDIR) != 0) { error = vfs_emptydir(vp); if (error != 0) { vput(vp); return (error); } } /* * If the jail of the calling thread lacks permission for this type of * file system, deny immediately. */ if (jailed(td->td_ucred) && !prison_allow(td->td_ucred, vfsp->vfc_prison_flag)) { vput(vp); return (EPERM); } /* * If the user is not root, ensure that they own the directory * onto which we are attempting to mount. */ error = VOP_GETATTR(vp, &va, td->td_ucred); if (error == 0 && va.va_uid != td->td_ucred->cr_uid) error = priv_check_cred(td->td_ucred, PRIV_VFS_ADMIN); if (error == 0) error = vinvalbuf(vp, V_SAVE, 0, 0); if (error == 0 && vp->v_type != VDIR) error = ENOTDIR; if (error == 0) { VI_LOCK(vp); if ((vp->v_iflag & VI_MOUNT) == 0 && vp->v_mountedhere == NULL) vp->v_iflag |= VI_MOUNT; else error = EBUSY; VI_UNLOCK(vp); } if (error != 0) { vput(vp); return (error); } vn_seqc_write_begin(vp); VOP_UNLOCK(vp); /* Allocate and initialize the filesystem. */ mp = vfs_mount_alloc(vp, vfsp, fspath, td->td_ucred); /* XXXMAC: pass to vfs_mount_alloc? */ mp->mnt_optnew = *optlist; /* Set the mount level flags. */ mp->mnt_flag = (fsflags & (MNT_UPDATEMASK | MNT_ROOTFS | MNT_RDONLY)); /* * Mount the filesystem. * XXX The final recipients of VFS_MOUNT just overwrite the ndp they * get. No freeing of cn_pnbuf. */ error1 = 0; if ((error = VFS_MOUNT(mp)) != 0 || (error1 = VFS_STATFS(mp, &mp->mnt_stat)) != 0 || (error1 = VFS_ROOT(mp, LK_EXCLUSIVE, &newdp)) != 0) { rootvp = NULL; if (error1 != 0) { error = error1; rootvp = vfs_cache_root_clear(mp); if (rootvp != NULL) { vhold(rootvp); vrele(rootvp); } if ((error1 = VFS_UNMOUNT(mp, 0)) != 0) printf("VFS_UNMOUNT returned %d\n", error1); } vfs_unbusy(mp); mp->mnt_vnodecovered = NULL; vfs_mount_destroy(mp); VI_LOCK(vp); vp->v_iflag &= ~VI_MOUNT; VI_UNLOCK(vp); if (rootvp != NULL) { vn_seqc_write_end(rootvp); vdrop(rootvp); } vn_seqc_write_end(vp); vrele(vp); return (error); } vn_seqc_write_begin(newdp); VOP_UNLOCK(newdp); if (mp->mnt_opt != NULL) vfs_freeopts(mp->mnt_opt); mp->mnt_opt = mp->mnt_optnew; *optlist = NULL; /* * Prevent external consumers of mount options from reading mnt_optnew. */ mp->mnt_optnew = NULL; MNT_ILOCK(mp); if ((mp->mnt_flag & MNT_ASYNC) != 0 && (mp->mnt_kern_flag & MNTK_NOASYNC) == 0) mp->mnt_kern_flag |= MNTK_ASYNC; else mp->mnt_kern_flag &= ~MNTK_ASYNC; MNT_IUNLOCK(mp); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); cache_purge(vp); VI_LOCK(vp); vp->v_iflag &= ~VI_MOUNT; VI_UNLOCK(vp); vp->v_mountedhere = mp; /* Place the new filesystem at the end of the mount list. */ mtx_lock(&mountlist_mtx); TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list); mtx_unlock(&mountlist_mtx); vfs_event_signal(NULL, VQ_MOUNT, 0); vn_lock(newdp, LK_EXCLUSIVE | LK_RETRY); VOP_UNLOCK(vp); EVENTHANDLER_DIRECT_INVOKE(vfs_mounted, mp, newdp, td); VOP_UNLOCK(newdp); mount_devctl_event("MOUNT", mp, false); mountcheckdirs(vp, newdp); vn_seqc_write_end(vp); vn_seqc_write_end(newdp); vrele(newdp); if ((mp->mnt_flag & MNT_RDONLY) == 0) vfs_allocate_syncvnode(mp); vfs_op_exit(mp); vfs_unbusy(mp); return (0); } /* * vfs_domount_update(): update of mounted file system */ static int vfs_domount_update( struct thread *td, /* Calling thread. */ struct vnode *vp, /* Mount point vnode. */ uint64_t fsflags, /* Flags common to all filesystems. */ struct vfsoptlist **optlist /* Options local to the filesystem. */ ) { struct export_args export; struct o2export_args o2export; struct vnode *rootvp; void *bufp; struct mount *mp; int error, export_error, i, len; uint64_t flag; gid_t *grps; ASSERT_VOP_ELOCKED(vp, __func__); KASSERT((fsflags & MNT_UPDATE) != 0, ("MNT_UPDATE should be here")); mp = vp->v_mount; if ((vp->v_vflag & VV_ROOT) == 0) { if (vfs_copyopt(*optlist, "export", &export, sizeof(export)) == 0) error = EXDEV; else error = EINVAL; vput(vp); return (error); } /* * We only allow the filesystem to be reloaded if it * is currently mounted read-only. */ flag = mp->mnt_flag; if ((fsflags & MNT_RELOAD) != 0 && (flag & MNT_RDONLY) == 0) { vput(vp); return (EOPNOTSUPP); /* Needs translation */ } /* * Only privileged root, or (if MNT_USER is set) the user that * did the original mount is permitted to update it. */ error = vfs_suser(mp, td); if (error != 0) { vput(vp); return (error); } if (vfs_busy(mp, MBF_NOWAIT)) { vput(vp); return (EBUSY); } VI_LOCK(vp); if ((vp->v_iflag & VI_MOUNT) != 0 || vp->v_mountedhere != NULL) { VI_UNLOCK(vp); vfs_unbusy(mp); vput(vp); return (EBUSY); } vp->v_iflag |= VI_MOUNT; VI_UNLOCK(vp); VOP_UNLOCK(vp); vfs_op_enter(mp); vn_seqc_write_begin(vp); rootvp = NULL; MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { MNT_IUNLOCK(mp); error = EBUSY; goto end; } mp->mnt_flag &= ~MNT_UPDATEMASK; mp->mnt_flag |= fsflags & (MNT_RELOAD | MNT_FORCE | MNT_UPDATE | MNT_SNAPSHOT | MNT_ROOTFS | MNT_UPDATEMASK | MNT_RDONLY); if ((mp->mnt_flag & MNT_ASYNC) == 0) mp->mnt_kern_flag &= ~MNTK_ASYNC; rootvp = vfs_cache_root_clear(mp); MNT_IUNLOCK(mp); mp->mnt_optnew = *optlist; vfs_mergeopts(mp->mnt_optnew, mp->mnt_opt); /* * Mount the filesystem. * XXX The final recipients of VFS_MOUNT just overwrite the ndp they * get. No freeing of cn_pnbuf. */ error = VFS_MOUNT(mp); export_error = 0; /* Process the export option. */ if (error == 0 && vfs_getopt(mp->mnt_optnew, "export", &bufp, &len) == 0) { /* Assume that there is only 1 ABI for each length. */ switch (len) { case (sizeof(struct oexport_args)): bzero(&o2export, sizeof(o2export)); /* FALLTHROUGH */ case (sizeof(o2export)): bcopy(bufp, &o2export, len); export.ex_flags = (uint64_t)o2export.ex_flags; export.ex_root = o2export.ex_root; export.ex_uid = o2export.ex_anon.cr_uid; export.ex_groups = NULL; export.ex_ngroups = o2export.ex_anon.cr_ngroups; if (export.ex_ngroups > 0) { if (export.ex_ngroups <= XU_NGROUPS) { export.ex_groups = malloc( export.ex_ngroups * sizeof(gid_t), M_TEMP, M_WAITOK); for (i = 0; i < export.ex_ngroups; i++) export.ex_groups[i] = o2export.ex_anon.cr_groups[i]; } else export_error = EINVAL; } else if (export.ex_ngroups < 0) export_error = EINVAL; export.ex_addr = o2export.ex_addr; export.ex_addrlen = o2export.ex_addrlen; export.ex_mask = o2export.ex_mask; export.ex_masklen = o2export.ex_masklen; export.ex_indexfile = o2export.ex_indexfile; export.ex_numsecflavors = o2export.ex_numsecflavors; if (export.ex_numsecflavors < MAXSECFLAVORS) { for (i = 0; i < export.ex_numsecflavors; i++) export.ex_secflavors[i] = o2export.ex_secflavors[i]; } else export_error = EINVAL; if (export_error == 0) export_error = vfs_export(mp, &export); free(export.ex_groups, M_TEMP); break; case (sizeof(export)): bcopy(bufp, &export, len); grps = NULL; if (export.ex_ngroups > 0) { if (export.ex_ngroups <= NGROUPS_MAX) { grps = malloc(export.ex_ngroups * sizeof(gid_t), M_TEMP, M_WAITOK); export_error = copyin(export.ex_groups, grps, export.ex_ngroups * sizeof(gid_t)); if (export_error == 0) export.ex_groups = grps; } else export_error = EINVAL; } else if (export.ex_ngroups == 0) export.ex_groups = NULL; else export_error = EINVAL; if (export_error == 0) export_error = vfs_export(mp, &export); free(grps, M_TEMP); break; default: export_error = EINVAL; break; } } MNT_ILOCK(mp); if (error == 0) { mp->mnt_flag &= ~(MNT_UPDATE | MNT_RELOAD | MNT_FORCE | MNT_SNAPSHOT); } else { /* * If we fail, restore old mount flags. MNT_QUOTA is special, * because it is not part of MNT_UPDATEMASK, but it could have * changed in the meantime if quotactl(2) was called. * All in all we want current value of MNT_QUOTA, not the old * one. */ mp->mnt_flag = (mp->mnt_flag & MNT_QUOTA) | (flag & ~MNT_QUOTA); } if ((mp->mnt_flag & MNT_ASYNC) != 0 && (mp->mnt_kern_flag & MNTK_NOASYNC) == 0) mp->mnt_kern_flag |= MNTK_ASYNC; else mp->mnt_kern_flag &= ~MNTK_ASYNC; MNT_IUNLOCK(mp); if (error != 0) goto end; mount_devctl_event("REMOUNT", mp, true); if (mp->mnt_opt != NULL) vfs_freeopts(mp->mnt_opt); mp->mnt_opt = mp->mnt_optnew; *optlist = NULL; (void)VFS_STATFS(mp, &mp->mnt_stat); /* * Prevent external consumers of mount options from reading * mnt_optnew. */ mp->mnt_optnew = NULL; if ((mp->mnt_flag & MNT_RDONLY) == 0) vfs_allocate_syncvnode(mp); else vfs_deallocate_syncvnode(mp); end: vfs_op_exit(mp); if (rootvp != NULL) { vn_seqc_write_end(rootvp); vrele(rootvp); } vn_seqc_write_end(vp); vfs_unbusy(mp); VI_LOCK(vp); vp->v_iflag &= ~VI_MOUNT; VI_UNLOCK(vp); vrele(vp); return (error != 0 ? error : export_error); } /* * vfs_domount(): actually attempt a filesystem mount. */ static int vfs_domount( struct thread *td, /* Calling thread. */ const char *fstype, /* Filesystem type. */ char *fspath, /* Mount path. */ uint64_t fsflags, /* Flags common to all filesystems. */ struct vfsoptlist **optlist /* Options local to the filesystem. */ ) { struct vfsconf *vfsp; struct nameidata nd; struct vnode *vp; char *pathbuf; int error; /* * Be ultra-paranoid about making sure the type and fspath * variables will fit in our mp buffers, including the * terminating NUL. */ if (strlen(fstype) >= MFSNAMELEN || strlen(fspath) >= MNAMELEN) return (ENAMETOOLONG); if (jailed(td->td_ucred) || usermount == 0) { if ((error = priv_check(td, PRIV_VFS_MOUNT)) != 0) return (error); } /* * Do not allow NFS export or MNT_SUIDDIR by unprivileged users. */ if (fsflags & MNT_EXPORTED) { error = priv_check(td, PRIV_VFS_MOUNT_EXPORTED); if (error) return (error); } if (fsflags & MNT_SUIDDIR) { error = priv_check(td, PRIV_VFS_MOUNT_SUIDDIR); if (error) return (error); } /* * Silently enforce MNT_NOSUID and MNT_USER for unprivileged users. */ if ((fsflags & (MNT_NOSUID | MNT_USER)) != (MNT_NOSUID | MNT_USER)) { if (priv_check(td, PRIV_VFS_MOUNT_NONUSER) != 0) fsflags |= MNT_NOSUID | MNT_USER; } /* Load KLDs before we lock the covered vnode to avoid reversals. */ vfsp = NULL; if ((fsflags & MNT_UPDATE) == 0) { /* Don't try to load KLDs if we're mounting the root. */ if (fsflags & MNT_ROOTFS) vfsp = vfs_byname(fstype); else vfsp = vfs_byname_kld(fstype, td, &error); if (vfsp == NULL) return (ENODEV); } /* * Get vnode to be covered or mount point's vnode in case of MNT_UPDATE. */ NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, fspath, td); error = namei(&nd); if (error != 0) return (error); NDFREE(&nd, NDF_ONLY_PNBUF); vp = nd.ni_vp; if ((fsflags & MNT_UPDATE) == 0) { if ((vp->v_vflag & VV_ROOT) != 0 && (fsflags & MNT_NOCOVER) != 0) { vput(vp); return (EBUSY); } pathbuf = malloc(MNAMELEN, M_TEMP, M_WAITOK); strcpy(pathbuf, fspath); error = vn_path_to_global_path(td, vp, pathbuf, MNAMELEN); if (error == 0) { error = vfs_domount_first(td, vfsp, pathbuf, vp, fsflags, optlist); } free(pathbuf, M_TEMP); } else error = vfs_domount_update(td, vp, fsflags, optlist); return (error); } /* * Unmount a filesystem. * * Note: unmount takes a path to the vnode mounted on as argument, not * special file (as before). */ #ifndef _SYS_SYSPROTO_H_ struct unmount_args { char *path; int flags; }; #endif /* ARGSUSED */ int sys_unmount(struct thread *td, struct unmount_args *uap) { return (kern_unmount(td, uap->path, uap->flags)); } int kern_unmount(struct thread *td, const char *path, int flags) { struct nameidata nd; struct mount *mp; char *pathbuf; int error, id0, id1; AUDIT_ARG_VALUE(flags); if (jailed(td->td_ucred) || usermount == 0) { error = priv_check(td, PRIV_VFS_UNMOUNT); if (error) return (error); } pathbuf = malloc(MNAMELEN, M_TEMP, M_WAITOK); error = copyinstr(path, pathbuf, MNAMELEN, NULL); if (error) { free(pathbuf, M_TEMP); return (error); } if (flags & MNT_BYFSID) { AUDIT_ARG_TEXT(pathbuf); /* Decode the filesystem ID. */ if (sscanf(pathbuf, "FSID:%d:%d", &id0, &id1) != 2) { free(pathbuf, M_TEMP); return (EINVAL); } mtx_lock(&mountlist_mtx); TAILQ_FOREACH_REVERSE(mp, &mountlist, mntlist, mnt_list) { if (mp->mnt_stat.f_fsid.val[0] == id0 && mp->mnt_stat.f_fsid.val[1] == id1) { vfs_ref(mp); break; } } mtx_unlock(&mountlist_mtx); } else { /* * Try to find global path for path argument. */ NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, pathbuf, td); if (namei(&nd) == 0) { NDFREE(&nd, NDF_ONLY_PNBUF); error = vn_path_to_global_path(td, nd.ni_vp, pathbuf, MNAMELEN); if (error == 0) vput(nd.ni_vp); } mtx_lock(&mountlist_mtx); TAILQ_FOREACH_REVERSE(mp, &mountlist, mntlist, mnt_list) { if (strcmp(mp->mnt_stat.f_mntonname, pathbuf) == 0) { vfs_ref(mp); break; } } mtx_unlock(&mountlist_mtx); } free(pathbuf, M_TEMP); if (mp == NULL) { /* * Previously we returned ENOENT for a nonexistent path and * EINVAL for a non-mountpoint. We cannot tell these apart * now, so in the !MNT_BYFSID case return the more likely * EINVAL for compatibility. */ return ((flags & MNT_BYFSID) ? ENOENT : EINVAL); } /* * Don't allow unmounting the root filesystem. */ if (mp->mnt_flag & MNT_ROOTFS) { vfs_rel(mp); return (EINVAL); } error = dounmount(mp, flags, td); return (error); } /* * Return error if any of the vnodes, ignoring the root vnode * and the syncer vnode, have non-zero usecount. * * This function is purely advisory - it can return false positives * and negatives. */ static int vfs_check_usecounts(struct mount *mp) { struct vnode *vp, *mvp; MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { if ((vp->v_vflag & VV_ROOT) == 0 && vp->v_type != VNON && vp->v_usecount != 0) { VI_UNLOCK(vp); MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); return (EBUSY); } VI_UNLOCK(vp); } return (0); } static void dounmount_cleanup(struct mount *mp, struct vnode *coveredvp, int mntkflags) { mtx_assert(MNT_MTX(mp), MA_OWNED); mp->mnt_kern_flag &= ~mntkflags; if ((mp->mnt_kern_flag & MNTK_MWAIT) != 0) { mp->mnt_kern_flag &= ~MNTK_MWAIT; wakeup(mp); } vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); if (coveredvp != NULL) { VOP_UNLOCK(coveredvp); vdrop(coveredvp); } vn_finished_write(mp); } /* * There are various reference counters associated with the mount point. * Normally it is permitted to modify them without taking the mnt ilock, * but this behavior can be temporarily disabled if stable value is needed * or callers are expected to block (e.g. to not allow new users during * forced unmount). */ void vfs_op_enter(struct mount *mp) { + struct mount_pcpu *mpcpu; int cpu; MNT_ILOCK(mp); mp->mnt_vfs_ops++; if (mp->mnt_vfs_ops > 1) { MNT_IUNLOCK(mp); return; } vfs_op_barrier_wait(mp); CPU_FOREACH(cpu) { - mp->mnt_ref += - zpcpu_replace_cpu(mp->mnt_ref_pcpu, 0, cpu); - mp->mnt_lockref += - zpcpu_replace_cpu(mp->mnt_lockref_pcpu, 0, cpu); - mp->mnt_writeopcount += - zpcpu_replace_cpu(mp->mnt_writeopcount_pcpu, 0, cpu); + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + + mp->mnt_ref += mpcpu->mntp_ref; + mpcpu->mntp_ref = 0; + + mp->mnt_lockref += mpcpu->mntp_lockref; + mpcpu->mntp_lockref = 0; + + mp->mnt_writeopcount += mpcpu->mntp_writeopcount; + mpcpu->mntp_writeopcount = 0; } if (mp->mnt_ref <= 0 || mp->mnt_lockref < 0 || mp->mnt_writeopcount < 0) panic("%s: invalid count(s) on mp %p: ref %d lockref %d writeopcount %d\n", __func__, mp, mp->mnt_ref, mp->mnt_lockref, mp->mnt_writeopcount); MNT_IUNLOCK(mp); vfs_assert_mount_counters(mp); } void vfs_op_exit_locked(struct mount *mp) { mtx_assert(MNT_MTX(mp), MA_OWNED); if (mp->mnt_vfs_ops <= 0) panic("%s: invalid vfs_ops count %d for mp %p\n", __func__, mp->mnt_vfs_ops, mp); mp->mnt_vfs_ops--; } void vfs_op_exit(struct mount *mp) { MNT_ILOCK(mp); vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); } struct vfs_op_barrier_ipi { struct mount *mp; struct smp_rendezvous_cpus_retry_arg srcra; }; static void vfs_op_action_func(void *arg) { struct vfs_op_barrier_ipi *vfsopipi; struct mount *mp; vfsopipi = __containerof(arg, struct vfs_op_barrier_ipi, srcra); mp = vfsopipi->mp; if (!vfs_op_thread_entered(mp)) smp_rendezvous_cpus_done(arg); } static void vfs_op_wait_func(void *arg, int cpu) { struct vfs_op_barrier_ipi *vfsopipi; struct mount *mp; - int *in_op; + struct mount_pcpu *mpcpu; vfsopipi = __containerof(arg, struct vfs_op_barrier_ipi, srcra); mp = vfsopipi->mp; - in_op = zpcpu_get_cpu(mp->mnt_thread_in_ops_pcpu, cpu); - while (atomic_load_int(in_op)) + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + while (atomic_load_int(&mpcpu->mntp_thread_in_ops)) cpu_spinwait(); } void vfs_op_barrier_wait(struct mount *mp) { struct vfs_op_barrier_ipi vfsopipi; vfsopipi.mp = mp; smp_rendezvous_cpus_retry(all_cpus, smp_no_rendezvous_barrier, vfs_op_action_func, smp_no_rendezvous_barrier, vfs_op_wait_func, &vfsopipi.srcra); } #ifdef DIAGNOSTIC void vfs_assert_mount_counters(struct mount *mp) { + struct mount_pcpu *mpcpu; int cpu; if (mp->mnt_vfs_ops == 0) return; CPU_FOREACH(cpu) { - if (*zpcpu_get_cpu(mp->mnt_ref_pcpu, cpu) != 0 || - *zpcpu_get_cpu(mp->mnt_lockref_pcpu, cpu) != 0 || - *zpcpu_get_cpu(mp->mnt_writeopcount_pcpu, cpu) != 0) + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + if (mpcpu->mntp_ref != 0 || + mpcpu->mntp_lockref != 0 || + mpcpu->mntp_writeopcount != 0) vfs_dump_mount_counters(mp); } } void vfs_dump_mount_counters(struct mount *mp) { - int cpu, *count; + struct mount_pcpu *mpcpu; int ref, lockref, writeopcount; + int cpu; printf("%s: mp %p vfs_ops %d\n", __func__, mp, mp->mnt_vfs_ops); printf(" ref : "); ref = mp->mnt_ref; CPU_FOREACH(cpu) { - count = zpcpu_get_cpu(mp->mnt_ref_pcpu, cpu); - printf("%d ", *count); - ref += *count; + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + printf("%d ", mpcpu->mntp_ref); + ref += mpcpu->mntp_ref; } printf("\n"); printf(" lockref : "); lockref = mp->mnt_lockref; CPU_FOREACH(cpu) { - count = zpcpu_get_cpu(mp->mnt_lockref_pcpu, cpu); - printf("%d ", *count); - lockref += *count; + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + printf("%d ", mpcpu->mntp_lockref); + lockref += mpcpu->mntp_lockref; } printf("\n"); printf("writeopcount: "); writeopcount = mp->mnt_writeopcount; CPU_FOREACH(cpu) { - count = zpcpu_get_cpu(mp->mnt_writeopcount_pcpu, cpu); - printf("%d ", *count); - writeopcount += *count; + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + printf("%d ", mpcpu->mntp_writeopcount); + writeopcount += mpcpu->mntp_writeopcount; } printf("\n"); printf("counter struct total\n"); printf("ref %-5d %-5d\n", mp->mnt_ref, ref); printf("lockref %-5d %-5d\n", mp->mnt_lockref, lockref); printf("writeopcount %-5d %-5d\n", mp->mnt_writeopcount, writeopcount); panic("invalid counts on struct mount"); } #endif int vfs_mount_fetch_counter(struct mount *mp, enum mount_counter which) { - int *base, *pcpu; + struct mount_pcpu *mpcpu; int cpu, sum; switch (which) { case MNT_COUNT_REF: - base = &mp->mnt_ref; - pcpu = mp->mnt_ref_pcpu; + sum = mp->mnt_ref; break; case MNT_COUNT_LOCKREF: - base = &mp->mnt_lockref; - pcpu = mp->mnt_lockref_pcpu; + sum = mp->mnt_lockref; break; case MNT_COUNT_WRITEOPCOUNT: - base = &mp->mnt_writeopcount; - pcpu = mp->mnt_writeopcount_pcpu; + sum = mp->mnt_writeopcount; break; } - sum = *base; CPU_FOREACH(cpu) { - sum += *zpcpu_get_cpu(pcpu, cpu); + mpcpu = vfs_mount_pcpu_remote(mp, cpu); + switch (which) { + case MNT_COUNT_REF: + sum += mpcpu->mntp_ref; + break; + case MNT_COUNT_LOCKREF: + sum += mpcpu->mntp_lockref; + break; + case MNT_COUNT_WRITEOPCOUNT: + sum += mpcpu->mntp_writeopcount; + break; + } } return (sum); } /* * Do the actual filesystem unmount. */ int dounmount(struct mount *mp, int flags, struct thread *td) { struct vnode *coveredvp, *rootvp; int error; uint64_t async_flag; int mnt_gen_r; if ((coveredvp = mp->mnt_vnodecovered) != NULL) { mnt_gen_r = mp->mnt_gen; VI_LOCK(coveredvp); vholdl(coveredvp); vn_lock(coveredvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_RETRY); /* * Check for mp being unmounted while waiting for the * covered vnode lock. */ if (coveredvp->v_mountedhere != mp || coveredvp->v_mountedhere->mnt_gen != mnt_gen_r) { VOP_UNLOCK(coveredvp); vdrop(coveredvp); vfs_rel(mp); return (EBUSY); } } /* * Only privileged root, or (if MNT_USER is set) the user that did the * original mount is permitted to unmount this filesystem. */ error = vfs_suser(mp, td); if (error != 0) { if (coveredvp != NULL) { VOP_UNLOCK(coveredvp); vdrop(coveredvp); } vfs_rel(mp); return (error); } vfs_op_enter(mp); vn_start_write(NULL, &mp, V_WAIT | V_MNTREF); MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 || (mp->mnt_flag & MNT_UPDATE) != 0 || !TAILQ_EMPTY(&mp->mnt_uppers)) { dounmount_cleanup(mp, coveredvp, 0); return (EBUSY); } mp->mnt_kern_flag |= MNTK_UNMOUNT; rootvp = vfs_cache_root_clear(mp); if (coveredvp != NULL) vn_seqc_write_begin(coveredvp); if (flags & MNT_NONBUSY) { MNT_IUNLOCK(mp); error = vfs_check_usecounts(mp); MNT_ILOCK(mp); if (error != 0) { vn_seqc_write_end(coveredvp); dounmount_cleanup(mp, coveredvp, MNTK_UNMOUNT); if (rootvp != NULL) { vn_seqc_write_end(rootvp); vrele(rootvp); } return (error); } } /* Allow filesystems to detect that a forced unmount is in progress. */ if (flags & MNT_FORCE) { mp->mnt_kern_flag |= MNTK_UNMOUNTF; MNT_IUNLOCK(mp); /* * Must be done after setting MNTK_UNMOUNTF and before * waiting for mnt_lockref to become 0. */ VFS_PURGE(mp); MNT_ILOCK(mp); } error = 0; if (mp->mnt_lockref) { mp->mnt_kern_flag |= MNTK_DRAINING; error = msleep(&mp->mnt_lockref, MNT_MTX(mp), PVFS, "mount drain", 0); } MNT_IUNLOCK(mp); KASSERT(mp->mnt_lockref == 0, ("%s: invalid lock refcount in the drain path @ %s:%d", __func__, __FILE__, __LINE__)); KASSERT(error == 0, ("%s: invalid return value for msleep in the drain path @ %s:%d", __func__, __FILE__, __LINE__)); /* * We want to keep the vnode around so that we can vn_seqc_write_end * after we are done with unmount. Downgrade our reference to a mere * hold count so that we don't interefere with anything. */ if (rootvp != NULL) { vhold(rootvp); vrele(rootvp); } if (mp->mnt_flag & MNT_EXPUBLIC) vfs_setpublicfs(NULL, NULL, NULL); vfs_periodic(mp, MNT_WAIT); MNT_ILOCK(mp); async_flag = mp->mnt_flag & MNT_ASYNC; mp->mnt_flag &= ~MNT_ASYNC; mp->mnt_kern_flag &= ~MNTK_ASYNC; MNT_IUNLOCK(mp); vfs_deallocate_syncvnode(mp); error = VFS_UNMOUNT(mp, flags); vn_finished_write(mp); /* * If we failed to flush the dirty blocks for this mount point, * undo all the cdir/rdir and rootvnode changes we made above. * Unless we failed to do so because the device is reporting that * it doesn't exist anymore. */ if (error && error != ENXIO) { MNT_ILOCK(mp); if ((mp->mnt_flag & MNT_RDONLY) == 0) { MNT_IUNLOCK(mp); vfs_allocate_syncvnode(mp); MNT_ILOCK(mp); } mp->mnt_kern_flag &= ~(MNTK_UNMOUNT | MNTK_UNMOUNTF); mp->mnt_flag |= async_flag; if ((mp->mnt_flag & MNT_ASYNC) != 0 && (mp->mnt_kern_flag & MNTK_NOASYNC) == 0) mp->mnt_kern_flag |= MNTK_ASYNC; if (mp->mnt_kern_flag & MNTK_MWAIT) { mp->mnt_kern_flag &= ~MNTK_MWAIT; wakeup(mp); } vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); if (coveredvp) { vn_seqc_write_end(coveredvp); VOP_UNLOCK(coveredvp); vdrop(coveredvp); } if (rootvp != NULL) { vn_seqc_write_end(rootvp); vdrop(rootvp); } return (error); } mtx_lock(&mountlist_mtx); TAILQ_REMOVE(&mountlist, mp, mnt_list); mtx_unlock(&mountlist_mtx); EVENTHANDLER_DIRECT_INVOKE(vfs_unmounted, mp, td); if (coveredvp != NULL) { coveredvp->v_mountedhere = NULL; vn_seqc_write_end(coveredvp); VOP_UNLOCK(coveredvp); vdrop(coveredvp); } mount_devctl_event("UNMOUNT", mp, false); if (rootvp != NULL) { vn_seqc_write_end(rootvp); vdrop(rootvp); } vfs_event_signal(NULL, VQ_UNMOUNT, 0); if (rootvnode != NULL && mp == rootvnode->v_mount) { vrele(rootvnode); rootvnode = NULL; } if (mp == rootdevmp) rootdevmp = NULL; vfs_mount_destroy(mp); return (0); } /* * Report errors during filesystem mounting. */ void vfs_mount_error(struct mount *mp, const char *fmt, ...) { struct vfsoptlist *moptlist = mp->mnt_optnew; va_list ap; int error, len; char *errmsg; error = vfs_getopt(moptlist, "errmsg", (void **)&errmsg, &len); if (error || errmsg == NULL || len <= 0) return; va_start(ap, fmt); vsnprintf(errmsg, (size_t)len, fmt, ap); va_end(ap); } void vfs_opterror(struct vfsoptlist *opts, const char *fmt, ...) { va_list ap; int error, len; char *errmsg; error = vfs_getopt(opts, "errmsg", (void **)&errmsg, &len); if (error || errmsg == NULL || len <= 0) return; va_start(ap, fmt); vsnprintf(errmsg, (size_t)len, fmt, ap); va_end(ap); } /* * --------------------------------------------------------------------- * Functions for querying mount options/arguments from filesystems. */ /* * Check that no unknown options are given */ int vfs_filteropt(struct vfsoptlist *opts, const char **legal) { struct vfsopt *opt; char errmsg[255]; const char **t, *p, *q; int ret = 0; TAILQ_FOREACH(opt, opts, link) { p = opt->name; q = NULL; if (p[0] == 'n' && p[1] == 'o') q = p + 2; for(t = global_opts; *t != NULL; t++) { if (strcmp(*t, p) == 0) break; if (q != NULL) { if (strcmp(*t, q) == 0) break; } } if (*t != NULL) continue; for(t = legal; *t != NULL; t++) { if (strcmp(*t, p) == 0) break; if (q != NULL) { if (strcmp(*t, q) == 0) break; } } if (*t != NULL) continue; snprintf(errmsg, sizeof(errmsg), "mount option <%s> is unknown", p); ret = EINVAL; } if (ret != 0) { TAILQ_FOREACH(opt, opts, link) { if (strcmp(opt->name, "errmsg") == 0) { strncpy((char *)opt->value, errmsg, opt->len); break; } } if (opt == NULL) printf("%s\n", errmsg); } return (ret); } /* * Get a mount option by its name. * * Return 0 if the option was found, ENOENT otherwise. * If len is non-NULL it will be filled with the length * of the option. If buf is non-NULL, it will be filled * with the address of the option. */ int vfs_getopt(struct vfsoptlist *opts, const char *name, void **buf, int *len) { struct vfsopt *opt; KASSERT(opts != NULL, ("vfs_getopt: caller passed 'opts' as NULL")); TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) == 0) { opt->seen = 1; if (len != NULL) *len = opt->len; if (buf != NULL) *buf = opt->value; return (0); } } return (ENOENT); } int vfs_getopt_pos(struct vfsoptlist *opts, const char *name) { struct vfsopt *opt; if (opts == NULL) return (-1); TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) == 0) { opt->seen = 1; return (opt->pos); } } return (-1); } int vfs_getopt_size(struct vfsoptlist *opts, const char *name, off_t *value) { char *opt_value, *vtp; quad_t iv; int error, opt_len; error = vfs_getopt(opts, name, (void **)&opt_value, &opt_len); if (error != 0) return (error); if (opt_len == 0 || opt_value == NULL) return (EINVAL); if (opt_value[0] == '\0' || opt_value[opt_len - 1] != '\0') return (EINVAL); iv = strtoq(opt_value, &vtp, 0); if (vtp == opt_value || (vtp[0] != '\0' && vtp[1] != '\0')) return (EINVAL); if (iv < 0) return (EINVAL); switch (vtp[0]) { case 't': case 'T': iv *= 1024; /* FALLTHROUGH */ case 'g': case 'G': iv *= 1024; /* FALLTHROUGH */ case 'm': case 'M': iv *= 1024; /* FALLTHROUGH */ case 'k': case 'K': iv *= 1024; case '\0': break; default: return (EINVAL); } *value = iv; return (0); } char * vfs_getopts(struct vfsoptlist *opts, const char *name, int *error) { struct vfsopt *opt; *error = 0; TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) != 0) continue; opt->seen = 1; if (opt->len == 0 || ((char *)opt->value)[opt->len - 1] != '\0') { *error = EINVAL; return (NULL); } return (opt->value); } *error = ENOENT; return (NULL); } int vfs_flagopt(struct vfsoptlist *opts, const char *name, uint64_t *w, uint64_t val) { struct vfsopt *opt; TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) == 0) { opt->seen = 1; if (w != NULL) *w |= val; return (1); } } if (w != NULL) *w &= ~val; return (0); } int vfs_scanopt(struct vfsoptlist *opts, const char *name, const char *fmt, ...) { va_list ap; struct vfsopt *opt; int ret; KASSERT(opts != NULL, ("vfs_getopt: caller passed 'opts' as NULL")); TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) != 0) continue; opt->seen = 1; if (opt->len == 0 || opt->value == NULL) return (0); if (((char *)opt->value)[opt->len - 1] != '\0') return (0); va_start(ap, fmt); ret = vsscanf(opt->value, fmt, ap); va_end(ap); return (ret); } return (0); } int vfs_setopt(struct vfsoptlist *opts, const char *name, void *value, int len) { struct vfsopt *opt; TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) != 0) continue; opt->seen = 1; if (opt->value == NULL) opt->len = len; else { if (opt->len != len) return (EINVAL); bcopy(value, opt->value, len); } return (0); } return (ENOENT); } int vfs_setopt_part(struct vfsoptlist *opts, const char *name, void *value, int len) { struct vfsopt *opt; TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) != 0) continue; opt->seen = 1; if (opt->value == NULL) opt->len = len; else { if (opt->len < len) return (EINVAL); opt->len = len; bcopy(value, opt->value, len); } return (0); } return (ENOENT); } int vfs_setopts(struct vfsoptlist *opts, const char *name, const char *value) { struct vfsopt *opt; TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) != 0) continue; opt->seen = 1; if (opt->value == NULL) opt->len = strlen(value) + 1; else if (strlcpy(opt->value, value, opt->len) >= opt->len) return (EINVAL); return (0); } return (ENOENT); } /* * Find and copy a mount option. * * The size of the buffer has to be specified * in len, if it is not the same length as the * mount option, EINVAL is returned. * Returns ENOENT if the option is not found. */ int vfs_copyopt(struct vfsoptlist *opts, const char *name, void *dest, int len) { struct vfsopt *opt; KASSERT(opts != NULL, ("vfs_copyopt: caller passed 'opts' as NULL")); TAILQ_FOREACH(opt, opts, link) { if (strcmp(name, opt->name) == 0) { opt->seen = 1; if (len != opt->len) return (EINVAL); bcopy(opt->value, dest, opt->len); return (0); } } return (ENOENT); } int __vfs_statfs(struct mount *mp, struct statfs *sbp) { /* * Filesystems only fill in part of the structure for updates, we * have to read the entirety first to get all content. */ if (sbp != &mp->mnt_stat) memcpy(sbp, &mp->mnt_stat, sizeof(*sbp)); /* * Set these in case the underlying filesystem fails to do so. */ sbp->f_version = STATFS_VERSION; sbp->f_namemax = NAME_MAX; sbp->f_flags = mp->mnt_flag & MNT_VISFLAGMASK; return (mp->mnt_op->vfs_statfs(mp, sbp)); } void vfs_mountedfrom(struct mount *mp, const char *from) { bzero(mp->mnt_stat.f_mntfromname, sizeof mp->mnt_stat.f_mntfromname); strlcpy(mp->mnt_stat.f_mntfromname, from, sizeof mp->mnt_stat.f_mntfromname); } /* * --------------------------------------------------------------------- * This is the api for building mount args and mounting filesystems from * inside the kernel. * * The API works by accumulation of individual args. First error is * latched. * * XXX: should be documented in new manpage kernel_mount(9) */ /* A memory allocation which must be freed when we are done */ struct mntaarg { SLIST_ENTRY(mntaarg) next; }; /* The header for the mount arguments */ struct mntarg { struct iovec *v; int len; int error; SLIST_HEAD(, mntaarg) list; }; /* * Add a boolean argument. * * flag is the boolean value. * name must start with "no". */ struct mntarg * mount_argb(struct mntarg *ma, int flag, const char *name) { KASSERT(name[0] == 'n' && name[1] == 'o', ("mount_argb(...,%s): name must start with 'no'", name)); return (mount_arg(ma, name + (flag ? 2 : 0), NULL, 0)); } /* * Add an argument printf style */ struct mntarg * mount_argf(struct mntarg *ma, const char *name, const char *fmt, ...) { va_list ap; struct mntaarg *maa; struct sbuf *sb; int len; if (ma == NULL) { ma = malloc(sizeof *ma, M_MOUNT, M_WAITOK | M_ZERO); SLIST_INIT(&ma->list); } if (ma->error) return (ma); ma->v = realloc(ma->v, sizeof *ma->v * (ma->len + 2), M_MOUNT, M_WAITOK); ma->v[ma->len].iov_base = (void *)(uintptr_t)name; ma->v[ma->len].iov_len = strlen(name) + 1; ma->len++; sb = sbuf_new_auto(); va_start(ap, fmt); sbuf_vprintf(sb, fmt, ap); va_end(ap); sbuf_finish(sb); len = sbuf_len(sb) + 1; maa = malloc(sizeof *maa + len, M_MOUNT, M_WAITOK | M_ZERO); SLIST_INSERT_HEAD(&ma->list, maa, next); bcopy(sbuf_data(sb), maa + 1, len); sbuf_delete(sb); ma->v[ma->len].iov_base = maa + 1; ma->v[ma->len].iov_len = len; ma->len++; return (ma); } /* * Add an argument which is a userland string. */ struct mntarg * mount_argsu(struct mntarg *ma, const char *name, const void *val, int len) { struct mntaarg *maa; char *tbuf; if (val == NULL) return (ma); if (ma == NULL) { ma = malloc(sizeof *ma, M_MOUNT, M_WAITOK | M_ZERO); SLIST_INIT(&ma->list); } if (ma->error) return (ma); maa = malloc(sizeof *maa + len, M_MOUNT, M_WAITOK | M_ZERO); SLIST_INSERT_HEAD(&ma->list, maa, next); tbuf = (void *)(maa + 1); ma->error = copyinstr(val, tbuf, len, NULL); return (mount_arg(ma, name, tbuf, -1)); } /* * Plain argument. * * If length is -1, treat value as a C string. */ struct mntarg * mount_arg(struct mntarg *ma, const char *name, const void *val, int len) { if (ma == NULL) { ma = malloc(sizeof *ma, M_MOUNT, M_WAITOK | M_ZERO); SLIST_INIT(&ma->list); } if (ma->error) return (ma); ma->v = realloc(ma->v, sizeof *ma->v * (ma->len + 2), M_MOUNT, M_WAITOK); ma->v[ma->len].iov_base = (void *)(uintptr_t)name; ma->v[ma->len].iov_len = strlen(name) + 1; ma->len++; ma->v[ma->len].iov_base = (void *)(uintptr_t)val; if (len < 0) ma->v[ma->len].iov_len = strlen(val) + 1; else ma->v[ma->len].iov_len = len; ma->len++; return (ma); } /* * Free a mntarg structure */ static void free_mntarg(struct mntarg *ma) { struct mntaarg *maa; while (!SLIST_EMPTY(&ma->list)) { maa = SLIST_FIRST(&ma->list); SLIST_REMOVE_HEAD(&ma->list, next); free(maa, M_MOUNT); } free(ma->v, M_MOUNT); free(ma, M_MOUNT); } /* * Mount a filesystem */ int kernel_mount(struct mntarg *ma, uint64_t flags) { struct uio auio; int error; KASSERT(ma != NULL, ("kernel_mount NULL ma")); KASSERT(ma->v != NULL, ("kernel_mount NULL ma->v")); KASSERT(!(ma->len & 1), ("kernel_mount odd ma->len (%d)", ma->len)); auio.uio_iov = ma->v; auio.uio_iovcnt = ma->len; auio.uio_segflg = UIO_SYSSPACE; error = ma->error; if (!error) error = vfs_donmount(curthread, flags, &auio); free_mntarg(ma); return (error); } /* * A printflike function to mount a filesystem. */ int kernel_vmount(int flags, ...) { struct mntarg *ma = NULL; va_list ap; const char *cp; const void *vp; int error; va_start(ap, flags); for (;;) { cp = va_arg(ap, const char *); if (cp == NULL) break; vp = va_arg(ap, const void *); ma = mount_arg(ma, cp, vp, (vp != NULL ? -1 : 0)); } va_end(ap); error = kernel_mount(ma, flags); return (error); } /* Map from mount options to printable formats. */ static struct mntoptnames optnames[] = { MNTOPT_NAMES }; static void mount_devctl_event_mntopt(struct sbuf *sb, const char *what, struct vfsoptlist *opts) { struct vfsopt *opt; if (opts == NULL || TAILQ_EMPTY(opts)) return; sbuf_printf(sb, " %s=\"", what); TAILQ_FOREACH(opt, opts, link) { if (opt->name[0] == '\0' || (opt->len > 0 && *(char *)opt->value == '\0')) continue; devctl_safe_quote_sb(sb, opt->name); if (opt->len > 0) { sbuf_putc(sb, '='); devctl_safe_quote_sb(sb, opt->value); } sbuf_putc(sb, ';'); } sbuf_putc(sb, '"'); } #define DEVCTL_LEN 1024 static void mount_devctl_event(const char *type, struct mount *mp, bool donew) { const uint8_t *cp; struct mntoptnames *fp; struct sbuf sb; struct statfs *sfp = &mp->mnt_stat; char *buf; buf = malloc(DEVCTL_LEN, M_MOUNT, M_NOWAIT); if (buf == NULL) return; sbuf_new(&sb, buf, DEVCTL_LEN, SBUF_FIXEDLEN); sbuf_cpy(&sb, "mount-point=\""); devctl_safe_quote_sb(&sb, sfp->f_mntonname); sbuf_cat(&sb, "\" mount-dev=\""); devctl_safe_quote_sb(&sb, sfp->f_mntfromname); sbuf_cat(&sb, "\" mount-type=\""); devctl_safe_quote_sb(&sb, sfp->f_fstypename); sbuf_cat(&sb, "\" fsid=0x"); cp = (const uint8_t *)&sfp->f_fsid.val[0]; for (int i = 0; i < sizeof(sfp->f_fsid); i++) sbuf_printf(&sb, "%02x", cp[i]); sbuf_printf(&sb, " owner=%u flags=\"", sfp->f_owner); for (fp = optnames; fp->o_opt != 0; fp++) { if ((mp->mnt_flag & fp->o_opt) != 0) { sbuf_cat(&sb, fp->o_name); sbuf_putc(&sb, ';'); } } sbuf_putc(&sb, '"'); mount_devctl_event_mntopt(&sb, "opt", mp->mnt_opt); if (donew) mount_devctl_event_mntopt(&sb, "optnew", mp->mnt_optnew); sbuf_finish(&sb); if (sbuf_error(&sb) == 0) devctl_notify("VFS", "FS", type, sbuf_data(&sb)); sbuf_delete(&sb); free(buf, M_MOUNT); } /* * Suspend write operations on all local writeable filesystems. Does * full sync of them in the process. * * Iterate over the mount points in reverse order, suspending most * recently mounted filesystems first. It handles a case where a * filesystem mounted from a md(4) vnode-backed device should be * suspended before the filesystem that owns the vnode. */ void suspend_all_fs(void) { struct mount *mp; int error; mtx_lock(&mountlist_mtx); TAILQ_FOREACH_REVERSE(mp, &mountlist, mntlist, mnt_list) { error = vfs_busy(mp, MBF_MNTLSTLOCK | MBF_NOWAIT); if (error != 0) continue; if ((mp->mnt_flag & (MNT_RDONLY | MNT_LOCAL)) != MNT_LOCAL || (mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { mtx_lock(&mountlist_mtx); vfs_unbusy(mp); continue; } error = vfs_write_suspend(mp, 0); if (error == 0) { MNT_ILOCK(mp); MPASS((mp->mnt_kern_flag & MNTK_SUSPEND_ALL) == 0); mp->mnt_kern_flag |= MNTK_SUSPEND_ALL; MNT_IUNLOCK(mp); mtx_lock(&mountlist_mtx); } else { printf("suspend of %s failed, error %d\n", mp->mnt_stat.f_mntonname, error); mtx_lock(&mountlist_mtx); vfs_unbusy(mp); } } mtx_unlock(&mountlist_mtx); } void resume_all_fs(void) { struct mount *mp; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if ((mp->mnt_kern_flag & MNTK_SUSPEND_ALL) == 0) continue; mtx_unlock(&mountlist_mtx); MNT_ILOCK(mp); MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) != 0); mp->mnt_kern_flag &= ~MNTK_SUSPEND_ALL; MNT_IUNLOCK(mp); vfs_write_resume(mp, 0); mtx_lock(&mountlist_mtx); vfs_unbusy(mp); } mtx_unlock(&mountlist_mtx); } Index: head/sys/kern/vfs_subr.c =================================================================== --- head/sys/kern/vfs_subr.c (revision 367534) +++ head/sys/kern/vfs_subr.c (revision 367535) @@ -1,6792 +1,6795 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 */ /* * External virtual filesystem routines */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_watchdog.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif static void delmntque(struct vnode *vp); static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo); static void syncer_shutdown(void *arg, int howto); static int vtryrecycle(struct vnode *vp); static void v_init_counters(struct vnode *); static void vgonel(struct vnode *); static bool vhold_recycle_free(struct vnode *); static void vfs_knllock(void *arg); static void vfs_knlunlock(void *arg); static void vfs_knl_assert_lock(void *arg, int what); static void destroy_vpollinfo(struct vpollinfo *vi); static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, daddr_t startlbn, daddr_t endlbn); static void vnlru_recalc(void); /* * These fences are intended for cases where some synchronization is * needed between access of v_iflags and lockless vnode refcount (v_holdcnt * and v_usecount) updates. Access to v_iflags is generally synchronized * by the interlock, but we have some internal assertions that check vnode * flags without acquiring the lock. Thus, these fences are INVARIANTS-only * for now. */ #ifdef INVARIANTS #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq() #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel() #else #define VNODE_REFCOUNT_FENCE_ACQ() #define VNODE_REFCOUNT_FENCE_REL() #endif /* * Number of vnodes in existence. Increased whenever getnewvnode() * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. */ static u_long __exclusive_cache_line numvnodes; SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "Number of vnodes in existence"); static counter_u64_t vnodes_created; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, "Number of vnodes created by getnewvnode"); /* * Conversion tables for conversion from vnode types to inode formats * and back. */ enum vtype iftovt_tab[16] = { VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON }; int vttoif_tab[10] = { 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT }; /* * List of allocates vnodes in the system. */ static TAILQ_HEAD(freelst, vnode) vnode_list; static struct vnode *vnode_list_free_marker; static struct vnode *vnode_list_reclaim_marker; /* * "Free" vnode target. Free vnodes are rarely completely free, but are * just ones that are cheap to recycle. Usually they are for files which * have been stat'd but not read; these usually have inode and namecache * data attached to them. This target is the preferred minimum size of a * sub-cache consisting mostly of such files. The system balances the size * of this sub-cache with its complement to try to prevent either from * thrashing while the other is relatively inactive. The targets express * a preference for the best balance. * * "Above" this target there are 2 further targets (watermarks) related * to recyling of free vnodes. In the best-operating case, the cache is * exactly full, the free list has size between vlowat and vhiwat above the * free target, and recycling from it and normal use maintains this state. * Sometimes the free list is below vlowat or even empty, but this state * is even better for immediate use provided the cache is not full. * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free * ones) to reach one of these states. The watermarks are currently hard- * coded as 4% and 9% of the available space higher. These and the default * of 25% for wantfreevnodes are too large if the memory size is large. * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim * whenever vnlru_proc() becomes active. */ static long wantfreevnodes; static long __exclusive_cache_line freevnodes; SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "Number of \"free\" vnodes"); static long freevnodes_old; static counter_u64_t recycles_count; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, "Number of vnodes recycled to meet vnode cache targets"); static counter_u64_t recycles_free_count; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, "Number of free vnodes recycled to meet vnode cache targets"); static counter_u64_t deferred_inact; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, "Number of times inactive processing was deferred"); /* To keep more than one thread at a time from running vfs_getnewfsid */ static struct mtx mntid_mtx; /* * Lock for any access to the following: * vnode_list * numvnodes * freevnodes */ static struct mtx __exclusive_cache_line vnode_list_mtx; /* Publicly exported FS */ struct nfs_public nfs_pub; static uma_zone_t buf_trie_zone; static smr_t buf_trie_smr; /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ static uma_zone_t vnode_zone; MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll"); __read_frequently smr_t vfs_smr; /* * The workitem queue. * * It is useful to delay writes of file data and filesystem metadata * for tens of seconds so that quickly created and deleted files need * not waste disk bandwidth being created and removed. To realize this, * we append vnodes to a "workitem" queue. When running with a soft * updates implementation, most pending metadata dependencies should * not wait for more than a few seconds. Thus, mounted on block devices * are delayed only about a half the time that file data is delayed. * Similarly, directory updates are more critical, so are only delayed * about a third the time that file data is delayed. Thus, there are * SYNCER_MAXDELAY queues that are processed round-robin at a rate of * one each second (driven off the filesystem syncer process). The * syncer_delayno variable indicates the next queue that is to be processed. * Items that need to be processed soon are placed in this queue: * * syncer_workitem_pending[syncer_delayno] * * A delay of fifteen seconds is done by placing the request fifteen * entries later in the queue: * * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] * */ static int syncer_delayno; static long syncer_mask; LIST_HEAD(synclist, bufobj); static struct synclist *syncer_workitem_pending; /* * The sync_mtx protects: * bo->bo_synclist * sync_vnode_count * syncer_delayno * syncer_state * syncer_workitem_pending * syncer_worklist_len * rushjob */ static struct mtx sync_mtx; static struct cv sync_wakeup; #define SYNCER_MAXDELAY 32 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ static int syncdelay = 30; /* max time to delay syncing data */ static int filedelay = 30; /* time to delay syncing files */ SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "Time to delay syncing files (in seconds)"); static int dirdelay = 29; /* time to delay syncing directories */ SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "Time to delay syncing directories (in seconds)"); static int metadelay = 28; /* time to delay syncing metadata */ SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "Time to delay syncing metadata (in seconds)"); static int rushjob; /* number of slots to run ASAP */ static int stat_rush_requests; /* number of times I/O speeded up */ SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "Number of times I/O speeded up (rush requests)"); #define VDBATCH_SIZE 8 struct vdbatch { u_int index; long freevnodes; struct mtx lock; struct vnode *tab[VDBATCH_SIZE]; }; DPCPU_DEFINE_STATIC(struct vdbatch, vd); static void vdbatch_dequeue(struct vnode *vp); /* * When shutting down the syncer, run it at four times normal speed. */ #define SYNCER_SHUTDOWN_SPEEDUP 4 static int sync_vnode_count; static int syncer_worklist_len; static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } syncer_state; /* Target for maximum number of vnodes. */ u_long desiredvnodes; static u_long gapvnodes; /* gap between wanted and desired */ static u_long vhiwat; /* enough extras after expansion */ static u_long vlowat; /* minimal extras before expansion */ static u_long vstir; /* nonzero to stir non-free vnodes */ static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ static u_long vnlru_read_freevnodes(void); /* * Note that no attempt is made to sanitize these parameters. */ static int sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) { u_long val; int error; val = desiredvnodes; error = sysctl_handle_long(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val == desiredvnodes) return (0); mtx_lock(&vnode_list_mtx); desiredvnodes = val; wantfreevnodes = desiredvnodes / 4; vnlru_recalc(); mtx_unlock(&vnode_list_mtx); /* * XXX There is no protection against multiple threads changing * desiredvnodes at the same time. Locking above only helps vnlru and * getnewvnode. */ vfs_hash_changesize(desiredvnodes); cache_changesize(desiredvnodes); return (0); } SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, "LU", "Target for maximum number of vnodes"); static int sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) { u_long val; int error; val = wantfreevnodes; error = sysctl_handle_long(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val == wantfreevnodes) return (0); mtx_lock(&vnode_list_mtx); wantfreevnodes = val; vnlru_recalc(); mtx_unlock(&vnode_list_mtx); return (0); } SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, "LU", "Target for minimum number of \"free\" vnodes"); SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); static int vnlru_nowhere; SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); static int sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) { struct vnode *vp; struct nameidata nd; char *buf; unsigned long ndflags; int error; if (req->newptr == NULL) return (EINVAL); if (req->newlen >= PATH_MAX) return (E2BIG); buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); error = SYSCTL_IN(req, buf, req->newlen); if (error != 0) goto out; buf[req->newlen] = '\0'; ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME; NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread); if ((error = namei(&nd)) != 0) goto out; vp = nd.ni_vp; if (VN_IS_DOOMED(vp)) { /* * This vnode is being recycled. Return != 0 to let the caller * know that the sysctl had no effect. Return EAGAIN because a * subsequent call will likely succeed (since namei will create * a new vnode if necessary) */ error = EAGAIN; goto putvnode; } counter_u64_add(recycles_count, 1); vgone(vp); putvnode: NDFREE(&nd, 0); out: free(buf, M_TEMP); return (error); } static int sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) { struct thread *td = curthread; struct vnode *vp; struct file *fp; int error; int fd; if (req->newptr == NULL) return (EBADF); error = sysctl_handle_int(oidp, &fd, 0, req); if (error != 0) return (error); error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); if (error != 0) return (error); vp = fp->f_vnode; error = vn_lock(vp, LK_EXCLUSIVE); if (error != 0) goto drop; counter_u64_add(recycles_count, 1); vgone(vp); VOP_UNLOCK(vp); drop: fdrop(fp, td); return (error); } SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_ftry_reclaim_vnode, "I", "Try to reclaim a vnode by its file descriptor"); /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ static int vnsz2log; /* * Support for the bufobj clean & dirty pctrie. */ static void * buf_trie_alloc(struct pctrie *ptree) { return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT)); } static void buf_trie_free(struct pctrie *ptree, void *node) { uma_zfree_smr(buf_trie_zone, node); } PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free, buf_trie_smr); /* * Initialize the vnode management data structures. * * Reevaluate the following cap on the number of vnodes after the physical * memory size exceeds 512GB. In the limit, as the physical memory size * grows, the ratio of the memory size in KB to vnodes approaches 64:1. */ #ifndef MAXVNODES_MAX #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ #endif static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); static struct vnode * vn_alloc_marker(struct mount *mp) { struct vnode *vp; vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); vp->v_type = VMARKER; vp->v_mount = mp; return (vp); } static void vn_free_marker(struct vnode *vp) { MPASS(vp->v_type == VMARKER); free(vp, M_VNODE_MARKER); } /* * Initialize a vnode as it first enters the zone. */ static int vnode_init(void *mem, int size, int flags) { struct vnode *vp; vp = mem; bzero(vp, size); /* * Setup locks. */ vp->v_vnlock = &vp->v_lock; mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); /* * By default, don't allow shared locks unless filesystems opt-in. */ lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, LK_NOSHARE | LK_IS_VNODE); /* * Initialize bufobj. */ bufobj_init(&vp->v_bufobj, vp); /* * Initialize namecache. */ cache_vnode_init(vp); /* * Initialize rangelocks. */ rangelock_init(&vp->v_rl); vp->v_dbatchcpu = NOCPU; /* * Check vhold_recycle_free for an explanation. */ vp->v_holdcnt = VHOLD_NO_SMR; vp->v_type = VNON; mtx_lock(&vnode_list_mtx); TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); mtx_unlock(&vnode_list_mtx); return (0); } /* * Free a vnode when it is cleared from the zone. */ static void vnode_fini(void *mem, int size) { struct vnode *vp; struct bufobj *bo; vp = mem; vdbatch_dequeue(vp); mtx_lock(&vnode_list_mtx); TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); mtx_unlock(&vnode_list_mtx); rangelock_destroy(&vp->v_rl); lockdestroy(vp->v_vnlock); mtx_destroy(&vp->v_interlock); bo = &vp->v_bufobj; rw_destroy(BO_LOCKPTR(bo)); } /* * Provide the size of NFS nclnode and NFS fh for calculation of the * vnode memory consumption. The size is specified directly to * eliminate dependency on NFS-private header. * * Other filesystems may use bigger or smaller (like UFS and ZFS) * private inode data, but the NFS-based estimation is ample enough. * Still, we care about differences in the size between 64- and 32-bit * platforms. * * Namecache structure size is heuristically * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. */ #ifdef _LP64 #define NFS_NCLNODE_SZ (528 + 64) #define NC_SZ 148 #else #define NFS_NCLNODE_SZ (360 + 32) #define NC_SZ 92 #endif static void vntblinit(void *dummy __unused) { struct vdbatch *vd; int cpu, physvnodes, virtvnodes; u_int i; /* * Desiredvnodes is a function of the physical memory size and the * kernel's heap size. Generally speaking, it scales with the * physical memory size. The ratio of desiredvnodes to the physical * memory size is 1:16 until desiredvnodes exceeds 98,304. * Thereafter, the * marginal ratio of desiredvnodes to the physical memory size is * 1:64. However, desiredvnodes is limited by the kernel's heap * size. The memory required by desiredvnodes vnodes and vm objects * must not exceed 1/10th of the kernel's heap size. */ physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); desiredvnodes = min(physvnodes, virtvnodes); if (desiredvnodes > MAXVNODES_MAX) { if (bootverbose) printf("Reducing kern.maxvnodes %lu -> %lu\n", desiredvnodes, MAXVNODES_MAX); desiredvnodes = MAXVNODES_MAX; } wantfreevnodes = desiredvnodes / 4; mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); TAILQ_INIT(&vnode_list); mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); /* * The lock is taken to appease WITNESS. */ mtx_lock(&vnode_list_mtx); vnlru_recalc(); mtx_unlock(&vnode_list_mtx); vnode_list_free_marker = vn_alloc_marker(NULL); TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); vnode_list_reclaim_marker = vn_alloc_marker(NULL); TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, vnode_init, vnode_fini, UMA_ALIGN_PTR, 0); uma_zone_set_smr(vnode_zone, vfs_smr); /* * Preallocate enough nodes to support one-per buf so that * we can not fail an insert. reassignbuf() callers can not * tolerate the insertion failure. */ buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_SMR); buf_trie_smr = uma_zone_get_smr(buf_trie_zone); uma_prealloc(buf_trie_zone, nbuf); vnodes_created = counter_u64_alloc(M_WAITOK); recycles_count = counter_u64_alloc(M_WAITOK); recycles_free_count = counter_u64_alloc(M_WAITOK); deferred_inact = counter_u64_alloc(M_WAITOK); /* * Initialize the filesystem syncer. */ syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, &syncer_mask); syncer_maxdelay = syncer_mask + 1; mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); cv_init(&sync_wakeup, "syncer"); for (i = 1; i <= sizeof(struct vnode); i <<= 1) vnsz2log++; vnsz2log--; CPU_FOREACH(cpu) { vd = DPCPU_ID_PTR((cpu), vd); bzero(vd, sizeof(*vd)); mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); } } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); /* * Mark a mount point as busy. Used to synchronize access and to delay * unmounting. Eventually, mountlist_mtx is not released on failure. * * vfs_busy() is a custom lock, it can block the caller. * vfs_busy() only sleeps if the unmount is active on the mount point. * For a mountpoint mp, vfs_busy-enforced lock is before lock of any * vnode belonging to mp. * * Lookup uses vfs_busy() to traverse mount points. * root fs var fs * / vnode lock A / vnode lock (/var) D * /var vnode lock B /log vnode lock(/var/log) E * vfs_busy lock C vfs_busy lock F * * Within each file system, the lock order is C->A->B and F->D->E. * * When traversing across mounts, the system follows that lock order: * * C->A->B * | * +->F->D->E * * The lookup() process for namei("/var") illustrates the process: * VOP_LOOKUP() obtains B while A is held * vfs_busy() obtains a shared lock on F while A and B are held * vput() releases lock on B * vput() releases lock on A * VFS_ROOT() obtains lock on D while shared lock on F is held * vfs_unbusy() releases shared lock on F * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. * Attempt to lock A (instead of vp_crossmp) while D is held would * violate the global order, causing deadlocks. * * dounmount() locks B while F is drained. */ int vfs_busy(struct mount *mp, int flags) { + struct mount_pcpu *mpcpu; MPASS((flags & ~MBF_MASK) == 0); CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); - if (vfs_op_thread_enter(mp)) { + if (vfs_op_thread_enter(mp, mpcpu)) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); - vfs_mp_count_add_pcpu(mp, ref, 1); - vfs_mp_count_add_pcpu(mp, lockref, 1); - vfs_op_thread_exit(mp); + vfs_mp_count_add_pcpu(mpcpu, ref, 1); + vfs_mp_count_add_pcpu(mpcpu, lockref, 1); + vfs_op_thread_exit(mp, mpcpu); if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); return (0); } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REF(mp); /* * If mount point is currently being unmounted, sleep until the * mount point fate is decided. If thread doing the unmounting fails, * it will clear MNTK_UNMOUNT flag before waking us up, indicating * that this mount point has survived the unmount attempt and vfs_busy * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE * flag in addition to MNTK_UNMOUNT, indicating that mount point is * about to be really destroyed. vfs_busy needs to release its * reference on the mount point in this case and return with ENOENT, * telling the caller that mount mount it tried to busy is no longer * valid. */ while (mp->mnt_kern_flag & MNTK_UNMOUNT) { if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { MNT_REL(mp); MNT_IUNLOCK(mp); CTR1(KTR_VFS, "%s: failed busying before sleeping", __func__); return (ENOENT); } if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); mp->mnt_kern_flag |= MNTK_MWAIT; msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); if (flags & MBF_MNTLSTLOCK) mtx_lock(&mountlist_mtx); MNT_ILOCK(mp); } if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); mp->mnt_lockref++; MNT_IUNLOCK(mp); return (0); } /* * Free a busy filesystem. */ void vfs_unbusy(struct mount *mp) { + struct mount_pcpu *mpcpu; int c; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); - if (vfs_op_thread_enter(mp)) { + if (vfs_op_thread_enter(mp, mpcpu)) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); - vfs_mp_count_sub_pcpu(mp, lockref, 1); - vfs_mp_count_sub_pcpu(mp, ref, 1); - vfs_op_thread_exit(mp); + vfs_mp_count_sub_pcpu(mpcpu, lockref, 1); + vfs_mp_count_sub_pcpu(mpcpu, ref, 1); + vfs_op_thread_exit(mp, mpcpu); return; } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REL(mp); c = --mp->mnt_lockref; if (mp->mnt_vfs_ops == 0) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); MNT_IUNLOCK(mp); return; } if (c < 0) vfs_dump_mount_counters(mp); if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); CTR1(KTR_VFS, "%s: waking up waiters", __func__); mp->mnt_kern_flag &= ~MNTK_DRAINING; wakeup(&mp->mnt_lockref); } MNT_IUNLOCK(mp); } /* * Lookup a mount point by filesystem identifier. */ struct mount * vfs_getvfs(fsid_t *fsid) { struct mount *mp; CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { vfs_ref(mp); mtx_unlock(&mountlist_mtx); return (mp); } } mtx_unlock(&mountlist_mtx); CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); return ((struct mount *) 0); } /* * Lookup a mount point by filesystem identifier, busying it before * returning. * * To avoid congestion on mountlist_mtx, implement simple direct-mapped * cache for popular filesystem identifiers. The cache is lockess, using * the fact that struct mount's are never freed. In worst case we may * get pointer to unmounted or even different filesystem, so we have to * check what we got, and go slow way if so. */ struct mount * vfs_busyfs(fsid_t *fsid) { #define FSID_CACHE_SIZE 256 typedef struct mount * volatile vmp_t; static vmp_t cache[FSID_CACHE_SIZE]; struct mount *mp; int error; uint32_t hash; CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); hash = fsid->val[0] ^ fsid->val[1]; hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); mp = cache[hash]; if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) goto slow; if (vfs_busy(mp, 0) != 0) { cache[hash] = NULL; goto slow; } if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) return (mp); else vfs_unbusy(mp); slow: mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { error = vfs_busy(mp, MBF_MNTLSTLOCK); if (error) { cache[hash] = NULL; mtx_unlock(&mountlist_mtx); return (NULL); } cache[hash] = mp; return (mp); } } CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); mtx_unlock(&mountlist_mtx); return ((struct mount *) 0); } /* * Check if a user can access privileged mount options. */ int vfs_suser(struct mount *mp, struct thread *td) { int error; if (jailed(td->td_ucred)) { /* * If the jail of the calling thread lacks permission for * this type of file system, deny immediately. */ if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) return (EPERM); /* * If the file system was mounted outside the jail of the * calling thread, deny immediately. */ if (prison_check(td->td_ucred, mp->mnt_cred) != 0) return (EPERM); } /* * If file system supports delegated administration, we don't check * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified * by the file system itself. * If this is not the user that did original mount, we check for * the PRIV_VFS_MOUNT_OWNER privilege. */ if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) return (error); } return (0); } /* * Get a new unique fsid. Try to make its val[0] unique, since this value * will be used to create fake device numbers for stat(). Also try (but * not so hard) make its val[0] unique mod 2^16, since some emulators only * support 16-bit device numbers. We end up with unique val[0]'s for the * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. * * Keep in mind that several mounts may be running in parallel. Starting * the search one past where the previous search terminated is both a * micro-optimization and a defense against returning the same fsid to * different mounts. */ void vfs_getnewfsid(struct mount *mp) { static uint16_t mntid_base; struct mount *nmp; fsid_t tfsid; int mtype; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); mtx_lock(&mntid_mtx); mtype = mp->mnt_vfc->vfc_typenum; tfsid.val[1] = mtype; mtype = (mtype & 0xFF) << 24; for (;;) { tfsid.val[0] = makedev(255, mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); mntid_base++; if ((nmp = vfs_getvfs(&tfsid)) == NULL) break; vfs_rel(nmp); } mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; mtx_unlock(&mntid_mtx); } /* * Knob to control the precision of file timestamps: * * 0 = seconds only; nanoseconds zeroed. * 1 = seconds and nanoseconds, accurate within 1/HZ. * 2 = seconds and nanoseconds, truncated to microseconds. * >=3 = seconds and nanoseconds, maximum precision. */ enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; static int timestamp_precision = TSP_USEC; SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, ×tamp_precision, 0, "File timestamp precision (0: seconds, " "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " "3+: sec + ns (max. precision))"); /* * Get a current timestamp. */ void vfs_timestamp(struct timespec *tsp) { struct timeval tv; switch (timestamp_precision) { case TSP_SEC: tsp->tv_sec = time_second; tsp->tv_nsec = 0; break; case TSP_HZ: getnanotime(tsp); break; case TSP_USEC: microtime(&tv); TIMEVAL_TO_TIMESPEC(&tv, tsp); break; case TSP_NSEC: default: nanotime(tsp); break; } } /* * Set vnode attributes to VNOVAL */ void vattr_null(struct vattr *vap) { vap->va_type = VNON; vap->va_size = VNOVAL; vap->va_bytes = VNOVAL; vap->va_mode = VNOVAL; vap->va_nlink = VNOVAL; vap->va_uid = VNOVAL; vap->va_gid = VNOVAL; vap->va_fsid = VNOVAL; vap->va_fileid = VNOVAL; vap->va_blocksize = VNOVAL; vap->va_rdev = VNOVAL; vap->va_atime.tv_sec = VNOVAL; vap->va_atime.tv_nsec = VNOVAL; vap->va_mtime.tv_sec = VNOVAL; vap->va_mtime.tv_nsec = VNOVAL; vap->va_ctime.tv_sec = VNOVAL; vap->va_ctime.tv_nsec = VNOVAL; vap->va_birthtime.tv_sec = VNOVAL; vap->va_birthtime.tv_nsec = VNOVAL; vap->va_flags = VNOVAL; vap->va_gen = VNOVAL; vap->va_vaflags = 0; } /* * Try to reduce the total number of vnodes. * * This routine (and its user) are buggy in at least the following ways: * - all parameters were picked years ago when RAM sizes were significantly * smaller * - it can pick vnodes based on pages used by the vm object, but filesystems * like ZFS don't use it making the pick broken * - since ZFS has its own aging policy it gets partially combated by this one * - a dedicated method should be provided for filesystems to let them decide * whether the vnode should be recycled * * This routine is called when we have too many vnodes. It attempts * to free vnodes and will potentially free vnodes that still * have VM backing store (VM backing store is typically the cause * of a vnode blowout so we want to do this). Therefore, this operation * is not considered cheap. * * A number of conditions may prevent a vnode from being reclaimed. * the buffer cache may have references on the vnode, a directory * vnode may still have references due to the namei cache representing * underlying files, or the vnode may be in active use. It is not * desirable to reuse such vnodes. These conditions may cause the * number of vnodes to reach some minimum value regardless of what * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. * * @param reclaim_nc_src Only reclaim directories with outgoing namecache * entries if this argument is strue * @param trigger Only reclaim vnodes with fewer than this many resident * pages. * @param target How many vnodes to reclaim. * @return The number of vnodes that were reclaimed. */ static int vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) { struct vnode *vp, *mvp; struct mount *mp; struct vm_object *object; u_long done; bool retried; mtx_assert(&vnode_list_mtx, MA_OWNED); retried = false; done = 0; mvp = vnode_list_reclaim_marker; restart: vp = mvp; while (done < target) { vp = TAILQ_NEXT(vp, v_vnodelist); if (__predict_false(vp == NULL)) break; if (__predict_false(vp->v_type == VMARKER)) continue; /* * If it's been deconstructed already, it's still * referenced, or it exceeds the trigger, skip it. * Also skip free vnodes. We are trying to make space * to expand the free list, not reduce it. */ if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) goto next_iter; if (vp->v_type == VBAD || vp->v_type == VNON) goto next_iter; object = atomic_load_ptr(&vp->v_object); if (object == NULL || object->resident_page_count > trigger) { goto next_iter; } /* * Handle races against vnode allocation. Filesystems lock the * vnode some time after it gets returned from getnewvnode, * despite type and hold count being manipulated earlier. * Resorting to checking v_mount restores guarantees present * before the global list was reworked to contain all vnodes. */ if (!VI_TRYLOCK(vp)) goto next_iter; if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { VI_UNLOCK(vp); goto next_iter; } if (vp->v_mount == NULL) { VI_UNLOCK(vp); goto next_iter; } vholdl(vp); VI_UNLOCK(vp); TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { vdrop(vp); goto next_iter_unlocked; } if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { vdrop(vp); vn_finished_write(mp); goto next_iter_unlocked; } VI_LOCK(vp); if (vp->v_usecount > 0 || (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || (vp->v_object != NULL && vp->v_object->resident_page_count > trigger)) { VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(mp); goto next_iter_unlocked; } counter_u64_add(recycles_count, 1); vgonel(vp); VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(mp); done++; next_iter_unlocked: if (should_yield()) kern_yield(PRI_USER); mtx_lock(&vnode_list_mtx); goto restart; next_iter: MPASS(vp->v_type != VMARKER); if (!should_yield()) continue; TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); kern_yield(PRI_USER); mtx_lock(&vnode_list_mtx); goto restart; } if (done == 0 && !retried) { TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); retried = true; goto restart; } return (done); } static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 0, "limit on vnode free requests per call to the vnlru_free routine"); /* * Attempt to reduce the free list by the requested amount. */ static int vnlru_free_locked(int count, struct vfsops *mnt_op) { struct vnode *vp, *mvp; struct mount *mp; int ocount; mtx_assert(&vnode_list_mtx, MA_OWNED); if (count > max_vnlru_free) count = max_vnlru_free; ocount = count; mvp = vnode_list_free_marker; vp = mvp; for (;;) { if (count == 0) { break; } vp = TAILQ_NEXT(vp, v_vnodelist); if (__predict_false(vp == NULL)) { TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); break; } if (__predict_false(vp->v_type == VMARKER)) continue; if (vp->v_holdcnt > 0) continue; /* * Don't recycle if our vnode is from different type * of mount point. Note that mp is type-safe, the * check does not reach unmapped address even if * vnode is reclaimed. */ if (mnt_op != NULL && (mp = vp->v_mount) != NULL && mp->mnt_op != mnt_op) { continue; } if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { continue; } if (!vhold_recycle_free(vp)) continue; TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); if (vtryrecycle(vp) == 0) count--; mtx_lock(&vnode_list_mtx); vp = mvp; } return (ocount - count); } void vnlru_free(int count, struct vfsops *mnt_op) { mtx_lock(&vnode_list_mtx); vnlru_free_locked(count, mnt_op); mtx_unlock(&vnode_list_mtx); } static void vnlru_recalc(void) { mtx_assert(&vnode_list_mtx, MA_OWNED); gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ vlowat = vhiwat / 2; } /* * Attempt to recycle vnodes in a context that is always safe to block. * Calling vlrurecycle() from the bowels of filesystem code has some * interesting deadlock problems. */ static struct proc *vnlruproc; static int vnlruproc_sig; /* * The main freevnodes counter is only updated when threads requeue their vnode * batches. CPUs are conditionally walked to compute a more accurate total. * * Limit how much of a slop are we willing to tolerate. Note: the actual value * at any given moment can still exceed slop, but it should not be by significant * margin in practice. */ #define VNLRU_FREEVNODES_SLOP 128 static __inline void vn_freevnodes_inc(void) { struct vdbatch *vd; critical_enter(); vd = DPCPU_PTR(vd); vd->freevnodes++; critical_exit(); } static __inline void vn_freevnodes_dec(void) { struct vdbatch *vd; critical_enter(); vd = DPCPU_PTR(vd); vd->freevnodes--; critical_exit(); } static u_long vnlru_read_freevnodes(void) { struct vdbatch *vd; long slop; int cpu; mtx_assert(&vnode_list_mtx, MA_OWNED); if (freevnodes > freevnodes_old) slop = freevnodes - freevnodes_old; else slop = freevnodes_old - freevnodes; if (slop < VNLRU_FREEVNODES_SLOP) return (freevnodes >= 0 ? freevnodes : 0); freevnodes_old = freevnodes; CPU_FOREACH(cpu) { vd = DPCPU_ID_PTR((cpu), vd); freevnodes_old += vd->freevnodes; } return (freevnodes_old >= 0 ? freevnodes_old : 0); } static bool vnlru_under(u_long rnumvnodes, u_long limit) { u_long rfreevnodes, space; if (__predict_false(rnumvnodes > desiredvnodes)) return (true); space = desiredvnodes - rnumvnodes; if (space < limit) { rfreevnodes = vnlru_read_freevnodes(); if (rfreevnodes > wantfreevnodes) space += rfreevnodes - wantfreevnodes; } return (space < limit); } static bool vnlru_under_unlocked(u_long rnumvnodes, u_long limit) { long rfreevnodes, space; if (__predict_false(rnumvnodes > desiredvnodes)) return (true); space = desiredvnodes - rnumvnodes; if (space < limit) { rfreevnodes = atomic_load_long(&freevnodes); if (rfreevnodes > wantfreevnodes) space += rfreevnodes - wantfreevnodes; } return (space < limit); } static void vnlru_kick(void) { mtx_assert(&vnode_list_mtx, MA_OWNED); if (vnlruproc_sig == 0) { vnlruproc_sig = 1; wakeup(vnlruproc); } } static void vnlru_proc(void) { u_long rnumvnodes, rfreevnodes, target; unsigned long onumvnodes; int done, force, trigger, usevnodes; bool reclaim_nc_src, want_reread; EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, SHUTDOWN_PRI_FIRST); force = 0; want_reread = false; for (;;) { kproc_suspend_check(vnlruproc); mtx_lock(&vnode_list_mtx); rnumvnodes = atomic_load_long(&numvnodes); if (want_reread) { force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; want_reread = false; } /* * If numvnodes is too large (due to desiredvnodes being * adjusted using its sysctl, or emergency growth), first * try to reduce it by discarding from the free list. */ if (rnumvnodes > desiredvnodes) { vnlru_free_locked(rnumvnodes - desiredvnodes, NULL); rnumvnodes = atomic_load_long(&numvnodes); } /* * Sleep if the vnode cache is in a good state. This is * when it is not over-full and has space for about a 4% * or 9% expansion (by growing its size or inexcessively * reducing its free list). Otherwise, try to reclaim * space for a 10% expansion. */ if (vstir && force == 0) { force = 1; vstir = 0; } if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { vnlruproc_sig = 0; wakeup(&vnlruproc_sig); msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); continue; } rfreevnodes = vnlru_read_freevnodes(); onumvnodes = rnumvnodes; /* * Calculate parameters for recycling. These are the same * throughout the loop to give some semblance of fairness. * The trigger point is to avoid recycling vnodes with lots * of resident pages. We aren't trying to free memory; we * are trying to recycle or at least free vnodes. */ if (rnumvnodes <= desiredvnodes) usevnodes = rnumvnodes - rfreevnodes; else usevnodes = rnumvnodes; if (usevnodes <= 0) usevnodes = 1; /* * The trigger value is is chosen to give a conservatively * large value to ensure that it alone doesn't prevent * making progress. The value can easily be so large that * it is effectively infinite in some congested and * misconfigured cases, and this is necessary. Normally * it is about 8 to 100 (pages), which is quite large. */ trigger = vm_cnt.v_page_count * 2 / usevnodes; if (force < 2) trigger = vsmalltrigger; reclaim_nc_src = force >= 3; target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); target = target / 10 + 1; done = vlrureclaim(reclaim_nc_src, trigger, target); mtx_unlock(&vnode_list_mtx); if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) uma_reclaim(UMA_RECLAIM_DRAIN); if (done == 0) { if (force == 0 || force == 1) { force = 2; continue; } if (force == 2) { force = 3; continue; } want_reread = true; force = 0; vnlru_nowhere++; tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); } else { want_reread = true; kern_yield(PRI_USER); } } } static struct kproc_desc vnlru_kp = { "vnlru", vnlru_proc, &vnlruproc }; SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp); /* * Routines having to do with the management of the vnode table. */ /* * Try to recycle a freed vnode. We abort if anyone picks up a reference * before we actually vgone(). This function must be called with the vnode * held to prevent the vnode from being returned to the free list midway * through vgone(). */ static int vtryrecycle(struct vnode *vp) { struct mount *vnmp; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); VNASSERT(vp->v_holdcnt, vp, ("vtryrecycle: Recycling vp %p without a reference.", vp)); /* * This vnode may found and locked via some other list, if so we * can't recycle it yet. */ if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { CTR2(KTR_VFS, "%s: impossible to recycle, vp %p lock is already held", __func__, vp); vdrop(vp); return (EWOULDBLOCK); } /* * Don't recycle if its filesystem is being suspended. */ if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { VOP_UNLOCK(vp); CTR2(KTR_VFS, "%s: impossible to recycle, cannot start the write for %p", __func__, vp); vdrop(vp); return (EBUSY); } /* * If we got this far, we need to acquire the interlock and see if * anyone picked up this vnode from another list. If not, we will * mark it with DOOMED via vgonel() so that anyone who does find it * will skip over it. */ VI_LOCK(vp); if (vp->v_usecount) { VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(vnmp); CTR2(KTR_VFS, "%s: impossible to recycle, %p is already referenced", __func__, vp); return (EBUSY); } if (!VN_IS_DOOMED(vp)) { counter_u64_add(recycles_free_count, 1); vgonel(vp); } VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(vnmp); return (0); } /* * Allocate a new vnode. * * The operation never returns an error. Returning an error was disabled * in r145385 (dated 2005) with the following comment: * * XXX Not all VFS_VGET/ffs_vget callers check returns. * * Given the age of this commit (almost 15 years at the time of writing this * comment) restoring the ability to fail requires a significant audit of * all codepaths. * * The routine can try to free a vnode or stall for up to 1 second waiting for * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. */ static u_long vn_alloc_cyclecount; static struct vnode * __noinline vn_alloc_hard(struct mount *mp) { u_long rnumvnodes, rfreevnodes; mtx_lock(&vnode_list_mtx); rnumvnodes = atomic_load_long(&numvnodes); if (rnumvnodes + 1 < desiredvnodes) { vn_alloc_cyclecount = 0; goto alloc; } rfreevnodes = vnlru_read_freevnodes(); if (vn_alloc_cyclecount++ >= rfreevnodes) { vn_alloc_cyclecount = 0; vstir = 1; } /* * Grow the vnode cache if it will not be above its target max * after growing. Otherwise, if the free list is nonempty, try * to reclaim 1 item from it before growing the cache (possibly * above its target max if the reclamation failed or is delayed). * Otherwise, wait for some space. In all cases, schedule * vnlru_proc() if we are getting short of space. The watermarks * should be chosen so that we never wait or even reclaim from * the free list to below its target minimum. */ if (vnlru_free_locked(1, NULL) > 0) goto alloc; if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { /* * Wait for space for a new vnode. */ vnlru_kick(); msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && vnlru_read_freevnodes() > 1) vnlru_free_locked(1, NULL); } alloc: rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; if (vnlru_under(rnumvnodes, vlowat)) vnlru_kick(); mtx_unlock(&vnode_list_mtx); return (uma_zalloc_smr(vnode_zone, M_WAITOK)); } static struct vnode * vn_alloc(struct mount *mp) { u_long rnumvnodes; if (__predict_false(vn_alloc_cyclecount != 0)) return (vn_alloc_hard(mp)); rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) { atomic_subtract_long(&numvnodes, 1); return (vn_alloc_hard(mp)); } return (uma_zalloc_smr(vnode_zone, M_WAITOK)); } static void vn_free(struct vnode *vp) { atomic_subtract_long(&numvnodes, 1); uma_zfree_smr(vnode_zone, vp); } /* * Return the next vnode from the free list. */ int getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, struct vnode **vpp) { struct vnode *vp; struct thread *td; struct lock_object *lo; CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); KASSERT(vops->registered, ("%s: not registered vector op %p\n", __func__, vops)); td = curthread; if (td->td_vp_reserved != NULL) { vp = td->td_vp_reserved; td->td_vp_reserved = NULL; } else { vp = vn_alloc(mp); } counter_u64_add(vnodes_created, 1); /* * Locks are given the generic name "vnode" when created. * Follow the historic practice of using the filesystem * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. * * Locks live in a witness group keyed on their name. Thus, * when a lock is renamed, it must also move from the witness * group of its old name to the witness group of its new name. * * The change only needs to be made when the vnode moves * from one filesystem type to another. We ensure that each * filesystem use a single static name pointer for its tag so * that we can compare pointers rather than doing a strcmp(). */ lo = &vp->v_vnlock->lock_object; #ifdef WITNESS if (lo->lo_name != tag) { #endif lo->lo_name = tag; #ifdef WITNESS WITNESS_DESTROY(lo); WITNESS_INIT(lo, tag); } #endif /* * By default, don't allow shared locks unless filesystems opt-in. */ vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; /* * Finalize various vnode identity bits. */ KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); vp->v_type = VNON; vp->v_op = vops; v_init_counters(vp); vp->v_bufobj.bo_ops = &buf_ops_bio; #ifdef DIAGNOSTIC if (mp == NULL && vops != &dead_vnodeops) printf("NULL mp in getnewvnode(9), tag %s\n", tag); #endif #ifdef MAC mac_vnode_init(vp); if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) mac_vnode_associate_singlelabel(mp, vp); #endif if (mp != NULL) { vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) vp->v_vflag |= VV_NOKNOTE; } /* * For the filesystems which do not use vfs_hash_insert(), * still initialize v_hash to have vfs_hash_index() useful. * E.g., nullfs uses vfs_hash_index() on the lower vnode for * its own hashing. */ vp->v_hash = (uintptr_t)vp >> vnsz2log; *vpp = vp; return (0); } void getnewvnode_reserve(void) { struct thread *td; td = curthread; MPASS(td->td_vp_reserved == NULL); td->td_vp_reserved = vn_alloc(NULL); } void getnewvnode_drop_reserve(void) { struct thread *td; td = curthread; if (td->td_vp_reserved != NULL) { vn_free(td->td_vp_reserved); td->td_vp_reserved = NULL; } } static void __noinline freevnode(struct vnode *vp) { struct bufobj *bo; /* * The vnode has been marked for destruction, so free it. * * The vnode will be returned to the zone where it will * normally remain until it is needed for another vnode. We * need to cleanup (or verify that the cleanup has already * been done) any residual data left from its current use * so as not to contaminate the freshly allocated vnode. */ CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); /* * Paired with vgone. */ vn_seqc_write_end_locked(vp); VNPASS(vp->v_seqc_users == 0, vp); bo = &vp->v_bufobj; VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, ("clean blk trie not empty")); VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, ("dirty blk trie not empty")); VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, ("Dangling rangelock waiters")); VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp, ("Leaked inactivation")); VI_UNLOCK(vp); #ifdef MAC mac_vnode_destroy(vp); #endif if (vp->v_pollinfo != NULL) { destroy_vpollinfo(vp->v_pollinfo); vp->v_pollinfo = NULL; } #ifdef INVARIANTS /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ vp->v_op = NULL; #endif vp->v_mountedhere = NULL; vp->v_unpcb = NULL; vp->v_rdev = NULL; vp->v_fifoinfo = NULL; vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; vp->v_irflag = 0; vp->v_iflag = 0; vp->v_vflag = 0; bo->bo_flag = 0; vn_free(vp); } /* * Delete from old mount point vnode list, if on one. */ static void delmntque(struct vnode *vp) { struct mount *mp; VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); mp = vp->v_mount; if (mp == NULL) return; MNT_ILOCK(mp); VI_LOCK(vp); vp->v_mount = NULL; VI_UNLOCK(vp); VNASSERT(mp->mnt_nvnodelistsize > 0, vp, ("bad mount point vnode list size")); TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); mp->mnt_nvnodelistsize--; MNT_REL(mp); MNT_IUNLOCK(mp); } static void insmntque_stddtr(struct vnode *vp, void *dtr_arg) { vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } /* * Insert into list of vnodes for the new mount point, if available. */ int insmntque1(struct vnode *vp, struct mount *mp, void (*dtr)(struct vnode *, void *), void *dtr_arg) { KASSERT(vp->v_mount == NULL, ("insmntque: vnode already on per mount vnode list")); VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); /* * We acquire the vnode interlock early to ensure that the * vnode cannot be recycled by another process releasing a * holdcnt on it before we get it on both the vnode list * and the active vnode list. The mount mutex protects only * manipulation of the vnode list and the vnode freelist * mutex protects only manipulation of the active vnode list. * Hence the need to hold the vnode interlock throughout. */ MNT_ILOCK(mp); VI_LOCK(vp); if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || mp->mnt_nvnodelistsize == 0)) && (vp->v_vflag & VV_FORCEINSMQ) == 0) { VI_UNLOCK(vp); MNT_IUNLOCK(mp); if (dtr != NULL) dtr(vp, dtr_arg); return (EBUSY); } vp->v_mount = mp; MNT_REF(mp); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, ("neg mount point vnode list size")); mp->mnt_nvnodelistsize++; VI_UNLOCK(vp); MNT_IUNLOCK(mp); return (0); } int insmntque(struct vnode *vp, struct mount *mp) { return (insmntque1(vp, mp, insmntque_stddtr, NULL)); } /* * Flush out and invalidate all buffers associated with a bufobj * Called with the underlying object locked. */ int bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) { int error; BO_LOCK(bo); if (flags & V_SAVE) { error = bufobj_wwait(bo, slpflag, slptimeo); if (error) { BO_UNLOCK(bo); return (error); } if (bo->bo_dirty.bv_cnt > 0) { BO_UNLOCK(bo); if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) return (error); /* * XXX We could save a lock/unlock if this was only * enabled under INVARIANTS */ BO_LOCK(bo); if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) panic("vinvalbuf: dirty bufs"); } } /* * If you alter this loop please notice that interlock is dropped and * reacquired in flushbuflist. Special care is needed to ensure that * no race conditions occur from this. */ do { error = flushbuflist(&bo->bo_clean, flags, bo, slpflag, slptimeo); if (error == 0 && !(flags & V_CLEANONLY)) error = flushbuflist(&bo->bo_dirty, flags, bo, slpflag, slptimeo); if (error != 0 && error != EAGAIN) { BO_UNLOCK(bo); return (error); } } while (error != 0); /* * Wait for I/O to complete. XXX needs cleaning up. The vnode can * have write I/O in-progress but if there is a VM object then the * VM object can also have read-I/O in-progress. */ do { bufobj_wwait(bo, 0, 0); if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { BO_UNLOCK(bo); vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); BO_LOCK(bo); } } while (bo->bo_numoutput > 0); BO_UNLOCK(bo); /* * Destroy the copy in the VM cache, too. */ if (bo->bo_object != NULL && (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { VM_OBJECT_WLOCK(bo->bo_object); vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? OBJPR_CLEANONLY : 0); VM_OBJECT_WUNLOCK(bo->bo_object); } #ifdef INVARIANTS BO_LOCK(bo); if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) panic("vinvalbuf: flush failed"); if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && bo->bo_dirty.bv_cnt > 0) panic("vinvalbuf: flush dirty failed"); BO_UNLOCK(bo); #endif return (0); } /* * Flush out and invalidate all buffers associated with a vnode. * Called with the underlying object locked. */ int vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) { CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); ASSERT_VOP_LOCKED(vp, "vinvalbuf"); if (vp->v_object != NULL && vp->v_object->handle != vp) return (0); return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); } /* * Flush out buffers on the specified list. * */ static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo) { struct buf *bp, *nbp; int retval, error; daddr_t lblkno; b_xflags_t xflags; ASSERT_BO_WLOCKED(bo); retval = 0; TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { /* * If we are flushing both V_NORMAL and V_ALT buffers then * do not skip any buffers. If we are flushing only V_NORMAL * buffers then skip buffers marked as BX_ALTDATA. If we are * flushing only V_ALT buffers then skip buffers not marked * as BX_ALTDATA. */ if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { continue; } if (nbp != NULL) { lblkno = nbp->b_lblkno; xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); } retval = EAGAIN; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), "flushbuf", slpflag, slptimeo); if (error) { BO_LOCK(bo); return (error != ENOLCK ? error : EAGAIN); } KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); /* * XXX Since there are no node locks for NFS, I * believe there is a slight chance that a delayed * write will occur while sleeping just above, so * check for it. */ if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && (flags & V_SAVE)) { bremfree(bp); bp->b_flags |= B_ASYNC; bwrite(bp); BO_LOCK(bo); return (EAGAIN); /* XXX: why not loop ? */ } bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); BO_LOCK(bo); if (nbp == NULL) break; nbp = gbincore(bo, lblkno); if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) != xflags) break; /* nbp invalid */ } return (retval); } int bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) { struct buf *bp; int error; daddr_t lblkno; ASSERT_BO_LOCKED(bo); for (lblkno = startn;;) { again: bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); if (bp == NULL || bp->b_lblkno >= endn || bp->b_lblkno < startn) break; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); if (error != 0) { BO_RLOCK(bo); if (error == ENOLCK) goto again; return (error); } KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); lblkno = bp->b_lblkno + 1; if ((bp->b_flags & B_MANAGED) == 0) bremfree(bp); bp->b_flags |= B_RELBUF; /* * In the VMIO case, use the B_NOREUSE flag to hint that the * pages backing each buffer in the range are unlikely to be * reused. Dirty buffers will have the hint applied once * they've been written. */ if ((bp->b_flags & B_VMIO) != 0) bp->b_flags |= B_NOREUSE; brelse(bp); BO_RLOCK(bo); } return (0); } /* * Truncate a file's buffer and pages to a specified length. This * is in lieu of the old vinvalbuf mechanism, which performed unneeded * sync activity. */ int vtruncbuf(struct vnode *vp, off_t length, int blksize) { struct buf *bp, *nbp; struct bufobj *bo; daddr_t startlbn; CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, vp, blksize, (uintmax_t)length); /* * Round up to the *next* lbn. */ startlbn = howmany(length, blksize); ASSERT_VOP_LOCKED(vp, "vtruncbuf"); bo = &vp->v_bufobj; restart_unlocked: BO_LOCK(bo); while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) ; if (length > 0) { restartsync: TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno > 0) continue; /* * Since we hold the vnode lock this should only * fail if we're racing with the buf daemon. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) goto restart_unlocked; VNASSERT((bp->b_flags & B_DELWRI), vp, ("buf(%p) on dirty queue without DELWRI", bp)); bremfree(bp); bawrite(bp); BO_LOCK(bo); goto restartsync; } } bufobj_wwait(bo, 0, 0); BO_UNLOCK(bo); vnode_pager_setsize(vp, length); return (0); } /* * Invalidate the cached pages of a file's buffer within the range of block * numbers [startlbn, endlbn). */ void v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, int blksize) { struct bufobj *bo; off_t start, end; ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); start = blksize * startlbn; end = blksize * endlbn; bo = &vp->v_bufobj; BO_LOCK(bo); MPASS(blksize == bo->bo_bsize); while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) ; BO_UNLOCK(bo); vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); } static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, daddr_t startlbn, daddr_t endlbn) { struct buf *bp, *nbp; bool anyfreed; ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); ASSERT_BO_LOCKED(bo); do { anyfreed = false; TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) { BO_LOCK(bo); return (EAGAIN); } bremfree(bp); bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = true; BO_LOCK(bo); if (nbp != NULL && (((nbp->b_xflags & BX_VNCLEAN) == 0) || nbp->b_vp != vp || (nbp->b_flags & B_DELWRI) != 0)) return (EAGAIN); } TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) { BO_LOCK(bo); return (EAGAIN); } bremfree(bp); bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = true; BO_LOCK(bo); if (nbp != NULL && (((nbp->b_xflags & BX_VNDIRTY) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI) == 0)) return (EAGAIN); } } while (anyfreed); return (0); } static void buf_vlist_remove(struct buf *bp) { struct bufv *bv; b_xflags_t flags; flags = bp->b_xflags; KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); ASSERT_BO_WLOCKED(bp->b_bufobj); KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 && (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN), ("%s: buffer %p has invalid queue state", __func__, bp)); if ((flags & BX_VNDIRTY) != 0) bv = &bp->b_bufobj->bo_dirty; else bv = &bp->b_bufobj->bo_clean; BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); bv->bv_cnt--; bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); } /* * Add the buffer to the sorted clean or dirty block list. * * NOTE: xflags is passed as a constant, optimizing this inline function! */ static void buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) { struct bufv *bv; struct buf *n; int error; ASSERT_BO_WLOCKED(bo); KASSERT((bo->bo_flag & BO_NOBUFS) == 0, ("buf_vlist_add: bo %p does not allow bufs", bo)); KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, ("dead bo %p", bo)); KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); bp->b_xflags |= xflags; if (xflags & BX_VNDIRTY) bv = &bo->bo_dirty; else bv = &bo->bo_clean; /* * Keep the list ordered. Optimize empty list insertion. Assume * we tend to grow at the tail so lookup_le should usually be cheaper * than _ge. */ if (bv->bv_cnt == 0 || bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); else TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); if (error) panic("buf_vlist_add: Preallocated nodes insufficient."); bv->bv_cnt++; } /* * Look up a buffer using the buffer tries. */ struct buf * gbincore(struct bufobj *bo, daddr_t lblkno) { struct buf *bp; ASSERT_BO_LOCKED(bo); bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); if (bp != NULL) return (bp); return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno)); } /* * Look up a buf using the buffer tries, without the bufobj lock. This relies * on SMR for safe lookup, and bufs being in a no-free zone to provide type * stability of the result. Like other lockless lookups, the found buf may * already be invalid by the time this function returns. */ struct buf * gbincore_unlocked(struct bufobj *bo, daddr_t lblkno) { struct buf *bp; ASSERT_BO_UNLOCKED(bo); bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno); if (bp != NULL) return (bp); return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno)); } /* * Associate a buffer with a vnode. */ void bgetvp(struct vnode *vp, struct buf *bp) { struct bufobj *bo; bo = &vp->v_bufobj; ASSERT_BO_WLOCKED(bo); VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, ("bgetvp: bp already attached! %p", bp)); vhold(vp); bp->b_vp = vp; bp->b_bufobj = bo; /* * Insert onto list for new vnode. */ buf_vlist_add(bp, bo, BX_VNCLEAN); } /* * Disassociate a buffer from a vnode. */ void brelvp(struct buf *bp) { struct bufobj *bo; struct vnode *vp; CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); /* * Delete from old vnode list, if on one. */ vp = bp->b_vp; /* XXX */ bo = bp->b_bufobj; BO_LOCK(bo); buf_vlist_remove(bp); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { bo->bo_flag &= ~BO_ONWORKLST; mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); } bp->b_vp = NULL; bp->b_bufobj = NULL; BO_UNLOCK(bo); vdrop(vp); } /* * Add an item to the syncer work queue. */ static void vn_syncer_add_to_worklist(struct bufobj *bo, int delay) { int slot; ASSERT_BO_WLOCKED(bo); mtx_lock(&sync_mtx); if (bo->bo_flag & BO_ONWORKLST) LIST_REMOVE(bo, bo_synclist); else { bo->bo_flag |= BO_ONWORKLST; syncer_worklist_len++; } if (delay > syncer_maxdelay - 2) delay = syncer_maxdelay - 2; slot = (syncer_delayno + delay) & syncer_mask; LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); mtx_unlock(&sync_mtx); } static int sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) { int error, len; mtx_lock(&sync_mtx); len = syncer_worklist_len - sync_vnode_count; mtx_unlock(&sync_mtx); error = SYSCTL_OUT(req, &len, sizeof(len)); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); static struct proc *updateproc; static void sched_sync(void); static struct kproc_desc up_kp = { "syncer", sched_sync, &updateproc }; SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); static int sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) { struct vnode *vp; struct mount *mp; *bo = LIST_FIRST(slp); if (*bo == NULL) return (0); vp = bo2vnode(*bo); if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) return (1); /* * We use vhold in case the vnode does not * successfully sync. vhold prevents the vnode from * going away when we unlock the sync_mtx so that * we can acquire the vnode interlock. */ vholdl(vp); mtx_unlock(&sync_mtx); VI_UNLOCK(vp); if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { vdrop(vp); mtx_lock(&sync_mtx); return (*bo == LIST_FIRST(slp)); } vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); (void) VOP_FSYNC(vp, MNT_LAZY, td); VOP_UNLOCK(vp); vn_finished_write(mp); BO_LOCK(*bo); if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { /* * Put us back on the worklist. The worklist * routine will remove us from our current * position and then add us back in at a later * position. */ vn_syncer_add_to_worklist(*bo, syncdelay); } BO_UNLOCK(*bo); vdrop(vp); mtx_lock(&sync_mtx); return (0); } static int first_printf = 1; /* * System filesystem synchronizer daemon. */ static void sched_sync(void) { struct synclist *next, *slp; struct bufobj *bo; long starttime; struct thread *td = curthread; int last_work_seen; int net_worklist_len; int syncer_final_iter; int error; last_work_seen = 0; syncer_final_iter = 0; syncer_state = SYNCER_RUNNING; starttime = time_uptime; td->td_pflags |= TDP_NORUNNINGBUF; EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, SHUTDOWN_PRI_LAST); mtx_lock(&sync_mtx); for (;;) { if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter == 0) { mtx_unlock(&sync_mtx); kproc_suspend_check(td->td_proc); mtx_lock(&sync_mtx); } net_worklist_len = syncer_worklist_len - sync_vnode_count; if (syncer_state != SYNCER_RUNNING && starttime != time_uptime) { if (first_printf) { printf("\nSyncing disks, vnodes remaining... "); first_printf = 0; } printf("%d ", net_worklist_len); } starttime = time_uptime; /* * Push files whose dirty time has expired. Be careful * of interrupt race on slp queue. * * Skip over empty worklist slots when shutting down. */ do { slp = &syncer_workitem_pending[syncer_delayno]; syncer_delayno += 1; if (syncer_delayno == syncer_maxdelay) syncer_delayno = 0; next = &syncer_workitem_pending[syncer_delayno]; /* * If the worklist has wrapped since the * it was emptied of all but syncer vnodes, * switch to the FINAL_DELAY state and run * for one more second. */ if (syncer_state == SYNCER_SHUTTING_DOWN && net_worklist_len == 0 && last_work_seen == syncer_delayno) { syncer_state = SYNCER_FINAL_DELAY; syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; } } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && syncer_worklist_len > 0); /* * Keep track of the last time there was anything * on the worklist other than syncer vnodes. * Return to the SHUTTING_DOWN state if any * new work appears. */ if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) last_work_seen = syncer_delayno; if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) syncer_state = SYNCER_SHUTTING_DOWN; while (!LIST_EMPTY(slp)) { error = sync_vnode(slp, &bo, td); if (error == 1) { LIST_REMOVE(bo, bo_synclist); LIST_INSERT_HEAD(next, bo, bo_synclist); continue; } if (first_printf == 0) { /* * Drop the sync mutex, because some watchdog * drivers need to sleep while patting */ mtx_unlock(&sync_mtx); wdog_kern_pat(WD_LASTVAL); mtx_lock(&sync_mtx); } } if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) syncer_final_iter--; /* * The variable rushjob allows the kernel to speed up the * processing of the filesystem syncer process. A rushjob * value of N tells the filesystem syncer to process the next * N seconds worth of work on its queue ASAP. Currently rushjob * is used by the soft update code to speed up the filesystem * syncer process when the incore state is getting so far * ahead of the disk that the kernel memory pool is being * threatened with exhaustion. */ if (rushjob > 0) { rushjob -= 1; continue; } /* * Just sleep for a short period of time between * iterations when shutting down to allow some I/O * to happen. * * If it has taken us less than a second to process the * current work, then wait. Otherwise start right over * again. We can still lose time if any single round * takes more than two seconds, but it does not really * matter as we are just trying to generally pace the * filesystem activity. */ if (syncer_state != SYNCER_RUNNING || time_uptime == starttime) { thread_lock(td); sched_prio(td, PPAUSE); thread_unlock(td); } if (syncer_state != SYNCER_RUNNING) cv_timedwait(&sync_wakeup, &sync_mtx, hz / SYNCER_SHUTDOWN_SPEEDUP); else if (time_uptime == starttime) cv_timedwait(&sync_wakeup, &sync_mtx, hz); } } /* * Request the syncer daemon to speed up its work. * We never push it to speed up more than half of its * normal turn time, otherwise it could take over the cpu. */ int speedup_syncer(void) { int ret = 0; mtx_lock(&sync_mtx); if (rushjob < syncdelay / 2) { rushjob += 1; stat_rush_requests += 1; ret = 1; } mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); return (ret); } /* * Tell the syncer to speed up its work and run though its work * list several times, then tell it to shut down. */ static void syncer_shutdown(void *arg, int howto) { if (howto & RB_NOSYNC) return; mtx_lock(&sync_mtx); syncer_state = SYNCER_SHUTTING_DOWN; rushjob = 0; mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); kproc_shutdown(arg, howto); } void syncer_suspend(void) { syncer_shutdown(updateproc, 0); } void syncer_resume(void) { mtx_lock(&sync_mtx); first_printf = 1; syncer_state = SYNCER_RUNNING; mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); kproc_resume(updateproc); } /* * Move the buffer between the clean and dirty lists of its vnode. */ void reassignbuf(struct buf *bp) { struct vnode *vp; struct bufobj *bo; int delay; #ifdef INVARIANTS struct bufv *bv; #endif vp = bp->b_vp; bo = bp->b_bufobj; KASSERT((bp->b_flags & B_PAGING) == 0, ("%s: cannot reassign paging buffer %p", __func__, bp)); CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); BO_LOCK(bo); buf_vlist_remove(bp); /* * If dirty, put on list of dirty buffers; otherwise insert onto list * of clean buffers. */ if (bp->b_flags & B_DELWRI) { if ((bo->bo_flag & BO_ONWORKLST) == 0) { switch (vp->v_type) { case VDIR: delay = dirdelay; break; case VCHR: delay = metadelay; break; default: delay = filedelay; } vn_syncer_add_to_worklist(bo, delay); } buf_vlist_add(bp, bo, BX_VNDIRTY); } else { buf_vlist_add(bp, bo, BX_VNCLEAN); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); bo->bo_flag &= ~BO_ONWORKLST; } } #ifdef INVARIANTS bv = &bo->bo_clean; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bv = &bo->bo_dirty; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); #endif BO_UNLOCK(bo); } static void v_init_counters(struct vnode *vp) { VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, vp, ("%s called for an initialized vnode", __FUNCTION__)); ASSERT_VI_UNLOCKED(vp, __FUNCTION__); refcount_init(&vp->v_holdcnt, 1); refcount_init(&vp->v_usecount, 1); } /* * Grab a particular vnode from the free list, increment its * reference count and lock it. VIRF_DOOMED is set if the vnode * is being destroyed. Only callers who specify LK_RETRY will * see doomed vnodes. If inactive processing was delayed in * vput try to do it here. * * usecount is manipulated using atomics without holding any locks. * * holdcnt can be manipulated using atomics without holding any locks, * except when transitioning 1<->0, in which case the interlock is held. * * Consumers which don't guarantee liveness of the vnode can use SMR to * try to get a reference. Note this operation can fail since the vnode * may be awaiting getting freed by the time they get to it. */ enum vgetstate vget_prep_smr(struct vnode *vp) { enum vgetstate vs; VFS_SMR_ASSERT_ENTERED(); if (refcount_acquire_if_not_zero(&vp->v_usecount)) { vs = VGET_USECOUNT; } else { if (vhold_smr(vp)) vs = VGET_HOLDCNT; else vs = VGET_NONE; } return (vs); } enum vgetstate vget_prep(struct vnode *vp) { enum vgetstate vs; if (refcount_acquire_if_not_zero(&vp->v_usecount)) { vs = VGET_USECOUNT; } else { vhold(vp); vs = VGET_HOLDCNT; } return (vs); } void vget_abort(struct vnode *vp, enum vgetstate vs) { switch (vs) { case VGET_USECOUNT: vrele(vp); break; case VGET_HOLDCNT: vdrop(vp); break; default: __assert_unreachable(); } } int vget(struct vnode *vp, int flags) { enum vgetstate vs; vs = vget_prep(vp); return (vget_finish(vp, flags, vs)); } int vget_finish(struct vnode *vp, int flags, enum vgetstate vs) { int error; if ((flags & LK_INTERLOCK) != 0) ASSERT_VI_LOCKED(vp, __func__); else ASSERT_VI_UNLOCKED(vp, __func__); VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); VNPASS(vp->v_holdcnt > 0, vp); VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); error = vn_lock(vp, flags); if (__predict_false(error != 0)) { vget_abort(vp, vs); CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, vp); return (error); } vget_finish_ref(vp, vs); return (0); } void vget_finish_ref(struct vnode *vp, enum vgetstate vs) { int old; VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); VNPASS(vp->v_holdcnt > 0, vp); VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); if (vs == VGET_USECOUNT) return; /* * We hold the vnode. If the usecount is 0 it will be utilized to keep * the vnode around. Otherwise someone else lended their hold count and * we have to drop ours. */ old = atomic_fetchadd_int(&vp->v_usecount, 1); VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); if (old != 0) { #ifdef INVARIANTS old = atomic_fetchadd_int(&vp->v_holdcnt, -1); VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); #else refcount_release(&vp->v_holdcnt); #endif } } void vref(struct vnode *vp) { enum vgetstate vs; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); vs = vget_prep(vp); vget_finish_ref(vp, vs); } void vrefact(struct vnode *vp) { CTR2(KTR_VFS, "%s: vp %p", __func__, vp); #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_usecount, 1); VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); #else refcount_acquire(&vp->v_usecount); #endif } void vlazy(struct vnode *vp) { struct mount *mp; VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); if ((vp->v_mflag & VMP_LAZYLIST) != 0) return; /* * We may get here for inactive routines after the vnode got doomed. */ if (VN_IS_DOOMED(vp)) return; mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); if ((vp->v_mflag & VMP_LAZYLIST) == 0) { vp->v_mflag |= VMP_LAZYLIST; TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize++; } mtx_unlock(&mp->mnt_listmtx); } /* * This routine is only meant to be called from vgonel prior to dooming * the vnode. */ static void vunlazy_gone(struct vnode *vp) { struct mount *mp; ASSERT_VOP_ELOCKED(vp, __func__); ASSERT_VI_LOCKED(vp, __func__); VNPASS(!VN_IS_DOOMED(vp), vp); if (vp->v_mflag & VMP_LAZYLIST) { mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); vp->v_mflag &= ~VMP_LAZYLIST; TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize--; mtx_unlock(&mp->mnt_listmtx); } } static void vdefer_inactive(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode without hold count", __func__)); if (VN_IS_DOOMED(vp)) { vdropl(vp); return; } if (vp->v_iflag & VI_DEFINACT) { VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); vdropl(vp); return; } if (vp->v_usecount > 0) { vp->v_iflag &= ~VI_OWEINACT; vdropl(vp); return; } vlazy(vp); vp->v_iflag |= VI_DEFINACT; VI_UNLOCK(vp); counter_u64_add(deferred_inact, 1); } static void vdefer_inactive_unlocked(struct vnode *vp) { VI_LOCK(vp); if ((vp->v_iflag & VI_OWEINACT) == 0) { vdropl(vp); return; } vdefer_inactive(vp); } enum vput_op { VRELE, VPUT, VUNREF }; /* * Handle ->v_usecount transitioning to 0. * * By releasing the last usecount we take ownership of the hold count which * provides liveness of the vnode, meaning we have to vdrop. * * For all vnodes we may need to perform inactive processing. It requires an * exclusive lock on the vnode, while it is legal to call here with only a * shared lock (or no locks). If locking the vnode in an expected manner fails, * inactive processing gets deferred to the syncer. * * XXX Some filesystems pass in an exclusively locked vnode and strongly depend * on the lock being held all the way until VOP_INACTIVE. This in particular * happens with UFS which adds half-constructed vnodes to the hash, where they * can be found by other code. */ static void vput_final(struct vnode *vp, enum vput_op func) { int error; bool want_unlock; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); VNPASS(vp->v_holdcnt > 0, vp); VI_LOCK(vp); /* * By the time we got here someone else might have transitioned * the count back to > 0. */ if (vp->v_usecount > 0) goto out; /* * If the vnode is doomed vgone already performed inactive processing * (if needed). */ if (VN_IS_DOOMED(vp)) goto out; if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) goto out; if (vp->v_iflag & VI_DOINGINACT) goto out; /* * Locking operations here will drop the interlock and possibly the * vnode lock, opening a window where the vnode can get doomed all the * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to * perform inactive. */ vp->v_iflag |= VI_OWEINACT; want_unlock = false; error = 0; switch (func) { case VRELE: switch (VOP_ISLOCKED(vp)) { case LK_EXCLUSIVE: break; case LK_EXCLOTHER: case 0: want_unlock = true; error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); VI_LOCK(vp); break; default: /* * The lock has at least one sharer, but we have no way * to conclude whether this is us. Play it safe and * defer processing. */ error = EAGAIN; break; } break; case VPUT: want_unlock = true; if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | LK_NOWAIT); VI_LOCK(vp); } break; case VUNREF: if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); VI_LOCK(vp); } break; } if (error == 0) { vinactive(vp); if (want_unlock) VOP_UNLOCK(vp); vdropl(vp); } else { vdefer_inactive(vp); } return; out: if (func == VPUT) VOP_UNLOCK(vp); vdropl(vp); } /* * Decrement ->v_usecount for a vnode. * * Releasing the last use count requires additional processing, see vput_final * above for details. * * Comment above each variant denotes lock state on entry and exit. */ /* * in: any * out: same as passed in */ void vrele(struct vnode *vp) { ASSERT_VI_UNLOCKED(vp, __func__); if (!refcount_release(&vp->v_usecount)) return; vput_final(vp, VRELE); } /* * in: locked * out: unlocked */ void vput(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, __func__); ASSERT_VI_UNLOCKED(vp, __func__); if (!refcount_release(&vp->v_usecount)) { VOP_UNLOCK(vp); return; } vput_final(vp, VPUT); } /* * in: locked * out: locked */ void vunref(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, __func__); ASSERT_VI_UNLOCKED(vp, __func__); if (!refcount_release(&vp->v_usecount)) return; vput_final(vp, VUNREF); } void vhold(struct vnode *vp) { int old; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); old = atomic_fetchadd_int(&vp->v_holdcnt, 1); VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, ("%s: wrong hold count %d", __func__, old)); if (old == 0) vn_freevnodes_dec(); } void vholdnz(struct vnode *vp) { CTR2(KTR_VFS, "%s: vp %p", __func__, vp); #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, ("%s: wrong hold count %d", __func__, old)); #else atomic_add_int(&vp->v_holdcnt, 1); #endif } /* * Grab a hold count unless the vnode is freed. * * Only use this routine if vfs smr is the only protection you have against * freeing the vnode. * * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag * is not set. After the flag is set the vnode becomes immutable to anyone but * the thread which managed to set the flag. * * It may be tempting to replace the loop with: * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); * if (count & VHOLD_NO_SMR) { * backpedal and error out; * } * * However, while this is more performant, it hinders debugging by eliminating * the previously mentioned invariant. */ bool vhold_smr(struct vnode *vp) { int count; VFS_SMR_ASSERT_ENTERED(); count = atomic_load_int(&vp->v_holdcnt); for (;;) { if (count & VHOLD_NO_SMR) { VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, ("non-zero hold count with flags %d\n", count)); return (false); } VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { if (count == 0) vn_freevnodes_dec(); return (true); } } } /* * Hold a free vnode for recycling. * * Note: vnode_init references this comment. * * Attempts to recycle only need the global vnode list lock and have no use for * SMR. * * However, vnodes get inserted into the global list before they get fully * initialized and stay there until UMA decides to free the memory. This in * particular means the target can be found before it becomes usable and after * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to * VHOLD_NO_SMR. * * Note: the vnode may gain more references after we transition the count 0->1. */ static bool vhold_recycle_free(struct vnode *vp) { int count; mtx_assert(&vnode_list_mtx, MA_OWNED); count = atomic_load_int(&vp->v_holdcnt); for (;;) { if (count & VHOLD_NO_SMR) { VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, ("non-zero hold count with flags %d\n", count)); return (false); } VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); if (count > 0) { return (false); } if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) { vn_freevnodes_dec(); return (true); } } } static void __noinline vdbatch_process(struct vdbatch *vd) { struct vnode *vp; int i; mtx_assert(&vd->lock, MA_OWNED); MPASS(curthread->td_pinned > 0); MPASS(vd->index == VDBATCH_SIZE); mtx_lock(&vnode_list_mtx); critical_enter(); freevnodes += vd->freevnodes; for (i = 0; i < VDBATCH_SIZE; i++) { vp = vd->tab[i]; TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); MPASS(vp->v_dbatchcpu != NOCPU); vp->v_dbatchcpu = NOCPU; } mtx_unlock(&vnode_list_mtx); vd->freevnodes = 0; bzero(vd->tab, sizeof(vd->tab)); vd->index = 0; critical_exit(); } static void vdbatch_enqueue(struct vnode *vp) { struct vdbatch *vd; ASSERT_VI_LOCKED(vp, __func__); VNASSERT(!VN_IS_DOOMED(vp), vp, ("%s: deferring requeue of a doomed vnode", __func__)); if (vp->v_dbatchcpu != NOCPU) { VI_UNLOCK(vp); return; } sched_pin(); vd = DPCPU_PTR(vd); mtx_lock(&vd->lock); MPASS(vd->index < VDBATCH_SIZE); MPASS(vd->tab[vd->index] == NULL); /* * A hack: we depend on being pinned so that we know what to put in * ->v_dbatchcpu. */ vp->v_dbatchcpu = curcpu; vd->tab[vd->index] = vp; vd->index++; VI_UNLOCK(vp); if (vd->index == VDBATCH_SIZE) vdbatch_process(vd); mtx_unlock(&vd->lock); sched_unpin(); } /* * This routine must only be called for vnodes which are about to be * deallocated. Supporting dequeue for arbitrary vndoes would require * validating that the locked batch matches. */ static void vdbatch_dequeue(struct vnode *vp) { struct vdbatch *vd; int i; short cpu; VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp, ("%s: called for a used vnode\n", __func__)); cpu = vp->v_dbatchcpu; if (cpu == NOCPU) return; vd = DPCPU_ID_PTR(cpu, vd); mtx_lock(&vd->lock); for (i = 0; i < vd->index; i++) { if (vd->tab[i] != vp) continue; vp->v_dbatchcpu = NOCPU; vd->index--; vd->tab[i] = vd->tab[vd->index]; vd->tab[vd->index] = NULL; break; } mtx_unlock(&vd->lock); /* * Either we dequeued the vnode above or the target CPU beat us to it. */ MPASS(vp->v_dbatchcpu == NOCPU); } /* * Drop the hold count of the vnode. If this is the last reference to * the vnode we place it on the free list unless it has been vgone'd * (marked VIRF_DOOMED) in which case we will free it. * * Because the vnode vm object keeps a hold reference on the vnode if * there is at least one resident non-cached page, the vnode cannot * leave the active list without the page cleanup done. */ static void vdrop_deactivate(struct vnode *vp) { struct mount *mp; ASSERT_VI_LOCKED(vp, __func__); /* * Mark a vnode as free: remove it from its active list * and put it up for recycling on the freelist. */ VNASSERT(!VN_IS_DOOMED(vp), vp, ("vdrop: returning doomed vnode")); VNASSERT(vp->v_op != NULL, vp, ("vdrop: vnode already reclaimed.")); VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, ("vnode with VI_OWEINACT set")); VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("vnode with VI_DEFINACT set")); if (vp->v_mflag & VMP_LAZYLIST) { mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST")); /* * Don't remove the vnode from the lazy list if another thread * has increased the hold count. It may have re-enqueued the * vnode to the lazy list and is now responsible for its * removal. */ if (vp->v_holdcnt == 0) { vp->v_mflag &= ~VMP_LAZYLIST; TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize--; } mtx_unlock(&mp->mnt_listmtx); } vdbatch_enqueue(vp); } static void __noinline vdropl_final(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNPASS(VN_IS_DOOMED(vp), vp); /* * Set the VHOLD_NO_SMR flag. * * We may be racing against vhold_smr. If they win we can just pretend * we never got this far, they will vdrop later. */ if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) { vn_freevnodes_inc(); VI_UNLOCK(vp); /* * We lost the aforementioned race. Any subsequent access is * invalid as they might have managed to vdropl on their own. */ return; } /* * Don't bump freevnodes as this one is going away. */ freevnode(vp); } void vdrop(struct vnode *vp) { ASSERT_VI_UNLOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (refcount_release_if_not_last(&vp->v_holdcnt)) return; VI_LOCK(vp); vdropl(vp); } void vdropl(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (!refcount_release(&vp->v_holdcnt)) { VI_UNLOCK(vp); return; } if (!VN_IS_DOOMED(vp)) { vn_freevnodes_inc(); vdrop_deactivate(vp); /* * Also unlocks the interlock. We can't assert on it as we * released our hold and by now the vnode might have been * freed. */ return; } vdropl_final(vp); } /* * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT * flags. DOINGINACT prevents us from recursing in calls to vinactive. */ static void vinactivef(struct vnode *vp) { struct vm_object *obj; ASSERT_VOP_ELOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, ("vinactive: recursed on VI_DOINGINACT")); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); vp->v_iflag |= VI_DOINGINACT; vp->v_iflag &= ~VI_OWEINACT; VI_UNLOCK(vp); /* * Before moving off the active list, we must be sure that any * modified pages are converted into the vnode's dirty * buffers, since these will no longer be checked once the * vnode is on the inactive list. * * The write-out of the dirty pages is asynchronous. At the * point that VOP_INACTIVE() is called, there could still be * pending I/O and dirty pages in the object. */ if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && vm_object_mightbedirty(obj)) { VM_OBJECT_WLOCK(obj); vm_object_page_clean(obj, 0, 0, 0); VM_OBJECT_WUNLOCK(obj); } VOP_INACTIVE(vp); VI_LOCK(vp); VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, ("vinactive: lost VI_DOINGINACT")); vp->v_iflag &= ~VI_DOINGINACT; } void vinactive(struct vnode *vp) { ASSERT_VOP_ELOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if ((vp->v_iflag & VI_OWEINACT) == 0) return; if (vp->v_iflag & VI_DOINGINACT) return; if (vp->v_usecount > 0) { vp->v_iflag &= ~VI_OWEINACT; return; } vinactivef(vp); } /* * Remove any vnodes in the vnode table belonging to mount point mp. * * If FORCECLOSE is not specified, there should not be any active ones, * return error if any are found (nb: this is a user error, not a * system error). If FORCECLOSE is specified, detach any active vnodes * that are found. * * If WRITECLOSE is set, only flush out regular file vnodes open for * writing. * * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. * * `rootrefs' specifies the base reference count for the root vnode * of this filesystem. The root vnode is considered busy if its * v_usecount exceeds this value. On a successful return, vflush(, td) * will call vrele() on the root vnode exactly rootrefs times. * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must * be zero. */ #ifdef DIAGNOSTIC static int busyprt = 0; /* print out busy vnodes */ SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); #endif int vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) { struct vnode *vp, *mvp, *rootvp = NULL; struct vattr vattr; int busy = 0, error; CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, rootrefs, flags); if (rootrefs > 0) { KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, ("vflush: bad args")); /* * Get the filesystem root vnode. We can vput() it * immediately, since with rootrefs > 0, it won't go away. */ if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", __func__, error); return (error); } vput(rootvp); } loop: MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { vholdl(vp); error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); if (error) { vdrop(vp); MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); goto loop; } /* * Skip over a vnodes marked VV_SYSTEM. */ if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { VOP_UNLOCK(vp); vdrop(vp); continue; } /* * If WRITECLOSE is set, flush out unlinked but still open * files (even if open only for reading) and regular file * vnodes open for writing. */ if (flags & WRITECLOSE) { if (vp->v_object != NULL) { VM_OBJECT_WLOCK(vp->v_object); vm_object_page_clean(vp->v_object, 0, 0, 0); VM_OBJECT_WUNLOCK(vp->v_object); } error = VOP_FSYNC(vp, MNT_WAIT, td); if (error != 0) { VOP_UNLOCK(vp); vdrop(vp); MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); return (error); } error = VOP_GETATTR(vp, &vattr, td->td_ucred); VI_LOCK(vp); if ((vp->v_type == VNON || (error == 0 && vattr.va_nlink > 0)) && (vp->v_writecount <= 0 || vp->v_type != VREG)) { VOP_UNLOCK(vp); vdropl(vp); continue; } } else VI_LOCK(vp); /* * With v_usecount == 0, all we need to do is clear out the * vnode data structures and we are done. * * If FORCECLOSE is set, forcibly close the vnode. */ if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { vgonel(vp); } else { busy++; #ifdef DIAGNOSTIC if (busyprt) vn_printf(vp, "vflush: busy vnode "); #endif } VOP_UNLOCK(vp); vdropl(vp); } if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { /* * If just the root vnode is busy, and if its refcount * is equal to `rootrefs', then go ahead and kill it. */ VI_LOCK(rootvp); KASSERT(busy > 0, ("vflush: not busy")); VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, ("vflush: usecount %d < rootrefs %d", rootvp->v_usecount, rootrefs)); if (busy == 1 && rootvp->v_usecount == rootrefs) { VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); vgone(rootvp); VOP_UNLOCK(rootvp); busy = 0; } else VI_UNLOCK(rootvp); } if (busy) { CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, busy); return (EBUSY); } for (; rootrefs > 0; rootrefs--) vrele(rootvp); return (0); } /* * Recycle an unused vnode to the front of the free list. */ int vrecycle(struct vnode *vp) { int recycled; VI_LOCK(vp); recycled = vrecyclel(vp); VI_UNLOCK(vp); return (recycled); } /* * vrecycle, with the vp interlock held. */ int vrecyclel(struct vnode *vp) { int recycled; ASSERT_VOP_ELOCKED(vp, __func__); ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); recycled = 0; if (vp->v_usecount == 0) { recycled = 1; vgonel(vp); } return (recycled); } /* * Eliminate all activity associated with a vnode * in preparation for reuse. */ void vgone(struct vnode *vp) { VI_LOCK(vp); vgonel(vp); VI_UNLOCK(vp); } static void notify_lowervp_vfs_dummy(struct mount *mp __unused, struct vnode *lowervp __unused) { } /* * Notify upper mounts about reclaimed or unlinked vnode. */ void vfs_notify_upper(struct vnode *vp, int event) { static struct vfsops vgonel_vfsops = { .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, }; struct mount *mp, *ump, *mmp; mp = vp->v_mount; if (mp == NULL) return; if (TAILQ_EMPTY(&mp->mnt_uppers)) return; mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); mmp->mnt_op = &vgonel_vfsops; mmp->mnt_kern_flag |= MNTK_MARKER; MNT_ILOCK(mp); mp->mnt_kern_flag |= MNTK_VGONE_UPPER; for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { ump = TAILQ_NEXT(ump, mnt_upper_link); continue; } TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); MNT_IUNLOCK(mp); switch (event) { case VFS_NOTIFY_UPPER_RECLAIM: VFS_RECLAIM_LOWERVP(ump, vp); break; case VFS_NOTIFY_UPPER_UNLINK: VFS_UNLINK_LOWERVP(ump, vp); break; default: KASSERT(0, ("invalid event %d", event)); break; } MNT_ILOCK(mp); ump = TAILQ_NEXT(mmp, mnt_upper_link); TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); } free(mmp, M_TEMP); mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; wakeup(&mp->mnt_uppers); } MNT_IUNLOCK(mp); } /* * vgone, with the vp interlock held. */ static void vgonel(struct vnode *vp) { struct thread *td; struct mount *mp; vm_object_t object; bool active, doinginact, oweinact; ASSERT_VOP_ELOCKED(vp, "vgonel"); ASSERT_VI_LOCKED(vp, "vgonel"); VNASSERT(vp->v_holdcnt, vp, ("vgonel: vp %p has no reference.", vp)); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); td = curthread; /* * Don't vgonel if we're already doomed. */ if (vp->v_irflag & VIRF_DOOMED) return; /* * Paired with freevnode. */ vn_seqc_write_begin_locked(vp); vunlazy_gone(vp); vp->v_irflag |= VIRF_DOOMED; /* * Check to see if the vnode is in use. If so, we have to * call VOP_CLOSE() and VOP_INACTIVE(). * * It could be that VOP_INACTIVE() requested reclamation, in * which case we should avoid recursion, so check * VI_DOINGINACT. This is not precise but good enough. */ active = vp->v_usecount > 0; oweinact = (vp->v_iflag & VI_OWEINACT) != 0; doinginact = (vp->v_iflag & VI_DOINGINACT) != 0; /* * If we need to do inactive VI_OWEINACT will be set. */ if (vp->v_iflag & VI_DEFINACT) { VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); vp->v_iflag &= ~VI_DEFINACT; vdropl(vp); } else { VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); VI_UNLOCK(vp); } cache_purge_vgone(vp); vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); /* * If purging an active vnode, it must be closed and * deactivated before being reclaimed. */ if (active) VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); if ((oweinact || active) && !doinginact) { VI_LOCK(vp); vinactivef(vp); VI_UNLOCK(vp); } if (vp->v_type == VSOCK) vfs_unp_reclaim(vp); /* * Clean out any buffers associated with the vnode. * If the flush fails, just toss the buffers. */ mp = NULL; if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) (void) vn_start_secondary_write(vp, &mp, V_WAIT); if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { while (vinvalbuf(vp, 0, 0, 0) != 0) ; } BO_LOCK(&vp->v_bufobj); KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && vp->v_bufobj.bo_dirty.bv_cnt == 0 && TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && vp->v_bufobj.bo_clean.bv_cnt == 0, ("vp %p bufobj not invalidated", vp)); /* * For VMIO bufobj, BO_DEAD is set later, or in * vm_object_terminate() after the object's page queue is * flushed. */ object = vp->v_bufobj.bo_object; if (object == NULL) vp->v_bufobj.bo_flag |= BO_DEAD; BO_UNLOCK(&vp->v_bufobj); /* * Handle the VM part. Tmpfs handles v_object on its own (the * OBJT_VNODE check). Nullfs or other bypassing filesystems * should not touch the object borrowed from the lower vnode * (the handle check). */ if (object != NULL && object->type == OBJT_VNODE && object->handle == vp) vnode_destroy_vobject(vp); /* * Reclaim the vnode. */ if (VOP_RECLAIM(vp)) panic("vgone: cannot reclaim"); if (mp != NULL) vn_finished_secondary_write(mp); VNASSERT(vp->v_object == NULL, vp, ("vop_reclaim left v_object vp=%p", vp)); /* * Clear the advisory locks and wake up waiting threads. */ (void)VOP_ADVLOCKPURGE(vp); vp->v_lockf = NULL; /* * Delete from old mount point vnode list. */ delmntque(vp); /* * Done with purge, reset to the standard lock and invalidate * the vnode. */ VI_LOCK(vp); vp->v_vnlock = &vp->v_lock; vp->v_op = &dead_vnodeops; vp->v_type = VBAD; } /* * Print out a description of a vnode. */ static const char * const typename[] = {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", "VMARKER"}; _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, "new hold count flag not added to vn_printf"); void vn_printf(struct vnode *vp, const char *fmt, ...) { va_list ap; char buf[256], buf2[16]; u_long flags; u_int holdcnt; va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf("%p: ", (void *)vp); printf("type %s\n", typename[vp->v_type]); holdcnt = atomic_load_int(&vp->v_holdcnt); printf(" usecount %d, writecount %d, refcount %d seqc users %d", vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS, vp->v_seqc_users); switch (vp->v_type) { case VDIR: printf(" mountedhere %p\n", vp->v_mountedhere); break; case VCHR: printf(" rdev %p\n", vp->v_rdev); break; case VSOCK: printf(" socket %p\n", vp->v_unpcb); break; case VFIFO: printf(" fifoinfo %p\n", vp->v_fifoinfo); break; default: printf("\n"); break; } buf[0] = '\0'; buf[1] = '\0'; if (holdcnt & VHOLD_NO_SMR) strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); printf(" hold count flags (%s)\n", buf + 1); buf[0] = '\0'; buf[1] = '\0'; if (vp->v_irflag & VIRF_DOOMED) strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); if (vp->v_irflag & VIRF_PGREAD) strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); flags = vp->v_irflag & ~(VIRF_DOOMED | VIRF_PGREAD); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_vflag & VV_ROOT) strlcat(buf, "|VV_ROOT", sizeof(buf)); if (vp->v_vflag & VV_ISTTY) strlcat(buf, "|VV_ISTTY", sizeof(buf)); if (vp->v_vflag & VV_NOSYNC) strlcat(buf, "|VV_NOSYNC", sizeof(buf)); if (vp->v_vflag & VV_ETERNALDEV) strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); if (vp->v_vflag & VV_CACHEDLABEL) strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); if (vp->v_vflag & VV_VMSIZEVNLOCK) strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); if (vp->v_vflag & VV_COPYONWRITE) strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); if (vp->v_vflag & VV_SYSTEM) strlcat(buf, "|VV_SYSTEM", sizeof(buf)); if (vp->v_vflag & VV_PROCDEP) strlcat(buf, "|VV_PROCDEP", sizeof(buf)); if (vp->v_vflag & VV_NOKNOTE) strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); if (vp->v_vflag & VV_DELETED) strlcat(buf, "|VV_DELETED", sizeof(buf)); if (vp->v_vflag & VV_MD) strlcat(buf, "|VV_MD", sizeof(buf)); if (vp->v_vflag & VV_FORCEINSMQ) strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); if (vp->v_vflag & VV_READLINK) strlcat(buf, "|VV_READLINK", sizeof(buf)); flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_iflag & VI_TEXT_REF) strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); if (vp->v_iflag & VI_MOUNT) strlcat(buf, "|VI_MOUNT", sizeof(buf)); if (vp->v_iflag & VI_DOINGINACT) strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); if (vp->v_iflag & VI_OWEINACT) strlcat(buf, "|VI_OWEINACT", sizeof(buf)); if (vp->v_iflag & VI_DEFINACT) strlcat(buf, "|VI_DEFINACT", sizeof(buf)); flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT | VI_OWEINACT | VI_DEFINACT); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_mflag & VMP_LAZYLIST) strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); flags = vp->v_mflag & ~(VMP_LAZYLIST); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } printf(" flags (%s)\n", buf + 1); if (mtx_owned(VI_MTX(vp))) printf(" VI_LOCKed"); if (vp->v_object != NULL) printf(" v_object %p ref %d pages %d " "cleanbuf %d dirtybuf %d\n", vp->v_object, vp->v_object->ref_count, vp->v_object->resident_page_count, vp->v_bufobj.bo_clean.bv_cnt, vp->v_bufobj.bo_dirty.bv_cnt); printf(" "); lockmgr_printinfo(vp->v_vnlock); if (vp->v_data != NULL) VOP_PRINT(vp); } #ifdef DDB /* * List all of the locked vnodes in the system. * Called when debugging the kernel. */ DB_SHOW_COMMAND(lockedvnods, lockedvnodes) { struct mount *mp; struct vnode *vp; /* * Note: because this is DDB, we can't obey the locking semantics * for these structures, which means we could catch an inconsistent * state and dereference a nasty pointer. Not much to be done * about that. */ db_printf("Locked vnodes\n"); TAILQ_FOREACH(mp, &mountlist, mnt_list) { TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) vn_printf(vp, "vnode "); } } } /* * Show details about the given vnode. */ DB_SHOW_COMMAND(vnode, db_show_vnode) { struct vnode *vp; if (!have_addr) return; vp = (struct vnode *)addr; vn_printf(vp, "vnode "); } /* * Show details about the given mount point. */ DB_SHOW_COMMAND(mount, db_show_mount) { struct mount *mp; struct vfsopt *opt; struct statfs *sp; struct vnode *vp; char buf[512]; uint64_t mflags; u_int flags; if (!have_addr) { /* No address given, print short info about all mount points. */ TAILQ_FOREACH(mp, &mountlist, mnt_list) { db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); if (db_pager_quit) break; } db_printf("\nMore info: show mount \n"); return; } mp = (struct mount *)addr; db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); buf[0] = '\0'; mflags = mp->mnt_flag; #define MNT_FLAG(flag) do { \ if (mflags & (flag)) { \ if (buf[0] != '\0') \ strlcat(buf, ", ", sizeof(buf)); \ strlcat(buf, (#flag) + 4, sizeof(buf)); \ mflags &= ~(flag); \ } \ } while (0) MNT_FLAG(MNT_RDONLY); MNT_FLAG(MNT_SYNCHRONOUS); MNT_FLAG(MNT_NOEXEC); MNT_FLAG(MNT_NOSUID); MNT_FLAG(MNT_NFS4ACLS); MNT_FLAG(MNT_UNION); MNT_FLAG(MNT_ASYNC); MNT_FLAG(MNT_SUIDDIR); MNT_FLAG(MNT_SOFTDEP); MNT_FLAG(MNT_NOSYMFOLLOW); MNT_FLAG(MNT_GJOURNAL); MNT_FLAG(MNT_MULTILABEL); MNT_FLAG(MNT_ACLS); MNT_FLAG(MNT_NOATIME); MNT_FLAG(MNT_NOCLUSTERR); MNT_FLAG(MNT_NOCLUSTERW); MNT_FLAG(MNT_SUJ); MNT_FLAG(MNT_EXRDONLY); MNT_FLAG(MNT_EXPORTED); MNT_FLAG(MNT_DEFEXPORTED); MNT_FLAG(MNT_EXPORTANON); MNT_FLAG(MNT_EXKERB); MNT_FLAG(MNT_EXPUBLIC); MNT_FLAG(MNT_LOCAL); MNT_FLAG(MNT_QUOTA); MNT_FLAG(MNT_ROOTFS); MNT_FLAG(MNT_USER); MNT_FLAG(MNT_IGNORE); MNT_FLAG(MNT_UPDATE); MNT_FLAG(MNT_DELEXPORT); MNT_FLAG(MNT_RELOAD); MNT_FLAG(MNT_FORCE); MNT_FLAG(MNT_SNAPSHOT); MNT_FLAG(MNT_BYFSID); #undef MNT_FLAG if (mflags != 0) { if (buf[0] != '\0') strlcat(buf, ", ", sizeof(buf)); snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "0x%016jx", mflags); } db_printf(" mnt_flag = %s\n", buf); buf[0] = '\0'; flags = mp->mnt_kern_flag; #define MNT_KERN_FLAG(flag) do { \ if (flags & (flag)) { \ if (buf[0] != '\0') \ strlcat(buf, ", ", sizeof(buf)); \ strlcat(buf, (#flag) + 5, sizeof(buf)); \ flags &= ~(flag); \ } \ } while (0) MNT_KERN_FLAG(MNTK_UNMOUNTF); MNT_KERN_FLAG(MNTK_ASYNC); MNT_KERN_FLAG(MNTK_SOFTDEP); MNT_KERN_FLAG(MNTK_DRAINING); MNT_KERN_FLAG(MNTK_REFEXPIRE); MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); MNT_KERN_FLAG(MNTK_SHARED_WRITES); MNT_KERN_FLAG(MNTK_NO_IOPF); MNT_KERN_FLAG(MNTK_VGONE_UPPER); MNT_KERN_FLAG(MNTK_VGONE_WAITER); MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); MNT_KERN_FLAG(MNTK_MARKER); MNT_KERN_FLAG(MNTK_USES_BCACHE); MNT_KERN_FLAG(MNTK_FPLOOKUP); MNT_KERN_FLAG(MNTK_NOASYNC); MNT_KERN_FLAG(MNTK_UNMOUNT); MNT_KERN_FLAG(MNTK_MWAIT); MNT_KERN_FLAG(MNTK_SUSPEND); MNT_KERN_FLAG(MNTK_SUSPEND2); MNT_KERN_FLAG(MNTK_SUSPENDED); MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); MNT_KERN_FLAG(MNTK_NOKNOTE); #undef MNT_KERN_FLAG if (flags != 0) { if (buf[0] != '\0') strlcat(buf, ", ", sizeof(buf)); snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "0x%08x", flags); } db_printf(" mnt_kern_flag = %s\n", buf); db_printf(" mnt_opt = "); opt = TAILQ_FIRST(mp->mnt_opt); if (opt != NULL) { db_printf("%s", opt->name); opt = TAILQ_NEXT(opt, link); while (opt != NULL) { db_printf(", %s", opt->name); opt = TAILQ_NEXT(opt, link); } } db_printf("\n"); sp = &mp->mnt_stat; db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); db_printf(" mnt_cred = { uid=%u ruid=%u", (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); if (jailed(mp->mnt_cred)) db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); db_printf(" }\n"); db_printf(" mnt_ref = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); db_printf(" mnt_gen = %d\n", mp->mnt_gen); db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); db_printf(" mnt_lazyvnodelistsize = %d\n", mp->mnt_lazyvnodelistsize); db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); db_printf(" mnt_lockref = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); db_printf(" mnt_secondary_accwrites = %d\n", mp->mnt_secondary_accwrites); db_printf(" mnt_gjprovider = %s\n", mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); db_printf("\n\nList of active vnodes\n"); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { vn_printf(vp, "vnode "); if (db_pager_quit) break; } } db_printf("\n\nList of inactive vnodes\n"); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { vn_printf(vp, "vnode "); if (db_pager_quit) break; } } } #endif /* DDB */ /* * Fill in a struct xvfsconf based on a struct vfsconf. */ static int vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) { struct xvfsconf xvfsp; bzero(&xvfsp, sizeof(xvfsp)); strcpy(xvfsp.vfc_name, vfsp->vfc_name); xvfsp.vfc_typenum = vfsp->vfc_typenum; xvfsp.vfc_refcount = vfsp->vfc_refcount; xvfsp.vfc_flags = vfsp->vfc_flags; /* * These are unused in userland, we keep them * to not break binary compatibility. */ xvfsp.vfc_vfsops = NULL; xvfsp.vfc_next = NULL; return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); } #ifdef COMPAT_FREEBSD32 struct xvfsconf32 { uint32_t vfc_vfsops; char vfc_name[MFSNAMELEN]; int32_t vfc_typenum; int32_t vfc_refcount; int32_t vfc_flags; uint32_t vfc_next; }; static int vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) { struct xvfsconf32 xvfsp; bzero(&xvfsp, sizeof(xvfsp)); strcpy(xvfsp.vfc_name, vfsp->vfc_name); xvfsp.vfc_typenum = vfsp->vfc_typenum; xvfsp.vfc_refcount = vfsp->vfc_refcount; xvfsp.vfc_flags = vfsp->vfc_flags; return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); } #endif /* * Top level filesystem related information gathering. */ static int sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) { struct vfsconf *vfsp; int error; error = 0; vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { #ifdef COMPAT_FREEBSD32 if (req->flags & SCTL_MASK32) error = vfsconf2x32(req, vfsp); else #endif error = vfsconf2x(req, vfsp); if (error) break; } vfsconf_sunlock(); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, "S,xvfsconf", "List of all configured filesystems"); #ifndef BURN_BRIDGES static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); static int vfs_sysctl(SYSCTL_HANDLER_ARGS) { int *name = (int *)arg1 - 1; /* XXX */ u_int namelen = arg2 + 1; /* XXX */ struct vfsconf *vfsp; log(LOG_WARNING, "userland calling deprecated sysctl, " "please rebuild world\n"); #if 1 || defined(COMPAT_PRELITE2) /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ if (namelen == 1) return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); #endif switch (name[1]) { case VFS_MAXTYPENUM: if (namelen != 2) return (ENOTDIR); return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); case VFS_CONF: if (namelen != 3) return (ENOTDIR); /* overloaded */ vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { if (vfsp->vfc_typenum == name[2]) break; } vfsconf_sunlock(); if (vfsp == NULL) return (EOPNOTSUPP); #ifdef COMPAT_FREEBSD32 if (req->flags & SCTL_MASK32) return (vfsconf2x32(req, vfsp)); else #endif return (vfsconf2x(req, vfsp)); } return (EOPNOTSUPP); } static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE, vfs_sysctl, "Generic filesystem"); #if 1 || defined(COMPAT_PRELITE2) static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) { int error; struct vfsconf *vfsp; struct ovfsconf ovfs; vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { bzero(&ovfs, sizeof(ovfs)); ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ strcpy(ovfs.vfc_name, vfsp->vfc_name); ovfs.vfc_index = vfsp->vfc_typenum; ovfs.vfc_refcount = vfsp->vfc_refcount; ovfs.vfc_flags = vfsp->vfc_flags; error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); if (error != 0) { vfsconf_sunlock(); return (error); } } vfsconf_sunlock(); return (0); } #endif /* 1 || COMPAT_PRELITE2 */ #endif /* !BURN_BRIDGES */ #define KINFO_VNODESLOP 10 #ifdef notyet /* * Dump vnode list (via sysctl). */ /* ARGSUSED */ static int sysctl_vnode(SYSCTL_HANDLER_ARGS) { struct xvnode *xvn; struct mount *mp; struct vnode *vp; int error, len, n; /* * Stale numvnodes access is not fatal here. */ req->lock = 0; len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; if (!req->oldptr) /* Make an estimate */ return (SYSCTL_OUT(req, 0, len)); error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); n = 0; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) continue; MNT_ILOCK(mp); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (n == len) break; vref(vp); xvn[n].xv_size = sizeof *xvn; xvn[n].xv_vnode = vp; xvn[n].xv_id = 0; /* XXX compat */ #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field XV_COPY(usecount); XV_COPY(writecount); XV_COPY(holdcnt); XV_COPY(mount); XV_COPY(numoutput); XV_COPY(type); #undef XV_COPY xvn[n].xv_flag = vp->v_vflag; switch (vp->v_type) { case VREG: case VDIR: case VLNK: break; case VBLK: case VCHR: if (vp->v_rdev == NULL) { vrele(vp); continue; } xvn[n].xv_dev = dev2udev(vp->v_rdev); break; case VSOCK: xvn[n].xv_socket = vp->v_socket; break; case VFIFO: xvn[n].xv_fifo = vp->v_fifoinfo; break; case VNON: case VBAD: default: /* shouldn't happen? */ vrele(vp); continue; } vrele(vp); ++n; } MNT_IUNLOCK(mp); mtx_lock(&mountlist_mtx); vfs_unbusy(mp); if (n == len) break; } mtx_unlock(&mountlist_mtx); error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); free(xvn, M_TEMP); return (error); } SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", ""); #endif static void unmount_or_warn(struct mount *mp) { int error; error = dounmount(mp, MNT_FORCE, curthread); if (error != 0) { printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); if (error == EBUSY) printf("BUSY)\n"); else printf("%d)\n", error); } } /* * Unmount all filesystems. The list is traversed in reverse order * of mounting to avoid dependencies. */ void vfs_unmountall(void) { struct mount *mp, *tmp; CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); /* * Since this only runs when rebooting, it is not interlocked. */ TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { vfs_ref(mp); /* * Forcibly unmounting "/dev" before "/" would prevent clean * unmount of the latter. */ if (mp == rootdevmp) continue; unmount_or_warn(mp); } if (rootdevmp != NULL) unmount_or_warn(rootdevmp); } static void vfs_deferred_inactive(struct vnode *vp, int lkflags) { ASSERT_VI_LOCKED(vp, __func__); VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); if ((vp->v_iflag & VI_OWEINACT) == 0) { vdropl(vp); return; } if (vn_lock(vp, lkflags) == 0) { VI_LOCK(vp); vinactive(vp); VOP_UNLOCK(vp); vdropl(vp); return; } vdefer_inactive_unlocked(vp); } static int vfs_periodic_inactive_filter(struct vnode *vp, void *arg) { return (vp->v_iflag & VI_DEFINACT); } static void __noinline vfs_periodic_inactive(struct mount *mp, int flags) { struct vnode *vp, *mvp; int lkflags; lkflags = LK_EXCLUSIVE | LK_INTERLOCK; if (flags != MNT_WAIT) lkflags |= LK_NOWAIT; MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { if ((vp->v_iflag & VI_DEFINACT) == 0) { VI_UNLOCK(vp); continue; } vp->v_iflag &= ~VI_DEFINACT; vfs_deferred_inactive(vp, lkflags); } } static inline bool vfs_want_msync(struct vnode *vp) { struct vm_object *obj; /* * This test may be performed without any locks held. * We rely on vm_object's type stability. */ if (vp->v_vflag & VV_NOSYNC) return (false); obj = vp->v_object; return (obj != NULL && vm_object_mightbedirty(obj)); } static int vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) { if (vp->v_vflag & VV_NOSYNC) return (false); if (vp->v_iflag & VI_DEFINACT) return (true); return (vfs_want_msync(vp)); } static void __noinline vfs_periodic_msync_inactive(struct mount *mp, int flags) { struct vnode *vp, *mvp; struct vm_object *obj; int lkflags, objflags; bool seen_defer; lkflags = LK_EXCLUSIVE | LK_INTERLOCK; if (flags != MNT_WAIT) { lkflags |= LK_NOWAIT; objflags = OBJPC_NOSYNC; } else { objflags = OBJPC_SYNC; } MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { seen_defer = false; if (vp->v_iflag & VI_DEFINACT) { vp->v_iflag &= ~VI_DEFINACT; seen_defer = true; } if (!vfs_want_msync(vp)) { if (seen_defer) vfs_deferred_inactive(vp, lkflags); else VI_UNLOCK(vp); continue; } if (vget(vp, lkflags) == 0) { obj = vp->v_object; if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { VM_OBJECT_WLOCK(obj); vm_object_page_clean(obj, 0, 0, objflags); VM_OBJECT_WUNLOCK(obj); } vput(vp); if (seen_defer) vdrop(vp); } else { if (seen_defer) vdefer_inactive_unlocked(vp); } } } void vfs_periodic(struct mount *mp, int flags) { CTR2(KTR_VFS, "%s: mp %p", __func__, mp); if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) vfs_periodic_inactive(mp, flags); else vfs_periodic_msync_inactive(mp, flags); } static void destroy_vpollinfo_free(struct vpollinfo *vi) { knlist_destroy(&vi->vpi_selinfo.si_note); mtx_destroy(&vi->vpi_lock); free(vi, M_VNODEPOLL); } static void destroy_vpollinfo(struct vpollinfo *vi) { knlist_clear(&vi->vpi_selinfo.si_note, 1); seldrain(&vi->vpi_selinfo); destroy_vpollinfo_free(vi); } /* * Initialize per-vnode helper structure to hold poll-related state. */ void v_addpollinfo(struct vnode *vp) { struct vpollinfo *vi; if (vp->v_pollinfo != NULL) return; vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO); mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, vfs_knlunlock, vfs_knl_assert_lock); VI_LOCK(vp); if (vp->v_pollinfo != NULL) { VI_UNLOCK(vp); destroy_vpollinfo_free(vi); return; } vp->v_pollinfo = vi; VI_UNLOCK(vp); } /* * Record a process's interest in events which might happen to * a vnode. Because poll uses the historic select-style interface * internally, this routine serves as both the ``check for any * pending events'' and the ``record my interest in future events'' * functions. (These are done together, while the lock is held, * to avoid race conditions.) */ int vn_pollrecord(struct vnode *vp, struct thread *td, int events) { v_addpollinfo(vp); mtx_lock(&vp->v_pollinfo->vpi_lock); if (vp->v_pollinfo->vpi_revents & events) { /* * This leaves events we are not interested * in available for the other process which * which presumably had requested them * (otherwise they would never have been * recorded). */ events &= vp->v_pollinfo->vpi_revents; vp->v_pollinfo->vpi_revents &= ~events; mtx_unlock(&vp->v_pollinfo->vpi_lock); return (events); } vp->v_pollinfo->vpi_events |= events; selrecord(td, &vp->v_pollinfo->vpi_selinfo); mtx_unlock(&vp->v_pollinfo->vpi_lock); return (0); } /* * Routine to create and manage a filesystem syncer vnode. */ #define sync_close ((int (*)(struct vop_close_args *))nullop) static int sync_fsync(struct vop_fsync_args *); static int sync_inactive(struct vop_inactive_args *); static int sync_reclaim(struct vop_reclaim_args *); static struct vop_vector sync_vnodeops = { .vop_bypass = VOP_EOPNOTSUPP, .vop_close = sync_close, /* close */ .vop_fsync = sync_fsync, /* fsync */ .vop_inactive = sync_inactive, /* inactive */ .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ .vop_reclaim = sync_reclaim, /* reclaim */ .vop_lock1 = vop_stdlock, /* lock */ .vop_unlock = vop_stdunlock, /* unlock */ .vop_islocked = vop_stdislocked, /* islocked */ }; VFS_VOP_VECTOR_REGISTER(sync_vnodeops); /* * Create a new filesystem syncer vnode for the specified mount point. */ void vfs_allocate_syncvnode(struct mount *mp) { struct vnode *vp; struct bufobj *bo; static long start, incr, next; int error; /* Allocate a new vnode */ error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); if (error != 0) panic("vfs_allocate_syncvnode: getnewvnode() failed"); vp->v_type = VNON; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vp->v_vflag |= VV_FORCEINSMQ; error = insmntque(vp, mp); if (error != 0) panic("vfs_allocate_syncvnode: insmntque() failed"); vp->v_vflag &= ~VV_FORCEINSMQ; VOP_UNLOCK(vp); /* * Place the vnode onto the syncer worklist. We attempt to * scatter them about on the list so that they will go off * at evenly distributed times even if all the filesystems * are mounted at once. */ next += incr; if (next == 0 || next > syncer_maxdelay) { start /= 2; incr /= 2; if (start == 0) { start = syncer_maxdelay / 2; incr = syncer_maxdelay; } next = start; } bo = &vp->v_bufobj; BO_LOCK(bo); vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ mtx_lock(&sync_mtx); sync_vnode_count++; if (mp->mnt_syncer == NULL) { mp->mnt_syncer = vp; vp = NULL; } mtx_unlock(&sync_mtx); BO_UNLOCK(bo); if (vp != NULL) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vgone(vp); vput(vp); } } void vfs_deallocate_syncvnode(struct mount *mp) { struct vnode *vp; mtx_lock(&sync_mtx); vp = mp->mnt_syncer; if (vp != NULL) mp->mnt_syncer = NULL; mtx_unlock(&sync_mtx); if (vp != NULL) vrele(vp); } /* * Do a lazy sync of the filesystem. */ static int sync_fsync(struct vop_fsync_args *ap) { struct vnode *syncvp = ap->a_vp; struct mount *mp = syncvp->v_mount; int error, save; struct bufobj *bo; /* * We only need to do something if this is a lazy evaluation. */ if (ap->a_waitfor != MNT_LAZY) return (0); /* * Move ourselves to the back of the sync list. */ bo = &syncvp->v_bufobj; BO_LOCK(bo); vn_syncer_add_to_worklist(bo, syncdelay); BO_UNLOCK(bo); /* * Walk the list of vnodes pushing all that are dirty and * not already on the sync list. */ if (vfs_busy(mp, MBF_NOWAIT) != 0) return (0); if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { vfs_unbusy(mp); return (0); } save = curthread_pflags_set(TDP_SYNCIO); /* * The filesystem at hand may be idle with free vnodes stored in the * batch. Return them instead of letting them stay there indefinitely. */ vfs_periodic(mp, MNT_NOWAIT); error = VFS_SYNC(mp, MNT_LAZY); curthread_pflags_restore(save); vn_finished_write(mp); vfs_unbusy(mp); return (error); } /* * The syncer vnode is no referenced. */ static int sync_inactive(struct vop_inactive_args *ap) { vgone(ap->a_vp); return (0); } /* * The syncer vnode is no longer needed and is being decommissioned. * * Modifications to the worklist must be protected by sync_mtx. */ static int sync_reclaim(struct vop_reclaim_args *ap) { struct vnode *vp = ap->a_vp; struct bufobj *bo; bo = &vp->v_bufobj; BO_LOCK(bo); mtx_lock(&sync_mtx); if (vp->v_mount->mnt_syncer == vp) vp->v_mount->mnt_syncer = NULL; if (bo->bo_flag & BO_ONWORKLST) { LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; sync_vnode_count--; bo->bo_flag &= ~BO_ONWORKLST; } mtx_unlock(&sync_mtx); BO_UNLOCK(bo); return (0); } int vn_need_pageq_flush(struct vnode *vp) { struct vm_object *obj; int need; MPASS(mtx_owned(VI_MTX(vp))); need = 0; if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && vm_object_mightbedirty(obj)) need = 1; return (need); } /* * Check if vnode represents a disk device */ bool vn_isdisk_error(struct vnode *vp, int *errp) { int error; if (vp->v_type != VCHR) { error = ENOTBLK; goto out; } error = 0; dev_lock(); if (vp->v_rdev == NULL) error = ENXIO; else if (vp->v_rdev->si_devsw == NULL) error = ENXIO; else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) error = ENOTBLK; dev_unlock(); out: *errp = error; return (error == 0); } bool vn_isdisk(struct vnode *vp) { int error; return (vn_isdisk_error(vp, &error)); } /* * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see * the comment above cache_fplookup for details. */ int vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred) { int error; VFS_SMR_ASSERT_ENTERED(); /* Check the owner. */ if (cred->cr_uid == file_uid) { if (file_mode & S_IXUSR) return (0); goto out_error; } /* Otherwise, check the groups (first match) */ if (groupmember(file_gid, cred)) { if (file_mode & S_IXGRP) return (0); goto out_error; } /* Otherwise, check everyone else. */ if (file_mode & S_IXOTH) return (0); out_error: /* * Permission check failed, but it is possible denial will get overwritten * (e.g., when root is traversing through a 700 directory owned by someone * else). * * vaccess() calls priv_check_cred which in turn can descent into MAC * modules overriding this result. It's quite unclear what semantics * are allowed for them to operate, thus for safety we don't call them * from within the SMR section. This also means if any such modules * are present, we have to let the regular lookup decide. */ error = priv_check_cred_vfs_lookup_nomac(cred); switch (error) { case 0: return (0); case EAGAIN: /* * MAC modules present. */ return (EAGAIN); case EPERM: return (EACCES); default: return (error); } } /* * Common filesystem object access control check routine. Accepts a * vnode's type, "mode", uid and gid, requested access mode, and credentials. * Returns 0 on success, or an errno on failure. */ int vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, accmode_t accmode, struct ucred *cred) { accmode_t dac_granted; accmode_t priv_granted; KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, ("invalid bit in accmode")); KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), ("VAPPEND without VWRITE")); /* * Look for a normal, non-privileged way to access the file/directory * as requested. If it exists, go with that. */ dac_granted = 0; /* Check the owner. */ if (cred->cr_uid == file_uid) { dac_granted |= VADMIN; if (file_mode & S_IXUSR) dac_granted |= VEXEC; if (file_mode & S_IRUSR) dac_granted |= VREAD; if (file_mode & S_IWUSR) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); goto privcheck; } /* Otherwise, check the groups (first match) */ if (groupmember(file_gid, cred)) { if (file_mode & S_IXGRP) dac_granted |= VEXEC; if (file_mode & S_IRGRP) dac_granted |= VREAD; if (file_mode & S_IWGRP) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); goto privcheck; } /* Otherwise, check everyone else. */ if (file_mode & S_IXOTH) dac_granted |= VEXEC; if (file_mode & S_IROTH) dac_granted |= VREAD; if (file_mode & S_IWOTH) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); privcheck: /* * Build a privilege mask to determine if the set of privileges * satisfies the requirements when combined with the granted mask * from above. For each privilege, if the privilege is required, * bitwise or the request type onto the priv_granted mask. */ priv_granted = 0; if (type == VDIR) { /* * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC * requests, instead of PRIV_VFS_EXEC. */ if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && !priv_check_cred(cred, PRIV_VFS_LOOKUP)) priv_granted |= VEXEC; } else { /* * Ensure that at least one execute bit is on. Otherwise, * a privileged user will always succeed, and we don't want * this to happen unless the file really is executable. */ if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && !priv_check_cred(cred, PRIV_VFS_EXEC)) priv_granted |= VEXEC; } if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && !priv_check_cred(cred, PRIV_VFS_READ)) priv_granted |= VREAD; if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && !priv_check_cred(cred, PRIV_VFS_WRITE)) priv_granted |= (VWRITE | VAPPEND); if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && !priv_check_cred(cred, PRIV_VFS_ADMIN)) priv_granted |= VADMIN; if ((accmode & (priv_granted | dac_granted)) == accmode) { return (0); } return ((accmode & VADMIN) ? EPERM : EACCES); } /* * Credential check based on process requesting service, and per-attribute * permissions. */ int extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, struct thread *td, accmode_t accmode) { /* * Kernel-invoked always succeeds. */ if (cred == NOCRED) return (0); /* * Do not allow privileged processes in jail to directly manipulate * system attributes. */ switch (attrnamespace) { case EXTATTR_NAMESPACE_SYSTEM: /* Potentially should be: return (EPERM); */ return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); case EXTATTR_NAMESPACE_USER: return (VOP_ACCESS(vp, accmode, cred, td)); default: return (EPERM); } } #ifdef DEBUG_VFS_LOCKS /* * This only exists to suppress warnings from unlocked specfs accesses. It is * no longer ok to have an unlocked VFS. */ #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \ (vp)->v_type == VCHR || (vp)->v_type == VBAD) int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "Drop into debugger on lock violation"); int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "Check for interlock across VOPs"); int vfs_badlock_print = 1; /* Print lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "Print lock violations"); int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 0, "Print vnode details on lock violations"); #ifdef KDB int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); #endif static void vfs_badlock(const char *msg, const char *str, struct vnode *vp) { #ifdef KDB if (vfs_badlock_backtrace) kdb_backtrace(); #endif if (vfs_badlock_vnode) vn_printf(vp, "vnode "); if (vfs_badlock_print) printf("%s: %p %s\n", str, (void *)vp, msg); if (vfs_badlock_ddb) kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); } void assert_vi_locked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is not locked but should be", str, vp); } void assert_vi_unlocked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is locked but should not be", str, vp); } void assert_vop_locked(struct vnode *vp, const char *str) { int locked; if (!IGNORE_LOCK(vp)) { locked = VOP_ISLOCKED(vp); if (locked == 0 || locked == LK_EXCLOTHER) vfs_badlock("is not locked but should be", str, vp); } } void assert_vop_unlocked(struct vnode *vp, const char *str) { if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) vfs_badlock("is locked but should not be", str, vp); } void assert_vop_elocked(struct vnode *vp, const char *str) { if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) vfs_badlock("is not exclusive locked but should be", str, vp); } #endif /* DEBUG_VFS_LOCKS */ void vop_rename_fail(struct vop_rename_args *ap) { if (ap->a_tvp != NULL) vput(ap->a_tvp); if (ap->a_tdvp == ap->a_tvp) vrele(ap->a_tdvp); else vput(ap->a_tdvp); vrele(ap->a_fdvp); vrele(ap->a_fvp); } void vop_rename_pre(void *ap) { struct vop_rename_args *a = ap; #ifdef DEBUG_VFS_LOCKS if (a->a_tvp) ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); /* Check the source (from). */ if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); /* Check the target. */ if (a->a_tvp) ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); #endif /* * It may be tempting to add vn_seqc_write_begin/end calls here and * in vop_rename_post but that's not going to work out since some * filesystems relookup vnodes mid-rename. This is probably a bug. * * For now filesystems are expected to do the relevant calls after they * decide what vnodes to operate on. */ if (a->a_tdvp != a->a_fdvp) vhold(a->a_fdvp); if (a->a_tvp != a->a_fvp) vhold(a->a_fvp); vhold(a->a_tdvp); if (a->a_tvp) vhold(a->a_tvp); } #ifdef DEBUG_VFS_LOCKS void vop_fplookup_vexec_debugpre(void *ap __unused) { VFS_SMR_ASSERT_ENTERED(); } void vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused) { VFS_SMR_ASSERT_ENTERED(); } void vop_strategy_debugpre(void *ap) { struct vop_strategy_args *a; struct buf *bp; a = ap; bp = a->a_bp; /* * Cluster ops lock their component buffers but not the IO container. */ if ((bp->b_flags & B_CLUSTER) != 0) return; if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { if (vfs_badlock_print) printf( "VOP_STRATEGY: bp is not locked but should be\n"); if (vfs_badlock_ddb) kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); } } void vop_lock_debugpre(void *ap) { struct vop_lock1_args *a = ap; if ((a->a_flags & LK_INTERLOCK) == 0) ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); else ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); } void vop_lock_debugpost(void *ap, int rc) { struct vop_lock1_args *a = ap; ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); } void vop_unlock_debugpre(void *ap) { struct vop_unlock_args *a = ap; ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); } void vop_need_inactive_debugpre(void *ap) { struct vop_need_inactive_args *a = ap; ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); } void vop_need_inactive_debugpost(void *ap, int rc) { struct vop_need_inactive_args *a = ap; ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); } #endif void vop_create_pre(void *ap) { struct vop_create_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_begin(dvp); } void vop_create_post(void *ap, int rc) { struct vop_create_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_end(dvp); if (!rc) VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); } void vop_whiteout_pre(void *ap) { struct vop_whiteout_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_begin(dvp); } void vop_whiteout_post(void *ap, int rc) { struct vop_whiteout_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_end(dvp); } void vop_deleteextattr_pre(void *ap) { struct vop_deleteextattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_begin(vp); } void vop_deleteextattr_post(void *ap, int rc) { struct vop_deleteextattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_end(vp); if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); } void vop_link_pre(void *ap) { struct vop_link_args *a; struct vnode *vp, *tdvp; a = ap; vp = a->a_vp; tdvp = a->a_tdvp; vn_seqc_write_begin(vp); vn_seqc_write_begin(tdvp); } void vop_link_post(void *ap, int rc) { struct vop_link_args *a; struct vnode *vp, *tdvp; a = ap; vp = a->a_vp; tdvp = a->a_tdvp; vn_seqc_write_end(vp); vn_seqc_write_end(tdvp); if (!rc) { VFS_KNOTE_LOCKED(vp, NOTE_LINK); VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE); } } void vop_mkdir_pre(void *ap) { struct vop_mkdir_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_begin(dvp); } void vop_mkdir_post(void *ap, int rc) { struct vop_mkdir_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_end(dvp); if (!rc) VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); } #ifdef DEBUG_VFS_LOCKS void vop_mkdir_debugpost(void *ap, int rc) { struct vop_mkdir_args *a; a = ap; if (!rc) cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp); } #endif void vop_mknod_pre(void *ap) { struct vop_mknod_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_begin(dvp); } void vop_mknod_post(void *ap, int rc) { struct vop_mknod_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_end(dvp); if (!rc) VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); } void vop_reclaim_post(void *ap, int rc) { struct vop_reclaim_args *a; struct vnode *vp; a = ap; vp = a->a_vp; ASSERT_VOP_IN_SEQC(vp); if (!rc) VFS_KNOTE_LOCKED(vp, NOTE_REVOKE); } void vop_remove_pre(void *ap) { struct vop_remove_args *a; struct vnode *dvp, *vp; a = ap; dvp = a->a_dvp; vp = a->a_vp; vn_seqc_write_begin(dvp); vn_seqc_write_begin(vp); } void vop_remove_post(void *ap, int rc) { struct vop_remove_args *a; struct vnode *dvp, *vp; a = ap; dvp = a->a_dvp; vp = a->a_vp; vn_seqc_write_end(dvp); vn_seqc_write_end(vp); if (!rc) { VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); VFS_KNOTE_LOCKED(vp, NOTE_DELETE); } } void vop_rename_post(void *ap, int rc) { struct vop_rename_args *a = ap; long hint; if (!rc) { hint = NOTE_WRITE; if (a->a_fdvp == a->a_tdvp) { if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) hint |= NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); } else { hint |= NOTE_EXTEND; if (a->a_fvp->v_type == VDIR) hint |= NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && a->a_tvp->v_type == VDIR) hint &= ~NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); } VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); if (a->a_tvp) VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); } if (a->a_tdvp != a->a_fdvp) vdrop(a->a_fdvp); if (a->a_tvp != a->a_fvp) vdrop(a->a_fvp); vdrop(a->a_tdvp); if (a->a_tvp) vdrop(a->a_tvp); } void vop_rmdir_pre(void *ap) { struct vop_rmdir_args *a; struct vnode *dvp, *vp; a = ap; dvp = a->a_dvp; vp = a->a_vp; vn_seqc_write_begin(dvp); vn_seqc_write_begin(vp); } void vop_rmdir_post(void *ap, int rc) { struct vop_rmdir_args *a; struct vnode *dvp, *vp; a = ap; dvp = a->a_dvp; vp = a->a_vp; vn_seqc_write_end(dvp); vn_seqc_write_end(vp); if (!rc) { VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK); VFS_KNOTE_LOCKED(vp, NOTE_DELETE); } } void vop_setattr_pre(void *ap) { struct vop_setattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_begin(vp); } void vop_setattr_post(void *ap, int rc) { struct vop_setattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_end(vp); if (!rc) VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); } void vop_setacl_pre(void *ap) { struct vop_setacl_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_begin(vp); } void vop_setacl_post(void *ap, int rc __unused) { struct vop_setacl_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_end(vp); } void vop_setextattr_pre(void *ap) { struct vop_setextattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_begin(vp); } void vop_setextattr_post(void *ap, int rc) { struct vop_setextattr_args *a; struct vnode *vp; a = ap; vp = a->a_vp; vn_seqc_write_end(vp); if (!rc) VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB); } void vop_symlink_pre(void *ap) { struct vop_symlink_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_begin(dvp); } void vop_symlink_post(void *ap, int rc) { struct vop_symlink_args *a; struct vnode *dvp; a = ap; dvp = a->a_dvp; vn_seqc_write_end(dvp); if (!rc) VFS_KNOTE_LOCKED(dvp, NOTE_WRITE); } void vop_open_post(void *ap, int rc) { struct vop_open_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); } void vop_close_post(void *ap, int rc) { struct vop_close_args *a = ap; if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ !VN_IS_DOOMED(a->a_vp))) { VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? NOTE_CLOSE_WRITE : NOTE_CLOSE); } } void vop_read_post(void *ap, int rc) { struct vop_read_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); } void vop_read_pgcache_post(void *ap, int rc) { struct vop_read_pgcache_args *a = ap; if (!rc) VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ); } void vop_readdir_post(void *ap, int rc) { struct vop_readdir_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); } static struct knlist fs_knlist; static void vfs_event_init(void *arg) { knlist_init_mtx(&fs_knlist, NULL); } /* XXX - correct order? */ SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); void vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) { KNOTE_UNLOCKED(&fs_knlist, event); } static int filt_fsattach(struct knote *kn); static void filt_fsdetach(struct knote *kn); static int filt_fsevent(struct knote *kn, long hint); struct filterops fs_filtops = { .f_isfd = 0, .f_attach = filt_fsattach, .f_detach = filt_fsdetach, .f_event = filt_fsevent }; static int filt_fsattach(struct knote *kn) { kn->kn_flags |= EV_CLEAR; knlist_add(&fs_knlist, kn, 0); return (0); } static void filt_fsdetach(struct knote *kn) { knlist_remove(&fs_knlist, kn, 0); } static int filt_fsevent(struct knote *kn, long hint) { kn->kn_fflags |= hint; return (kn->kn_fflags != 0); } static int sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) { struct vfsidctl vc; int error; struct mount *mp; error = SYSCTL_IN(req, &vc, sizeof(vc)); if (error) return (error); if (vc.vc_vers != VFS_CTL_VERS1) return (EINVAL); mp = vfs_getvfs(&vc.vc_fsid); if (mp == NULL) return (ENOENT); /* ensure that a specific sysctl goes to the right filesystem. */ if (strcmp(vc.vc_fstypename, "*") != 0 && strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { vfs_rel(mp); return (EINVAL); } VCTLTOREQ(&vc, req); error = VFS_SYSCTL(mp, vc.vc_op, req); vfs_rel(mp); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); /* * Function to initialize a va_filerev field sensibly. * XXX: Wouldn't a random number make a lot more sense ?? */ u_quad_t init_va_filerev(void) { struct bintime bt; getbinuptime(&bt); return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); } static int filt_vfsread(struct knote *kn, long hint); static int filt_vfswrite(struct knote *kn, long hint); static int filt_vfsvnode(struct knote *kn, long hint); static void filt_vfsdetach(struct knote *kn); static struct filterops vfsread_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsread }; static struct filterops vfswrite_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfswrite }; static struct filterops vfsvnode_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsvnode }; static void vfs_knllock(void *arg) { struct vnode *vp = arg; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } static void vfs_knlunlock(void *arg) { struct vnode *vp = arg; VOP_UNLOCK(vp); } static void vfs_knl_assert_lock(void *arg, int what) { #ifdef DEBUG_VFS_LOCKS struct vnode *vp = arg; if (what == LA_LOCKED) ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); else ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); #endif } int vfs_kqfilter(struct vop_kqfilter_args *ap) { struct vnode *vp = ap->a_vp; struct knote *kn = ap->a_kn; struct knlist *knl; switch (kn->kn_filter) { case EVFILT_READ: kn->kn_fop = &vfsread_filtops; break; case EVFILT_WRITE: kn->kn_fop = &vfswrite_filtops; break; case EVFILT_VNODE: kn->kn_fop = &vfsvnode_filtops; break; default: return (EINVAL); } kn->kn_hook = (caddr_t)vp; v_addpollinfo(vp); if (vp->v_pollinfo == NULL) return (ENOMEM); knl = &vp->v_pollinfo->vpi_selinfo.si_note; vhold(vp); knlist_add(knl, kn, 0); return (0); } /* * Detach knote from vnode */ static void filt_vfsdetach(struct knote *kn) { struct vnode *vp = (struct vnode *)kn->kn_hook; KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); vdrop(vp); } /*ARGSUSED*/ static int filt_vfsread(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; struct vattr va; int res; /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { VI_LOCK(vp); kn->kn_flags |= (EV_EOF | EV_ONESHOT); VI_UNLOCK(vp); return (1); } if (VOP_GETATTR(vp, &va, curthread->td_ucred)) return (0); VI_LOCK(vp); kn->kn_data = va.va_size - kn->kn_fp->f_offset; res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; VI_UNLOCK(vp); return (res); } /*ARGSUSED*/ static int filt_vfswrite(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; VI_LOCK(vp); /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) kn->kn_flags |= (EV_EOF | EV_ONESHOT); kn->kn_data = 0; VI_UNLOCK(vp); return (1); } static int filt_vfsvnode(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; int res; VI_LOCK(vp); if (kn->kn_sfflags & hint) kn->kn_fflags |= hint; if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { kn->kn_flags |= EV_EOF; VI_UNLOCK(vp); return (1); } res = (kn->kn_fflags != 0); VI_UNLOCK(vp); return (res); } /* * Returns whether the directory is empty or not. * If it is empty, the return value is 0; otherwise * the return value is an error value (which may * be ENOTEMPTY). */ int vfs_emptydir(struct vnode *vp) { struct uio uio; struct iovec iov; struct dirent *dirent, *dp, *endp; int error, eof; error = 0; eof = 0; ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); iov.iov_base = dirent; iov.iov_len = sizeof(struct dirent); uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = 0; uio.uio_resid = sizeof(struct dirent); uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = curthread; while (eof == 0 && error == 0) { error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, NULL, NULL); if (error != 0) break; endp = (void *)((uint8_t *)dirent + sizeof(struct dirent) - uio.uio_resid); for (dp = dirent; dp < endp; dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { if (dp->d_type == DT_WHT) continue; if (dp->d_namlen == 0) continue; if (dp->d_type != DT_DIR && dp->d_type != DT_UNKNOWN) { error = ENOTEMPTY; break; } if (dp->d_namlen > 2) { error = ENOTEMPTY; break; } if (dp->d_namlen == 1 && dp->d_name[0] != '.') { error = ENOTEMPTY; break; } if (dp->d_namlen == 2 && dp->d_name[1] != '.') { error = ENOTEMPTY; break; } uio.uio_resid = sizeof(struct dirent); } } free(dirent, M_TEMP); return (error); } int vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) { int error; if (dp->d_reclen > ap->a_uio->uio_resid) return (ENAMETOOLONG); error = uiomove(dp, dp->d_reclen, ap->a_uio); if (error) { if (ap->a_ncookies != NULL) { if (ap->a_cookies != NULL) free(ap->a_cookies, M_TEMP); ap->a_cookies = NULL; *ap->a_ncookies = 0; } return (error); } if (ap->a_ncookies == NULL) return (0); KASSERT(ap->a_cookies, ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); *ap->a_cookies = realloc(*ap->a_cookies, (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); (*ap->a_cookies)[*ap->a_ncookies] = off; *ap->a_ncookies += 1; return (0); } /* * The purpose of this routine is to remove granularity from accmode_t, * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, * VADMIN and VAPPEND. * * If it returns 0, the caller is supposed to continue with the usual * access checks using 'accmode' as modified by this routine. If it * returns nonzero value, the caller is supposed to return that value * as errno. * * Note that after this routine runs, accmode may be zero. */ int vfs_unixify_accmode(accmode_t *accmode) { /* * There is no way to specify explicit "deny" rule using * file mode or POSIX.1e ACLs. */ if (*accmode & VEXPLICIT_DENY) { *accmode = 0; return (0); } /* * None of these can be translated into usual access bits. * Also, the common case for NFSv4 ACLs is to not contain * either of these bits. Caller should check for VWRITE * on the containing directory instead. */ if (*accmode & (VDELETE_CHILD | VDELETE)) return (EPERM); if (*accmode & VADMIN_PERMS) { *accmode &= ~VADMIN_PERMS; *accmode |= VADMIN; } /* * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL * or VSYNCHRONIZE using file mode or POSIX.1e ACL. */ *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); return (0); } /* * Clear out a doomed vnode (if any) and replace it with a new one as long * as the fs is not being unmounted. Return the root vnode to the caller. */ static int __noinline vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) { struct vnode *vp; int error; restart: if (mp->mnt_rootvnode != NULL) { MNT_ILOCK(mp); vp = mp->mnt_rootvnode; if (vp != NULL) { if (!VN_IS_DOOMED(vp)) { vrefact(vp); MNT_IUNLOCK(mp); error = vn_lock(vp, flags); if (error == 0) { *vpp = vp; return (0); } vrele(vp); goto restart; } /* * Clear the old one. */ mp->mnt_rootvnode = NULL; } MNT_IUNLOCK(mp); if (vp != NULL) { vfs_op_barrier_wait(mp); vrele(vp); } } error = VFS_CACHEDROOT(mp, flags, vpp); if (error != 0) return (error); if (mp->mnt_vfs_ops == 0) { MNT_ILOCK(mp); if (mp->mnt_vfs_ops != 0) { MNT_IUNLOCK(mp); return (0); } if (mp->mnt_rootvnode == NULL) { vrefact(*vpp); mp->mnt_rootvnode = *vpp; } else { if (mp->mnt_rootvnode != *vpp) { if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { panic("%s: mismatch between vnode returned " " by VFS_CACHEDROOT and the one cached " " (%p != %p)", __func__, *vpp, mp->mnt_rootvnode); } } } MNT_IUNLOCK(mp); } return (0); } int vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) { + struct mount_pcpu *mpcpu; struct vnode *vp; int error; - if (!vfs_op_thread_enter(mp)) + if (!vfs_op_thread_enter(mp, mpcpu)) return (vfs_cache_root_fallback(mp, flags, vpp)); vp = atomic_load_ptr(&mp->mnt_rootvnode); if (vp == NULL || VN_IS_DOOMED(vp)) { - vfs_op_thread_exit(mp); + vfs_op_thread_exit(mp, mpcpu); return (vfs_cache_root_fallback(mp, flags, vpp)); } vrefact(vp); - vfs_op_thread_exit(mp); + vfs_op_thread_exit(mp, mpcpu); error = vn_lock(vp, flags); if (error != 0) { vrele(vp); return (vfs_cache_root_fallback(mp, flags, vpp)); } *vpp = vp; return (0); } struct vnode * vfs_cache_root_clear(struct mount *mp) { struct vnode *vp; /* * ops > 0 guarantees there is nobody who can see this vnode */ MPASS(mp->mnt_vfs_ops > 0); vp = mp->mnt_rootvnode; if (vp != NULL) vn_seqc_write_begin(vp); mp->mnt_rootvnode = NULL; return (vp); } void vfs_cache_root_set(struct mount *mp, struct vnode *vp) { MPASS(mp->mnt_vfs_ops > 0); vrefact(vp); mp->mnt_rootvnode = vp; } /* * These are helper functions for filesystems to traverse all * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. * * This interface replaces MNT_VNODE_FOREACH. */ struct vnode * __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) { struct vnode *vp; if (should_yield()) kern_yield(PRI_USER); MNT_ILOCK(mp); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; vp = TAILQ_NEXT(vp, v_nmntvnodes)) { /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) continue; VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VI_UNLOCK(vp); continue; } break; } if (vp == NULL) { __mnt_vnode_markerfree_all(mvp, mp); /* MNT_IUNLOCK(mp); -- done in above function */ mtx_assert(MNT_MTX(mp), MA_NOTOWNED); return (NULL); } TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); MNT_IUNLOCK(mp); return (vp); } struct vnode * __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) { struct vnode *vp; *mvp = vn_alloc_marker(mp); MNT_ILOCK(mp); MNT_REF(mp); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) continue; VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VI_UNLOCK(vp); continue; } break; } if (vp == NULL) { MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; return (NULL); } TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); MNT_IUNLOCK(mp); return (vp); } void __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) { if (*mvp == NULL) { MNT_IUNLOCK(mp); return; } mtx_assert(MNT_MTX(mp), MA_OWNED); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; } /* * These are helper functions for filesystems to traverse their * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h */ static void mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) { KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); MNT_ILOCK(mp); MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; } /* * Relock the mp mount vnode list lock with the vp vnode interlock in the * conventional lock order during mnt_vnode_next_lazy iteration. * * On entry, the mount vnode list lock is held and the vnode interlock is not. * The list lock is dropped and reacquired. On success, both locks are held. * On failure, the mount vnode list lock is held but the vnode interlock is * not, and the procedure may have yielded. */ static bool mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, struct vnode *vp) { VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, ("%s: bad marker", __func__)); VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, ("%s: inappropriate vnode", __func__)); ASSERT_VI_UNLOCKED(vp, __func__); mtx_assert(&mp->mnt_listmtx, MA_OWNED); TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); /* * Note we may be racing against vdrop which transitioned the hold * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, * if we are the only user after we get the interlock we will just * vdrop. */ vhold(vp); mtx_unlock(&mp->mnt_listmtx); VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); goto out_lost; } VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); /* * There is nothing to do if we are the last user. */ if (!refcount_release_if_not_last(&vp->v_holdcnt)) goto out_lost; mtx_lock(&mp->mnt_listmtx); return (true); out_lost: vdropl(vp); maybe_yield(); mtx_lock(&mp->mnt_listmtx); return (false); } static struct vnode * mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { struct vnode *vp; mtx_assert(&mp->mnt_listmtx, MA_OWNED); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); restart: vp = TAILQ_NEXT(*mvp, v_lazylist); while (vp != NULL) { if (vp->v_type == VMARKER) { vp = TAILQ_NEXT(vp, v_lazylist); continue; } /* * See if we want to process the vnode. Note we may encounter a * long string of vnodes we don't care about and hog the list * as a result. Check for it and requeue the marker. */ VNPASS(!VN_IS_DOOMED(vp), vp); if (!cb(vp, cbarg)) { if (!should_yield()) { vp = TAILQ_NEXT(vp, v_lazylist); continue; } TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); kern_yield(PRI_USER); mtx_lock(&mp->mnt_listmtx); goto restart; } /* * Try-lock because this is the wrong lock order. */ if (!VI_TRYLOCK(vp) && !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) goto restart; KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); KASSERT(vp->v_mount == mp || vp->v_mount == NULL, ("alien vnode on the lazy list %p %p", vp, mp)); VNPASS(vp->v_mount == mp, vp); VNPASS(!VN_IS_DOOMED(vp), vp); break; } TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); /* Check if we are done */ if (vp == NULL) { mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); return (NULL); } TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); ASSERT_VI_LOCKED(vp, "lazy iter"); return (vp); } struct vnode * __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { if (should_yield()) kern_yield(PRI_USER); mtx_lock(&mp->mnt_listmtx); return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); } struct vnode * __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { struct vnode *vp; if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) return (NULL); *mvp = vn_alloc_marker(mp); MNT_ILOCK(mp); MNT_REF(mp); MNT_IUNLOCK(mp); mtx_lock(&mp->mnt_listmtx); vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); if (vp == NULL) { mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); return (NULL); } TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); } void __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) { if (*mvp == NULL) return; mtx_lock(&mp->mnt_listmtx); TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); } int vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) { if ((cnp->cn_flags & NOEXECCHECK) != 0) { cnp->cn_flags &= ~NOEXECCHECK; return (0); } return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread)); } /* * Do not use this variant unless you have means other than the hold count * to prevent the vnode from getting freed. */ void vn_seqc_write_begin_unheld_locked(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNPASS(vp->v_seqc_users >= 0, vp); vp->v_seqc_users++; if (vp->v_seqc_users == 1) seqc_sleepable_write_begin(&vp->v_seqc); } void vn_seqc_write_begin_locked(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNPASS(vp->v_holdcnt > 0, vp); vn_seqc_write_begin_unheld_locked(vp); } void vn_seqc_write_begin(struct vnode *vp) { VI_LOCK(vp); vn_seqc_write_begin_locked(vp); VI_UNLOCK(vp); } void vn_seqc_write_begin_unheld(struct vnode *vp) { VI_LOCK(vp); vn_seqc_write_begin_unheld_locked(vp); VI_UNLOCK(vp); } void vn_seqc_write_end_locked(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNPASS(vp->v_seqc_users > 0, vp); vp->v_seqc_users--; if (vp->v_seqc_users == 0) seqc_sleepable_write_end(&vp->v_seqc); } void vn_seqc_write_end(struct vnode *vp) { VI_LOCK(vp); vn_seqc_write_end_locked(vp); VI_UNLOCK(vp); } Index: head/sys/kern/vfs_vnops.c =================================================================== --- head/sys/kern/vfs_vnops.c (revision 367534) +++ head/sys/kern/vfs_vnops.c (revision 367535) @@ -1,3318 +1,3320 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Copyright (c) 2012 Konstantin Belousov * Copyright (c) 2013, 2014 The FreeBSD Foundation * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif static fo_rdwr_t vn_read; static fo_rdwr_t vn_write; static fo_rdwr_t vn_io_fault; static fo_truncate_t vn_truncate; static fo_ioctl_t vn_ioctl; static fo_poll_t vn_poll; static fo_kqfilter_t vn_kqfilter; static fo_stat_t vn_statfile; static fo_close_t vn_closefile; static fo_mmap_t vn_mmap; static fo_fallocate_t vn_fallocate; struct fileops vnops = { .fo_read = vn_io_fault, .fo_write = vn_io_fault, .fo_truncate = vn_truncate, .fo_ioctl = vn_ioctl, .fo_poll = vn_poll, .fo_kqfilter = vn_kqfilter, .fo_stat = vn_statfile, .fo_close = vn_closefile, .fo_chmod = vn_chmod, .fo_chown = vn_chown, .fo_sendfile = vn_sendfile, .fo_seek = vn_seek, .fo_fill_kinfo = vn_fill_kinfo, .fo_mmap = vn_mmap, .fo_fallocate = vn_fallocate, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; const u_int io_hold_cnt = 16; static int vn_io_fault_enable = 1; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN, &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); static int vn_io_fault_prefault = 0; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN, &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); static int vn_io_pgcache_read_enable = 1; SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN, &vn_io_pgcache_read_enable, 0, "Enable copying from page cache for reads, avoiding fs"); static u_long vn_io_faults_cnt; SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); static int vfs_allow_read_dir = 0; SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, &vfs_allow_read_dir, 0, "Enable read(2) of directory by root for filesystems that support it"); /* * Returns true if vn_io_fault mode of handling the i/o request should * be used. */ static bool do_vn_io_fault(struct vnode *vp, struct uio *uio) { struct mount *mp; return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && (mp = vp->v_mount) != NULL && (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); } /* * Structure used to pass arguments to vn_io_fault1(), to do either * file- or vnode-based I/O calls. */ struct vn_io_fault_args { enum { VN_IO_FAULT_FOP, VN_IO_FAULT_VOP } kind; struct ucred *cred; int flags; union { struct fop_args_tag { struct file *fp; fo_rdwr_t *doio; } fop_args; struct vop_args_tag { struct vnode *vp; } vop_args; } args; }; static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td); int vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) { struct thread *td = ndp->ni_cnd.cn_thread; return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); } static uint64_t open2nameif(int fmode, u_int vn_open_flags) { uint64_t res; res = ISOPEN | LOCKLEAF; if ((fmode & O_BENEATH) != 0) res |= BENEATH; if ((fmode & O_RESOLVE_BENEATH) != 0) res |= RBENEATH; if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0) res |= AUDITVNODE1; if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0) res |= NOCAPCHECK; return (res); } /* * Common code for vnode open operations via a name lookup. * Lookup the vnode and invoke VOP_CREATE if needed. * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. * * Note that this does NOT free nameidata for the successful case, * due to the NDINIT being done elsewhere. */ int vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, struct ucred *cred, struct file *fp) { struct vnode *vp; struct mount *mp; struct thread *td = ndp->ni_cnd.cn_thread; struct vattr vat; struct vattr *vap = &vat; int fmode, error; restart: fmode = *flagp; if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | O_EXCL | O_DIRECTORY)) return (EINVAL); else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { ndp->ni_cnd.cn_nameiop = CREATE; ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); /* * Set NOCACHE to avoid flushing the cache when * rolling in many files at once. */ ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE; if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) ndp->ni_cnd.cn_flags |= FOLLOW; if ((vn_open_flags & VN_OPEN_INVFS) == 0) bwillwrite(); if ((error = namei(ndp)) != 0) return (error); if (ndp->ni_vp == NULL) { VATTR_NULL(vap); vap->va_type = VREG; vap->va_mode = cmode; if (fmode & O_EXCL) vap->va_vaflags |= VA_EXCLUSIVE; if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { NDFREE(ndp, NDF_ONLY_PNBUF); vput(ndp->ni_dvp); if ((error = vn_start_write(NULL, &mp, V_XSLEEP | PCATCH)) != 0) return (error); NDREINIT(ndp); goto restart; } if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) ndp->ni_cnd.cn_flags |= MAKEENTRY; #ifdef MAC error = mac_vnode_check_create(cred, ndp->ni_dvp, &ndp->ni_cnd, vap); if (error == 0) #endif error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, &ndp->ni_cnd, vap); vput(ndp->ni_dvp); vn_finished_write(mp); if (error) { NDFREE(ndp, NDF_ONLY_PNBUF); return (error); } fmode &= ~O_TRUNC; vp = ndp->ni_vp; } else { if (ndp->ni_dvp == ndp->ni_vp) vrele(ndp->ni_dvp); else vput(ndp->ni_dvp); ndp->ni_dvp = NULL; vp = ndp->ni_vp; if (fmode & O_EXCL) { error = EEXIST; goto bad; } if (vp->v_type == VDIR) { error = EISDIR; goto bad; } fmode &= ~O_CREAT; } } else { ndp->ni_cnd.cn_nameiop = LOOKUP; ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags); ndp->ni_cnd.cn_flags |= (fmode & O_NOFOLLOW) != 0 ? NOFOLLOW : FOLLOW; if ((fmode & FWRITE) == 0) ndp->ni_cnd.cn_flags |= LOCKSHARED; if ((error = namei(ndp)) != 0) return (error); vp = ndp->ni_vp; } error = vn_open_vnode(vp, fmode, cred, td, fp); if (error) goto bad; *flagp = fmode; return (0); bad: NDFREE(ndp, NDF_ONLY_PNBUF); vput(vp); *flagp = fmode; ndp->ni_vp = NULL; return (error); } static int vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) { struct flock lf; int error, lock_flags, type; ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) return (0); KASSERT(fp != NULL, ("open with flock requires fp")); if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) return (EOPNOTSUPP); lock_flags = VOP_ISLOCKED(vp); VOP_UNLOCK(vp); lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; type = F_FLOCK; if ((fmode & FNONBLOCK) == 0) type |= F_WAIT; error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); if (error == 0) fp->f_flag |= FHASLOCK; vn_lock(vp, lock_flags | LK_RETRY); if (error == 0 && VN_IS_DOOMED(vp)) error = ENOENT; return (error); } /* * Common code for vnode open operations once a vnode is located. * Check permissions, and call the VOP_OPEN routine. */ int vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, struct thread *td, struct file *fp) { accmode_t accmode; int error; if (vp->v_type == VLNK) return (EMLINK); if (vp->v_type == VSOCK) return (EOPNOTSUPP); if (vp->v_type != VDIR && fmode & O_DIRECTORY) return (ENOTDIR); accmode = 0; if (fmode & (FWRITE | O_TRUNC)) { if (vp->v_type == VDIR) return (EISDIR); accmode |= VWRITE; } if (fmode & FREAD) accmode |= VREAD; if (fmode & FEXEC) accmode |= VEXEC; if ((fmode & O_APPEND) && (fmode & FWRITE)) accmode |= VAPPEND; #ifdef MAC if (fmode & O_CREAT) accmode |= VCREAT; if (fmode & O_VERIFY) accmode |= VVERIFY; error = mac_vnode_check_open(cred, vp, accmode); if (error) return (error); accmode &= ~(VCREAT | VVERIFY); #endif if ((fmode & O_CREAT) == 0 && accmode != 0) { error = VOP_ACCESS(vp, accmode, cred, td); if (error != 0) return (error); } if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) vn_lock(vp, LK_UPGRADE | LK_RETRY); error = VOP_OPEN(vp, fmode, cred, td, fp); if (error != 0) return (error); error = vn_open_vnode_advlock(vp, fmode, fp); if (error == 0 && (fmode & FWRITE) != 0) { error = VOP_ADD_WRITECOUNT(vp, 1); if (error == 0) { CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", __func__, vp, vp->v_writecount); } } /* * Error from advlock or VOP_ADD_WRITECOUNT() still requires * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). * Arrange for that by having fdrop() to use vn_closefile(). */ if (error != 0) { fp->f_flag |= FOPENFAILED; fp->f_vnode = vp; if (fp->f_ops == &badfileops) { fp->f_type = DTYPE_VNODE; fp->f_ops = &vnops; } vref(vp); } ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); return (error); } /* * Check for write permissions on the specified vnode. * Prototype text segments cannot be written. * It is racy. */ int vn_writechk(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, "vn_writechk"); /* * If there's shared text associated with * the vnode, try to free it up once. If * we fail, we can't allow writing. */ if (VOP_IS_TEXT(vp)) return (ETXTBSY); return (0); } /* * Vnode close call */ static int vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td, bool keep_ref) { struct mount *mp; int error, lock_flags; if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && MNT_EXTENDED_SHARED(vp->v_mount)) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, lock_flags | LK_RETRY); AUDIT_ARG_VNODE1(vp); if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { VOP_ADD_WRITECOUNT_CHECKED(vp, -1); CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", __func__, vp, vp->v_writecount); } error = VOP_CLOSE(vp, flags, file_cred, td); if (keep_ref) VOP_UNLOCK(vp); else vput(vp); vn_finished_write(mp); return (error); } int vn_close(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td) { return (vn_close1(vp, flags, file_cred, td, false)); } /* * Heuristic to detect sequential operation. */ static int sequential_heuristic(struct uio *uio, struct file *fp) { enum uio_rw rw; ASSERT_VOP_LOCKED(fp->f_vnode, __func__); rw = uio->uio_rw; if (fp->f_flag & FRDAHEAD) return (fp->f_seqcount[rw] << IO_SEQSHIFT); /* * Offset 0 is handled specially. open() sets f_seqcount to 1 so * that the first I/O is normally considered to be slightly * sequential. Seeking to offset 0 doesn't change sequentiality * unless previous seeks have reduced f_seqcount to 0, in which * case offset 0 is not special. */ if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || uio->uio_offset == fp->f_nextoff[rw]) { /* * f_seqcount is in units of fixed-size blocks so that it * depends mainly on the amount of sequential I/O and not * much on the number of sequential I/O's. The fixed size * of 16384 is hard-coded here since it is (not quite) just * a magic size that works well here. This size is more * closely related to the best I/O size for real disks than * to any block size used by software. */ if (uio->uio_resid >= IO_SEQMAX * 16384) fp->f_seqcount[rw] = IO_SEQMAX; else { fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); if (fp->f_seqcount[rw] > IO_SEQMAX) fp->f_seqcount[rw] = IO_SEQMAX; } return (fp->f_seqcount[rw] << IO_SEQSHIFT); } /* Not sequential. Quickly draw-down sequentiality. */ if (fp->f_seqcount[rw] > 1) fp->f_seqcount[rw] = 1; else fp->f_seqcount[rw] = 0; return (0); } /* * Package up an I/O request on a vnode into a uio and do it. */ int vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, struct ucred *file_cred, ssize_t *aresid, struct thread *td) { struct uio auio; struct iovec aiov; struct mount *mp; struct ucred *cred; void *rl_cookie; struct vn_io_fault_args args; int error, lock_flags; if (offset < 0 && vp->v_type != VCHR) return (EINVAL); auio.uio_iov = &aiov; auio.uio_iovcnt = 1; aiov.iov_base = base; aiov.iov_len = len; auio.uio_resid = len; auio.uio_offset = offset; auio.uio_segflg = segflg; auio.uio_rw = rw; auio.uio_td = td; error = 0; if ((ioflg & IO_NODELOCKED) == 0) { if ((ioflg & IO_RANGELOCKED) == 0) { if (rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, offset, offset + len); } else { rl_cookie = vn_rangelock_wlock(vp, offset, offset + len); } } else rl_cookie = NULL; mp = NULL; if (rw == UIO_WRITE) { if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto out; if (MNT_SHARED_WRITES(mp) || ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; } else lock_flags = LK_SHARED; vn_lock(vp, lock_flags | LK_RETRY); } else rl_cookie = NULL; ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); #ifdef MAC if ((ioflg & IO_NOMACCHECK) == 0) { if (rw == UIO_READ) error = mac_vnode_check_read(active_cred, file_cred, vp); else error = mac_vnode_check_write(active_cred, file_cred, vp); } #endif if (error == 0) { if (file_cred != NULL) cred = file_cred; else cred = active_cred; if (do_vn_io_fault(vp, &auio)) { args.kind = VN_IO_FAULT_VOP; args.cred = cred; args.flags = ioflg; args.args.vop_args.vp = vp; error = vn_io_fault1(vp, &auio, &args, td); } else if (rw == UIO_READ) { error = VOP_READ(vp, &auio, ioflg, cred); } else /* if (rw == UIO_WRITE) */ { error = VOP_WRITE(vp, &auio, ioflg, cred); } } if (aresid) *aresid = auio.uio_resid; else if (auio.uio_resid && error == 0) error = EIO; if ((ioflg & IO_NODELOCKED) == 0) { VOP_UNLOCK(vp); if (mp != NULL) vn_finished_write(mp); } out: if (rl_cookie != NULL) vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * Package up an I/O request on a vnode into a uio and do it. The I/O * request is split up into smaller chunks and we try to avoid saturating * the buffer cache while potentially holding a vnode locked, so we * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() * to give other processes a chance to lock the vnode (either other processes * core'ing the same binary, or unrelated processes scanning the directory). */ int vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, struct ucred *file_cred, size_t *aresid, struct thread *td) { int error = 0; ssize_t iaresid; do { int chunk; /* * Force `offset' to a multiple of MAXBSIZE except possibly * for the first chunk, so that filesystems only need to * write full blocks except possibly for the first and last * chunks. */ chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; if (chunk > len) chunk = len; if (rw != UIO_READ && vp->v_type == VREG) bwillwrite(); iaresid = 0; error = vn_rdwr(rw, vp, base, chunk, offset, segflg, ioflg, active_cred, file_cred, &iaresid, td); len -= chunk; /* aresid calc already includes length */ if (error) break; offset += chunk; base = (char *)base + chunk; kern_yield(PRI_USER); } while (len); if (aresid) *aresid = len + iaresid; return (error); } #if OFF_MAX <= LONG_MAX off_t foffset_lock(struct file *fp, int flags) { volatile short *flagsp; off_t res; short state; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); if ((flags & FOF_NOLOCK) != 0) return (atomic_load_long(&fp->f_offset)); /* * According to McKusick the vn lock was protecting f_offset here. * It is now protected by the FOFFSET_LOCKED flag. */ flagsp = &fp->f_vnread_flags; if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) return (atomic_load_long(&fp->f_offset)); sleepq_lock(&fp->f_vnread_flags); state = atomic_load_16(flagsp); for (;;) { if ((state & FOFFSET_LOCKED) == 0) { if (!atomic_fcmpset_acq_16(flagsp, &state, FOFFSET_LOCKED)) continue; break; } if ((state & FOFFSET_LOCK_WAITING) == 0) { if (!atomic_fcmpset_acq_16(flagsp, &state, state | FOFFSET_LOCK_WAITING)) continue; } DROP_GIANT(); sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); sleepq_wait(&fp->f_vnread_flags, PUSER -1); PICKUP_GIANT(); sleepq_lock(&fp->f_vnread_flags); state = atomic_load_16(flagsp); } res = atomic_load_long(&fp->f_offset); sleepq_release(&fp->f_vnread_flags); return (res); } void foffset_unlock(struct file *fp, off_t val, int flags) { volatile short *flagsp; short state; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); if ((flags & FOF_NOUPDATE) == 0) atomic_store_long(&fp->f_offset, val); if ((flags & FOF_NEXTOFF_R) != 0) fp->f_nextoff[UIO_READ] = val; if ((flags & FOF_NEXTOFF_W) != 0) fp->f_nextoff[UIO_WRITE] = val; if ((flags & FOF_NOLOCK) != 0) return; flagsp = &fp->f_vnread_flags; state = atomic_load_16(flagsp); if ((state & FOFFSET_LOCK_WAITING) == 0 && atomic_cmpset_rel_16(flagsp, state, 0)) return; sleepq_lock(&fp->f_vnread_flags); MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); fp->f_vnread_flags = 0; sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); sleepq_release(&fp->f_vnread_flags); } #else off_t foffset_lock(struct file *fp, int flags) { struct mtx *mtxp; off_t res; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOLOCK) == 0) { while (fp->f_vnread_flags & FOFFSET_LOCKED) { fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; msleep(&fp->f_vnread_flags, mtxp, PUSER -1, "vofflock", 0); } fp->f_vnread_flags |= FOFFSET_LOCKED; } res = fp->f_offset; mtx_unlock(mtxp); return (res); } void foffset_unlock(struct file *fp, off_t val, int flags) { struct mtx *mtxp; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOUPDATE) == 0) fp->f_offset = val; if ((flags & FOF_NEXTOFF_R) != 0) fp->f_nextoff[UIO_READ] = val; if ((flags & FOF_NEXTOFF_W) != 0) fp->f_nextoff[UIO_WRITE] = val; if ((flags & FOF_NOLOCK) == 0) { KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, ("Lost FOFFSET_LOCKED")); if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) wakeup(&fp->f_vnread_flags); fp->f_vnread_flags = 0; } mtx_unlock(mtxp); } #endif void foffset_lock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) uio->uio_offset = foffset_lock(fp, flags); } void foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) foffset_unlock(fp, uio->uio_offset, flags); } static int get_advice(struct file *fp, struct uio *uio) { struct mtx *mtxp; int ret; ret = POSIX_FADV_NORMAL; if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) return (ret); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if (fp->f_advice != NULL && uio->uio_offset >= fp->f_advice->fa_start && uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) ret = fp->f_advice->fa_advice; mtx_unlock(mtxp); return (ret); } int vn_read_from_obj(struct vnode *vp, struct uio *uio) { vm_object_t obj; vm_page_t ma[io_hold_cnt + 2]; off_t off, vsz; ssize_t resid; int error, i, j; obj = vp->v_object; MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2)); MPASS(obj != NULL); MPASS(obj->type == OBJT_VNODE); /* * Depends on type stability of vm_objects. */ vm_object_pip_add(obj, 1); if ((obj->flags & OBJ_DEAD) != 0) { /* * Note that object might be already reused from the * vnode, and the OBJ_DEAD flag cleared. This is fine, * we recheck for DOOMED vnode state after all pages * are busied, and retract then. * * But we check for OBJ_DEAD to ensure that we do not * busy pages while vm_object_terminate_pages() * processes the queue. */ error = EJUSTRETURN; goto out_pip; } resid = uio->uio_resid; off = uio->uio_offset; for (i = 0; resid > 0; i++) { MPASS(i < io_hold_cnt + 2); ma[i] = vm_page_grab_unlocked(obj, atop(off), VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOWAIT); if (ma[i] == NULL) break; /* * Skip invalid pages. Valid mask can be partial only * at EOF, and we clip later. */ if (vm_page_none_valid(ma[i])) { vm_page_sunbusy(ma[i]); break; } resid -= PAGE_SIZE; off += PAGE_SIZE; } if (i == 0) { error = EJUSTRETURN; goto out_pip; } /* * Check VIRF_DOOMED after we busied our pages. Since * vgonel() terminates the vnode' vm_object, it cannot * process past pages busied by us. */ if (VN_IS_DOOMED(vp)) { error = EJUSTRETURN; goto out; } resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1); if (resid > uio->uio_resid) resid = uio->uio_resid; /* * Unlocked read of vnp_size is safe because truncation cannot * pass busied page. But we load vnp_size into a local * variable so that possible concurrent extension does not * break calculation. */ #if defined(__powerpc__) && !defined(__powerpc64__) vsz = obj->un_pager.vnp.vnp_size; #else vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size); #endif if (uio->uio_offset + resid > vsz) resid = vsz - uio->uio_offset; error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio); out: for (j = 0; j < i; j++) { if (error == 0) vm_page_reference(ma[j]); vm_page_sunbusy(ma[j]); } out_pip: vm_object_pip_wakeup(obj); if (error != 0) return (error); return (uio->uio_resid == 0 ? 0 : EJUSTRETURN); } /* * File table vnode read routine. */ static int vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct vnode *vp; off_t orig_offset; int error, ioflag; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; ioflag = 0; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; /* * Try to read from page cache. VIRF_DOOMED check is racy but * allows us to avoid unneeded work outright. */ if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() && (vp->v_irflag & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) { error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred); if (error == 0) { fp->f_nextoff[UIO_READ] = uio->uio_offset; return (0); } if (error != EJUSTRETURN) return (error); } advice = get_advice(fp, uio); vn_lock(vp, LK_SHARED | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* Disable read-ahead for random I/O. */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_READ(vp, uio, ioflag, fp->f_cred); fp->f_nextoff[UIO_READ] = uio->uio_offset; VOP_UNLOCK(vp); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * read(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); return (error); } /* * File table vnode write routine. */ static int vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct vnode *vp; struct mount *mp; off_t orig_offset; int error, ioflag, lock_flags; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; if (vp->v_type == VREG) bwillwrite(); ioflag = IO_UNIT; if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) ioflag |= IO_APPEND; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; if ((fp->f_flag & O_FSYNC) || (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) ioflag |= IO_SYNC; mp = NULL; if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto unlock; advice = get_advice(fp, uio); if (MNT_SHARED_WRITES(mp) || (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { lock_flags = LK_SHARED; } else { lock_flags = LK_EXCLUSIVE; } vn_lock(vp, lock_flags | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* XXX: Is this correct? */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); fp->f_nextoff[UIO_WRITE] = uio->uio_offset; VOP_UNLOCK(vp); if (vp->v_type != VCHR) vn_finished_write(mp); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * write(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); unlock: return (error); } /* * The vn_io_fault() is a wrapper around vn_read() and vn_write() to * prevent the following deadlock: * * Assume that the thread A reads from the vnode vp1 into userspace * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is * currently not resident, then system ends up with the call chain * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) * which establishes lock order vp1->vn_lock, then vp2->vn_lock. * If, at the same time, thread B reads from vnode vp2 into buffer buf2 * backed by the pages of vnode vp1, and some page in buf2 is not * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. * * To prevent the lock order reversal and deadlock, vn_io_fault() does * not allow page faults to happen during VOP_READ() or VOP_WRITE(). * Instead, it first tries to do the whole range i/o with pagefaults * disabled. If all pages in the i/o buffer are resident and mapped, * VOP will succeed (ignoring the genuine filesystem errors). * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do * i/o in chunks, with all pages in the chunk prefaulted and held * using vm_fault_quick_hold_pages(). * * Filesystems using this deadlock avoidance scheme should use the * array of the held pages from uio, saved in the curthread->td_ma, * instead of doing uiomove(). A helper function * vn_io_fault_uiomove() converts uiomove request into * uiomove_fromphys() over td_ma array. * * Since vnode locks do not cover the whole i/o anymore, rangelocks * make the current i/o request atomic with respect to other i/os and * truncations. */ /* * Decode vn_io_fault_args and perform the corresponding i/o. */ static int vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, struct thread *td) { int error, save; error = 0; save = vm_fault_disable_pagefaults(); switch (args->kind) { case VN_IO_FAULT_FOP: error = (args->args.fop_args.doio)(args->args.fop_args.fp, uio, args->cred, args->flags, td); break; case VN_IO_FAULT_VOP: if (uio->uio_rw == UIO_READ) { error = VOP_READ(args->args.vop_args.vp, uio, args->flags, args->cred); } else if (uio->uio_rw == UIO_WRITE) { error = VOP_WRITE(args->args.vop_args.vp, uio, args->flags, args->cred); } break; default: panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind, uio->uio_rw); } vm_fault_enable_pagefaults(save); return (error); } static int vn_io_fault_touch(char *base, const struct uio *uio) { int r; r = fubyte(base); if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) return (EFAULT); return (0); } static int vn_io_fault_prefault_user(const struct uio *uio) { char *base; const struct iovec *iov; size_t len; ssize_t resid; int error, i; KASSERT(uio->uio_segflg == UIO_USERSPACE, ("vn_io_fault_prefault userspace")); error = i = 0; iov = uio->uio_iov; resid = uio->uio_resid; base = iov->iov_base; len = iov->iov_len; while (resid > 0) { error = vn_io_fault_touch(base, uio); if (error != 0) break; if (len < PAGE_SIZE) { if (len != 0) { error = vn_io_fault_touch(base + len - 1, uio); if (error != 0) break; resid -= len; } if (++i >= uio->uio_iovcnt) break; iov = uio->uio_iov + i; base = iov->iov_base; len = iov->iov_len; } else { len -= PAGE_SIZE; base += PAGE_SIZE; resid -= PAGE_SIZE; } } return (error); } /* * Common code for vn_io_fault(), agnostic to the kind of i/o request. * Uses vn_io_fault_doio() to make the call to an actual i/o function. * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request * into args and call vn_io_fault1() to handle faults during the user * mode buffer accesses. */ static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td) { vm_page_t ma[io_hold_cnt + 2]; struct uio *uio_clone, short_uio; struct iovec short_iovec[1]; vm_page_t *prev_td_ma; vm_prot_t prot; vm_offset_t addr, end; size_t len, resid; ssize_t adv; int error, cnt, saveheld, prev_td_ma_cnt; if (vn_io_fault_prefault) { error = vn_io_fault_prefault_user(uio); if (error != 0) return (error); /* Or ignore ? */ } prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; /* * The UFS follows IO_UNIT directive and replays back both * uio_offset and uio_resid if an error is encountered during the * operation. But, since the iovec may be already advanced, * uio is still in an inconsistent state. * * Cache a copy of the original uio, which is advanced to the redo * point using UIO_NOCOPY below. */ uio_clone = cloneuio(uio); resid = uio->uio_resid; short_uio.uio_segflg = UIO_USERSPACE; short_uio.uio_rw = uio->uio_rw; short_uio.uio_td = uio->uio_td; error = vn_io_fault_doio(args, uio, td); if (error != EFAULT) goto out; atomic_add_long(&vn_io_faults_cnt, 1); uio_clone->uio_segflg = UIO_NOCOPY; uiomove(NULL, resid - uio->uio_resid, uio_clone); uio_clone->uio_segflg = uio->uio_segflg; saveheld = curthread_pflags_set(TDP_UIOHELD); prev_td_ma = td->td_ma; prev_td_ma_cnt = td->td_ma_cnt; while (uio_clone->uio_resid != 0) { len = uio_clone->uio_iov->iov_len; if (len == 0) { KASSERT(uio_clone->uio_iovcnt >= 1, ("iovcnt underflow")); uio_clone->uio_iov++; uio_clone->uio_iovcnt--; continue; } if (len > ptoa(io_hold_cnt)) len = ptoa(io_hold_cnt); addr = (uintptr_t)uio_clone->uio_iov->iov_base; end = round_page(addr + len); if (end < addr) { error = EFAULT; break; } cnt = atop(end - trunc_page(addr)); /* * A perfectly misaligned address and length could cause * both the start and the end of the chunk to use partial * page. +2 accounts for such a situation. */ cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, addr, len, prot, ma, io_hold_cnt + 2); if (cnt == -1) { error = EFAULT; break; } short_uio.uio_iov = &short_iovec[0]; short_iovec[0].iov_base = (void *)addr; short_uio.uio_iovcnt = 1; short_uio.uio_resid = short_iovec[0].iov_len = len; short_uio.uio_offset = uio_clone->uio_offset; td->td_ma = ma; td->td_ma_cnt = cnt; error = vn_io_fault_doio(args, &short_uio, td); vm_page_unhold_pages(ma, cnt); adv = len - short_uio.uio_resid; uio_clone->uio_iov->iov_base = (char *)uio_clone->uio_iov->iov_base + adv; uio_clone->uio_iov->iov_len -= adv; uio_clone->uio_resid -= adv; uio_clone->uio_offset += adv; uio->uio_resid -= adv; uio->uio_offset += adv; if (error != 0 || adv == 0) break; } td->td_ma = prev_td_ma; td->td_ma_cnt = prev_td_ma_cnt; curthread_pflags_restore(saveheld); out: free(uio_clone, M_IOV); return (error); } static int vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { fo_rdwr_t *doio; struct vnode *vp; void *rl_cookie; struct vn_io_fault_args args; int error; doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; vp = fp->f_vnode; /* * The ability to read(2) on a directory has historically been * allowed for all users, but this can and has been the source of * at least one security issue in the past. As such, it is now hidden * away behind a sysctl for those that actually need it to use it, and * restricted to root when it's turned on to make it relatively safe to * leave on for longer sessions of need. */ if (vp->v_type == VDIR) { KASSERT(uio->uio_rw == UIO_READ, ("illegal write attempted on a directory")); if (!vfs_allow_read_dir) return (EISDIR); if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) return (EISDIR); } foffset_lock_uio(fp, uio, flags); if (do_vn_io_fault(vp, uio)) { args.kind = VN_IO_FAULT_FOP; args.args.fop_args.fp = fp; args.args.fop_args.doio = doio; args.cred = active_cred; args.flags = flags | FOF_OFFSET; if (uio->uio_rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } else if ((fp->f_flag & O_APPEND) != 0 || (flags & FOF_OFFSET) == 0) { /* For appenders, punt and lock the whole range. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); } else { rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } error = vn_io_fault1(vp, uio, &args, td); vn_rangelock_unlock(vp, rl_cookie); } else { error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); } foffset_unlock_uio(fp, uio, flags); return (error); } /* * Helper function to perform the requested uiomove operation using * the held pages for io->uio_iov[0].iov_base buffer instead of * copyin/copyout. Access to the pages with uiomove_fromphys() * instead of iov_base prevents page faults that could occur due to * pmap_collect() invalidating the mapping created by * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or * object cleanup revoking the write access from page mappings. * * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() * instead of plain uiomove(). */ int vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) { struct uio transp_uio; struct iovec transp_iov[1]; struct thread *td; size_t adv; int error, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove(data, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); transp_iov[0].iov_base = data; transp_uio.uio_iov = &transp_iov[0]; transp_uio.uio_iovcnt = 1; if (xfersize > uio->uio_resid) xfersize = uio->uio_resid; transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; transp_uio.uio_offset = 0; transp_uio.uio_segflg = UIO_SYSSPACE; /* * Since transp_iov points to data, and td_ma page array * corresponds to original uio->uio_iov, we need to invert the * direction of the i/o operation as passed to * uiomove_fromphys(). */ switch (uio->uio_rw) { case UIO_WRITE: transp_uio.uio_rw = UIO_READ; break; case UIO_READ: transp_uio.uio_rw = UIO_WRITE; break; } transp_uio.uio_td = uio->uio_td; error = uiomove_fromphys(td->td_ma, ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, xfersize, &transp_uio); adv = xfersize - transp_uio.uio_resid; pgadv = (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; uio->uio_iov->iov_len -= adv; uio->uio_resid -= adv; uio->uio_offset += adv; return (error); } int vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, struct uio *uio) { struct thread *td; vm_offset_t iov_base; int cnt, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove_fromphys(ma, offset, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; iov_base = (vm_offset_t)uio->uio_iov->iov_base; switch (uio->uio_rw) { case UIO_WRITE: pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, offset, cnt); break; case UIO_READ: pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, cnt); break; } pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)(iov_base + cnt); uio->uio_iov->iov_len -= cnt; uio->uio_resid -= cnt; uio->uio_offset += cnt; return (0); } /* * File table truncate routine. */ static int vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct mount *mp; struct vnode *vp; void *rl_cookie; int error; vp = fp->f_vnode; /* * Lock the whole range for truncation. Otherwise split i/o * might happen partly before and partly after the truncation. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error) goto out1; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); if (vp->v_type == VDIR) { error = EISDIR; goto out; } #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error) goto out; #endif error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, fp->f_cred); out: VOP_UNLOCK(vp); vn_finished_write(mp); out1: vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * Truncate a file that is already locked. */ int vn_truncate_locked(struct vnode *vp, off_t length, bool sync, struct ucred *cred) { struct vattr vattr; int error; error = VOP_ADD_WRITECOUNT(vp, 1); if (error == 0) { VATTR_NULL(&vattr); vattr.va_size = length; if (sync) vattr.va_vaflags |= VA_SYNC; error = VOP_SETATTR(vp, &vattr, cred); VOP_ADD_WRITECOUNT_CHECKED(vp, -1); } return (error); } /* * File table vnode stat routine. */ static int vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct vnode *vp = fp->f_vnode; int error; vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_STAT(vp, sb, active_cred, fp->f_cred, td); VOP_UNLOCK(vp); return (error); } /* * File table vnode ioctl routine. */ static int vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { struct vattr vattr; struct vnode *vp; struct fiobmap2_arg *bmarg; int error; vp = fp->f_vnode; switch (vp->v_type) { case VDIR: case VREG: switch (com) { case FIONREAD: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, active_cred); VOP_UNLOCK(vp); if (error == 0) *(int *)data = vattr.va_size - fp->f_offset; return (error); case FIOBMAP2: bmarg = (struct fiobmap2_arg *)data; vn_lock(vp, LK_SHARED | LK_RETRY); #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_BMAP(vp, bmarg->bn, NULL, &bmarg->bn, &bmarg->runp, &bmarg->runb); VOP_UNLOCK(vp); return (error); case FIONBIO: case FIOASYNC: return (0); default: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); } break; case VCHR: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); default: return (ENOTTY); } } /* * File table vnode poll routine. */ static int vn_poll(struct file *fp, int events, struct ucred *active_cred, struct thread *td) { struct vnode *vp; int error; vp = fp->f_vnode; #if defined(MAC) || defined(AUDIT) if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); VOP_UNLOCK(vp); if (error != 0) return (error); } #endif error = VOP_POLL(vp, events, fp->f_cred, td); return (error); } /* * Acquire the requested lock and then check for validity. LK_RETRY * permits vn_lock to return doomed vnodes. */ static int __noinline _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, int error) { KASSERT((flags & LK_RETRY) == 0 || error == 0, ("vn_lock: error %d incompatible with flags %#x", error, flags)); if (error == 0) VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); if ((flags & LK_RETRY) == 0) { if (error == 0) { VOP_UNLOCK(vp); error = ENOENT; } return (error); } /* * LK_RETRY case. * * Nothing to do if we got the lock. */ if (error == 0) return (0); /* * Interlock was dropped by the call in _vn_lock. */ flags &= ~LK_INTERLOCK; do { error = VOP_LOCK1(vp, flags, file, line); } while (error != 0); return (0); } int _vn_lock(struct vnode *vp, int flags, const char *file, int line) { int error; VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock: no locktype (%d passed)", flags)); VNPASS(vp->v_holdcnt > 0, vp); error = VOP_LOCK1(vp, flags, file, line); if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) return (_vn_lock_fallback(vp, flags, file, line, error)); return (0); } /* * File table vnode close routine. */ static int vn_closefile(struct file *fp, struct thread *td) { struct vnode *vp; struct flock lf; int error; bool ref; vp = fp->f_vnode; fp->f_ops = &badfileops; ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); if (__predict_false(ref)) { lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = F_UNLCK; (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); vrele(vp); } return (error); } /* * Preparing to start a filesystem write operation. If the operation is * permitted, then we bump the count of operations in progress and * proceed. If a suspend request is in progress, we wait until the * suspension is over, and then proceed. */ static int vn_start_write_refed(struct mount *mp, int flags, bool mplocked) { + struct mount_pcpu *mpcpu; int error, mflags; if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && - vfs_op_thread_enter(mp)) { + vfs_op_thread_enter(mp, mpcpu)) { MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); - vfs_mp_count_add_pcpu(mp, writeopcount, 1); - vfs_op_thread_exit(mp); + vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1); + vfs_op_thread_exit(mp, mpcpu); return (0); } if (mplocked) mtx_assert(MNT_MTX(mp), MA_OWNED); else MNT_ILOCK(mp); error = 0; /* * Check on status of suspension. */ if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || mp->mnt_susp_owner != curthread) { mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0) | (PUSER - 1); while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { if (flags & V_NOWAIT) { error = EWOULDBLOCK; goto unlock; } error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); if (error) goto unlock; } } if (flags & V_XSLEEP) goto unlock; mp->mnt_writeopcount++; unlock: if (error != 0 || (flags & V_XSLEEP) != 0) MNT_REL(mp); MNT_IUNLOCK(mp); return (error); } int vn_start_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); error = 0; /* * If a vnode is provided, get and return the mount point that * to which it will write. */ if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ if (vp == NULL && (flags & V_MNTREF) == 0) vfs_ref(mp); return (vn_start_write_refed(mp, flags, false)); } /* * Secondary suspension. Used by operations such as vop_inactive * routines that are needed by the higher level functions. These * are allowed to proceed until all the higher level functions have * completed (indicated by mnt_writeopcount dropping to zero). At that * time, these operations are halted until the suspension is over. */ int vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); retry: if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } /* * If we are not suspended or have not yet reached suspended * mode, then let the operation proceed. */ if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL && (flags & V_MNTREF) == 0) MNT_REF(mp); if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { mp->mnt_secondary_writes++; mp->mnt_secondary_accwrites++; MNT_IUNLOCK(mp); return (0); } if (flags & V_NOWAIT) { MNT_REL(mp); MNT_IUNLOCK(mp); return (EWOULDBLOCK); } /* * Wait for the suspension to finish. */ error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), "suspfs", 0); vfs_rel(mp); if (error == 0) goto retry; return (error); } /* * Filesystem write operation has completed. If we are suspending and this * operation is the last one, notify the suspender that the suspension is * now in effect. */ void vn_finished_write(struct mount *mp) { + struct mount_pcpu *mpcpu; int c; if (mp == NULL) return; - if (vfs_op_thread_enter(mp)) { - vfs_mp_count_sub_pcpu(mp, writeopcount, 1); - vfs_mp_count_sub_pcpu(mp, ref, 1); - vfs_op_thread_exit(mp); + if (vfs_op_thread_enter(mp, mpcpu)) { + vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1); + vfs_mp_count_sub_pcpu(mpcpu, ref, 1); + vfs_op_thread_exit(mp, mpcpu); return; } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REL(mp); c = --mp->mnt_writeopcount; if (mp->mnt_vfs_ops == 0) { MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); MNT_IUNLOCK(mp); return; } if (c < 0) vfs_dump_mount_counters(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) wakeup(&mp->mnt_writeopcount); MNT_IUNLOCK(mp); } /* * Filesystem secondary write operation has completed. If we are * suspending and this operation is the last one, notify the suspender * that the suspension is now in effect. */ void vn_finished_secondary_write(struct mount *mp) { if (mp == NULL) return; MNT_ILOCK(mp); MNT_REL(mp); mp->mnt_secondary_writes--; if (mp->mnt_secondary_writes < 0) panic("vn_finished_secondary_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_secondary_writes <= 0) wakeup(&mp->mnt_secondary_writes); MNT_IUNLOCK(mp); } /* * Request a filesystem to suspend write operations. */ int vfs_write_suspend(struct mount *mp, int flags) { int error; vfs_op_enter(mp); MNT_ILOCK(mp); vfs_assert_mount_counters(mp); if (mp->mnt_susp_owner == curthread) { vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); return (EALREADY); } while (mp->mnt_kern_flag & MNTK_SUSPEND) msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); /* * Unmount holds a write reference on the mount point. If we * own busy reference and drain for writers, we deadlock with * the reference draining in the unmount path. Callers of * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if * vfs_busy() reference is owned and caller is not in the * unmount context. */ if ((flags & VS_SKIP_UNMOUNT) != 0 && (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); return (EBUSY); } mp->mnt_kern_flag |= MNTK_SUSPEND; mp->mnt_susp_owner = curthread; if (mp->mnt_writeopcount > 0) (void) msleep(&mp->mnt_writeopcount, MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); else MNT_IUNLOCK(mp); if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { vfs_write_resume(mp, 0); /* vfs_write_resume does vfs_op_exit() for us */ } return (error); } /* * Request a filesystem to resume write operations. */ void vfs_write_resume(struct mount *mp, int flags) { MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | MNTK_SUSPENDED); mp->mnt_susp_owner = NULL; wakeup(&mp->mnt_writeopcount); wakeup(&mp->mnt_flag); curthread->td_pflags &= ~TDP_IGNSUSP; if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); mp->mnt_writeopcount++; } MNT_IUNLOCK(mp); if ((flags & VR_NO_SUSPCLR) == 0) VFS_SUSP_CLEAN(mp); vfs_op_exit(mp); } else if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); vn_start_write_refed(mp, 0, true); } else { MNT_IUNLOCK(mp); } } /* * Helper loop around vfs_write_suspend() for filesystem unmount VFS * methods. */ int vfs_write_suspend_umnt(struct mount *mp) { int error; KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, ("vfs_write_suspend_umnt: recursed")); /* dounmount() already called vn_start_write(). */ for (;;) { vn_finished_write(mp); error = vfs_write_suspend(mp, 0); if (error != 0) { vn_start_write(NULL, &mp, V_WAIT); return (error); } MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) break; MNT_IUNLOCK(mp); vn_start_write(NULL, &mp, V_WAIT); } mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); wakeup(&mp->mnt_flag); MNT_IUNLOCK(mp); curthread->td_pflags |= TDP_IGNSUSP; return (0); } /* * Implement kqueues for files by translating it to vnode operation. */ static int vn_kqfilter(struct file *fp, struct knote *kn) { return (VOP_KQFILTER(fp->f_vnode, kn)); } /* * Simplified in-kernel wrapper calls for extended attribute access. * Both calls pass in a NULL credential, authorizing as "kernel" access. * Set IO_NODELOCKED in ioflg if the vnode is already locked. */ int vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int *buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; int error; iov.iov_len = *buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = *buflen; if ((ioflg & IO_NODELOCKED) == 0) vn_lock(vp, LK_SHARED | LK_RETRY); ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute retrieval as kernel */ error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) VOP_UNLOCK(vp); if (error == 0) { *buflen = *buflen - auio.uio_resid; } return (error); } /* * XXX failure mode if partially written? */ int vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; struct mount *mp; int error; iov.iov_len = buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_WRITE; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = buflen; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute setting as kernel */ error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp); } return (error); } int vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, struct thread *td) { struct mount *mp; int error; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute removal as kernel */ error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); if (error == EOPNOTSUPP) error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp); } return (error); } static int vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, struct vnode **rvp) { return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); } int vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) { return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, lkflags, rvp)); } int vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, int lkflags, struct vnode **rvp) { struct mount *mp; int ltype, error; ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); mp = vp->v_mount; ltype = VOP_ISLOCKED(vp); KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, ("vn_vget_ino: vp not locked")); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { vfs_ref(mp); VOP_UNLOCK(vp); error = vfs_busy(mp, 0); vn_lock(vp, ltype | LK_RETRY); vfs_rel(mp); if (error != 0) return (ENOENT); if (VN_IS_DOOMED(vp)) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(vp); error = alloc(mp, alloc_arg, lkflags, rvp); vfs_unbusy(mp); if (error != 0 || *rvp != vp) vn_lock(vp, ltype | LK_RETRY); if (VN_IS_DOOMED(vp)) { if (error == 0) { if (*rvp == vp) vunref(vp); else vput(*rvp); } error = ENOENT; } return (error); } int vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, struct thread *td) { if (vp->v_type != VREG || td == NULL) return (0); if ((uoff_t)uio->uio_offset + uio->uio_resid > lim_cur(td, RLIMIT_FSIZE)) { PROC_LOCK(td->td_proc); kern_psignal(td->td_proc, SIGXFSZ); PROC_UNLOCK(td->td_proc); return (EFBIG); } return (0); } int vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp); #endif return (setfmode(td, active_cred, vp, mode)); } int vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp); #endif return (setfown(td, active_cred, vp, uid, gid)); } void vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) { vm_object_t object; if ((object = vp->v_object) == NULL) return; VM_OBJECT_WLOCK(object); vm_object_page_remove(object, start, end, 0); VM_OBJECT_WUNLOCK(object); } int vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) { struct vattr va; daddr_t bn, bnp; uint64_t bsize; off_t noff; int error; KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, ("Wrong command %lu", cmd)); if (vn_lock(vp, LK_SHARED) != 0) return (EBADF); if (vp->v_type != VREG) { error = ENOTTY; goto unlock; } error = VOP_GETATTR(vp, &va, cred); if (error != 0) goto unlock; noff = *off; if (noff >= va.va_size) { error = ENXIO; goto unlock; } bsize = vp->v_mount->mnt_stat.f_iosize; for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - noff % bsize) { error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); if (error == EOPNOTSUPP) { error = ENOTTY; goto unlock; } if ((bnp == -1 && cmd == FIOSEEKHOLE) || (bnp != -1 && cmd == FIOSEEKDATA)) { noff = bn * bsize; if (noff < *off) noff = *off; goto unlock; } } if (noff > va.va_size) noff = va.va_size; /* noff == va.va_size. There is an implicit hole at the end of file. */ if (cmd == FIOSEEKDATA) error = ENXIO; unlock: VOP_UNLOCK(vp); if (error == 0) *off = noff; return (error); } int vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct ucred *cred; struct vnode *vp; struct vattr vattr; off_t foffset, size; int error, noneg; cred = td->td_ucred; vp = fp->f_vnode; foffset = foffset_lock(fp, 0); noneg = (vp->v_type != VCHR); error = 0; switch (whence) { case L_INCR: if (noneg && (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, cred); VOP_UNLOCK(vp); if (error) break; /* * If the file references a disk device, then fetch * the media size and use that to determine the ending * offset. */ if (vattr.va_size == 0 && vp->v_type == VCHR && fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) vattr.va_size = size; if (noneg && (vattr.va_size > OFF_MAX || (offset > 0 && vattr.va_size > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += vattr.va_size; break; case L_SET: break; case SEEK_DATA: error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); if (error == ENOTTY) error = EINVAL; break; case SEEK_HOLE: error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); if (error == ENOTTY) error = EINVAL; break; default: error = EINVAL; } if (error == 0 && noneg && offset < 0) error = EINVAL; if (error != 0) goto drop; VFS_KNOTE_UNLOCKED(vp, 0); td->td_uretoff.tdu_off = offset; drop: foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } int vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *td) { int error; /* * Grant permission if the caller is the owner of the file, or * the super-user, or has ACL_WRITE_ATTRIBUTES permission on * on the file. If the time pointer is null, then write * permission on the file is also sufficient. * * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES * will be allowed to set the times [..] to the current * server time. */ error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) error = VOP_ACCESS(vp, VWRITE, cred, td); return (error); } int vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) { struct vnode *vp; int error; if (fp->f_type == DTYPE_FIFO) kif->kf_type = KF_TYPE_FIFO; else kif->kf_type = KF_TYPE_VNODE; vp = fp->f_vnode; vref(vp); FILEDESC_SUNLOCK(fdp); error = vn_fill_kinfo_vnode(vp, kif); vrele(vp); FILEDESC_SLOCK(fdp); return (error); } static inline void vn_fill_junk(struct kinfo_file *kif) { size_t len, olen; /* * Simulate vn_fullpath returning changing values for a given * vp during e.g. coredump. */ len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; olen = strlen(kif->kf_path); if (len < olen) strcpy(&kif->kf_path[len - 1], "$"); else for (; olen < len; olen++) strcpy(&kif->kf_path[olen], "A"); } int vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) { struct vattr va; char *fullpath, *freepath; int error; kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); freepath = NULL; fullpath = "-"; error = vn_fullpath(vp, &fullpath, &freepath); if (error == 0) { strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); } if (freepath != NULL) free(freepath, M_TEMP); KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, vn_fill_junk(kif); ); /* * Retrieve vnode attributes. */ va.va_fsid = VNOVAL; va.va_rdev = NODEV; vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &va, curthread->td_ucred); VOP_UNLOCK(vp); if (error != 0) return (error); if (va.va_fsid != VNOVAL) kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; else kif->kf_un.kf_file.kf_file_fsid = vp->v_mount->mnt_stat.f_fsid.val[0]; kif->kf_un.kf_file.kf_file_fsid_freebsd11 = kif->kf_un.kf_file.kf_file_fsid; /* truncate */ kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); kif->kf_un.kf_file.kf_file_size = va.va_size; kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; kif->kf_un.kf_file.kf_file_rdev_freebsd11 = kif->kf_un.kf_file.kf_file_rdev; /* truncate */ return (0); } int vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { #ifdef HWPMC_HOOKS struct pmckern_map_in pkm; #endif struct mount *mp; struct vnode *vp; vm_object_t object; vm_prot_t maxprot; boolean_t writecounted; int error; #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) /* * POSIX shared-memory objects are defined to have * kernel persistence, and are not defined to support * read(2)/write(2) -- or even open(2). Thus, we can * use MAP_ASYNC to trade on-disk coherence for speed. * The shm_open(3) library routine turns on the FPOSIXSHM * flag to request this behavior. */ if ((fp->f_flag & FPOSIXSHM) != 0) flags |= MAP_NOSYNC; #endif vp = fp->f_vnode; /* * Ensure that file and memory protections are * compatible. Note that we only worry about * writability if mapping is shared; in this case, * current and max prot are dictated by the open file. * XXX use the vnode instead? Problem is: what * credentials do we use for determination? What if * proc does a setuid? */ mp = vp->v_mount; if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { maxprot = VM_PROT_NONE; if ((prot & VM_PROT_EXECUTE) != 0) return (EACCES); } else maxprot = VM_PROT_EXECUTE; if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_READ; else if ((prot & VM_PROT_READ) != 0) return (EACCES); /* * If we are sharing potential changes via MAP_SHARED and we * are trying to get write permission although we opened it * without asking for it, bail out. */ if ((flags & MAP_SHARED) != 0) { if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; else if ((prot & VM_PROT_WRITE) != 0) return (EACCES); } else { maxprot |= VM_PROT_WRITE; cap_maxprot |= VM_PROT_WRITE; } maxprot &= cap_maxprot; /* * For regular files and shared memory, POSIX requires that * the value of foff be a legitimate offset within the data * object. In particular, negative offsets are invalid. * Blocking negative offsets and overflows here avoids * possible wraparound or user-level access into reserved * ranges of the data object later. In contrast, POSIX does * not dictate how offsets are used by device drivers, so in * the case of a device mapping a negative offset is passed * on. */ if ( #ifdef _LP64 size > OFF_MAX || #endif foff > OFF_MAX - size) return (EINVAL); writecounted = FALSE; error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, &foff, &object, &writecounted); if (error != 0) return (error); error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, foff, writecounted, td); if (error != 0) { /* * If this mapping was accounted for in the vnode's * writecount, then undo that now. */ if (writecounted) vm_pager_release_writecount(object, 0, size); vm_object_deallocate(object); } #ifdef HWPMC_HOOKS /* Inform hwpmc(4) if an executable is being mapped. */ if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { pkm.pm_file = vp; pkm.pm_address = (uintptr_t) *addr; PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); } } #endif return (error); } void vn_fsid(struct vnode *vp, struct vattr *va) { fsid_t *f; f = &vp->v_mount->mnt_stat.f_fsid; va->va_fsid = (uint32_t)f->val[1]; va->va_fsid <<= sizeof(f->val[1]) * NBBY; va->va_fsid += (uint32_t)f->val[0]; } int vn_fsync_buf(struct vnode *vp, int waitfor) { struct buf *bp, *nbp; struct bufobj *bo; struct mount *mp; int error, maxretry; error = 0; maxretry = 10000; /* large, arbitrarily chosen */ mp = NULL; if (vp->v_type == VCHR) { VI_LOCK(vp); mp = vp->v_rdev->si_mountpt; VI_UNLOCK(vp); } bo = &vp->v_bufobj; BO_LOCK(bo); loop1: /* * MARK/SCAN initialization to avoid infinite loops. */ TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { bp->b_vflags &= ~BV_SCANNED; bp->b_error = 0; } /* * Flush all dirty buffers associated with a vnode. */ loop2: TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if ((bp->b_vflags & BV_SCANNED) != 0) continue; bp->b_vflags |= BV_SCANNED; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { if (waitfor != MNT_WAIT) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, BO_LOCKPTR(bo)) != 0) { BO_LOCK(bo); goto loop1; } BO_LOCK(bo); } BO_UNLOCK(bo); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); if ((bp->b_flags & B_DELWRI) == 0) panic("fsync: not dirty"); if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { vfs_bio_awrite(bp); } else { bremfree(bp); bawrite(bp); } if (maxretry < 1000) pause("dirty", hz < 1000 ? 1 : hz / 1000); BO_LOCK(bo); goto loop2; } /* * If synchronous the caller expects us to completely resolve all * dirty buffers in the system. Wait for in-progress I/O to * complete (which could include background bitmap writes), then * retry if dirty blocks still exist. */ if (waitfor == MNT_WAIT) { bufobj_wwait(bo, 0, 0); if (bo->bo_dirty.bv_cnt > 0) { /* * If we are unable to write any of these buffers * then we fail now rather than trying endlessly * to write them out. */ TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) if ((error = bp->b_error) != 0) break; if ((mp != NULL && mp->mnt_secondary_writes > 0) || (error == 0 && --maxretry >= 0)) goto loop1; if (error == 0) error = EAGAIN; } } BO_UNLOCK(bo); if (error != 0) vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); return (error); } /* * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() * or vn_generic_copy_file_range() after rangelocking the byte ranges, * to do the actual copy. * vn_generic_copy_file_range() is factored out, so it can be called * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from * different file systems. */ int vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) { int error; size_t len; uint64_t uval; len = *lenp; *lenp = 0; /* For error returns. */ error = 0; /* Do some sanity checks on the arguments. */ if (invp->v_type == VDIR || outvp->v_type == VDIR) error = EISDIR; else if (*inoffp < 0 || *outoffp < 0 || invp->v_type != VREG || outvp->v_type != VREG) error = EINVAL; if (error != 0) goto out; /* Ensure offset + len does not wrap around. */ uval = *inoffp; uval += len; if (uval > INT64_MAX) len = INT64_MAX - *inoffp; uval = *outoffp; uval += len; if (uval > INT64_MAX) len = INT64_MAX - *outoffp; if (len == 0) goto out; /* * If the two vnode are for the same file system, call * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() * which can handle copies across multiple file systems. */ *lenp = len; if (invp->v_mount == outvp->v_mount) error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, lenp, flags, incred, outcred, fsize_td); else error = vn_generic_copy_file_range(invp, inoffp, outvp, outoffp, lenp, flags, incred, outcred, fsize_td); out: return (error); } /* * Test len bytes of data starting at dat for all bytes == 0. * Return true if all bytes are zero, false otherwise. * Expects dat to be well aligned. */ static bool mem_iszero(void *dat, int len) { int i; const u_int *p; const char *cp; for (p = dat; len > 0; len -= sizeof(*p), p++) { if (len >= sizeof(*p)) { if (*p != 0) return (false); } else { cp = (const char *)p; for (i = 0; i < len; i++, cp++) if (*cp != '\0') return (false); } } return (true); } /* * Look for a hole in the output file and, if found, adjust *outoffp * and *xferp to skip past the hole. * *xferp is the entire hole length to be written and xfer2 is how many bytes * to be written as 0's upon return. */ static off_t vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, off_t *dataoffp, off_t *holeoffp, struct ucred *cred) { int error; off_t delta; if (*holeoffp == 0 || *holeoffp <= *outoffp) { *dataoffp = *outoffp; error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, curthread); if (error == 0) { *holeoffp = *dataoffp; error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, curthread); } if (error != 0 || *holeoffp == *dataoffp) { /* * Since outvp is unlocked, it may be possible for * another thread to do a truncate(), lseek(), write() * creating a hole at startoff between the above * VOP_IOCTL() calls, if the other thread does not do * rangelocking. * If that happens, *holeoffp == *dataoffp and finding * the hole has failed, so disable vn_skip_hole(). */ *holeoffp = -1; /* Disable use of vn_skip_hole(). */ return (xfer2); } KASSERT(*dataoffp >= *outoffp, ("vn_skip_hole: dataoff=%jd < outoff=%jd", (intmax_t)*dataoffp, (intmax_t)*outoffp)); KASSERT(*holeoffp > *dataoffp, ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", (intmax_t)*holeoffp, (intmax_t)*dataoffp)); } /* * If there is a hole before the data starts, advance *outoffp and * *xferp past the hole. */ if (*dataoffp > *outoffp) { delta = *dataoffp - *outoffp; if (delta >= *xferp) { /* Entire *xferp is a hole. */ *outoffp += *xferp; *xferp = 0; return (0); } *xferp -= delta; *outoffp += delta; xfer2 = MIN(xfer2, *xferp); } /* * If a hole starts before the end of this xfer2, reduce this xfer2 so * that the write ends at the start of the hole. * *holeoffp should always be greater than *outoffp, but for the * non-INVARIANTS case, check this to make sure xfer2 remains a sane * value. */ if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) xfer2 = *holeoffp - *outoffp; return (xfer2); } /* * Write an xfer sized chunk to outvp in blksize blocks from dat. * dat is a maximum of blksize in length and can be written repeatedly in * the chunk. * If growfile == true, just grow the file via vn_truncate_locked() instead * of doing actual writes. * If checkhole == true, a hole is being punched, so skip over any hole * already in the output file. */ static int vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, u_long blksize, bool growfile, bool checkhole, struct ucred *cred) { struct mount *mp; off_t dataoff, holeoff, xfer2; int error, lckf; /* * Loop around doing writes of blksize until write has been completed. * Lock/unlock on each loop iteration so that a bwillwrite() can be * done for each iteration, since the xfer argument can be very * large if there is a large hole to punch in the output file. */ error = 0; holeoff = 0; do { xfer2 = MIN(xfer, blksize); if (checkhole) { /* * Punching a hole. Skip writing if there is * already a hole in the output file. */ xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, &dataoff, &holeoff, cred); if (xfer == 0) break; if (holeoff < 0) checkhole = false; KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", (intmax_t)xfer2)); } bwillwrite(); mp = NULL; error = vn_start_write(outvp, &mp, V_WAIT); if (error != 0) break; if (growfile) { error = vn_lock(outvp, LK_EXCLUSIVE); if (error == 0) { error = vn_truncate_locked(outvp, outoff + xfer, false, cred); VOP_UNLOCK(outvp); } } else { if (MNT_SHARED_WRITES(mp)) lckf = LK_SHARED; else lckf = LK_EXCLUSIVE; error = vn_lock(outvp, lckf); if (error == 0) { error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, outoff, UIO_SYSSPACE, IO_NODELOCKED, curthread->td_ucred, cred, NULL, curthread); outoff += xfer2; xfer -= xfer2; } VOP_UNLOCK(outvp); } if (mp != NULL) vn_finished_write(mp); } while (!growfile && xfer > 0 && error == 0); return (error); } /* * Copy a byte range of one file to another. This function can handle the * case where invp and outvp are on different file systems. * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there * is no better file system specific way to do it. */ int vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) { struct vattr va; struct mount *mp; struct uio io; off_t startoff, endoff, xfer, xfer2; u_long blksize; int error, interrupted; bool cantseek, readzeros, eof, lastblock; ssize_t aresid; size_t copylen, len, rem, savlen; char *dat; long holein, holeout; holein = holeout = 0; savlen = len = *lenp; error = 0; interrupted = 0; dat = NULL; error = vn_lock(invp, LK_SHARED); if (error != 0) goto out; if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) holein = 0; VOP_UNLOCK(invp); mp = NULL; error = vn_start_write(outvp, &mp, V_WAIT); if (error == 0) error = vn_lock(outvp, LK_EXCLUSIVE); if (error == 0) { /* * If fsize_td != NULL, do a vn_rlimit_fsize() call, * now that outvp is locked. */ if (fsize_td != NULL) { io.uio_offset = *outoffp; io.uio_resid = len; error = vn_rlimit_fsize(outvp, &io, fsize_td); if (error != 0) error = EFBIG; } if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) holeout = 0; /* * Holes that are past EOF do not need to be written as a block * of zero bytes. So, truncate the output file as far as * possible and then use va.va_size to decide if writing 0 * bytes is necessary in the loop below. */ if (error == 0) error = VOP_GETATTR(outvp, &va, outcred); if (error == 0 && va.va_size > *outoffp && va.va_size <= *outoffp + len) { #ifdef MAC error = mac_vnode_check_write(curthread->td_ucred, outcred, outvp); if (error == 0) #endif error = vn_truncate_locked(outvp, *outoffp, false, outcred); if (error == 0) va.va_size = *outoffp; } VOP_UNLOCK(outvp); } if (mp != NULL) vn_finished_write(mp); if (error != 0) goto out; /* * Set the blksize to the larger of the hole sizes for invp and outvp. * If hole sizes aren't available, set the blksize to the larger * f_iosize of invp and outvp. * This code expects the hole sizes and f_iosizes to be powers of 2. * This value is clipped at 4Kbytes and 1Mbyte. */ blksize = MAX(holein, holeout); /* Clip len to end at an exact multiple of hole size. */ if (blksize > 1) { rem = *inoffp % blksize; if (rem > 0) rem = blksize - rem; if (len - rem > blksize) len = savlen = rounddown(len - rem, blksize) + rem; } if (blksize <= 1) blksize = MAX(invp->v_mount->mnt_stat.f_iosize, outvp->v_mount->mnt_stat.f_iosize); if (blksize < 4096) blksize = 4096; else if (blksize > 1024 * 1024) blksize = 1024 * 1024; dat = malloc(blksize, M_TEMP, M_WAITOK); /* * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA * to find holes. Otherwise, just scan the read block for all 0s * in the inner loop where the data copying is done. * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may * support holes on the server, but do not support FIOSEEKHOLE. */ eof = false; while (len > 0 && error == 0 && !eof && interrupted == 0) { endoff = 0; /* To shut up compilers. */ cantseek = true; startoff = *inoffp; copylen = len; /* * Find the next data area. If there is just a hole to EOF, * FIOSEEKDATA should fail and then we drop down into the * inner loop and create the hole on the outvp file. * (I do not know if any file system will report a hole to * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA * will fail for those file systems.) * * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, * the code just falls through to the inner copy loop. */ error = EINVAL; if (holein > 0) error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, incred, curthread); if (error == 0) { endoff = startoff; error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, incred, curthread); /* * Since invp is unlocked, it may be possible for * another thread to do a truncate(), lseek(), write() * creating a hole at startoff between the above * VOP_IOCTL() calls, if the other thread does not do * rangelocking. * If that happens, startoff == endoff and finding * the hole has failed, so set an error. */ if (error == 0 && startoff == endoff) error = EINVAL; /* Any error. Reset to 0. */ } if (error == 0) { if (startoff > *inoffp) { /* Found hole before data block. */ xfer = MIN(startoff - *inoffp, len); if (*outoffp < va.va_size) { /* Must write 0s to punch hole. */ xfer2 = MIN(va.va_size - *outoffp, xfer); memset(dat, 0, MIN(xfer2, blksize)); error = vn_write_outvp(outvp, dat, *outoffp, xfer2, blksize, false, holeout > 0, outcred); } if (error == 0 && *outoffp + xfer > va.va_size && xfer == len) /* Grow last block. */ error = vn_write_outvp(outvp, dat, *outoffp, xfer, blksize, true, false, outcred); if (error == 0) { *inoffp += xfer; *outoffp += xfer; len -= xfer; if (len < savlen) interrupted = sig_intr(); } } copylen = MIN(len, endoff - startoff); cantseek = false; } else { cantseek = true; startoff = *inoffp; copylen = len; error = 0; } xfer = blksize; if (cantseek) { /* * Set first xfer to end at a block boundary, so that * holes are more likely detected in the loop below via * the for all bytes 0 method. */ xfer -= (*inoffp % blksize); } /* Loop copying the data block. */ while (copylen > 0 && error == 0 && !eof && interrupted == 0) { if (copylen < xfer) xfer = copylen; error = vn_lock(invp, LK_SHARED); if (error != 0) goto out; error = vn_rdwr(UIO_READ, invp, dat, xfer, startoff, UIO_SYSSPACE, IO_NODELOCKED, curthread->td_ucred, incred, &aresid, curthread); VOP_UNLOCK(invp); lastblock = false; if (error == 0 && aresid > 0) { /* Stop the copy at EOF on the input file. */ xfer -= aresid; eof = true; lastblock = true; } if (error == 0) { /* * Skip the write for holes past the initial EOF * of the output file, unless this is the last * write of the output file at EOF. */ readzeros = cantseek ? mem_iszero(dat, xfer) : false; if (xfer == len) lastblock = true; if (!cantseek || *outoffp < va.va_size || lastblock || !readzeros) error = vn_write_outvp(outvp, dat, *outoffp, xfer, blksize, readzeros && lastblock && *outoffp >= va.va_size, false, outcred); if (error == 0) { *inoffp += xfer; startoff += xfer; *outoffp += xfer; copylen -= xfer; len -= xfer; if (len < savlen) interrupted = sig_intr(); } } xfer = blksize; } } out: *lenp = savlen - len; free(dat, M_TEMP); return (error); } static int vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) { struct mount *mp; struct vnode *vp; off_t olen, ooffset; int error; #ifdef AUDIT int audited_vnode1 = 0; #endif vp = fp->f_vnode; if (vp->v_type != VREG) return (ENODEV); /* Allocating blocks may take a long time, so iterate. */ for (;;) { olen = len; ooffset = offset; bwillwrite(); mp = NULL; error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error != 0) break; error = vn_lock(vp, LK_EXCLUSIVE); if (error != 0) { vn_finished_write(mp); break; } #ifdef AUDIT if (!audited_vnode1) { AUDIT_ARG_VNODE1(vp); audited_vnode1 = 1; } #endif #ifdef MAC error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); if (error == 0) #endif error = VOP_ALLOCATE(vp, &offset, &len); VOP_UNLOCK(vp); vn_finished_write(mp); if (olen + ooffset != offset + len) { panic("offset + len changed from %jx/%jx to %jx/%jx", ooffset, olen, offset, len); } if (error != 0 || len == 0) break; KASSERT(olen > len, ("Iteration did not make progress?")); maybe_yield(); } return (error); } Index: head/sys/sys/mount.h =================================================================== --- head/sys/sys/mount.h (revision 367534) +++ head/sys/sys/mount.h (revision 367535) @@ -1,1145 +1,1158 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1991, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)mount.h 8.21 (Berkeley) 5/20/95 * $FreeBSD$ */ #ifndef _SYS_MOUNT_H_ #define _SYS_MOUNT_H_ #include #include #ifdef _KERNEL #include #include #include #include #include #endif /* * NOTE: When changing statfs structure, mount structure, MNT_* flags or * MNTK_* flags also update DDB show mount command in vfs_subr.c. */ typedef struct fsid { int32_t val[2]; } fsid_t; /* filesystem id type */ #define fsidcmp(a, b) memcmp((a), (b), sizeof(fsid_t)) /* * File identifier. * These are unique per filesystem on a single machine. * * Note that the offset of fid_data is 4 bytes, so care must be taken to avoid * undefined behavior accessing unaligned fields within an embedded struct. */ #define MAXFIDSZ 16 struct fid { u_short fid_len; /* length of data in bytes */ u_short fid_data0; /* force longword alignment */ char fid_data[MAXFIDSZ]; /* data (variable length) */ }; /* * filesystem statistics */ #define MFSNAMELEN 16 /* length of type name including null */ #define MNAMELEN 1024 /* size of on/from name bufs */ #define STATFS_VERSION 0x20140518 /* current version number */ struct statfs { uint32_t f_version; /* structure version number */ uint32_t f_type; /* type of filesystem */ uint64_t f_flags; /* copy of mount exported flags */ uint64_t f_bsize; /* filesystem fragment size */ uint64_t f_iosize; /* optimal transfer block size */ uint64_t f_blocks; /* total data blocks in filesystem */ uint64_t f_bfree; /* free blocks in filesystem */ int64_t f_bavail; /* free blocks avail to non-superuser */ uint64_t f_files; /* total file nodes in filesystem */ int64_t f_ffree; /* free nodes avail to non-superuser */ uint64_t f_syncwrites; /* count of sync writes since mount */ uint64_t f_asyncwrites; /* count of async writes since mount */ uint64_t f_syncreads; /* count of sync reads since mount */ uint64_t f_asyncreads; /* count of async reads since mount */ uint64_t f_spare[10]; /* unused spare */ uint32_t f_namemax; /* maximum filename length */ uid_t f_owner; /* user that mounted the filesystem */ fsid_t f_fsid; /* filesystem id */ char f_charspare[80]; /* spare string space */ char f_fstypename[MFSNAMELEN]; /* filesystem type name */ char f_mntfromname[MNAMELEN]; /* mounted filesystem */ char f_mntonname[MNAMELEN]; /* directory on which mounted */ }; #if defined(_WANT_FREEBSD11_STATFS) || defined(_KERNEL) #define FREEBSD11_STATFS_VERSION 0x20030518 /* current version number */ struct freebsd11_statfs { uint32_t f_version; /* structure version number */ uint32_t f_type; /* type of filesystem */ uint64_t f_flags; /* copy of mount exported flags */ uint64_t f_bsize; /* filesystem fragment size */ uint64_t f_iosize; /* optimal transfer block size */ uint64_t f_blocks; /* total data blocks in filesystem */ uint64_t f_bfree; /* free blocks in filesystem */ int64_t f_bavail; /* free blocks avail to non-superuser */ uint64_t f_files; /* total file nodes in filesystem */ int64_t f_ffree; /* free nodes avail to non-superuser */ uint64_t f_syncwrites; /* count of sync writes since mount */ uint64_t f_asyncwrites; /* count of async writes since mount */ uint64_t f_syncreads; /* count of sync reads since mount */ uint64_t f_asyncreads; /* count of async reads since mount */ uint64_t f_spare[10]; /* unused spare */ uint32_t f_namemax; /* maximum filename length */ uid_t f_owner; /* user that mounted the filesystem */ fsid_t f_fsid; /* filesystem id */ char f_charspare[80]; /* spare string space */ char f_fstypename[16]; /* filesystem type name */ char f_mntfromname[88]; /* mounted filesystem */ char f_mntonname[88]; /* directory on which mounted */ }; #endif /* _WANT_FREEBSD11_STATFS || _KERNEL */ #ifdef _KERNEL #define OMFSNAMELEN 16 /* length of fs type name, including null */ #define OMNAMELEN (88 - 2 * sizeof(long)) /* size of on/from name bufs */ /* XXX getfsstat.2 is out of date with write and read counter changes here. */ /* XXX statfs.2 is out of date with read counter changes here. */ struct ostatfs { long f_spare2; /* placeholder */ long f_bsize; /* fundamental filesystem block size */ long f_iosize; /* optimal transfer block size */ long f_blocks; /* total data blocks in filesystem */ long f_bfree; /* free blocks in fs */ long f_bavail; /* free blocks avail to non-superuser */ long f_files; /* total file nodes in filesystem */ long f_ffree; /* free file nodes in fs */ fsid_t f_fsid; /* filesystem id */ uid_t f_owner; /* user that mounted the filesystem */ int f_type; /* type of filesystem */ int f_flags; /* copy of mount exported flags */ long f_syncwrites; /* count of sync writes since mount */ long f_asyncwrites; /* count of async writes since mount */ char f_fstypename[OMFSNAMELEN]; /* fs type name */ char f_mntonname[OMNAMELEN]; /* directory on which mounted */ long f_syncreads; /* count of sync reads since mount */ long f_asyncreads; /* count of async reads since mount */ short f_spares1; /* unused spare */ char f_mntfromname[OMNAMELEN];/* mounted filesystem */ short f_spares2; /* unused spare */ /* * XXX on machines where longs are aligned to 8-byte boundaries, there * is an unnamed int32_t here. This spare was after the apparent end * of the struct until we bit off the read counters from f_mntonname. */ long f_spare[2]; /* unused spare */ }; TAILQ_HEAD(vnodelst, vnode); /* Mount options list */ TAILQ_HEAD(vfsoptlist, vfsopt); struct vfsopt { TAILQ_ENTRY(vfsopt) link; char *name; void *value; int len; int pos; int seen; }; +struct mount_pcpu { + int mntp_thread_in_ops; + int mntp_ref; + int mntp_lockref; + int mntp_writeopcount; +}; + +_Static_assert(sizeof(struct mount_pcpu) == 16, + "the struct is allocated from pcpu 16 zone"); + /* * Structure per mounted filesystem. Each mounted filesystem has an * array of operations and an instance record. The filesystems are * put on a doubly linked list. * * Lock reference: * l - mnt_listmtx * m - mountlist_mtx * i - interlock * v - vnode freelist mutex * * Unmarked fields are considered stable as long as a ref is held. * */ struct mount { - struct mtx mnt_mtx; /* mount structure interlock */ + int mnt_vfs_ops; /* (i) pending vfs ops */ + int mnt_kern_flag; /* (i) kernel only flags */ + uint64_t mnt_flag; /* (i) flags shared with user */ + struct mount_pcpu *mnt_pcpu; /* per-CPU data */ + struct vnode *mnt_rootvnode; + struct vnode *mnt_vnodecovered; /* vnode we mounted on */ + struct vfsops *mnt_op; /* operations on fs */ + struct vfsconf *mnt_vfc; /* configuration info */ + struct mtx __aligned(CACHE_LINE_SIZE) mnt_mtx; /* mount structure interlock */ int mnt_gen; /* struct mount generation */ #define mnt_startzero mnt_list TAILQ_ENTRY(mount) mnt_list; /* (m) mount list */ - struct vfsops *mnt_op; /* operations on fs */ - struct vfsconf *mnt_vfc; /* configuration info */ - struct vnode *mnt_vnodecovered; /* vnode we mounted on */ struct vnode *mnt_syncer; /* syncer vnode */ int mnt_ref; /* (i) Reference count */ struct vnodelst mnt_nvnodelist; /* (i) list of vnodes */ int mnt_nvnodelistsize; /* (i) # of vnodes */ int mnt_writeopcount; /* (i) write syscalls pending */ - int mnt_kern_flag; /* (i) kernel only flags */ - uint64_t mnt_flag; /* (i) flags shared with user */ struct vfsoptlist *mnt_opt; /* current mount options */ struct vfsoptlist *mnt_optnew; /* new options passed to fs */ int mnt_maxsymlinklen; /* max size of short symlink */ struct statfs mnt_stat; /* cache of filesystem stats */ struct ucred *mnt_cred; /* credentials of mounter */ void * mnt_data; /* private data */ time_t mnt_time; /* last time written*/ int mnt_iosize_max; /* max size for clusters, etc */ struct netexport *mnt_export; /* export list */ struct label *mnt_label; /* MAC label for the fs */ u_int mnt_hashseed; /* Random seed for vfs_hash */ int mnt_lockref; /* (i) Lock reference count */ int mnt_secondary_writes; /* (i) # of secondary writes */ int mnt_secondary_accwrites;/* (i) secondary wr. starts */ struct thread *mnt_susp_owner; /* (i) thread owning suspension */ #define mnt_endzero mnt_gjprovider char *mnt_gjprovider; /* gjournal provider name */ struct mtx mnt_listmtx; struct vnodelst mnt_lazyvnodelist; /* (l) list of lazy vnodes */ int mnt_lazyvnodelistsize; /* (l) # of lazy vnodes */ struct lock mnt_explock; /* vfs_export walkers lock */ TAILQ_ENTRY(mount) mnt_upper_link; /* (m) we in the all uppers */ TAILQ_HEAD(, mount) mnt_uppers; /* (m) upper mounts over us*/ - int __aligned(CACHE_LINE_SIZE) mnt_vfs_ops;/* (i) pending vfs ops */ - int *mnt_thread_in_ops_pcpu; - int *mnt_ref_pcpu; - int *mnt_lockref_pcpu; - int *mnt_writeopcount_pcpu; - struct vnode *mnt_rootvnode; }; /* * Definitions for MNT_VNODE_FOREACH_ALL. */ struct vnode *__mnt_vnode_next_all(struct vnode **mvp, struct mount *mp); struct vnode *__mnt_vnode_first_all(struct vnode **mvp, struct mount *mp); void __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp); #define MNT_VNODE_FOREACH_ALL(vp, mp, mvp) \ for (vp = __mnt_vnode_first_all(&(mvp), (mp)); \ (vp) != NULL; vp = __mnt_vnode_next_all(&(mvp), (mp))) #define MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp) \ do { \ MNT_ILOCK(mp); \ __mnt_vnode_markerfree_all(&(mvp), (mp)); \ /* MNT_IUNLOCK(mp); -- done in above function */ \ mtx_assert(MNT_MTX(mp), MA_NOTOWNED); \ } while (0) /* * Definitions for MNT_VNODE_FOREACH_LAZY. */ typedef int mnt_lazy_cb_t(struct vnode *, void *); struct vnode *__mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg); struct vnode *__mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg); void __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp); #define MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, cb, cbarg) \ for (vp = __mnt_vnode_first_lazy(&(mvp), (mp), (cb), (cbarg)); \ (vp) != NULL; \ vp = __mnt_vnode_next_lazy(&(mvp), (mp), (cb), (cbarg))) #define MNT_VNODE_FOREACH_LAZY_ABORT(mp, mvp) \ __mnt_vnode_markerfree_lazy(&(mvp), (mp)) #define MNT_ILOCK(mp) mtx_lock(&(mp)->mnt_mtx) #define MNT_ITRYLOCK(mp) mtx_trylock(&(mp)->mnt_mtx) #define MNT_IUNLOCK(mp) mtx_unlock(&(mp)->mnt_mtx) #define MNT_MTX(mp) (&(mp)->mnt_mtx) #define MNT_REF(mp) do { \ mtx_assert(MNT_MTX(mp), MA_OWNED); \ mp->mnt_ref++; \ } while (0) #define MNT_REL(mp) do { \ mtx_assert(MNT_MTX(mp), MA_OWNED); \ (mp)->mnt_ref--; \ if ((mp)->mnt_vfs_ops && (mp)->mnt_ref < 0) \ vfs_dump_mount_counters(mp); \ if ((mp)->mnt_ref == 0 && (mp)->mnt_vfs_ops) \ wakeup((mp)); \ } while (0) #endif /* _KERNEL */ #if defined(_WANT_MNTOPTNAMES) || defined(_KERNEL) struct mntoptnames { uint64_t o_opt; const char *o_name; }; #define MNTOPT_NAMES \ { MNT_ASYNC, "asynchronous" }, \ { MNT_EXPORTED, "NFS exported" }, \ { MNT_LOCAL, "local" }, \ { MNT_NOATIME, "noatime" }, \ { MNT_NOEXEC, "noexec" }, \ { MNT_NOSUID, "nosuid" }, \ { MNT_NOSYMFOLLOW, "nosymfollow" }, \ { MNT_QUOTA, "with quotas" }, \ { MNT_RDONLY, "read-only" }, \ { MNT_SYNCHRONOUS, "synchronous" }, \ { MNT_UNION, "union" }, \ { MNT_NOCLUSTERR, "noclusterr" }, \ { MNT_NOCLUSTERW, "noclusterw" }, \ { MNT_SUIDDIR, "suiddir" }, \ { MNT_SOFTDEP, "soft-updates" }, \ { MNT_SUJ, "journaled soft-updates" }, \ { MNT_MULTILABEL, "multilabel" }, \ { MNT_ACLS, "acls" }, \ { MNT_NFS4ACLS, "nfsv4acls" }, \ { MNT_GJOURNAL, "gjournal" }, \ { MNT_AUTOMOUNTED, "automounted" }, \ { MNT_VERIFIED, "verified" }, \ { MNT_UNTRUSTED, "untrusted" }, \ { MNT_NOCOVER, "nocover" }, \ { MNT_EMPTYDIR, "emptydir" }, \ { MNT_UPDATE, "update" }, \ { MNT_DELEXPORT, "delexport" }, \ { MNT_RELOAD, "reload" }, \ { MNT_FORCE, "force" }, \ { MNT_SNAPSHOT, "snapshot" }, \ { 0, NULL } #endif /* * User specifiable flags, stored in mnt_flag. */ #define MNT_RDONLY 0x0000000000000001ULL /* read only filesystem */ #define MNT_SYNCHRONOUS 0x0000000000000002ULL /* fs written synchronously */ #define MNT_NOEXEC 0x0000000000000004ULL /* can't exec from filesystem */ #define MNT_NOSUID 0x0000000000000008ULL /* don't honor setuid fs bits */ #define MNT_NFS4ACLS 0x0000000000000010ULL /* enable NFS version 4 ACLs */ #define MNT_UNION 0x0000000000000020ULL /* union with underlying fs */ #define MNT_ASYNC 0x0000000000000040ULL /* fs written asynchronously */ #define MNT_SUIDDIR 0x0000000000100000ULL /* special SUID dir handling */ #define MNT_SOFTDEP 0x0000000000200000ULL /* using soft updates */ #define MNT_NOSYMFOLLOW 0x0000000000400000ULL /* do not follow symlinks */ #define MNT_GJOURNAL 0x0000000002000000ULL /* GEOM journal support enabled */ #define MNT_MULTILABEL 0x0000000004000000ULL /* MAC support for objects */ #define MNT_ACLS 0x0000000008000000ULL /* ACL support enabled */ #define MNT_NOATIME 0x0000000010000000ULL /* dont update file access time */ #define MNT_NOCLUSTERR 0x0000000040000000ULL /* disable cluster read */ #define MNT_NOCLUSTERW 0x0000000080000000ULL /* disable cluster write */ #define MNT_SUJ 0x0000000100000000ULL /* using journaled soft updates */ #define MNT_AUTOMOUNTED 0x0000000200000000ULL /* mounted by automountd(8) */ #define MNT_UNTRUSTED 0x0000000800000000ULL /* filesys metadata untrusted */ /* * NFS export related mount flags. */ #define MNT_EXRDONLY 0x0000000000000080ULL /* exported read only */ #define MNT_EXPORTED 0x0000000000000100ULL /* filesystem is exported */ #define MNT_DEFEXPORTED 0x0000000000000200ULL /* exported to the world */ #define MNT_EXPORTANON 0x0000000000000400ULL /* anon uid mapping for all */ #define MNT_EXKERB 0x0000000000000800ULL /* exported with Kerberos */ #define MNT_EXPUBLIC 0x0000000020000000ULL /* public export (WebNFS) */ #define MNT_EXTLS 0x0000004000000000ULL /* require TLS */ #define MNT_EXTLSCERT 0x0000008000000000ULL /* require TLS with client cert */ #define MNT_EXTLSCERTUSER 0x0000010000000000ULL /* require TLS with user cert */ /* * Flags set by internal operations, * but visible to the user. * XXX some of these are not quite right.. (I've never seen the root flag set) */ #define MNT_LOCAL 0x0000000000001000ULL /* filesystem is stored locally */ #define MNT_QUOTA 0x0000000000002000ULL /* quotas are enabled on fs */ #define MNT_ROOTFS 0x0000000000004000ULL /* identifies the root fs */ #define MNT_USER 0x0000000000008000ULL /* mounted by a user */ #define MNT_IGNORE 0x0000000000800000ULL /* do not show entry in df */ #define MNT_VERIFIED 0x0000000400000000ULL /* filesystem is verified */ /* * Mask of flags that are visible to statfs(). * XXX I think that this could now become (~(MNT_CMDFLAGS)) * but the 'mount' program may need changing to handle this. */ #define MNT_VISFLAGMASK (MNT_RDONLY | MNT_SYNCHRONOUS | MNT_NOEXEC | \ MNT_NOSUID | MNT_UNION | MNT_SUJ | \ MNT_ASYNC | MNT_EXRDONLY | MNT_EXPORTED | \ MNT_DEFEXPORTED | MNT_EXPORTANON| MNT_EXKERB | \ MNT_LOCAL | MNT_USER | MNT_QUOTA | \ MNT_ROOTFS | MNT_NOATIME | MNT_NOCLUSTERR| \ MNT_NOCLUSTERW | MNT_SUIDDIR | MNT_SOFTDEP | \ MNT_IGNORE | MNT_EXPUBLIC | MNT_NOSYMFOLLOW | \ MNT_GJOURNAL | MNT_MULTILABEL | MNT_ACLS | \ MNT_NFS4ACLS | MNT_AUTOMOUNTED | MNT_VERIFIED | \ MNT_UNTRUSTED) /* Mask of flags that can be updated. */ #define MNT_UPDATEMASK (MNT_NOSUID | MNT_NOEXEC | \ MNT_SYNCHRONOUS | MNT_UNION | MNT_ASYNC | \ MNT_NOATIME | \ MNT_NOSYMFOLLOW | MNT_IGNORE | \ MNT_NOCLUSTERR | MNT_NOCLUSTERW | MNT_SUIDDIR | \ MNT_ACLS | MNT_USER | MNT_NFS4ACLS | \ MNT_AUTOMOUNTED | MNT_UNTRUSTED) /* * External filesystem command modifier flags. * Unmount can use the MNT_FORCE flag. * XXX: These are not STATES and really should be somewhere else. * XXX: MNT_BYFSID and MNT_NONBUSY collide with MNT_ACLS and MNT_MULTILABEL, * but because MNT_ACLS and MNT_MULTILABEL are only used for mount(2), * and MNT_BYFSID and MNT_NONBUSY are only used for unmount(2), * it's harmless. */ #define MNT_UPDATE 0x0000000000010000ULL /* not real mount, just update */ #define MNT_DELEXPORT 0x0000000000020000ULL /* delete export host lists */ #define MNT_RELOAD 0x0000000000040000ULL /* reload filesystem data */ #define MNT_FORCE 0x0000000000080000ULL /* force unmount or readonly */ #define MNT_SNAPSHOT 0x0000000001000000ULL /* snapshot the filesystem */ #define MNT_NONBUSY 0x0000000004000000ULL /* check vnode use counts. */ #define MNT_BYFSID 0x0000000008000000ULL /* specify filesystem by ID. */ #define MNT_NOCOVER 0x0000001000000000ULL /* Do not cover a mount point */ #define MNT_EMPTYDIR 0x0000002000000000ULL /* Only mount on empty dir */ #define MNT_CMDFLAGS (MNT_UPDATE | MNT_DELEXPORT | MNT_RELOAD | \ MNT_FORCE | MNT_SNAPSHOT | MNT_NONBUSY | \ MNT_BYFSID | MNT_NOCOVER | MNT_EMPTYDIR) /* * Internal filesystem control flags stored in mnt_kern_flag. * * MNTK_UNMOUNT locks the mount entry so that name lookup cannot * proceed past the mount point. This keeps the subtree stable during * mounts and unmounts. When non-forced unmount flushes all vnodes * from the mp queue, the MNTK_UNMOUNT flag prevents insmntque() from * queueing new vnodes. * * MNTK_UNMOUNTF permits filesystems to detect a forced unmount while * dounmount() is still waiting to lock the mountpoint. This allows * the filesystem to cancel operations that might otherwise deadlock * with the unmount attempt (used by NFS). */ #define MNTK_UNMOUNTF 0x00000001 /* forced unmount in progress */ #define MNTK_ASYNC 0x00000002 /* filtered async flag */ #define MNTK_SOFTDEP 0x00000004 /* async disabled by softdep */ #define MNTK_NOMSYNC 0x00000008 /* don't do msync */ #define MNTK_DRAINING 0x00000010 /* lock draining is happening */ #define MNTK_REFEXPIRE 0x00000020 /* refcount expiring is happening */ #define MNTK_EXTENDED_SHARED 0x00000040 /* Allow shared locking for more ops */ #define MNTK_SHARED_WRITES 0x00000080 /* Allow shared locking for writes */ #define MNTK_NO_IOPF 0x00000100 /* Disallow page faults during reads and writes. Filesystem shall properly handle i/o state on EFAULT. */ #define MNTK_VGONE_UPPER 0x00000200 #define MNTK_VGONE_WAITER 0x00000400 #define MNTK_LOOKUP_EXCL_DOTDOT 0x00000800 #define MNTK_MARKER 0x00001000 #define MNTK_UNMAPPED_BUFS 0x00002000 #define MNTK_USES_BCACHE 0x00004000 /* FS uses the buffer cache. */ #define MNTK_TEXT_REFS 0x00008000 /* Keep use ref for text */ #define MNTK_VMSETSIZE_BUG 0x00010000 #define MNTK_UNIONFS 0x00020000 /* A hack for F_ISUNIONSTACK */ #define MNTK_FPLOOKUP 0x00040000 /* fast path lookup is supported */ #define MNTK_SUSPEND_ALL 0x00080000 /* Suspended by all-fs suspension */ #define MNTK_NOASYNC 0x00800000 /* disable async */ #define MNTK_UNMOUNT 0x01000000 /* unmount in progress */ #define MNTK_MWAIT 0x02000000 /* waiting for unmount to finish */ #define MNTK_SUSPEND 0x08000000 /* request write suspension */ #define MNTK_SUSPEND2 0x04000000 /* block secondary writes */ #define MNTK_SUSPENDED 0x10000000 /* write operations are suspended */ #define MNTK_NULL_NOCACHE 0x20000000 /* auto disable cache for nullfs mounts over this fs */ #define MNTK_LOOKUP_SHARED 0x40000000 /* FS supports shared lock lookups */ #define MNTK_NOKNOTE 0x80000000 /* Don't send KNOTEs from VOP hooks */ #ifdef _KERNEL static inline int MNT_SHARED_WRITES(struct mount *mp) { return (mp != NULL && (mp->mnt_kern_flag & MNTK_SHARED_WRITES) != 0); } static inline int MNT_EXTENDED_SHARED(struct mount *mp) { return (mp != NULL && (mp->mnt_kern_flag & MNTK_EXTENDED_SHARED) != 0); } #endif /* * Sysctl CTL_VFS definitions. * * Second level identifier specifies which filesystem. Second level * identifier VFS_VFSCONF returns information about all filesystems. * Second level identifier VFS_GENERIC is non-terminal. */ #define VFS_VFSCONF 0 /* get configured filesystems */ #define VFS_GENERIC 0 /* generic filesystem information */ /* * Third level identifiers for VFS_GENERIC are given below; third * level identifiers for specific filesystems are given in their * mount specific header files. */ #define VFS_MAXTYPENUM 1 /* int: highest defined filesystem type */ #define VFS_CONF 2 /* struct: vfsconf for filesystem given as next argument */ /* * Flags for various system call interfaces. * * waitfor flags to vfs_sync() and getfsstat() */ #define MNT_WAIT 1 /* synchronously wait for I/O to complete */ #define MNT_NOWAIT 2 /* start all I/O, but do not wait for it */ #define MNT_LAZY 3 /* push data not written by filesystem syncer */ #define MNT_SUSPEND 4 /* Suspend file system after sync */ /* * Generic file handle */ struct fhandle { fsid_t fh_fsid; /* Filesystem id of mount point */ struct fid fh_fid; /* Filesys specific id */ }; typedef struct fhandle fhandle_t; /* * Old export arguments without security flavor list */ struct oexport_args { int ex_flags; /* export related flags */ uid_t ex_root; /* mapping for root uid */ struct xucred ex_anon; /* mapping for anonymous user */ struct sockaddr *ex_addr; /* net address to which exported */ u_char ex_addrlen; /* and the net address length */ struct sockaddr *ex_mask; /* mask of valid bits in saddr */ u_char ex_masklen; /* and the smask length */ char *ex_indexfile; /* index file for WebNFS URLs */ }; /* * Not quite so old export arguments with 32bit ex_flags and xucred ex_anon. */ #define MAXSECFLAVORS 5 struct o2export_args { int ex_flags; /* export related flags */ uid_t ex_root; /* mapping for root uid */ struct xucred ex_anon; /* mapping for anonymous user */ struct sockaddr *ex_addr; /* net address to which exported */ u_char ex_addrlen; /* and the net address length */ struct sockaddr *ex_mask; /* mask of valid bits in saddr */ u_char ex_masklen; /* and the smask length */ char *ex_indexfile; /* index file for WebNFS URLs */ int ex_numsecflavors; /* security flavor count */ int ex_secflavors[MAXSECFLAVORS]; /* list of security flavors */ }; /* * Export arguments for local filesystem mount calls. */ struct export_args { uint64_t ex_flags; /* export related flags */ uid_t ex_root; /* mapping for root uid */ uid_t ex_uid; /* mapping for anonymous user */ int ex_ngroups; gid_t *ex_groups; struct sockaddr *ex_addr; /* net address to which exported */ u_char ex_addrlen; /* and the net address length */ struct sockaddr *ex_mask; /* mask of valid bits in saddr */ u_char ex_masklen; /* and the smask length */ char *ex_indexfile; /* index file for WebNFS URLs */ int ex_numsecflavors; /* security flavor count */ int ex_secflavors[MAXSECFLAVORS]; /* list of security flavors */ }; /* * Structure holding information for a publicly exported filesystem * (WebNFS). Currently the specs allow just for one such filesystem. */ struct nfs_public { int np_valid; /* Do we hold valid information */ fhandle_t np_handle; /* Filehandle for pub fs (internal) */ struct mount *np_mount; /* Mountpoint of exported fs */ char *np_index; /* Index file */ }; /* * Filesystem configuration information. One of these exists for each * type of filesystem supported by the kernel. These are searched at * mount time to identify the requested filesystem. * * XXX: Never change the first two arguments! */ struct vfsconf { u_int vfc_version; /* ABI version number */ char vfc_name[MFSNAMELEN]; /* filesystem type name */ struct vfsops *vfc_vfsops; /* filesystem operations vector */ struct vfsops *vfc_vfsops_sd; /* ... signal-deferred */ int vfc_typenum; /* historic filesystem type number */ int vfc_refcount; /* number mounted of this type */ int vfc_flags; /* permanent flags */ int vfc_prison_flag; /* prison allow.mount.* flag */ struct vfsoptdecl *vfc_opts; /* mount options */ TAILQ_ENTRY(vfsconf) vfc_list; /* list of vfscons */ }; /* Userland version of the struct vfsconf. */ struct xvfsconf { struct vfsops *vfc_vfsops; /* filesystem operations vector */ char vfc_name[MFSNAMELEN]; /* filesystem type name */ int vfc_typenum; /* historic filesystem type number */ int vfc_refcount; /* number mounted of this type */ int vfc_flags; /* permanent flags */ struct vfsconf *vfc_next; /* next in list */ }; #ifndef BURN_BRIDGES struct ovfsconf { void *vfc_vfsops; char vfc_name[32]; int vfc_index; int vfc_refcount; int vfc_flags; }; #endif /* * NB: these flags refer to IMPLEMENTATION properties, not properties of * any actual mounts; i.e., it does not make sense to change the flags. */ #define VFCF_STATIC 0x00010000 /* statically compiled into kernel */ #define VFCF_NETWORK 0x00020000 /* may get data over the network */ #define VFCF_READONLY 0x00040000 /* writes are not implemented */ #define VFCF_SYNTHETIC 0x00080000 /* data does not represent real files */ #define VFCF_LOOPBACK 0x00100000 /* aliases some other mounted FS */ #define VFCF_UNICODE 0x00200000 /* stores file names as Unicode */ #define VFCF_JAIL 0x00400000 /* can be mounted from within a jail */ #define VFCF_DELEGADMIN 0x00800000 /* supports delegated administration */ #define VFCF_SBDRY 0x01000000 /* Stop at Boundary: defer stop requests to kernel->user (AST) transition */ typedef uint32_t fsctlop_t; struct vfsidctl { int vc_vers; /* should be VFSIDCTL_VERS1 (below) */ fsid_t vc_fsid; /* fsid to operate on */ char vc_fstypename[MFSNAMELEN]; /* type of fs 'nfs' or '*' */ fsctlop_t vc_op; /* operation VFS_CTL_* (below) */ void *vc_ptr; /* pointer to data structure */ size_t vc_len; /* sizeof said structure */ u_int32_t vc_spare[12]; /* spare (must be zero) */ }; /* vfsidctl API version. */ #define VFS_CTL_VERS1 0x01 /* * New style VFS sysctls, do not reuse/conflict with the namespace for * private sysctls. * All "global" sysctl ops have the 33rd bit set: * 0x...1.... * Private sysctl ops should have the 33rd bit unset. */ #define VFS_CTL_QUERY 0x00010001 /* anything wrong? (vfsquery) */ #define VFS_CTL_TIMEO 0x00010002 /* set timeout for vfs notification */ #define VFS_CTL_NOLOCKS 0x00010003 /* disable file locking */ struct vfsquery { u_int32_t vq_flags; u_int32_t vq_spare[31]; }; /* vfsquery flags */ #define VQ_NOTRESP 0x0001 /* server down */ #define VQ_NEEDAUTH 0x0002 /* server bad auth */ #define VQ_LOWDISK 0x0004 /* we're low on space */ #define VQ_MOUNT 0x0008 /* new filesystem arrived */ #define VQ_UNMOUNT 0x0010 /* filesystem has left */ #define VQ_DEAD 0x0020 /* filesystem is dead, needs force unmount */ #define VQ_ASSIST 0x0040 /* filesystem needs assistance from external program */ #define VQ_NOTRESPLOCK 0x0080 /* server lockd down */ #define VQ_FLAG0100 0x0100 /* placeholder */ #define VQ_FLAG0200 0x0200 /* placeholder */ #define VQ_FLAG0400 0x0400 /* placeholder */ #define VQ_FLAG0800 0x0800 /* placeholder */ #define VQ_FLAG1000 0x1000 /* placeholder */ #define VQ_FLAG2000 0x2000 /* placeholder */ #define VQ_FLAG4000 0x4000 /* placeholder */ #define VQ_FLAG8000 0x8000 /* placeholder */ #ifdef _KERNEL /* Point a sysctl request at a vfsidctl's data. */ #define VCTLTOREQ(vc, req) \ do { \ (req)->newptr = (vc)->vc_ptr; \ (req)->newlen = (vc)->vc_len; \ (req)->newidx = 0; \ } while (0) #endif struct iovec; struct uio; #ifdef _KERNEL /* * vfs_busy specific flags and mask. */ #define MBF_NOWAIT 0x01 #define MBF_MNTLSTLOCK 0x02 #define MBF_MASK (MBF_NOWAIT | MBF_MNTLSTLOCK) #ifdef MALLOC_DECLARE MALLOC_DECLARE(M_MOUNT); MALLOC_DECLARE(M_STATFS); #endif extern int maxvfsconf; /* highest defined filesystem type */ TAILQ_HEAD(vfsconfhead, vfsconf); extern struct vfsconfhead vfsconf; /* * Operations supported on mounted filesystem. */ struct mount_args; struct nameidata; struct sysctl_req; struct mntarg; /* * N.B., vfs_cmount is the ancient vfsop invoked by the old mount(2) syscall. * The new way is vfs_mount. * * vfs_cmount implementations typically translate arguments from their * respective old per-FS structures into the key-value list supported by * nmount(2), then use kernel_mount(9) to mimic nmount(2) from kernelspace. * * Filesystems with mounters that use nmount(2) do not need to and should not * implement vfs_cmount. Hopefully a future cleanup can remove vfs_cmount and * mount(2) entirely. */ typedef int vfs_cmount_t(struct mntarg *ma, void *data, uint64_t flags); typedef int vfs_unmount_t(struct mount *mp, int mntflags); typedef int vfs_root_t(struct mount *mp, int flags, struct vnode **vpp); typedef int vfs_quotactl_t(struct mount *mp, int cmds, uid_t uid, void *arg); typedef int vfs_statfs_t(struct mount *mp, struct statfs *sbp); typedef int vfs_sync_t(struct mount *mp, int waitfor); typedef int vfs_vget_t(struct mount *mp, ino_t ino, int flags, struct vnode **vpp); typedef int vfs_fhtovp_t(struct mount *mp, struct fid *fhp, int flags, struct vnode **vpp); typedef int vfs_checkexp_t(struct mount *mp, struct sockaddr *nam, uint64_t *extflagsp, struct ucred **credanonp, int *numsecflavors, int *secflavors); typedef int vfs_init_t(struct vfsconf *); typedef int vfs_uninit_t(struct vfsconf *); typedef int vfs_extattrctl_t(struct mount *mp, int cmd, struct vnode *filename_vp, int attrnamespace, const char *attrname); typedef int vfs_mount_t(struct mount *mp); typedef int vfs_sysctl_t(struct mount *mp, fsctlop_t op, struct sysctl_req *req); typedef void vfs_susp_clean_t(struct mount *mp); typedef void vfs_notify_lowervp_t(struct mount *mp, struct vnode *lowervp); typedef void vfs_purge_t(struct mount *mp); struct vfsops { vfs_mount_t *vfs_mount; vfs_cmount_t *vfs_cmount; vfs_unmount_t *vfs_unmount; vfs_root_t *vfs_root; vfs_root_t *vfs_cachedroot; vfs_quotactl_t *vfs_quotactl; vfs_statfs_t *vfs_statfs; vfs_sync_t *vfs_sync; vfs_vget_t *vfs_vget; vfs_fhtovp_t *vfs_fhtovp; vfs_checkexp_t *vfs_checkexp; vfs_init_t *vfs_init; vfs_uninit_t *vfs_uninit; vfs_extattrctl_t *vfs_extattrctl; vfs_sysctl_t *vfs_sysctl; vfs_susp_clean_t *vfs_susp_clean; vfs_notify_lowervp_t *vfs_reclaim_lowervp; vfs_notify_lowervp_t *vfs_unlink_lowervp; vfs_purge_t *vfs_purge; vfs_mount_t *vfs_spare[6]; /* spares for ABI compat */ }; vfs_statfs_t __vfs_statfs; #define VFS_MOUNT(MP) ({ \ int _rc; \ \ TSRAW(curthread, TS_ENTER, "VFS_MOUNT", (MP)->mnt_vfc->vfc_name);\ _rc = (*(MP)->mnt_op->vfs_mount)(MP); \ TSRAW(curthread, TS_EXIT, "VFS_MOUNT", (MP)->mnt_vfc->vfc_name);\ _rc; }) #define VFS_UNMOUNT(MP, FORCE) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_unmount)(MP, FORCE); \ _rc; }) #define VFS_ROOT(MP, FLAGS, VPP) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_root)(MP, FLAGS, VPP); \ _rc; }) #define VFS_CACHEDROOT(MP, FLAGS, VPP) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_cachedroot)(MP, FLAGS, VPP); \ _rc; }) #define VFS_QUOTACTL(MP, C, U, A) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_quotactl)(MP, C, U, A); \ _rc; }) #define VFS_STATFS(MP, SBP) ({ \ int _rc; \ \ _rc = __vfs_statfs((MP), (SBP)); \ _rc; }) #define VFS_SYNC(MP, WAIT) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_sync)(MP, WAIT); \ _rc; }) #define VFS_VGET(MP, INO, FLAGS, VPP) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_vget)(MP, INO, FLAGS, VPP); \ _rc; }) #define VFS_FHTOVP(MP, FIDP, FLAGS, VPP) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_fhtovp)(MP, FIDP, FLAGS, VPP); \ _rc; }) #define VFS_CHECKEXP(MP, NAM, EXFLG, CRED, NUMSEC, SEC) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_checkexp)(MP, NAM, EXFLG, CRED, NUMSEC,\ SEC); \ _rc; }) #define VFS_EXTATTRCTL(MP, C, FN, NS, N) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_extattrctl)(MP, C, FN, NS, N); \ _rc; }) #define VFS_SYSCTL(MP, OP, REQ) ({ \ int _rc; \ \ _rc = (*(MP)->mnt_op->vfs_sysctl)(MP, OP, REQ); \ _rc; }) #define VFS_SUSP_CLEAN(MP) do { \ if (*(MP)->mnt_op->vfs_susp_clean != NULL) { \ (*(MP)->mnt_op->vfs_susp_clean)(MP); \ } \ } while (0) #define VFS_RECLAIM_LOWERVP(MP, VP) do { \ if (*(MP)->mnt_op->vfs_reclaim_lowervp != NULL) { \ (*(MP)->mnt_op->vfs_reclaim_lowervp)((MP), (VP)); \ } \ } while (0) #define VFS_UNLINK_LOWERVP(MP, VP) do { \ if (*(MP)->mnt_op->vfs_unlink_lowervp != NULL) { \ (*(MP)->mnt_op->vfs_unlink_lowervp)((MP), (VP)); \ } \ } while (0) #define VFS_PURGE(MP) do { \ if (*(MP)->mnt_op->vfs_purge != NULL) { \ (*(MP)->mnt_op->vfs_purge)(MP); \ } \ } while (0) #define VFS_KNOTE_LOCKED(vp, hint) do \ { \ if (((vp)->v_vflag & VV_NOKNOTE) == 0) \ VN_KNOTE((vp), (hint), KNF_LISTLOCKED); \ } while (0) #define VFS_KNOTE_UNLOCKED(vp, hint) do \ { \ if (((vp)->v_vflag & VV_NOKNOTE) == 0) \ VN_KNOTE((vp), (hint), 0); \ } while (0) #define VFS_NOTIFY_UPPER_RECLAIM 1 #define VFS_NOTIFY_UPPER_UNLINK 2 #include /* * Version numbers. */ #define VFS_VERSION_00 0x19660120 #define VFS_VERSION_01 0x20121030 #define VFS_VERSION_02 0x20180504 #define VFS_VERSION VFS_VERSION_02 #define VFS_SET(vfsops, fsname, flags) \ static struct vfsconf fsname ## _vfsconf = { \ .vfc_version = VFS_VERSION, \ .vfc_name = #fsname, \ .vfc_vfsops = &vfsops, \ .vfc_typenum = -1, \ .vfc_flags = flags, \ }; \ static moduledata_t fsname ## _mod = { \ #fsname, \ vfs_modevent, \ & fsname ## _vfsconf \ }; \ DECLARE_MODULE(fsname, fsname ## _mod, SI_SUB_VFS, SI_ORDER_MIDDLE) /* * exported vnode operations */ int dounmount(struct mount *, int, struct thread *); int kernel_mount(struct mntarg *ma, uint64_t flags); int kernel_vmount(int flags, ...); struct mntarg *mount_arg(struct mntarg *ma, const char *name, const void *val, int len); struct mntarg *mount_argb(struct mntarg *ma, int flag, const char *name); struct mntarg *mount_argf(struct mntarg *ma, const char *name, const char *fmt, ...); struct mntarg *mount_argsu(struct mntarg *ma, const char *name, const void *val, int len); void statfs_scale_blocks(struct statfs *sf, long max_size); struct vfsconf *vfs_byname(const char *); struct vfsconf *vfs_byname_kld(const char *, struct thread *td, int *); void vfs_mount_destroy(struct mount *); void vfs_event_signal(fsid_t *, u_int32_t, intptr_t); void vfs_freeopts(struct vfsoptlist *opts); void vfs_deleteopt(struct vfsoptlist *opts, const char *name); int vfs_buildopts(struct uio *auio, struct vfsoptlist **options); int vfs_flagopt(struct vfsoptlist *opts, const char *name, uint64_t *w, uint64_t val); int vfs_getopt(struct vfsoptlist *, const char *, void **, int *); int vfs_getopt_pos(struct vfsoptlist *opts, const char *name); int vfs_getopt_size(struct vfsoptlist *opts, const char *name, off_t *value); char *vfs_getopts(struct vfsoptlist *, const char *, int *error); int vfs_copyopt(struct vfsoptlist *, const char *, void *, int); int vfs_filteropt(struct vfsoptlist *, const char **legal); void vfs_opterror(struct vfsoptlist *opts, const char *fmt, ...); int vfs_scanopt(struct vfsoptlist *opts, const char *name, const char *fmt, ...); int vfs_setopt(struct vfsoptlist *opts, const char *name, void *value, int len); int vfs_setopt_part(struct vfsoptlist *opts, const char *name, void *value, int len); int vfs_setopts(struct vfsoptlist *opts, const char *name, const char *value); int vfs_setpublicfs /* set publicly exported fs */ (struct mount *, struct netexport *, struct export_args *); void vfs_periodic(struct mount *, int); int vfs_busy(struct mount *, int); int vfs_export /* process mount export info */ (struct mount *, struct export_args *); void vfs_allocate_syncvnode(struct mount *); void vfs_deallocate_syncvnode(struct mount *); int vfs_donmount(struct thread *td, uint64_t fsflags, struct uio *fsoptions); void vfs_getnewfsid(struct mount *); struct cdev *vfs_getrootfsid(struct mount *); struct mount *vfs_getvfs(fsid_t *); /* return vfs given fsid */ struct mount *vfs_busyfs(fsid_t *); int vfs_modevent(module_t, int, void *); void vfs_mount_error(struct mount *, const char *, ...); void vfs_mountroot(void); /* mount our root filesystem */ void vfs_mountedfrom(struct mount *, const char *from); void vfs_notify_upper(struct vnode *, int); void vfs_ref(struct mount *); void vfs_rel(struct mount *); struct mount *vfs_mount_alloc(struct vnode *, struct vfsconf *, const char *, struct ucred *); int vfs_suser(struct mount *, struct thread *); void vfs_unbusy(struct mount *); void vfs_unmountall(void); extern TAILQ_HEAD(mntlist, mount) mountlist; /* mounted filesystem list */ extern struct mtx_padalign mountlist_mtx; extern struct nfs_public nfs_pub; extern struct sx vfsconf_sx; #define vfsconf_lock() sx_xlock(&vfsconf_sx) #define vfsconf_unlock() sx_xunlock(&vfsconf_sx) #define vfsconf_slock() sx_slock(&vfsconf_sx) #define vfsconf_sunlock() sx_sunlock(&vfsconf_sx) struct vnode *mntfs_allocvp(struct mount *, struct vnode *); void mntfs_freevp(struct vnode *); /* * Declarations for these vfs default operations are located in * kern/vfs_default.c. They will be automatically used to replace * null entries in VFS ops tables when registering a new filesystem * type in the global table. */ vfs_root_t vfs_stdroot; vfs_quotactl_t vfs_stdquotactl; vfs_statfs_t vfs_stdstatfs; vfs_sync_t vfs_stdsync; vfs_sync_t vfs_stdnosync; vfs_vget_t vfs_stdvget; vfs_fhtovp_t vfs_stdfhtovp; vfs_checkexp_t vfs_stdcheckexp; vfs_init_t vfs_stdinit; vfs_uninit_t vfs_stduninit; vfs_extattrctl_t vfs_stdextattrctl; vfs_sysctl_t vfs_stdsysctl; void syncer_suspend(void); void syncer_resume(void); struct vnode *vfs_cache_root_clear(struct mount *); void vfs_cache_root_set(struct mount *, struct vnode *); void vfs_op_barrier_wait(struct mount *); void vfs_op_enter(struct mount *); void vfs_op_exit_locked(struct mount *); void vfs_op_exit(struct mount *); #ifdef DIAGNOSTIC void vfs_assert_mount_counters(struct mount *); void vfs_dump_mount_counters(struct mount *); #else #define vfs_assert_mount_counters(mp) do { } while (0) #define vfs_dump_mount_counters(mp) do { } while (0) #endif enum mount_counter { MNT_COUNT_REF, MNT_COUNT_LOCKREF, MNT_COUNT_WRITEOPCOUNT }; int vfs_mount_fetch_counter(struct mount *, enum mount_counter); void suspend_all_fs(void); void resume_all_fs(void); /* * Code transitioning mnt_vfs_ops to > 0 issues IPIs until it observes - * all CPUs not executing code enclosed by mnt_thread_in_ops_pcpu. + * all CPUs not executing code enclosed by thread_in_ops_pcpu variable. * * This provides an invariant that by the time the last CPU is observed not * executing, everyone else entering will see the counter > 0 and exit. * * Note there is no barrier between vfs_ops and the rest of the code in the * section. It is not necessary as the writer has to wait for everyone to drain * before making any changes or only make changes safe while the section is * executed. */ +#define vfs_mount_pcpu(mp) zpcpu_get(mp->mnt_pcpu) +#define vfs_mount_pcpu_remote(mp, cpu) zpcpu_get_cpu(mp->mnt_pcpu, cpu) + #define vfs_op_thread_entered(mp) ({ \ MPASS(curthread->td_critnest > 0); \ - *zpcpu_get(mp->mnt_thread_in_ops_pcpu) == 1; \ + struct mount_pcpu *_mpcpu = vfs_mount_pcpu(mp); \ + _mpcpu->mntp_thread_in_ops == 1; \ }) -#define vfs_op_thread_enter_crit(mp) ({ \ +#define vfs_op_thread_enter_crit(mp, _mpcpu) ({ \ bool _retval_crit = true; \ MPASS(curthread->td_critnest > 0); \ - MPASS(!vfs_op_thread_entered(mp)); \ - zpcpu_set_protected(mp->mnt_thread_in_ops_pcpu, 1); \ + _mpcpu = vfs_mount_pcpu(mp); \ + MPASS(mpcpu->mntp_thread_in_ops == 0); \ + _mpcpu->mntp_thread_in_ops = 1; \ __compiler_membar(); \ if (__predict_false(mp->mnt_vfs_ops > 0)) { \ - vfs_op_thread_exit_crit(mp); \ + vfs_op_thread_exit_crit(mp, _mpcpu); \ _retval_crit = false; \ } \ _retval_crit; \ }) -#define vfs_op_thread_enter(mp) ({ \ +#define vfs_op_thread_enter(mp, _mpcpu) ({ \ bool _retval; \ critical_enter(); \ - _retval = vfs_op_thread_enter_crit(mp); \ + _retval = vfs_op_thread_enter_crit(mp, _mpcpu); \ if (__predict_false(!_retval)) \ critical_exit(); \ _retval; \ }) -#define vfs_op_thread_exit_crit(mp) do { \ - MPASS(vfs_op_thread_entered(mp)); \ +#define vfs_op_thread_exit_crit(mp, _mpcpu) do { \ + MPASS(_mpcpu == vfs_mount_pcpu(mp)); \ + MPASS(_mpcpu->mntp_thread_in_ops == 1); \ __compiler_membar(); \ - zpcpu_set_protected(mp->mnt_thread_in_ops_pcpu, 0); \ + _mpcpu->mntp_thread_in_ops = 0; \ } while (0) -#define vfs_op_thread_exit(mp) do { \ - vfs_op_thread_exit_crit(mp); \ +#define vfs_op_thread_exit(mp, _mpcpu) do { \ + vfs_op_thread_exit_crit(mp, _mpcpu); \ critical_exit(); \ } while (0) -#define vfs_mp_count_add_pcpu(mp, count, val) do { \ - MPASS(vfs_op_thread_entered(mp)); \ - zpcpu_add_protected(mp->mnt_##count##_pcpu, val); \ +#define vfs_mp_count_add_pcpu(_mpcpu, count, val) do { \ + MPASS(_mpcpu->mntp_thread_in_ops == 1); \ + _mpcpu->mntp_##count += val; \ } while (0) -#define vfs_mp_count_sub_pcpu(mp, count, val) do { \ - MPASS(vfs_op_thread_entered(mp)); \ - zpcpu_sub_protected(mp->mnt_##count##_pcpu, val); \ +#define vfs_mp_count_sub_pcpu(_mpcpu, count, val) do { \ + MPASS(_mpcpu->mntp_thread_in_ops == 1); \ + _mpcpu->mntp_##count -= val; \ } while (0) #else /* !_KERNEL */ #include struct stat; __BEGIN_DECLS int fhlink(struct fhandle *, const char *); int fhlinkat(struct fhandle *, int, const char *); int fhopen(const struct fhandle *, int); int fhreadlink(struct fhandle *, char *, size_t); int fhstat(const struct fhandle *, struct stat *); int fhstatfs(const struct fhandle *, struct statfs *); int fstatfs(int, struct statfs *); int getfh(const char *, fhandle_t *); int getfhat(int, char *, struct fhandle *, int); int getfsstat(struct statfs *, long, int); int getmntinfo(struct statfs **, int); int lgetfh(const char *, fhandle_t *); int mount(const char *, const char *, int, void *); int nmount(struct iovec *, unsigned int, int); int statfs(const char *, struct statfs *); int unmount(const char *, int); /* C library stuff */ int getvfsbyname(const char *, struct xvfsconf *); __END_DECLS #endif /* _KERNEL */ #endif /* !_SYS_MOUNT_H_ */