diff --git a/sys/kern/kern_rangelock.c b/sys/kern/kern_rangelock.c index a9d8c342c267..186a42caebf0 100644 --- a/sys/kern/kern_rangelock.c +++ b/sys/kern/kern_rangelock.c @@ -1,330 +1,361 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2009 Konstantin Belousov * 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 unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 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 +#include #include #include #include #include #include -#include +#include +#include #include +/* + * Implementation of range locks based on the paper + * https://doi.org/10.1145/3342195.3387533 + * arXiv:2006.12144v1 [cs.OS] 22 Jun 2020 + * Scalable Range Locks for Scalable Address Spaces and Beyond + * by Alex Kogan, Dave Dice, and Shady Issa + */ + +static struct rl_q_entry *rl_e_unmark(const struct rl_q_entry *e); + +/* + * rl_q_next links all granted ranges in the lock. We cannot free an + * rl_q_entry while in the smr section, and cannot reuse rl_q_next + * linkage since other threads might follow it even after CAS removed + * the range. Use rl_q_free for local list of ranges to remove after + * the smr section is dropped. + */ struct rl_q_entry { - TAILQ_ENTRY(rl_q_entry) rl_q_link; + struct rl_q_entry *rl_q_next; + struct rl_q_entry *rl_q_free; off_t rl_q_start, rl_q_end; int rl_q_flags; +#ifdef INVARIANTS + struct thread *rl_q_owner; +#endif }; static uma_zone_t rl_entry_zone; +static smr_t rl_smr; static void rangelock_sys_init(void) { - rl_entry_zone = uma_zcreate("rl_entry", sizeof(struct rl_q_entry), - NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); + NULL, NULL, NULL, NULL, UMA_ALIGNOF(struct rl_q_entry), + UMA_ZONE_SMR); + rl_smr = uma_zone_get_smr(rl_entry_zone); } -SYSINIT(vfs, SI_SUB_LOCK, SI_ORDER_ANY, rangelock_sys_init, NULL); +SYSINIT(rl, SI_SUB_LOCK, SI_ORDER_ANY, rangelock_sys_init, NULL); static struct rl_q_entry * -rlqentry_alloc(void) +rlqentry_alloc(vm_ooffset_t start, vm_ooffset_t end, int flags) { + struct rl_q_entry *e; + + e = uma_zalloc_smr(rl_entry_zone, M_WAITOK); + e->rl_q_next = NULL; + e->rl_q_free = NULL; + e->rl_q_start = start; + e->rl_q_end = end; + e->rl_q_flags = flags; +#ifdef INVARIANTS + e->rl_q_owner = curthread; +#endif + return (e); +} - return (uma_zalloc(rl_entry_zone, M_WAITOK)); +void +rangelock_init(struct rangelock *lock) +{ + lock->sleepers = false; + atomic_store_ptr(&lock->head, NULL); } void -rlqentry_free(struct rl_q_entry *rleq) +rangelock_destroy(struct rangelock *lock) { + struct rl_q_entry *e, *ep; - uma_zfree(rl_entry_zone, rleq); + MPASS(!lock->sleepers); + for (e = (struct rl_q_entry *)atomic_load_ptr(&lock->head); + e != NULL; e = rl_e_unmark(ep)) { + ep = atomic_load_ptr(&e->rl_q_next); + uma_zfree_smr(rl_entry_zone, e); + } } -void -rangelock_init(struct rangelock *lock) +static bool +rl_e_is_marked(const struct rl_q_entry *e) { + return (((uintptr_t)e & 1) != 0); +} - TAILQ_INIT(&lock->rl_waiters); - lock->rl_currdep = NULL; +static struct rl_q_entry * +rl_e_unmark(const struct rl_q_entry *e) +{ + MPASS(rl_e_is_marked(e)); + return ((struct rl_q_entry *)((uintptr_t)e & ~1)); } -void -rangelock_destroy(struct rangelock *lock) +static struct rl_q_entry * +rl_q_load(struct rl_q_entry **p) { + return ((struct rl_q_entry *)atomic_load_acq_ptr((uintptr_t *)p)); +} - KASSERT(TAILQ_EMPTY(&lock->rl_waiters), ("Dangling waiters")); +void +rangelock_unlock(struct rangelock *lock, void *cookie) +{ + struct rl_q_entry *e; + + e = cookie; + MPASS(lock != NULL && e != NULL); + MPASS(!rl_e_is_marked(rl_q_load(&e->rl_q_next))); + MPASS(e->rl_q_owner == curthread); + + sleepq_lock(&lock->sleepers); +#ifdef INVARIANTS + int r = atomic_testandset_long((uintptr_t *)&e->rl_q_next, 0); + MPASS(r == 0); +#else + atomic_set_ptr((uintptr_t *)&e->rl_q_next, 1); +#endif + lock->sleepers = false; + sleepq_broadcast(&lock->sleepers, SLEEPQ_SLEEP, 0, 0); + sleepq_release(&lock->sleepers); } /* - * Two entries are compatible if their ranges do not overlap, or both - * entries are for read. + * result: -1 if e1 before e2 + * 1 if e1 after e2 + * 0 if e1 and e2 overlap */ static int -ranges_overlap(const struct rl_q_entry *e1, - const struct rl_q_entry *e2) +rl_e_compare(const struct rl_q_entry *e1, const struct rl_q_entry *e2) { - - if (e1->rl_q_start < e2->rl_q_end && e1->rl_q_end > e2->rl_q_start) + if (e1 == NULL) return (1); + if (e1->rl_q_start >= e2->rl_q_end) + return (1); + if (e2->rl_q_start >= e1->rl_q_end) + return (-1); return (0); } -/* - * Recalculate the lock->rl_currdep after an unlock. - */ static void -rangelock_calc_block(struct rangelock *lock) +rl_insert_sleep(struct rangelock *lock) { - struct rl_q_entry *entry, *nextentry, *entry1; - - for (entry = lock->rl_currdep; entry != NULL; entry = nextentry) { - nextentry = TAILQ_NEXT(entry, rl_q_link); - if (entry->rl_q_flags & RL_LOCK_READ) { - /* Reads must not overlap with granted writes. */ - for (entry1 = TAILQ_FIRST(&lock->rl_waiters); - !(entry1->rl_q_flags & RL_LOCK_READ); - entry1 = TAILQ_NEXT(entry1, rl_q_link)) { - if (ranges_overlap(entry, entry1)) - goto out; - } - } else { - /* Write must not overlap with any granted locks. */ - for (entry1 = TAILQ_FIRST(&lock->rl_waiters); - entry1 != entry; - entry1 = TAILQ_NEXT(entry1, rl_q_link)) { - if (ranges_overlap(entry, entry1)) - goto out; - } - - /* Move grantable write locks to the front. */ - TAILQ_REMOVE(&lock->rl_waiters, entry, rl_q_link); - TAILQ_INSERT_HEAD(&lock->rl_waiters, entry, rl_q_link); - } - - /* Grant this lock. */ - entry->rl_q_flags |= RL_LOCK_GRANTED; - wakeup(entry); - } -out: - lock->rl_currdep = entry; + smr_exit(rl_smr); + DROP_GIANT(); + lock->sleepers = true; + sleepq_add(&lock->sleepers, NULL, "rangelk", 0, 0); + sleepq_wait(&lock->sleepers, PRI_USER); + PICKUP_GIANT(); + smr_enter(rl_smr); } -static void -rangelock_unlock_locked(struct rangelock *lock, struct rl_q_entry *entry, - struct mtx *ilk, bool do_calc_block) +static bool +rl_q_cas(struct rl_q_entry **prev, struct rl_q_entry *old, + struct rl_q_entry *new) { - - MPASS(lock != NULL && entry != NULL && ilk != NULL); - mtx_assert(ilk, MA_OWNED); - - if (!do_calc_block) { - /* - * This is the case where rangelock_enqueue() has been called - * with trylock == true and just inserted this entry in the - * queue. - * If rl_currdep is this entry, rl_currdep needs to - * be set to the next entry in the rl_waiters list. - * However, since this entry is the last entry in the - * list, the next entry is NULL. - */ - if (lock->rl_currdep == entry) { - KASSERT(TAILQ_NEXT(lock->rl_currdep, rl_q_link) == NULL, - ("rangelock_enqueue: next entry not NULL")); - lock->rl_currdep = NULL; - } - } else - KASSERT(entry != lock->rl_currdep, ("stuck currdep")); - - TAILQ_REMOVE(&lock->rl_waiters, entry, rl_q_link); - if (do_calc_block) - rangelock_calc_block(lock); - mtx_unlock(ilk); - if (curthread->td_rlqe == NULL) - curthread->td_rlqe = entry; - else - rlqentry_free(entry); + return (atomic_cmpset_rel_ptr((uintptr_t *)prev, (uintptr_t)old, + (uintptr_t)new) != 0); } -void -rangelock_unlock(struct rangelock *lock, void *cookie, struct mtx *ilk) +static bool +rl_insert(struct rangelock *lock, struct rl_q_entry *e, bool trylock, + struct rl_q_entry **free) { + struct rl_q_entry *cur, *next, **prev; + int r; + +again: + prev = &lock->head; + if (rl_q_load(prev) == NULL && rl_q_cas(prev, NULL, e)) + return (true); + + for (cur = rl_q_load(prev);;) { + if (rl_e_is_marked(cur)) + goto again; + + if (cur != NULL) { + next = rl_q_load(&cur->rl_q_next); + if (rl_e_is_marked(next)) { + next = rl_e_unmark(next); + if (rl_q_cas(prev, cur, next)) { +#ifdef INVARIANTS + cur->rl_q_owner = NULL; +#endif + cur->rl_q_free = *free; + *free = cur; + } + cur = next; + continue; + } + } - MPASS(lock != NULL && cookie != NULL && ilk != NULL); - - mtx_lock(ilk); - rangelock_unlock_locked(lock, cookie, ilk, true); -} - -/* - * Unlock the sub-range of granted lock. - */ -void * -rangelock_unlock_range(struct rangelock *lock, void *cookie, off_t start, - off_t end, struct mtx *ilk) -{ - struct rl_q_entry *entry; - - MPASS(lock != NULL && cookie != NULL && ilk != NULL); - entry = cookie; - KASSERT(entry->rl_q_flags & RL_LOCK_GRANTED, - ("Unlocking non-granted lock")); - KASSERT(entry->rl_q_start == start, ("wrong start")); - KASSERT(entry->rl_q_end >= end, ("wrong end")); - - mtx_lock(ilk); - if (entry->rl_q_end == end) { - rangelock_unlock_locked(lock, cookie, ilk, true); - return (NULL); + r = rl_e_compare(cur, e); + if (r == -1) { + prev = &cur->rl_q_next; + cur = rl_q_load(prev); + } else if (r == 0) { + sleepq_lock(&lock->sleepers); + if (__predict_false(rl_e_is_marked(rl_q_load( + &cur->rl_q_next)))) { + sleepq_release(&lock->sleepers); + continue; + } + if (trylock) { + sleepq_release(&lock->sleepers); + return (false); + } + rl_insert_sleep(lock); + /* e is still valid */ + goto again; + } else /* r == 1 */ { + e->rl_q_next = cur; + if (rl_q_cas(prev, cur, e)) { + atomic_thread_fence_acq(); + return (true); + } + /* Reset rl_q_next in case we hit fast path. */ + e->rl_q_next = NULL; + cur = rl_q_load(prev); + } } - entry->rl_q_end = end; - rangelock_calc_block(lock); - mtx_unlock(ilk); - return (cookie); } -/* - * Add the lock request to the queue of the pending requests for - * rangelock. Sleep until the request can be granted unless trylock == true. - */ -static void * -rangelock_enqueue(struct rangelock *lock, off_t start, off_t end, int mode, - struct mtx *ilk, bool trylock) +static struct rl_q_entry * +rangelock_lock_int(struct rangelock *lock, struct rl_q_entry *e, + bool trylock) { - struct rl_q_entry *entry; - struct thread *td; - - MPASS(lock != NULL && ilk != NULL); - - td = curthread; - if (td->td_rlqe != NULL) { - entry = td->td_rlqe; - td->td_rlqe = NULL; - } else - entry = rlqentry_alloc(); - MPASS(entry != NULL); - entry->rl_q_flags = mode; - entry->rl_q_start = start; - entry->rl_q_end = end; - - mtx_lock(ilk); - /* - * XXXKIB TODO. Check that a thread does not try to enqueue a - * lock that is incompatible with another request from the same - * thread. - */ - - TAILQ_INSERT_TAIL(&lock->rl_waiters, entry, rl_q_link); - /* - * If rl_currdep == NULL, there is no entry waiting for a conflicting - * range to be resolved, so set rl_currdep to this entry. If there is - * no conflicting entry for this entry, rl_currdep will be set back to - * NULL by rangelock_calc_block(). - */ - if (lock->rl_currdep == NULL) - lock->rl_currdep = entry; - rangelock_calc_block(lock); - while (!(entry->rl_q_flags & RL_LOCK_GRANTED)) { - if (trylock) { - /* - * For this case, the range is not actually locked - * yet, but removal from the list requires the same - * steps, except for not doing a rangelock_calc_block() - * call, since rangelock_calc_block() was called above. - */ - rangelock_unlock_locked(lock, entry, ilk, false); - return (NULL); - } - msleep(entry, ilk, 0, "range", 0); + struct rl_q_entry *free, *x, *xp; + bool res; + + free = NULL; + smr_enter(rl_smr); + res = rl_insert(lock, e, trylock, &free); + smr_exit(rl_smr); + MPASS(trylock || res); + if (!res) { + e->rl_q_free = free; + free = e; + e = NULL; + } + for (x = free; x != NULL; x = xp) { + MPASS(!rl_e_is_marked(x)); + xp = x->rl_q_free; + MPASS(!rl_e_is_marked(xp)); + uma_zfree_smr(rl_entry_zone, x); } - mtx_unlock(ilk); - return (entry); + return (e); } void * -rangelock_rlock(struct rangelock *lock, off_t start, off_t end, struct mtx *ilk) +rangelock_rlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) { + struct rl_q_entry *e; - return (rangelock_enqueue(lock, start, end, RL_LOCK_READ, ilk, false)); + e = rlqentry_alloc(start, end, RL_LOCK_READ); + return (rangelock_lock_int(lock, e, false)); } void * -rangelock_tryrlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk) +rangelock_tryrlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) { + struct rl_q_entry *e; - return (rangelock_enqueue(lock, start, end, RL_LOCK_READ, ilk, true)); + e = rlqentry_alloc(start, end, RL_LOCK_READ); + return (rangelock_lock_int(lock, e, true)); } void * -rangelock_wlock(struct rangelock *lock, off_t start, off_t end, struct mtx *ilk) +rangelock_wlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) { + struct rl_q_entry *e; - return (rangelock_enqueue(lock, start, end, RL_LOCK_WRITE, ilk, false)); + e = rlqentry_alloc(start, end, RL_LOCK_WRITE); + return (rangelock_lock_int(lock, e, true)); } void * -rangelock_trywlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk) +rangelock_trywlock(struct rangelock *lock, vm_ooffset_t start, vm_ooffset_t end) { + struct rl_q_entry *e; - return (rangelock_enqueue(lock, start, end, RL_LOCK_WRITE, ilk, true)); + e = rlqentry_alloc(start, end, RL_LOCK_WRITE); + return (rangelock_lock_int(lock, e, true)); } #ifdef INVARIANT_SUPPORT void _rangelock_cookie_assert(void *cookie, int what, const char *file, int line) { - struct rl_q_entry *entry; - int flags; - - MPASS(cookie != NULL); - entry = cookie; - flags = entry->rl_q_flags; - switch (what) { - case RCA_LOCKED: - if ((flags & RL_LOCK_GRANTED) == 0) - panic("rangelock not held @ %s:%d\n", file, line); - break; - case RCA_RLOCKED: - if ((flags & (RL_LOCK_GRANTED | RL_LOCK_READ)) != - (RL_LOCK_GRANTED | RL_LOCK_READ)) - panic("rangelock not rlocked @ %s:%d\n", file, line); - break; - case RCA_WLOCKED: - if ((flags & (RL_LOCK_GRANTED | RL_LOCK_WRITE)) != - (RL_LOCK_GRANTED | RL_LOCK_WRITE)) - panic("rangelock not wlocked @ %s:%d\n", file, line); - break; - default: - panic("Unknown rangelock assertion: %d @ %s:%d", what, file, - line); - } } #endif /* INVARIANT_SUPPORT */ + +#include "opt_ddb.h" +#ifdef DDB +#include + +DB_SHOW_COMMAND(rangelock, db_show_rangelock) +{ + struct rangelock *lock; + struct rl_q_entry *e, *x; + + if (!have_addr) { + db_printf("show rangelock addr\n"); + return; + } + + lock = (struct rangelock *)addr; + db_printf("rangelock %p sleepers %d\n", lock, lock->sleepers); + for (e = lock->head;;) { + x = rl_e_is_marked(e) ? rl_e_unmark(e) : e; + if (x == NULL) + break; + db_printf(" entry %p marked %d %d start %#jx end %#jx " + "flags %x next %p", + e, rl_e_is_marked(e), rl_e_is_marked(x->rl_q_next), + x->rl_q_start, x->rl_q_end, x->rl_q_flags, x->rl_q_next); +#ifdef INVARIANTS + db_printf(" owner %p (%d)", x->rl_q_owner, + x->rl_q_owner != NULL ? x->rl_q_owner->td_tid : -1); +#endif + db_printf("\n"); + e = x->rl_q_next; + } +} + +#endif /* DDB */ diff --git a/sys/kern/kern_thread.c b/sys/kern/kern_thread.c index 00f99516773c..c951e7297c89 100644 --- a/sys/kern/kern_thread.c +++ b/sys/kern/kern_thread.c @@ -1,1799 +1,1798 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (C) 2001 Julian Elischer . * 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(s), this list of conditions and the following disclaimer as * the first lines of this file unmodified other than the possible * addition of one or more copyright notices. * 2. Redistributions in binary form must reproduce the above copyright * notice(s), this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``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 COPYRIGHT HOLDER(S) 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 "opt_witness.h" #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 #ifdef HWPMC_HOOKS #include #endif #include #include #include #include #include #include #include #include /* * Asserts below verify the stability of struct thread and struct proc * layout, as exposed by KBI to modules. On head, the KBI is allowed * to drift, change to the structures must be accompanied by the * assert update. * * On the stable branches after KBI freeze, conditions must not be * violated. Typically new fields are moved to the end of the * structures. */ #ifdef __amd64__ _Static_assert(offsetof(struct thread, td_flags) == 0x108, "struct thread KBI td_flags"); _Static_assert(offsetof(struct thread, td_pflags) == 0x114, "struct thread KBI td_pflags"); _Static_assert(offsetof(struct thread, td_frame) == 0x4b8, "struct thread KBI td_frame"); _Static_assert(offsetof(struct thread, td_emuldata) == 0x6c0, "struct thread KBI td_emuldata"); _Static_assert(offsetof(struct proc, p_flag) == 0xb8, "struct proc KBI p_flag"); _Static_assert(offsetof(struct proc, p_pid) == 0xc4, "struct proc KBI p_pid"); _Static_assert(offsetof(struct proc, p_filemon) == 0x3c8, "struct proc KBI p_filemon"); _Static_assert(offsetof(struct proc, p_comm) == 0x3e0, "struct proc KBI p_comm"); _Static_assert(offsetof(struct proc, p_emuldata) == 0x4d0, "struct proc KBI p_emuldata"); #endif #ifdef __i386__ _Static_assert(offsetof(struct thread, td_flags) == 0x9c, "struct thread KBI td_flags"); _Static_assert(offsetof(struct thread, td_pflags) == 0xa8, "struct thread KBI td_pflags"); _Static_assert(offsetof(struct thread, td_frame) == 0x318, "struct thread KBI td_frame"); _Static_assert(offsetof(struct thread, td_emuldata) == 0x35c, "struct thread KBI td_emuldata"); _Static_assert(offsetof(struct proc, p_flag) == 0x6c, "struct proc KBI p_flag"); _Static_assert(offsetof(struct proc, p_pid) == 0x78, "struct proc KBI p_pid"); _Static_assert(offsetof(struct proc, p_filemon) == 0x270, "struct proc KBI p_filemon"); _Static_assert(offsetof(struct proc, p_comm) == 0x284, "struct proc KBI p_comm"); _Static_assert(offsetof(struct proc, p_emuldata) == 0x318, "struct proc KBI p_emuldata"); #endif SDT_PROVIDER_DECLARE(proc); SDT_PROBE_DEFINE(proc, , , lwp__exit); /* * thread related storage. */ static uma_zone_t thread_zone; struct thread_domain_data { struct thread *tdd_zombies; int tdd_reapticks; } __aligned(CACHE_LINE_SIZE); static struct thread_domain_data thread_domain_data[MAXMEMDOM]; static struct task thread_reap_task; static struct callout thread_reap_callout; static void thread_zombie(struct thread *); static void thread_reap(void); static void thread_reap_all(void); static void thread_reap_task_cb(void *, int); static void thread_reap_callout_cb(void *); static void thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary); static void thread_free_batched(struct thread *td); static __exclusive_cache_line struct mtx tid_lock; static bitstr_t *tid_bitmap; static MALLOC_DEFINE(M_TIDHASH, "tidhash", "thread hash"); static int maxthread; SYSCTL_INT(_kern, OID_AUTO, maxthread, CTLFLAG_RDTUN, &maxthread, 0, "Maximum number of threads"); static __exclusive_cache_line int nthreads; static LIST_HEAD(tidhashhead, thread) *tidhashtbl; static u_long tidhash; static u_long tidhashlock; static struct rwlock *tidhashtbl_lock; #define TIDHASH(tid) (&tidhashtbl[(tid) & tidhash]) #define TIDHASHLOCK(tid) (&tidhashtbl_lock[(tid) & tidhashlock]) EVENTHANDLER_LIST_DEFINE(thread_ctor); EVENTHANDLER_LIST_DEFINE(thread_dtor); EVENTHANDLER_LIST_DEFINE(thread_init); EVENTHANDLER_LIST_DEFINE(thread_fini); static bool thread_count_inc_try(void) { int nthreads_new; nthreads_new = atomic_fetchadd_int(&nthreads, 1) + 1; if (nthreads_new >= maxthread - 100) { if (priv_check_cred(curthread->td_ucred, PRIV_MAXPROC) != 0 || nthreads_new >= maxthread) { atomic_subtract_int(&nthreads, 1); return (false); } } return (true); } static bool thread_count_inc(void) { static struct timeval lastfail; static int curfail; thread_reap(); if (thread_count_inc_try()) { return (true); } thread_reap_all(); if (thread_count_inc_try()) { return (true); } if (ppsratecheck(&lastfail, &curfail, 1)) { printf("maxthread limit exceeded by uid %u " "(pid %d); consider increasing kern.maxthread\n", curthread->td_ucred->cr_ruid, curproc->p_pid); } return (false); } static void thread_count_sub(int n) { atomic_subtract_int(&nthreads, n); } static void thread_count_dec(void) { thread_count_sub(1); } static lwpid_t tid_alloc(void) { static lwpid_t trytid; lwpid_t tid; mtx_lock(&tid_lock); /* * It is an invariant that the bitmap is big enough to hold maxthread * IDs. If we got to this point there has to be at least one free. */ if (trytid >= maxthread) trytid = 0; bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); if (tid == -1) { KASSERT(trytid != 0, ("unexpectedly ran out of IDs")); trytid = 0; bit_ffc_at(tid_bitmap, trytid, maxthread, &tid); KASSERT(tid != -1, ("unexpectedly ran out of IDs")); } bit_set(tid_bitmap, tid); trytid = tid + 1; mtx_unlock(&tid_lock); return (tid + NO_PID); } static void tid_free_locked(lwpid_t rtid) { lwpid_t tid; mtx_assert(&tid_lock, MA_OWNED); KASSERT(rtid >= NO_PID, ("%s: invalid tid %d\n", __func__, rtid)); tid = rtid - NO_PID; KASSERT(bit_test(tid_bitmap, tid) != 0, ("thread ID %d not allocated\n", rtid)); bit_clear(tid_bitmap, tid); } static void tid_free(lwpid_t rtid) { mtx_lock(&tid_lock); tid_free_locked(rtid); mtx_unlock(&tid_lock); } static void tid_free_batch(lwpid_t *batch, int n) { int i; mtx_lock(&tid_lock); for (i = 0; i < n; i++) { tid_free_locked(batch[i]); } mtx_unlock(&tid_lock); } /* * Batching for thread reapping. */ struct tidbatch { lwpid_t tab[16]; int n; }; static void tidbatch_prep(struct tidbatch *tb) { tb->n = 0; } static void tidbatch_add(struct tidbatch *tb, struct thread *td) { KASSERT(tb->n < nitems(tb->tab), ("%s: count too high %d", __func__, tb->n)); tb->tab[tb->n] = td->td_tid; tb->n++; } static void tidbatch_process(struct tidbatch *tb) { KASSERT(tb->n <= nitems(tb->tab), ("%s: count too high %d", __func__, tb->n)); if (tb->n == nitems(tb->tab)) { tid_free_batch(tb->tab, tb->n); tb->n = 0; } } static void tidbatch_final(struct tidbatch *tb) { KASSERT(tb->n <= nitems(tb->tab), ("%s: count too high %d", __func__, tb->n)); if (tb->n != 0) { tid_free_batch(tb->tab, tb->n); } } /* * Batching thread count free, for consistency */ struct tdcountbatch { int n; }; static void tdcountbatch_prep(struct tdcountbatch *tb) { tb->n = 0; } static void tdcountbatch_add(struct tdcountbatch *tb, struct thread *td __unused) { tb->n++; } static void tdcountbatch_process(struct tdcountbatch *tb) { if (tb->n == 32) { thread_count_sub(tb->n); tb->n = 0; } } static void tdcountbatch_final(struct tdcountbatch *tb) { if (tb->n != 0) { thread_count_sub(tb->n); } } /* * Prepare a thread for use. */ static int thread_ctor(void *mem, int size, void *arg, int flags) { struct thread *td; td = (struct thread *)mem; TD_SET_STATE(td, TDS_INACTIVE); td->td_lastcpu = td->td_oncpu = NOCPU; /* * Note that td_critnest begins life as 1 because the thread is not * running and is thereby implicitly waiting to be on the receiving * end of a context switch. */ td->td_critnest = 1; td->td_lend_user_pri = PRI_MAX; #ifdef AUDIT audit_thread_alloc(td); #endif #ifdef KDTRACE_HOOKS kdtrace_thread_ctor(td); #endif umtx_thread_alloc(td); MPASS(td->td_sel == NULL); return (0); } /* * Reclaim a thread after use. */ static void thread_dtor(void *mem, int size, void *arg) { struct thread *td; td = (struct thread *)mem; #ifdef INVARIANTS /* Verify that this thread is in a safe state to free. */ switch (TD_GET_STATE(td)) { case TDS_INHIBITED: case TDS_RUNNING: case TDS_CAN_RUN: case TDS_RUNQ: /* * We must never unlink a thread that is in one of * these states, because it is currently active. */ panic("bad state for thread unlinking"); /* NOTREACHED */ case TDS_INACTIVE: break; default: panic("bad thread state"); /* NOTREACHED */ } #endif #ifdef AUDIT audit_thread_free(td); #endif #ifdef KDTRACE_HOOKS kdtrace_thread_dtor(td); #endif /* Free all OSD associated to this thread. */ osd_thread_exit(td); ast_kclear(td); seltdfini(td); } /* * Initialize type-stable parts of a thread (when newly created). */ static int thread_init(void *mem, int size, int flags) { struct thread *td; td = (struct thread *)mem; td->td_allocdomain = vm_phys_domain(vtophys(td)); td->td_sleepqueue = sleepq_alloc(); td->td_turnstile = turnstile_alloc(); td->td_rlqe = NULL; EVENTHANDLER_DIRECT_INVOKE(thread_init, td); umtx_thread_init(td); td->td_kstack = 0; td->td_sel = NULL; return (0); } /* * Tear down type-stable parts of a thread (just before being discarded). */ static void thread_fini(void *mem, int size) { struct thread *td; td = (struct thread *)mem; EVENTHANDLER_DIRECT_INVOKE(thread_fini, td); - rlqentry_free(td->td_rlqe); turnstile_free(td->td_turnstile); sleepq_free(td->td_sleepqueue); umtx_thread_fini(td); MPASS(td->td_sel == NULL); } /* * For a newly created process, * link up all the structures and its initial threads etc. * called from: * {arch}/{arch}/machdep.c {arch}_init(), init386() etc. * proc_dtor() (should go away) * proc_init() */ void proc_linkup0(struct proc *p, struct thread *td) { TAILQ_INIT(&p->p_threads); /* all threads in proc */ proc_linkup(p, td); } void proc_linkup(struct proc *p, struct thread *td) { sigqueue_init(&p->p_sigqueue, p); p->p_ksi = ksiginfo_alloc(M_WAITOK); if (p->p_ksi != NULL) { /* XXX p_ksi may be null if ksiginfo zone is not ready */ p->p_ksi->ksi_flags = KSI_EXT | KSI_INS; } LIST_INIT(&p->p_mqnotifier); p->p_numthreads = 0; thread_link(td, p); } static void ast_suspend(struct thread *td, int tda __unused) { struct proc *p; p = td->td_proc; /* * We need to check to see if we have to exit or wait due to a * single threading requirement or some other STOP condition. */ PROC_LOCK(p); thread_suspend_check(0); PROC_UNLOCK(p); } extern int max_threads_per_proc; /* * Initialize global thread allocation resources. */ void threadinit(void) { u_long i; lwpid_t tid0; /* * Place an upper limit on threads which can be allocated. * * Note that other factors may make the de facto limit much lower. * * Platform limits are somewhat arbitrary but deemed "more than good * enough" for the foreseable future. */ if (maxthread == 0) { #ifdef _LP64 maxthread = MIN(maxproc * max_threads_per_proc, 1000000); #else maxthread = MIN(maxproc * max_threads_per_proc, 100000); #endif } mtx_init(&tid_lock, "TID lock", NULL, MTX_DEF); tid_bitmap = bit_alloc(maxthread, M_TIDHASH, M_WAITOK); /* * Handle thread0. */ thread_count_inc(); tid0 = tid_alloc(); if (tid0 != THREAD0_TID) panic("tid0 %d != %d\n", tid0, THREAD0_TID); /* * Thread structures are specially aligned so that (at least) the * 5 lower bits of a pointer to 'struct thead' must be 0. These bits * are used by synchronization primitives to store flags in pointers to * such structures. */ thread_zone = uma_zcreate("THREAD", sched_sizeof_thread(), thread_ctor, thread_dtor, thread_init, thread_fini, UMA_ALIGN_CACHE_AND_MASK(32 - 1), UMA_ZONE_NOFREE); tidhashtbl = hashinit(maxproc / 2, M_TIDHASH, &tidhash); tidhashlock = (tidhash + 1) / 64; if (tidhashlock > 0) tidhashlock--; tidhashtbl_lock = malloc(sizeof(*tidhashtbl_lock) * (tidhashlock + 1), M_TIDHASH, M_WAITOK | M_ZERO); for (i = 0; i < tidhashlock + 1; i++) rw_init(&tidhashtbl_lock[i], "tidhash"); TASK_INIT(&thread_reap_task, 0, thread_reap_task_cb, NULL); callout_init(&thread_reap_callout, 1); callout_reset(&thread_reap_callout, 5 * hz, thread_reap_callout_cb, NULL); ast_register(TDA_SUSPEND, ASTR_ASTF_REQUIRED, 0, ast_suspend); } /* * Place an unused thread on the zombie list. */ void thread_zombie(struct thread *td) { struct thread_domain_data *tdd; struct thread *ztd; tdd = &thread_domain_data[td->td_allocdomain]; ztd = atomic_load_ptr(&tdd->tdd_zombies); for (;;) { td->td_zombie = ztd; if (atomic_fcmpset_rel_ptr((uintptr_t *)&tdd->tdd_zombies, (uintptr_t *)&ztd, (uintptr_t)td)) break; continue; } } /* * Release a thread that has exited after cpu_throw(). */ void thread_stash(struct thread *td) { atomic_subtract_rel_int(&td->td_proc->p_exitthreads, 1); thread_zombie(td); } /* * Reap zombies from passed domain. */ static void thread_reap_domain(struct thread_domain_data *tdd) { struct thread *itd, *ntd; struct tidbatch tidbatch; struct credbatch credbatch; struct limbatch limbatch; struct tdcountbatch tdcountbatch; /* * Reading upfront is pessimal if followed by concurrent atomic_swap, * but most of the time the list is empty. */ if (tdd->tdd_zombies == NULL) return; itd = (struct thread *)atomic_swap_ptr((uintptr_t *)&tdd->tdd_zombies, (uintptr_t)NULL); if (itd == NULL) return; /* * Multiple CPUs can get here, the race is fine as ticks is only * advisory. */ tdd->tdd_reapticks = ticks; tidbatch_prep(&tidbatch); credbatch_prep(&credbatch); limbatch_prep(&limbatch); tdcountbatch_prep(&tdcountbatch); while (itd != NULL) { ntd = itd->td_zombie; EVENTHANDLER_DIRECT_INVOKE(thread_dtor, itd); tidbatch_add(&tidbatch, itd); credbatch_add(&credbatch, itd); limbatch_add(&limbatch, itd); tdcountbatch_add(&tdcountbatch, itd); thread_free_batched(itd); tidbatch_process(&tidbatch); credbatch_process(&credbatch); limbatch_process(&limbatch); tdcountbatch_process(&tdcountbatch); itd = ntd; } tidbatch_final(&tidbatch); credbatch_final(&credbatch); limbatch_final(&limbatch); tdcountbatch_final(&tdcountbatch); } /* * Reap zombies from all domains. */ static void thread_reap_all(void) { struct thread_domain_data *tdd; int i, domain; domain = PCPU_GET(domain); for (i = 0; i < vm_ndomains; i++) { tdd = &thread_domain_data[(i + domain) % vm_ndomains]; thread_reap_domain(tdd); } } /* * Reap zombies from local domain. */ static void thread_reap(void) { struct thread_domain_data *tdd; int domain; domain = PCPU_GET(domain); tdd = &thread_domain_data[domain]; thread_reap_domain(tdd); } static void thread_reap_task_cb(void *arg __unused, int pending __unused) { thread_reap_all(); } static void thread_reap_callout_cb(void *arg __unused) { struct thread_domain_data *tdd; int i, cticks, lticks; bool wantreap; wantreap = false; cticks = atomic_load_int(&ticks); for (i = 0; i < vm_ndomains; i++) { tdd = &thread_domain_data[i]; lticks = tdd->tdd_reapticks; if (tdd->tdd_zombies != NULL && (u_int)(cticks - lticks) > 5 * hz) { wantreap = true; break; } } if (wantreap) taskqueue_enqueue(taskqueue_thread, &thread_reap_task); callout_reset(&thread_reap_callout, 5 * hz, thread_reap_callout_cb, NULL); } /* * Calling this function guarantees that any thread that exited before * the call is reaped when the function returns. By 'exited' we mean * a thread removed from the process linkage with thread_unlink(). * Practically this means that caller must lock/unlock corresponding * process lock before the call, to synchronize with thread_exit(). */ void thread_reap_barrier(void) { struct task *t; /* * First do context switches to each CPU to ensure that all * PCPU pc_deadthreads are moved to zombie list. */ quiesce_all_cpus("", PDROP); /* * Second, fire the task in the same thread as normal * thread_reap() is done, to serialize reaping. */ t = malloc(sizeof(*t), M_TEMP, M_WAITOK); TASK_INIT(t, 0, thread_reap_task_cb, t); taskqueue_enqueue(taskqueue_thread, t); taskqueue_drain(taskqueue_thread, t); free(t, M_TEMP); } /* * Allocate a thread. */ struct thread * thread_alloc(int pages) { struct thread *td; lwpid_t tid; if (!thread_count_inc()) { return (NULL); } tid = tid_alloc(); td = uma_zalloc(thread_zone, M_WAITOK); KASSERT(td->td_kstack == 0, ("thread_alloc got thread with kstack")); if (!vm_thread_new(td, pages)) { uma_zfree(thread_zone, td); tid_free(tid); thread_count_dec(); return (NULL); } td->td_tid = tid; bzero(&td->td_sa.args, sizeof(td->td_sa.args)); kasan_thread_alloc(td); kmsan_thread_alloc(td); cpu_thread_alloc(td); EVENTHANDLER_DIRECT_INVOKE(thread_ctor, td); return (td); } int thread_recycle(struct thread *td, int pages) { if (td->td_kstack == 0 || td->td_kstack_pages != pages) { if (td->td_kstack != 0) vm_thread_dispose(td); if (!vm_thread_new(td, pages)) return (ENOMEM); cpu_thread_alloc(td); } kasan_thread_alloc(td); kmsan_thread_alloc(td); return (0); } /* * Deallocate a thread. */ static void thread_free_batched(struct thread *td) { lock_profile_thread_exit(td); if (td->td_cpuset) cpuset_rel(td->td_cpuset); td->td_cpuset = NULL; cpu_thread_free(td); if (td->td_kstack != 0) vm_thread_dispose(td); callout_drain(&td->td_slpcallout); /* * Freeing handled by the caller. */ td->td_tid = -1; kmsan_thread_free(td); uma_zfree(thread_zone, td); } void thread_free(struct thread *td) { lwpid_t tid; EVENTHANDLER_DIRECT_INVOKE(thread_dtor, td); tid = td->td_tid; thread_free_batched(td); tid_free(tid); thread_count_dec(); } void thread_cow_get_proc(struct thread *newtd, struct proc *p) { PROC_LOCK_ASSERT(p, MA_OWNED); newtd->td_realucred = crcowget(p->p_ucred); newtd->td_ucred = newtd->td_realucred; newtd->td_limit = lim_hold(p->p_limit); newtd->td_cowgen = p->p_cowgen; } void thread_cow_get(struct thread *newtd, struct thread *td) { MPASS(td->td_realucred == td->td_ucred); newtd->td_realucred = crcowget(td->td_realucred); newtd->td_ucred = newtd->td_realucred; newtd->td_limit = lim_hold(td->td_limit); newtd->td_cowgen = td->td_cowgen; } void thread_cow_free(struct thread *td) { if (td->td_realucred != NULL) crcowfree(td); if (td->td_limit != NULL) lim_free(td->td_limit); } void thread_cow_update(struct thread *td) { struct proc *p; struct ucred *oldcred; struct plimit *oldlimit; p = td->td_proc; PROC_LOCK(p); oldcred = crcowsync(); oldlimit = lim_cowsync(); td->td_cowgen = p->p_cowgen; PROC_UNLOCK(p); if (oldcred != NULL) crfree(oldcred); if (oldlimit != NULL) lim_free(oldlimit); } void thread_cow_synced(struct thread *td) { struct proc *p; p = td->td_proc; PROC_LOCK_ASSERT(p, MA_OWNED); MPASS(td->td_cowgen != p->p_cowgen); MPASS(td->td_ucred == p->p_ucred); MPASS(td->td_limit == p->p_limit); td->td_cowgen = p->p_cowgen; } /* * Discard the current thread and exit from its context. * Always called with scheduler locked. * * Because we can't free a thread while we're operating under its context, * push the current thread into our CPU's deadthread holder. This means * we needn't worry about someone else grabbing our context before we * do a cpu_throw(). */ void thread_exit(void) { uint64_t runtime, new_switchtime; struct thread *td; struct thread *td2; struct proc *p; td = curthread; p = td->td_proc; PROC_SLOCK_ASSERT(p, MA_OWNED); mtx_assert(&Giant, MA_NOTOWNED); PROC_LOCK_ASSERT(p, MA_OWNED); KASSERT(p != NULL, ("thread exiting without a process")); CTR3(KTR_PROC, "thread_exit: thread %p (pid %ld, %s)", td, (long)p->p_pid, td->td_name); SDT_PROBE0(proc, , , lwp__exit); KASSERT(TAILQ_EMPTY(&td->td_sigqueue.sq_list), ("signal pending")); MPASS(td->td_realucred == td->td_ucred); /* * drop FPU & debug register state storage, or any other * architecture specific resources that * would not be on a new untouched process. */ cpu_thread_exit(td); /* * The last thread is left attached to the process * So that the whole bundle gets recycled. Skip * all this stuff if we never had threads. * EXIT clears all sign of other threads when * it goes to single threading, so the last thread always * takes the short path. */ if (p->p_flag & P_HADTHREADS) { if (p->p_numthreads > 1) { atomic_add_int(&td->td_proc->p_exitthreads, 1); thread_unlink(td); td2 = FIRST_THREAD_IN_PROC(p); sched_exit_thread(td2, td); /* * The test below is NOT true if we are the * sole exiting thread. P_STOPPED_SINGLE is unset * in exit1() after it is the only survivor. */ if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { if (p->p_numthreads == p->p_suspcount) { thread_lock(p->p_singlethread); thread_unsuspend_one(p->p_singlethread, p, false); } } PCPU_SET(deadthread, td); } else { /* * The last thread is exiting.. but not through exit() */ panic ("thread_exit: Last thread exiting on its own"); } } #ifdef HWPMC_HOOKS /* * If this thread is part of a process that is being tracked by hwpmc(4), * inform the module of the thread's impending exit. */ if (PMC_PROC_IS_USING_PMCS(td->td_proc)) { PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT, NULL); } else if (PMC_SYSTEM_SAMPLING_ACTIVE()) PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_THR_EXIT_LOG, NULL); #endif PROC_UNLOCK(p); PROC_STATLOCK(p); thread_lock(td); PROC_SUNLOCK(p); /* Do the same timestamp bookkeeping that mi_switch() would do. */ new_switchtime = cpu_ticks(); runtime = new_switchtime - PCPU_GET(switchtime); td->td_runtime += runtime; td->td_incruntime += runtime; PCPU_SET(switchtime, new_switchtime); PCPU_SET(switchticks, ticks); VM_CNT_INC(v_swtch); /* Save our resource usage in our process. */ td->td_ru.ru_nvcsw++; ruxagg_locked(p, td); rucollect(&p->p_ru, &td->td_ru); PROC_STATUNLOCK(p); TD_SET_STATE(td, TDS_INACTIVE); #ifdef WITNESS witness_thread_exit(td); #endif CTR1(KTR_PROC, "thread_exit: cpu_throw() thread %p", td); sched_throw(td); panic("I'm a teapot!"); /* NOTREACHED */ } /* * Do any thread specific cleanups that may be needed in wait() * called with Giant, proc and schedlock not held. */ void thread_wait(struct proc *p) { struct thread *td; mtx_assert(&Giant, MA_NOTOWNED); KASSERT(p->p_numthreads == 1, ("multiple threads in thread_wait()")); KASSERT(p->p_exitthreads == 0, ("p_exitthreads leaking")); td = FIRST_THREAD_IN_PROC(p); /* Lock the last thread so we spin until it exits cpu_throw(). */ thread_lock(td); thread_unlock(td); lock_profile_thread_exit(td); cpuset_rel(td->td_cpuset); td->td_cpuset = NULL; cpu_thread_clean(td); thread_cow_free(td); callout_drain(&td->td_slpcallout); thread_reap(); /* check for zombie threads etc. */ } /* * Link a thread to a process. * set up anything that needs to be initialized for it to * be used by the process. */ void thread_link(struct thread *td, struct proc *p) { /* * XXX This can't be enabled because it's called for proc0 before * its lock has been created. * PROC_LOCK_ASSERT(p, MA_OWNED); */ TD_SET_STATE(td, TDS_INACTIVE); td->td_proc = p; td->td_flags = TDF_INMEM; LIST_INIT(&td->td_contested); LIST_INIT(&td->td_lprof[0]); LIST_INIT(&td->td_lprof[1]); #ifdef EPOCH_TRACE SLIST_INIT(&td->td_epochs); #endif sigqueue_init(&td->td_sigqueue, p); callout_init(&td->td_slpcallout, 1); TAILQ_INSERT_TAIL(&p->p_threads, td, td_plist); p->p_numthreads++; } /* * Called from: * thread_exit() */ void thread_unlink(struct thread *td) { struct proc *p = td->td_proc; PROC_LOCK_ASSERT(p, MA_OWNED); #ifdef EPOCH_TRACE MPASS(SLIST_EMPTY(&td->td_epochs)); #endif TAILQ_REMOVE(&p->p_threads, td, td_plist); p->p_numthreads--; /* could clear a few other things here */ /* Must NOT clear links to proc! */ } static int calc_remaining(struct proc *p, int mode) { int remaining; PROC_LOCK_ASSERT(p, MA_OWNED); PROC_SLOCK_ASSERT(p, MA_OWNED); if (mode == SINGLE_EXIT) remaining = p->p_numthreads; else if (mode == SINGLE_BOUNDARY) remaining = p->p_numthreads - p->p_boundary_count; else if (mode == SINGLE_NO_EXIT || mode == SINGLE_ALLPROC) remaining = p->p_numthreads - p->p_suspcount; else panic("calc_remaining: wrong mode %d", mode); return (remaining); } static int remain_for_mode(int mode) { return (mode == SINGLE_ALLPROC ? 0 : 1); } static void weed_inhib(int mode, struct thread *td2, struct proc *p) { PROC_LOCK_ASSERT(p, MA_OWNED); PROC_SLOCK_ASSERT(p, MA_OWNED); THREAD_LOCK_ASSERT(td2, MA_OWNED); /* * Since the thread lock is dropped by the scheduler we have * to retry to check for races. */ restart: switch (mode) { case SINGLE_EXIT: if (TD_IS_SUSPENDED(td2)) { thread_unsuspend_one(td2, p, true); thread_lock(td2); goto restart; } if (TD_CAN_ABORT(td2)) { sleepq_abort(td2, EINTR); return; } break; case SINGLE_BOUNDARY: case SINGLE_NO_EXIT: if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_BOUNDARY) == 0) { thread_unsuspend_one(td2, p, false); thread_lock(td2); goto restart; } if (TD_CAN_ABORT(td2)) { sleepq_abort(td2, ERESTART); return; } break; case SINGLE_ALLPROC: /* * ALLPROC suspend tries to avoid spurious EINTR for * threads sleeping interruptable, by suspending the * thread directly, similarly to sig_suspend_threads(). * Since such sleep is not neccessary performed at the user * boundary, TDF_ALLPROCSUSP is used to avoid immediate * un-suspend. */ if (TD_IS_SUSPENDED(td2) && (td2->td_flags & TDF_ALLPROCSUSP) == 0) { thread_unsuspend_one(td2, p, false); thread_lock(td2); goto restart; } if (TD_CAN_ABORT(td2)) { td2->td_flags |= TDF_ALLPROCSUSP; sleepq_abort(td2, ERESTART); return; } break; default: break; } thread_unlock(td2); } /* * Enforce single-threading. * * Returns 1 if the caller must abort (another thread is waiting to * exit the process or similar). Process is locked! * Returns 0 when you are successfully the only thread running. * A process has successfully single threaded in the suspend mode when * There are no threads in user mode. Threads in the kernel must be * allowed to continue until they get to the user boundary. They may even * copy out their return values and data before suspending. They may however be * accelerated in reaching the user boundary as we will wake up * any sleeping threads that are interruptable. (PCATCH). */ int thread_single(struct proc *p, int mode) { struct thread *td; struct thread *td2; int remaining; td = curthread; KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, ("invalid mode %d", mode)); /* * If allowing non-ALLPROC singlethreading for non-curproc * callers, calc_remaining() and remain_for_mode() should be * adjusted to also account for td->td_proc != p. For now * this is not implemented because it is not used. */ KASSERT((mode == SINGLE_ALLPROC && td->td_proc != p) || (mode != SINGLE_ALLPROC && td->td_proc == p), ("mode %d proc %p curproc %p", mode, p, td->td_proc)); mtx_assert(&Giant, MA_NOTOWNED); PROC_LOCK_ASSERT(p, MA_OWNED); /* * Is someone already single threading? * Or may be singlethreading is not needed at all. */ if (mode == SINGLE_ALLPROC) { while ((p->p_flag & P_STOPPED_SINGLE) != 0) { if ((p->p_flag2 & P2_WEXIT) != 0) return (1); msleep(&p->p_flag, &p->p_mtx, PCATCH, "thrsgl", 0); } if ((p->p_flag & (P_STOPPED_SIG | P_TRACED)) != 0 || (p->p_flag2 & P2_WEXIT) != 0) return (1); } else if ((p->p_flag & P_HADTHREADS) == 0) return (0); if (p->p_singlethread != NULL && p->p_singlethread != td) return (1); if (mode == SINGLE_EXIT) { p->p_flag |= P_SINGLE_EXIT; p->p_flag &= ~P_SINGLE_BOUNDARY; } else { p->p_flag &= ~P_SINGLE_EXIT; if (mode == SINGLE_BOUNDARY) p->p_flag |= P_SINGLE_BOUNDARY; else p->p_flag &= ~P_SINGLE_BOUNDARY; } if (mode == SINGLE_ALLPROC) p->p_flag |= P_TOTAL_STOP; p->p_flag |= P_STOPPED_SINGLE; PROC_SLOCK(p); p->p_singlethread = td; remaining = calc_remaining(p, mode); while (remaining != remain_for_mode(mode)) { if (P_SHOULDSTOP(p) != P_STOPPED_SINGLE) goto stopme; FOREACH_THREAD_IN_PROC(p, td2) { if (td2 == td) continue; thread_lock(td2); ast_sched_locked(td2, TDA_SUSPEND); if (TD_IS_INHIBITED(td2)) { weed_inhib(mode, td2, p); #ifdef SMP } else if (TD_IS_RUNNING(td2)) { forward_signal(td2); thread_unlock(td2); #endif } else thread_unlock(td2); } remaining = calc_remaining(p, mode); /* * Maybe we suspended some threads.. was it enough? */ if (remaining == remain_for_mode(mode)) break; stopme: /* * Wake us up when everyone else has suspended. * In the mean time we suspend as well. */ thread_suspend_switch(td, p); remaining = calc_remaining(p, mode); } if (mode == SINGLE_EXIT) { /* * Convert the process to an unthreaded process. The * SINGLE_EXIT is called by exit1() or execve(), in * both cases other threads must be retired. */ KASSERT(p->p_numthreads == 1, ("Unthreading with >1 threads")); p->p_singlethread = NULL; p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_HADTHREADS); /* * Wait for any remaining threads to exit cpu_throw(). */ while (p->p_exitthreads != 0) { PROC_SUNLOCK(p); PROC_UNLOCK(p); sched_relinquish(td); PROC_LOCK(p); PROC_SLOCK(p); } } else if (mode == SINGLE_BOUNDARY) { /* * Wait until all suspended threads are removed from * the processors. The thread_suspend_check() * increments p_boundary_count while it is still * running, which makes it possible for the execve() * to destroy vmspace while our other threads are * still using the address space. * * We lock the thread, which is only allowed to * succeed after context switch code finished using * the address space. */ FOREACH_THREAD_IN_PROC(p, td2) { if (td2 == td) continue; thread_lock(td2); KASSERT((td2->td_flags & TDF_BOUNDARY) != 0, ("td %p not on boundary", td2)); KASSERT(TD_IS_SUSPENDED(td2), ("td %p is not suspended", td2)); thread_unlock(td2); } } PROC_SUNLOCK(p); return (0); } bool thread_suspend_check_needed(void) { struct proc *p; struct thread *td; td = curthread; p = td->td_proc; PROC_LOCK_ASSERT(p, MA_OWNED); return (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) != 0 && (td->td_dbgflags & TDB_SUSPEND) != 0)); } /* * Called in from locations that can safely check to see * whether we have to suspend or at least throttle for a * single-thread event (e.g. fork). * * Such locations include userret(). * If the "return_instead" argument is non zero, the thread must be able to * accept 0 (caller may continue), or 1 (caller must abort) as a result. * * The 'return_instead' argument tells the function if it may do a * thread_exit() or suspend, or whether the caller must abort and back * out instead. * * If the thread that set the single_threading request has set the * P_SINGLE_EXIT bit in the process flags then this call will never return * if 'return_instead' is false, but will exit. * * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 *---------------+--------------------+--------------------- * 0 | returns 0 | returns 0 or 1 * | when ST ends | immediately *---------------+--------------------+--------------------- * 1 | thread exits | returns 1 * | | immediately * 0 = thread_exit() or suspension ok, * other = return error instead of stopping the thread. * * While a full suspension is under effect, even a single threading * thread would be suspended if it made this call (but it shouldn't). * This call should only be made from places where * thread_exit() would be safe as that may be the outcome unless * return_instead is set. */ int thread_suspend_check(int return_instead) { struct thread *td; struct proc *p; td = curthread; p = td->td_proc; mtx_assert(&Giant, MA_NOTOWNED); PROC_LOCK_ASSERT(p, MA_OWNED); while (thread_suspend_check_needed()) { if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { KASSERT(p->p_singlethread != NULL, ("singlethread not set")); /* * The only suspension in action is a * single-threading. Single threader need not stop. * It is safe to access p->p_singlethread unlocked * because it can only be set to our address by us. */ if (p->p_singlethread == td) return (0); /* Exempt from stopping. */ } if ((p->p_flag & P_SINGLE_EXIT) && return_instead) return (EINTR); /* Should we goto user boundary if we didn't come from there? */ if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && (p->p_flag & P_SINGLE_BOUNDARY) && return_instead) return (ERESTART); /* * Ignore suspend requests if they are deferred. */ if ((td->td_flags & TDF_SBDRY) != 0) { KASSERT(return_instead, ("TDF_SBDRY set for unsafe thread_suspend_check")); KASSERT((td->td_flags & (TDF_SEINTR | TDF_SERESTART)) != (TDF_SEINTR | TDF_SERESTART), ("both TDF_SEINTR and TDF_SERESTART")); return (TD_SBDRY_INTR(td) ? TD_SBDRY_ERRNO(td) : 0); } /* * If the process is waiting for us to exit, * this thread should just suicide. * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. */ if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { PROC_UNLOCK(p); /* * Allow Linux emulation layer to do some work * before thread suicide. */ if (__predict_false(p->p_sysent->sv_thread_detach != NULL)) (p->p_sysent->sv_thread_detach)(td); umtx_thread_exit(td); kern_thr_exit(td); panic("stopped thread did not exit"); } PROC_SLOCK(p); thread_stopped(p); if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { if (p->p_numthreads == p->p_suspcount + 1) { thread_lock(p->p_singlethread); thread_unsuspend_one(p->p_singlethread, p, false); } } PROC_UNLOCK(p); thread_lock(td); /* * When a thread suspends, it just * gets taken off all queues. */ thread_suspend_one(td); if (return_instead == 0) { p->p_boundary_count++; td->td_flags |= TDF_BOUNDARY; } PROC_SUNLOCK(p); mi_switch(SW_INVOL | SWT_SUSPEND); PROC_LOCK(p); } return (0); } /* * Check for possible stops and suspensions while executing a * casueword or similar transiently failing operation. * * The sleep argument controls whether the function can handle a stop * request itself or it should return ERESTART and the request is * proceed at the kernel/user boundary in ast. * * Typically, when retrying due to casueword(9) failure (rv == 1), we * should handle the stop requests there, with exception of cases when * the thread owns a kernel resource, for instance busied the umtx * key, or when functions return immediately if thread_check_susp() * returned non-zero. On the other hand, retrying the whole lock * operation, we better not stop there but delegate the handling to * ast. * * If the request is for thread termination P_SINGLE_EXIT, we cannot * handle it at all, and simply return EINTR. */ int thread_check_susp(struct thread *td, bool sleep) { struct proc *p; int error; /* * The check for TDA_SUSPEND is racy, but it is enough to * eventually break the lockstep loop. */ if (!td_ast_pending(td, TDA_SUSPEND)) return (0); error = 0; p = td->td_proc; PROC_LOCK(p); if (p->p_flag & P_SINGLE_EXIT) error = EINTR; else if (P_SHOULDSTOP(p) || ((p->p_flag & P_TRACED) && (td->td_dbgflags & TDB_SUSPEND))) error = sleep ? thread_suspend_check(0) : ERESTART; PROC_UNLOCK(p); return (error); } void thread_suspend_switch(struct thread *td, struct proc *p) { KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); PROC_LOCK_ASSERT(p, MA_OWNED); PROC_SLOCK_ASSERT(p, MA_OWNED); /* * We implement thread_suspend_one in stages here to avoid * dropping the proc lock while the thread lock is owned. */ if (p == td->td_proc) { thread_stopped(p); p->p_suspcount++; } PROC_UNLOCK(p); thread_lock(td); ast_unsched_locked(td, TDA_SUSPEND); TD_SET_SUSPENDED(td); sched_sleep(td, 0); PROC_SUNLOCK(p); DROP_GIANT(); mi_switch(SW_VOL | SWT_SUSPEND); PICKUP_GIANT(); PROC_LOCK(p); PROC_SLOCK(p); } void thread_suspend_one(struct thread *td) { struct proc *p; p = td->td_proc; PROC_SLOCK_ASSERT(p, MA_OWNED); THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(!TD_IS_SUSPENDED(td), ("already suspended")); p->p_suspcount++; ast_unsched_locked(td, TDA_SUSPEND); TD_SET_SUSPENDED(td); sched_sleep(td, 0); } static void thread_unsuspend_one(struct thread *td, struct proc *p, bool boundary) { THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); TD_CLR_SUSPENDED(td); td->td_flags &= ~TDF_ALLPROCSUSP; if (td->td_proc == p) { PROC_SLOCK_ASSERT(p, MA_OWNED); p->p_suspcount--; if (boundary && (td->td_flags & TDF_BOUNDARY) != 0) { td->td_flags &= ~TDF_BOUNDARY; p->p_boundary_count--; } } setrunnable(td, 0); } void thread_run_flash(struct thread *td) { struct proc *p; p = td->td_proc; PROC_LOCK_ASSERT(p, MA_OWNED); if (TD_ON_SLEEPQ(td)) sleepq_remove_nested(td); else thread_lock(td); THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(TD_IS_SUSPENDED(td), ("Thread not suspended")); TD_CLR_SUSPENDED(td); PROC_SLOCK(p); MPASS(p->p_suspcount > 0); p->p_suspcount--; PROC_SUNLOCK(p); setrunnable(td, 0); } /* * Allow all threads blocked by single threading to continue running. */ void thread_unsuspend(struct proc *p) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); PROC_SLOCK_ASSERT(p, MA_OWNED); if (!P_SHOULDSTOP(p)) { FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (TD_IS_SUSPENDED(td)) thread_unsuspend_one(td, p, true); else thread_unlock(td); } } else if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE && p->p_numthreads == p->p_suspcount) { /* * Stopping everything also did the job for the single * threading request. Now we've downgraded to single-threaded, * let it continue. */ if (p->p_singlethread->td_proc == p) { thread_lock(p->p_singlethread); thread_unsuspend_one(p->p_singlethread, p, false); } } } /* * End the single threading mode.. */ void thread_single_end(struct proc *p, int mode) { struct thread *td; KASSERT(mode == SINGLE_EXIT || mode == SINGLE_BOUNDARY || mode == SINGLE_ALLPROC || mode == SINGLE_NO_EXIT, ("invalid mode %d", mode)); PROC_LOCK_ASSERT(p, MA_OWNED); KASSERT((mode == SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) != 0) || (mode != SINGLE_ALLPROC && (p->p_flag & P_TOTAL_STOP) == 0), ("mode %d does not match P_TOTAL_STOP", mode)); KASSERT(mode == SINGLE_ALLPROC || p->p_singlethread == curthread, ("thread_single_end from other thread %p %p", curthread, p->p_singlethread)); KASSERT(mode != SINGLE_BOUNDARY || (p->p_flag & P_SINGLE_BOUNDARY) != 0, ("mis-matched SINGLE_BOUNDARY flags %x", p->p_flag)); p->p_flag &= ~(P_STOPPED_SINGLE | P_SINGLE_EXIT | P_SINGLE_BOUNDARY | P_TOTAL_STOP); PROC_SLOCK(p); p->p_singlethread = NULL; /* * If there are other threads they may now run, * unless of course there is a blanket 'stop order' * on the process. The single threader must be allowed * to continue however as this is a bad place to stop. */ if (p->p_numthreads != remain_for_mode(mode) && !P_SHOULDSTOP(p)) { FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (TD_IS_SUSPENDED(td)) thread_unsuspend_one(td, p, true); else thread_unlock(td); } } KASSERT(mode != SINGLE_BOUNDARY || p->p_boundary_count == 0, ("inconsistent boundary count %d", p->p_boundary_count)); PROC_SUNLOCK(p); wakeup(&p->p_flag); } /* * Locate a thread by number and return with proc lock held. * * thread exit establishes proc -> tidhash lock ordering, but lookup * takes tidhash first and needs to return locked proc. * * The problem is worked around by relying on type-safety of both * structures and doing the work in 2 steps: * - tidhash-locked lookup which saves both thread and proc pointers * - proc-locked verification that the found thread still matches */ static bool tdfind_hash(lwpid_t tid, pid_t pid, struct proc **pp, struct thread **tdp) { #define RUN_THRESH 16 struct proc *p; struct thread *td; int run; bool locked; run = 0; rw_rlock(TIDHASHLOCK(tid)); locked = true; LIST_FOREACH(td, TIDHASH(tid), td_hash) { if (td->td_tid != tid) { run++; continue; } p = td->td_proc; if (pid != -1 && p->p_pid != pid) { td = NULL; break; } if (run > RUN_THRESH) { if (rw_try_upgrade(TIDHASHLOCK(tid))) { LIST_REMOVE(td, td_hash); LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); rw_wunlock(TIDHASHLOCK(tid)); locked = false; break; } } break; } if (locked) rw_runlock(TIDHASHLOCK(tid)); if (td == NULL) return (false); *pp = p; *tdp = td; return (true); } struct thread * tdfind(lwpid_t tid, pid_t pid) { struct proc *p; struct thread *td; td = curthread; if (td->td_tid == tid) { if (pid != -1 && td->td_proc->p_pid != pid) return (NULL); PROC_LOCK(td->td_proc); return (td); } for (;;) { if (!tdfind_hash(tid, pid, &p, &td)) return (NULL); PROC_LOCK(p); if (td->td_tid != tid) { PROC_UNLOCK(p); continue; } if (td->td_proc != p) { PROC_UNLOCK(p); continue; } if (p->p_state == PRS_NEW) { PROC_UNLOCK(p); return (NULL); } return (td); } } void tidhash_add(struct thread *td) { rw_wlock(TIDHASHLOCK(td->td_tid)); LIST_INSERT_HEAD(TIDHASH(td->td_tid), td, td_hash); rw_wunlock(TIDHASHLOCK(td->td_tid)); } void tidhash_remove(struct thread *td) { rw_wlock(TIDHASHLOCK(td->td_tid)); LIST_REMOVE(td, td_hash); rw_wunlock(TIDHASHLOCK(td->td_tid)); } diff --git a/sys/kern/uipc_shm.c b/sys/kern/uipc_shm.c index 7672ded459df..49b5b56dee17 100644 --- a/sys/kern/uipc_shm.c +++ b/sys/kern/uipc_shm.c @@ -1,2204 +1,2204 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson * Copyright 2020 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * 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. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Support for shared swap-backed anonymous memory objects via * shm_open(2), shm_rename(2), and shm_unlink(2). * While most of the implementation is here, vm_mmap.c contains * mapping logic changes. * * posixshmcontrol(1) allows users to inspect the state of the memory * objects. Per-uid swap resource limit controls total amount of * memory that user can consume for anonymous objects, including * shared. */ #include #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct shm_mapping { char *sm_path; Fnv32_t sm_fnv; struct shmfd *sm_shmfd; LIST_ENTRY(shm_mapping) sm_link; }; static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor"); static LIST_HEAD(, shm_mapping) *shm_dictionary; static struct sx shm_dict_lock; static struct mtx shm_timestamp_lock; static u_long shm_hash; static struct unrhdr64 shm_ino_unr; static dev_t shm_dev_ino; #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash]) static void shm_init(void *arg); static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd); static struct shmfd *shm_lookup(char *path, Fnv32_t fnv); static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred); static void shm_doremove(struct shm_mapping *map); static int shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie); static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie); static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out); static int shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags); static fo_rdwr_t shm_read; static fo_rdwr_t shm_write; static fo_truncate_t shm_truncate; static fo_ioctl_t shm_ioctl; static fo_stat_t shm_stat; static fo_close_t shm_close; static fo_chmod_t shm_chmod; static fo_chown_t shm_chown; static fo_seek_t shm_seek; static fo_fill_kinfo_t shm_fill_kinfo; static fo_mmap_t shm_mmap; static fo_get_seals_t shm_get_seals; static fo_add_seals_t shm_add_seals; static fo_fallocate_t shm_fallocate; static fo_fspacectl_t shm_fspacectl; /* File descriptor operations. */ struct fileops shm_ops = { .fo_read = shm_read, .fo_write = shm_write, .fo_truncate = shm_truncate, .fo_ioctl = shm_ioctl, .fo_poll = invfo_poll, .fo_kqfilter = invfo_kqfilter, .fo_stat = shm_stat, .fo_close = shm_close, .fo_chmod = shm_chmod, .fo_chown = shm_chown, .fo_sendfile = vn_sendfile, .fo_seek = shm_seek, .fo_fill_kinfo = shm_fill_kinfo, .fo_mmap = shm_mmap, .fo_get_seals = shm_get_seals, .fo_add_seals = shm_add_seals, .fo_fallocate = shm_fallocate, .fo_fspacectl = shm_fspacectl, .fo_cmp = file_kcmp_generic, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE, }; FEATURE(posix_shm, "POSIX shared memory"); static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, ""); static int largepage_reclaim_tries = 1; SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries, CTLFLAG_RWTUN, &largepage_reclaim_tries, 0, "Number of contig reclaims before giving up for default alloc policy"); #define shm_rangelock_unlock(shmfd, cookie) \ - rangelock_unlock(&(shmfd)->shm_rl, (cookie), &(shmfd)->shm_mtx) + rangelock_unlock(&(shmfd)->shm_rl, (cookie)) #define shm_rangelock_rlock(shmfd, start, end) \ - rangelock_rlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) + rangelock_rlock(&(shmfd)->shm_rl, (start), (end)) #define shm_rangelock_tryrlock(shmfd, start, end) \ - rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) + rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end)) #define shm_rangelock_wlock(shmfd, start, end) \ - rangelock_wlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx) + rangelock_wlock(&(shmfd)->shm_rl, (start), (end)) static int uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio) { vm_page_t m; vm_pindex_t idx; size_t tlen; int error, offset, rv; idx = OFF_TO_IDX(uio->uio_offset); offset = uio->uio_offset & PAGE_MASK; tlen = MIN(PAGE_SIZE - offset, len); rv = vm_page_grab_valid_unlocked(&m, obj, idx, VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT); if (rv == VM_PAGER_OK) goto found; /* * Read I/O without either a corresponding resident page or swap * page: use zero_region. This is intended to avoid instantiating * pages on read from a sparse region. */ VM_OBJECT_WLOCK(obj); m = vm_page_lookup(obj, idx); if (uio->uio_rw == UIO_READ && m == NULL && !vm_pager_has_page(obj, idx, NULL, NULL)) { VM_OBJECT_WUNLOCK(obj); return (uiomove(__DECONST(void *, zero_region), tlen, uio)); } /* * Although the tmpfs vnode lock is held here, it is * nonetheless safe to sleep waiting for a free page. The * pageout daemon does not need to acquire the tmpfs vnode * lock to page out tobj's pages because tobj is a OBJT_SWAP * type object. */ rv = vm_page_grab_valid(&m, obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY); if (rv != VM_PAGER_OK) { VM_OBJECT_WUNLOCK(obj); if (bootverbose) { printf("uiomove_object: vm_obj %p idx %jd " "pager error %d\n", obj, idx, rv); } return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO); } VM_OBJECT_WUNLOCK(obj); found: error = uiomove_fromphys(&m, offset, tlen, uio); if (uio->uio_rw == UIO_WRITE && error == 0) vm_page_set_dirty(m); vm_page_activate(m); vm_page_sunbusy(m); return (error); } int uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio) { ssize_t resid; size_t len; int error; error = 0; while ((resid = uio->uio_resid) > 0) { if (obj_size <= uio->uio_offset) break; len = MIN(obj_size - uio->uio_offset, resid); if (len == 0) break; error = uiomove_object_page(obj, len, uio); if (error != 0 || resid == uio->uio_resid) break; } return (error); } static u_long count_largepages[MAXPAGESIZES]; static int shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx, int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last) { vm_page_t m __diagused; int psind; psind = object->un_pager.phys.data_val; if (psind == 0 || pidx >= object->size) return (VM_PAGER_FAIL); *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE); /* * We only busy the first page in the superpage run. It is * useless to busy whole run since we only remove full * superpage, and it takes too long to busy e.g. 512 * 512 == * 262144 pages constituing 1G amd64 superage. */ m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT); MPASS(m != NULL); *last = *first + atop(pagesizes[psind]) - 1; return (VM_PAGER_OK); } static boolean_t shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex, int *before, int *after) { int psind; psind = object->un_pager.phys.data_val; if (psind == 0 || pindex >= object->size) return (FALSE); if (before != NULL) { *before = pindex - rounddown2(pindex, pagesizes[psind] / PAGE_SIZE); } if (after != NULL) { *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) - pindex; } return (TRUE); } static void shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred) { } static void shm_largepage_phys_dtor(vm_object_t object) { int psind; psind = object->un_pager.phys.data_val; if (psind != 0) { atomic_subtract_long(&count_largepages[psind], object->size / (pagesizes[psind] / PAGE_SIZE)); vm_wire_sub(object->size); } else { KASSERT(object->size == 0, ("largepage phys obj %p not initialized bit size %#jx > 0", object, (uintmax_t)object->size)); } } static const struct phys_pager_ops shm_largepage_phys_ops = { .phys_pg_populate = shm_largepage_phys_populate, .phys_pg_haspage = shm_largepage_phys_haspage, .phys_pg_ctor = shm_largepage_phys_ctor, .phys_pg_dtor = shm_largepage_phys_dtor, }; bool shm_largepage(struct shmfd *shmfd) { return (shmfd->shm_object->type == OBJT_PHYS); } static void shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size) { struct shmfd *shm; vm_size_t c; swap_pager_freespace(obj, start, size, &c); if (c == 0) return; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; KASSERT(shm->shm_pages >= c, ("shm %p pages %jd free %jd", shm, (uintmax_t)shm->shm_pages, (uintmax_t)c)); shm->shm_pages -= c; } static void shm_page_inserted(vm_object_t obj, vm_page_t m) { struct shmfd *shm; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) shm->shm_pages += 1; } static void shm_page_removed(vm_object_t obj, vm_page_t m) { struct shmfd *shm; shm = obj->un_pager.swp.swp_priv; if (shm == NULL) return; if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) { KASSERT(shm->shm_pages >= 1, ("shm %p pages %jd free 1", shm, (uintmax_t)shm->shm_pages)); shm->shm_pages -= 1; } } static struct pagerops shm_swap_pager_ops = { .pgo_kvme_type = KVME_TYPE_SWAP, .pgo_freespace = shm_pager_freespace, .pgo_page_inserted = shm_page_inserted, .pgo_page_removed = shm_page_removed, }; static int shmfd_pager_type = -1; static int shm_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct shmfd *shmfd; off_t foffset; int error; shmfd = fp->f_data; foffset = foffset_lock(fp, 0); error = 0; switch (whence) { case L_INCR: if (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset)) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) { error = EOVERFLOW; break; } offset += shmfd->shm_size; break; case L_SET: break; default: error = EINVAL; } if (error == 0) { if (offset < 0 || offset > shmfd->shm_size) error = EINVAL; else td->td_uretoff.tdu_off = offset; } foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } static int shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif foffset_lock_uio(fp, uio, flags); rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset, uio->uio_offset + uio->uio_resid); error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); shm_rangelock_unlock(shmfd, rl_cookie); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct shmfd *shmfd; void *rl_cookie; int error; off_t size; shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0) return (EINVAL); foffset_lock_uio(fp, uio, flags); if (uio->uio_resid > OFF_MAX - uio->uio_offset) { /* * Overflow is only an error if we're supposed to expand on * write. Otherwise, we'll just truncate the write to the * size of the file, which can only grow up to OFF_MAX. */ if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) { foffset_unlock_uio(fp, uio, flags); return (EFBIG); } size = shmfd->shm_size; } else { size = uio->uio_offset + uio->uio_resid; } if ((flags & FOF_OFFSET) == 0) rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); else rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, size); if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; } else { error = 0; if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 && size > shmfd->shm_size) { error = shm_dotruncate_cookie(shmfd, size, rl_cookie); } if (error == 0) error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio); } shm_rangelock_unlock(shmfd, rl_cookie); foffset_unlock_uio(fp, uio, flags); return (error); } static int shm_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif return (shm_dotruncate(shmfd, length)); } int shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; struct shm_largepage_conf *conf; void *rl_cookie; shmfd = fp->f_data; switch (com) { case FIONBIO: case FIOASYNC: /* * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work, * just like it would on an unlinked regular file */ return (0); case FIOSSHMLPGCNF: if (!shm_largepage(shmfd)) return (ENOTTY); conf = data; if (shmfd->shm_lp_psind != 0 && conf->psind != shmfd->shm_lp_psind) return (EINVAL); if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES || pagesizes[conf->psind] == 0) return (EINVAL); if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT && conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT && conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD) return (EINVAL); rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); shmfd->shm_lp_psind = conf->psind; shmfd->shm_lp_alloc_policy = conf->alloc_policy; shmfd->shm_object->un_pager.phys.data_val = conf->psind; shm_rangelock_unlock(shmfd, rl_cookie); return (0); case FIOGSHMLPGCNF: if (!shm_largepage(shmfd)) return (ENOTTY); conf = data; rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX); conf->psind = shmfd->shm_lp_psind; conf->alloc_policy = shmfd->shm_lp_alloc_policy; shm_rangelock_unlock(shmfd, rl_cookie); return (0); default: return (ENOTTY); } } static int shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred) { struct shmfd *shmfd; #ifdef MAC int error; #endif shmfd = fp->f_data; #ifdef MAC error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd); if (error) return (error); #endif /* * Attempt to return sanish values for fstat() on a memory file * descriptor. */ bzero(sb, sizeof(*sb)); sb->st_blksize = PAGE_SIZE; sb->st_size = shmfd->shm_size; mtx_lock(&shm_timestamp_lock); sb->st_atim = shmfd->shm_atime; sb->st_ctim = shmfd->shm_ctime; sb->st_mtim = shmfd->shm_mtime; sb->st_birthtim = shmfd->shm_birthtime; sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */ sb->st_uid = shmfd->shm_uid; sb->st_gid = shmfd->shm_gid; mtx_unlock(&shm_timestamp_lock); sb->st_dev = shm_dev_ino; sb->st_ino = shmfd->shm_ino; sb->st_nlink = shmfd->shm_object->ref_count; if (shm_largepage(shmfd)) { sb->st_blocks = shmfd->shm_object->size / (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT); } else { sb->st_blocks = shmfd->shm_pages; } return (0); } static int shm_close(struct file *fp, struct thread *td) { struct shmfd *shmfd; shmfd = fp->f_data; fp->f_data = NULL; shm_drop(shmfd); return (0); } static int shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) { int error; char *path; const char *pr_path; size_t pr_pathlen; path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK); pr_path = td->td_ucred->cr_prison->pr_path; /* Construct a full pathname for jailed callers. */ pr_pathlen = strcmp(pr_path, "/") == 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN); error = copyinstr(userpath_in, path + pr_pathlen, MAXPATHLEN - pr_pathlen, NULL); if (error != 0) goto out; #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(path); #endif /* Require paths to start with a '/' character. */ if (path[pr_pathlen] != '/') { error = EINVAL; goto out; } *path_out = path; out: if (error != 0) free(path, M_SHMFD); return (error); } static int shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base, int end) { vm_page_t m; int rv; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(base >= 0, ("%s: base %d", __func__, base)); KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base, end)); retry: m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT); if (m != NULL) { MPASS(vm_page_all_valid(m)); } else if (vm_pager_has_page(object, idx, NULL, NULL)) { m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL); if (m == NULL) goto retry; vm_object_pip_add(object, 1); VM_OBJECT_WUNLOCK(object); rv = vm_pager_get_pages(object, &m, 1, NULL, NULL); VM_OBJECT_WLOCK(object); vm_object_pip_wakeup(object); if (rv == VM_PAGER_OK) { /* * Since the page was not resident, and therefore not * recently accessed, immediately enqueue it for * asynchronous laundering. The current operation is * not regarded as an access. */ vm_page_launder(m); } else { vm_page_free(m); VM_OBJECT_WUNLOCK(object); return (EIO); } } if (m != NULL) { pmap_zero_page_area(m, base, end - base); KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid", __func__, m)); vm_page_set_dirty(m); vm_page_xunbusy(m); } return (0); } static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_pindex_t nobjsize; vm_ooffset_t delta; int base, error; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); if (length == shmfd->shm_size) return (0); nobjsize = OFF_TO_IDX(length + PAGE_MASK); /* Are we shrinking? If so, trim the end. */ if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); /* * Disallow any requests to shrink the size if this * object is mapped into the kernel. */ if (shmfd->shm_kmappings > 0) return (EBUSY); /* * Zero the truncated part of the last page. */ base = length & PAGE_MASK; if (base != 0) { error = shm_partial_page_invalidate(object, OFF_TO_IDX(length), base, PAGE_SIZE); if (error) return (error); } delta = IDX_TO_OFF(object->size - nobjsize); if (nobjsize < object->size) vm_object_page_remove(object, nobjsize, object->size, 0); /* Free the swap accounted for shm */ swap_release_by_cred(delta, object->cred); object->charge -= delta; } else { if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); /* Try to reserve additional swap space. */ delta = IDX_TO_OFF(nobjsize - object->size); if (!swap_reserve_by_cred(delta, object->cred)) return (ENOMEM); object->charge += delta; } shmfd->shm_size = length; mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_ctime); shmfd->shm_mtime = shmfd->shm_ctime; mtx_unlock(&shm_timestamp_lock); object->size = nobjsize; return (0); } static int shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie) { vm_object_t object; vm_page_t m; vm_pindex_t newobjsz; vm_pindex_t oldobjsz __unused; int aflags, error, i, psind, try; KASSERT(length >= 0, ("shm_dotruncate: length < 0")); object = shmfd->shm_object; VM_OBJECT_ASSERT_WLOCKED(object); rangelock_cookie_assert(rl_cookie, RA_WLOCKED); oldobjsz = object->size; newobjsz = OFF_TO_IDX(length); if (length == shmfd->shm_size) return (0); psind = shmfd->shm_lp_psind; if (psind == 0 && length != 0) return (EINVAL); if ((length & (pagesizes[psind] - 1)) != 0) return (EINVAL); if (length < shmfd->shm_size) { if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0) return (EPERM); if (shmfd->shm_kmappings > 0) return (EBUSY); return (ENOTSUP); /* Pages are unmanaged. */ #if 0 vm_object_page_remove(object, newobjsz, oldobjsz, 0); object->size = newobjsz; shmfd->shm_size = length; return (0); #endif } if ((shmfd->shm_seals & F_SEAL_GROW) != 0) return (EPERM); aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO; if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT) aflags |= VM_ALLOC_WAITFAIL; try = 0; /* * Extend shmfd and object, keeping all already fully * allocated large pages intact even on error, because dropped * object lock might allowed mapping of them. */ while (object->size < newobjsz) { m = vm_page_alloc_contig(object, object->size, aflags, pagesizes[psind] / PAGE_SIZE, 0, ~0, pagesizes[psind], 0, VM_MEMATTR_DEFAULT); if (m == NULL) { VM_OBJECT_WUNLOCK(object); if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT || (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_DEFAULT && try >= largepage_reclaim_tries)) { VM_OBJECT_WLOCK(object); return (ENOMEM); } error = vm_page_reclaim_contig(aflags, pagesizes[psind] / PAGE_SIZE, 0, ~0, pagesizes[psind], 0); if (error == ENOMEM) error = vm_wait_intr(object); if (error != 0) { VM_OBJECT_WLOCK(object); return (error); } try++; VM_OBJECT_WLOCK(object); continue; } try = 0; for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) { if ((m[i].flags & PG_ZERO) == 0) pmap_zero_page(&m[i]); vm_page_valid(&m[i]); vm_page_xunbusy(&m[i]); } object->size += OFF_TO_IDX(pagesizes[psind]); shmfd->shm_size += pagesizes[psind]; atomic_add_long(&count_largepages[psind], 1); vm_wire_add(atop(pagesizes[psind])); } return (0); } static int shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie) { int error; VM_OBJECT_WLOCK(shmfd->shm_object); error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd, length, rl_cookie) : shm_dotruncate_locked(shmfd, length, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); return (error); } int shm_dotruncate(struct shmfd *shmfd, off_t length) { void *rl_cookie; int error; rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); error = shm_dotruncate_cookie(shmfd, length, rl_cookie); shm_rangelock_unlock(shmfd, rl_cookie); return (error); } /* * shmfd object management including creation and reference counting * routines. */ struct shmfd * shm_alloc(struct ucred *ucred, mode_t mode, bool largepage) { struct shmfd *shmfd; vm_object_t obj; shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO); shmfd->shm_size = 0; shmfd->shm_uid = ucred->cr_uid; shmfd->shm_gid = ucred->cr_gid; shmfd->shm_mode = mode; if (largepage) { shmfd->shm_object = phys_pager_allocate(NULL, &shm_largepage_phys_ops, NULL, shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT; } else { obj = vm_pager_allocate(shmfd_pager_type, NULL, shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred); VM_OBJECT_WLOCK(obj); obj->un_pager.swp.swp_priv = shmfd; VM_OBJECT_WUNLOCK(obj); shmfd->shm_object = obj; } KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate")); vfs_timestamp(&shmfd->shm_birthtime); shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime = shmfd->shm_birthtime; shmfd->shm_ino = alloc_unr64(&shm_ino_unr); refcount_init(&shmfd->shm_refs, 1); mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF); rangelock_init(&shmfd->shm_rl); #ifdef MAC mac_posixshm_init(shmfd); mac_posixshm_create(ucred, shmfd); #endif return (shmfd); } struct shmfd * shm_hold(struct shmfd *shmfd) { refcount_acquire(&shmfd->shm_refs); return (shmfd); } void shm_drop(struct shmfd *shmfd) { vm_object_t obj; if (refcount_release(&shmfd->shm_refs)) { #ifdef MAC mac_posixshm_destroy(shmfd); #endif rangelock_destroy(&shmfd->shm_rl); mtx_destroy(&shmfd->shm_mtx); obj = shmfd->shm_object; if (!shm_largepage(shmfd)) { VM_OBJECT_WLOCK(obj); obj->un_pager.swp.swp_priv = NULL; VM_OBJECT_WUNLOCK(obj); } vm_object_deallocate(obj); free(shmfd, M_SHMFD); } } /* * Determine if the credentials have sufficient permissions for a * specified combination of FREAD and FWRITE. */ int shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags) { accmode_t accmode; int error; accmode = 0; if (flags & FREAD) accmode |= VREAD; if (flags & FWRITE) accmode |= VWRITE; mtx_lock(&shm_timestamp_lock); error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, accmode, ucred); mtx_unlock(&shm_timestamp_lock); return (error); } static void shm_init(void *arg) { char name[32]; int i; mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF); sx_init(&shm_dict_lock, "shm dictionary"); shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash); new_unrhdr64(&shm_ino_unr, 1); shm_dev_ino = devfs_alloc_cdp_inode(); KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized")); shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops, OBJT_SWAP); MPASS(shmfd_pager_type != -1); for (i = 1; i < MAXPAGESIZES; i++) { if (pagesizes[i] == 0) break; #define M (1024 * 1024) #define G (1024 * M) if (pagesizes[i] >= G) snprintf(name, sizeof(name), "%luG", pagesizes[i] / G); else if (pagesizes[i] >= M) snprintf(name, sizeof(name), "%luM", pagesizes[i] / M); else snprintf(name, sizeof(name), "%lu", pagesizes[i]); #undef G #undef M SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages), OID_AUTO, name, CTLFLAG_RD, &count_largepages[i], "number of non-transient largepages allocated"); } } SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL); /* * Remove all shared memory objects that belong to a prison. */ void shm_remove_prison(struct prison *pr) { struct shm_mapping *shmm, *tshmm; u_long i; sx_xlock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) { if (shmm->sm_shmfd->shm_object->cred && shmm->sm_shmfd->shm_object->cred->cr_prison == pr) shm_doremove(shmm); } } sx_xunlock(&shm_dict_lock); } /* * Dictionary management. We maintain an in-kernel dictionary to map * paths to shmfd objects. We use the FNV hash on the path to store * the mappings in a hash table. */ static struct shmfd * shm_lookup(char *path, Fnv32_t fnv) { struct shm_mapping *map; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) return (map->sm_shmfd); } return (NULL); } static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd) { struct shm_mapping *map; map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK); map->sm_path = path; map->sm_fnv = fnv; map->sm_shmfd = shm_hold(shmfd); shmfd->shm_path = path; LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link); } static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred) { struct shm_mapping *map; int error; LIST_FOREACH(map, SHM_HASH(fnv), sm_link) { if (map->sm_fnv != fnv) continue; if (strcmp(map->sm_path, path) == 0) { #ifdef MAC error = mac_posixshm_check_unlink(ucred, map->sm_shmfd); if (error) return (error); #endif error = shm_access(map->sm_shmfd, ucred, FREAD | FWRITE); if (error) return (error); shm_doremove(map); return (0); } } return (ENOENT); } static void shm_doremove(struct shm_mapping *map) { map->sm_shmfd->shm_path = NULL; LIST_REMOVE(map, sm_link); shm_drop(map->sm_shmfd); free(map->sm_path, M_SHMFD); free(map, M_SHMFD); } int kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode, int shmflags, struct filecaps *fcaps, const char *name __unused) { struct pwddesc *pdp; struct shmfd *shmfd; struct file *fp; char *path; void *rl_cookie; Fnv32_t fnv; mode_t cmode; int error, fd, initial_seals; bool largepage; if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE | SHM_LARGEPAGE)) != 0) return (EINVAL); initial_seals = F_SEAL_SEAL; if ((shmflags & SHM_ALLOW_SEALING) != 0) initial_seals &= ~F_SEAL_SEAL; AUDIT_ARG_FFLAGS(flags); AUDIT_ARG_MODE(mode); if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR) return (EINVAL); if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0) return (EINVAL); largepage = (shmflags & SHM_LARGEPAGE) != 0; if (largepage && !PMAP_HAS_LARGEPAGES) return (ENOTTY); /* * Currently only F_SEAL_SEAL may be set when creating or opening shmfd. * If the decision is made later to allow additional seals, care must be * taken below to ensure that the seals are properly set if the shmfd * already existed -- this currently assumes that only F_SEAL_SEAL can * be set and doesn't take further precautions to ensure the validity of * the seals being added with respect to current mappings. */ if ((initial_seals & ~F_SEAL_SEAL) != 0) return (EINVAL); if (userpath != SHM_ANON) { error = shm_copyin_path(td, userpath, &path); if (error != 0) return (error); #ifdef CAPABILITY_MODE /* * shm_open(2) is only allowed for anonymous objects. */ if (CAP_TRACING(td)) ktrcapfail(CAPFAIL_NAMEI, path); if (IN_CAPABILITY_MODE(td)) { free(path, M_SHMFD); return (ECAPMODE); } #endif AUDIT_ARG_UPATH1_CANON(path); } else { path = NULL; } pdp = td->td_proc->p_pd; cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS; /* * shm_open(2) created shm should always have O_CLOEXEC set, as mandated * by POSIX. We allow it to be unset here so that an in-kernel * interface may be written as a thin layer around shm, optionally not * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally * in sys_shm_open() to keep this implementation compliant. */ error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps); if (error) { free(path, M_SHMFD); return (error); } /* A SHM_ANON path pointer creates an anonymous object. */ if (userpath == SHM_ANON) { /* A read-only anonymous object is pointless. */ if ((flags & O_ACCMODE) == O_RDONLY) { fdclose(td, fp, fd); fdrop(fp, td); return (EINVAL); } shmfd = shm_alloc(td->td_ucred, cmode, largepage); shmfd->shm_seals = initial_seals; shmfd->shm_flags = shmflags; } else { fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); shmfd = shm_lookup(path, fnv); if (shmfd == NULL) { /* Object does not yet exist, create it if requested. */ if (flags & O_CREAT) { #ifdef MAC error = mac_posixshm_check_create(td->td_ucred, path); if (error == 0) { #endif shmfd = shm_alloc(td->td_ucred, cmode, largepage); shmfd->shm_seals = initial_seals; shmfd->shm_flags = shmflags; shm_insert(path, fnv, shmfd); #ifdef MAC } #endif } else { free(path, M_SHMFD); error = ENOENT; } } else { rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); /* * kern_shm_open() likely shouldn't ever error out on * trying to set a seal that already exists, unlike * F_ADD_SEALS. This would break terribly as * shm_open(2) actually sets F_SEAL_SEAL to maintain * historical behavior where the underlying file could * not be sealed. */ initial_seals &= ~shmfd->shm_seals; /* * Object already exists, obtain a new * reference if requested and permitted. */ free(path, M_SHMFD); /* * initial_seals can't set additional seals if we've * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set, * then we've already removed that one from * initial_seals. This is currently redundant as we * only allow setting F_SEAL_SEAL at creation time, but * it's cheap to check and decreases the effort required * to allow additional seals. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 && initial_seals != 0) error = EPERM; else if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) error = EEXIST; else if (shmflags != 0 && shmflags != shmfd->shm_flags) error = EINVAL; else { #ifdef MAC error = mac_posixshm_check_open(td->td_ucred, shmfd, FFLAGS(flags & O_ACCMODE)); if (error == 0) #endif error = shm_access(shmfd, td->td_ucred, FFLAGS(flags & O_ACCMODE)); } /* * Truncate the file back to zero length if * O_TRUNC was specified and the object was * opened with read/write. */ if (error == 0 && (flags & (O_ACCMODE | O_TRUNC)) == (O_RDWR | O_TRUNC)) { VM_OBJECT_WLOCK(shmfd->shm_object); #ifdef MAC error = mac_posixshm_check_truncate( td->td_ucred, fp->f_cred, shmfd); if (error == 0) #endif error = shm_dotruncate_locked(shmfd, 0, rl_cookie); VM_OBJECT_WUNLOCK(shmfd->shm_object); } if (error == 0) { /* * Currently we only allow F_SEAL_SEAL to be * set initially. As noted above, this would * need to be reworked should that change. */ shmfd->shm_seals |= initial_seals; shm_hold(shmfd); } shm_rangelock_unlock(shmfd, rl_cookie); } sx_xunlock(&shm_dict_lock); if (error) { fdclose(td, fp, fd); fdrop(fp, td); return (error); } } finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops); td->td_retval[0] = fd; fdrop(fp, td); return (0); } /* System calls. */ #ifdef COMPAT_FREEBSD12 int freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap) { return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode, NULL)); } #endif int sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap) { char *path; Fnv32_t fnv; int error; error = shm_copyin_path(td, uap->path, &path); if (error != 0) return (error); AUDIT_ARG_UPATH1_CANON(path); fnv = fnv_32_str(path, FNV1_32_INIT); sx_xlock(&shm_dict_lock); error = shm_remove(path, fnv, td->td_ucred); sx_xunlock(&shm_dict_lock); free(path, M_SHMFD); return (error); } int sys_shm_rename(struct thread *td, struct shm_rename_args *uap) { char *path_from = NULL, *path_to = NULL; Fnv32_t fnv_from, fnv_to; struct shmfd *fd_from; struct shmfd *fd_to; int error; int flags; flags = uap->flags; AUDIT_ARG_FFLAGS(flags); /* * Make sure the user passed only valid flags. * If you add a new flag, please add a new term here. */ if ((flags & ~( SHM_RENAME_NOREPLACE | SHM_RENAME_EXCHANGE )) != 0) { error = EINVAL; goto out; } /* * EXCHANGE and NOREPLACE don't quite make sense together. Let's * force the user to choose one or the other. */ if ((flags & SHM_RENAME_NOREPLACE) != 0 && (flags & SHM_RENAME_EXCHANGE) != 0) { error = EINVAL; goto out; } /* Renaming to or from anonymous makes no sense */ if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) { error = EINVAL; goto out; } error = shm_copyin_path(td, uap->path_from, &path_from); if (error != 0) goto out; error = shm_copyin_path(td, uap->path_to, &path_to); if (error != 0) goto out; AUDIT_ARG_UPATH1_CANON(path_from); AUDIT_ARG_UPATH2_CANON(path_to); /* Rename with from/to equal is a no-op */ if (strcmp(path_from, path_to) == 0) goto out; fnv_from = fnv_32_str(path_from, FNV1_32_INIT); fnv_to = fnv_32_str(path_to, FNV1_32_INIT); sx_xlock(&shm_dict_lock); fd_from = shm_lookup(path_from, fnv_from); if (fd_from == NULL) { error = ENOENT; goto out_locked; } fd_to = shm_lookup(path_to, fnv_to); if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) { error = EEXIST; goto out_locked; } /* * Unconditionally prevents shm_remove from invalidating the 'from' * shm's state. */ shm_hold(fd_from); error = shm_remove(path_from, fnv_from, td->td_ucred); /* * One of my assumptions failed if ENOENT (e.g. locking didn't * protect us) */ KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s", path_from)); if (error != 0) { shm_drop(fd_from); goto out_locked; } /* * If we are exchanging, we need to ensure the shm_remove below * doesn't invalidate the dest shm's state. */ if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) shm_hold(fd_to); /* * NOTE: if path_to is not already in the hash, c'est la vie; * it simply means we have nothing already at path_to to unlink. * That is the ENOENT case. * * If we somehow don't have access to unlink this guy, but * did for the shm at path_from, then relink the shm to path_from * and abort with EACCES. * * All other errors: that is weird; let's relink and abort the * operation. */ error = shm_remove(path_to, fnv_to, td->td_ucred); if (error != 0 && error != ENOENT) { shm_insert(path_from, fnv_from, fd_from); shm_drop(fd_from); /* Don't free path_from now, since the hash references it */ path_from = NULL; goto out_locked; } error = 0; shm_insert(path_to, fnv_to, fd_from); /* Don't free path_to now, since the hash references it */ path_to = NULL; /* We kept a ref when we removed, and incremented again in insert */ shm_drop(fd_from); KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_from->shm_refs)); if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) { shm_insert(path_from, fnv_from, fd_to); path_from = NULL; shm_drop(fd_to); KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n", fd_to->shm_refs)); } out_locked: sx_xunlock(&shm_dict_lock); out: free(path_from, M_SHMFD); free(path_to, M_SHMFD); return (error); } static int shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags, vm_ooffset_t foff, struct thread *td) { struct vmspace *vms; vm_map_entry_t next_entry, prev_entry; vm_offset_t align, mask, maxaddr; int docow, error, rv, try; bool curmap; if (shmfd->shm_lp_psind == 0) return (EINVAL); /* MAP_PRIVATE is disabled */ if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL | MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0) return (EINVAL); vms = td->td_proc->p_vmspace; curmap = map == &vms->vm_map; if (curmap) { error = kern_mmap_racct_check(td, map, size); if (error != 0) return (error); } docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT; docow |= MAP_INHERIT_SHARE; if ((flags & MAP_NOCORE) != 0) docow |= MAP_DISABLE_COREDUMP; mask = pagesizes[shmfd->shm_lp_psind] - 1; if ((foff & mask) != 0) return (EINVAL); maxaddr = vm_map_max(map); if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR) maxaddr = MAP_32BIT_MAX_ADDR; if (size == 0 || (size & mask) != 0 || (*addr != 0 && ((*addr & mask) != 0 || *addr + size < *addr || *addr + size > maxaddr))) return (EINVAL); align = flags & MAP_ALIGNMENT_MASK; if (align == 0) { align = pagesizes[shmfd->shm_lp_psind]; } else if (align == MAP_ALIGNED_SUPER) { /* * MAP_ALIGNED_SUPER is only supported on superpage sizes, * i.e., [1, VM_NRESERVLEVEL]. shmfd->shm_lp_psind < 1 is * handled above. */ if ( #if VM_NRESERVLEVEL > 0 shmfd->shm_lp_psind > VM_NRESERVLEVEL #else shmfd->shm_lp_psind > 1 #endif ) return (EINVAL); align = pagesizes[shmfd->shm_lp_psind]; } else { align >>= MAP_ALIGNMENT_SHIFT; align = 1ULL << align; /* Also handles overflow. */ if (align < pagesizes[shmfd->shm_lp_psind]) return (EINVAL); } vm_map_lock(map); if ((flags & MAP_FIXED) == 0) { try = 1; if (curmap && (*addr == 0 || (*addr >= round_page((vm_offset_t)vms->vm_taddr) && *addr < round_page((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA))))) { *addr = roundup2((vm_offset_t)vms->vm_daddr + lim_max(td, RLIMIT_DATA), pagesizes[shmfd->shm_lp_psind]); } again: rv = vm_map_find_aligned(map, addr, size, maxaddr, align); if (rv != KERN_SUCCESS) { if (try == 1) { try = 2; *addr = vm_map_min(map); if ((*addr & mask) != 0) *addr = (*addr + mask) & mask; goto again; } goto fail1; } } else if ((flags & MAP_EXCL) == 0) { rv = vm_map_delete(map, *addr, *addr + size); if (rv != KERN_SUCCESS) goto fail1; } else { error = ENOSPC; if (vm_map_lookup_entry(map, *addr, &prev_entry)) goto fail; next_entry = vm_map_entry_succ(prev_entry); if (next_entry->start < *addr + size) goto fail; } rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size, prot, max_prot, docow); fail1: error = vm_mmap_to_errno(rv); fail: vm_map_unlock(map); return (error); } static int shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { struct shmfd *shmfd; vm_prot_t maxprot; int error; bool writecnt; void *rl_cookie; shmfd = fp->f_data; maxprot = VM_PROT_NONE; rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize); /* FREAD should always be set. */ if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_EXECUTE | VM_PROT_READ; /* * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared * mapping with a write seal applied. Private mappings are always * writeable. */ if ((flags & MAP_SHARED) == 0) { cap_maxprot |= VM_PROT_WRITE; maxprot |= VM_PROT_WRITE; writecnt = false; } else { if ((fp->f_flag & FWRITE) != 0 && (shmfd->shm_seals & F_SEAL_WRITE) == 0) maxprot |= VM_PROT_WRITE; /* * Any mappings from a writable descriptor may be upgraded to * VM_PROT_WRITE with mprotect(2), unless a write-seal was * applied between the open and subsequent mmap(2). We want to * reject application of a write seal as long as any such * mapping exists so that the seal cannot be trivially bypassed. */ writecnt = (maxprot & VM_PROT_WRITE) != 0; if (!writecnt && (prot & VM_PROT_WRITE) != 0) { error = EACCES; goto out; } } maxprot &= cap_maxprot; /* See comment in vn_mmap(). */ if ( #ifdef _LP64 objsize > OFF_MAX || #endif foff > OFF_MAX - objsize) { error = EINVAL; goto out; } #ifdef MAC error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags); if (error != 0) goto out; #endif mtx_lock(&shm_timestamp_lock); vfs_timestamp(&shmfd->shm_atime); mtx_unlock(&shm_timestamp_lock); vm_object_reference(shmfd->shm_object); if (shm_largepage(shmfd)) { writecnt = false; error = shm_mmap_large(shmfd, map, addr, objsize, prot, maxprot, flags, foff, td); } else { if (writecnt) { vm_pager_update_writecount(shmfd->shm_object, 0, objsize); } error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags, shmfd->shm_object, foff, writecnt, td); } if (error != 0) { if (writecnt) vm_pager_release_writecount(shmfd->shm_object, 0, objsize); vm_object_deallocate(shmfd->shm_object); } out: shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); /* * SUSv4 says that x bits of permission need not be affected. * Be consistent with our shm_open there. */ #ifdef MAC error = mac_posixshm_check_setmode(active_cred, shmfd, mode); if (error != 0) goto out; #endif error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid, VADMIN, active_cred); if (error != 0) goto out; shmfd->shm_mode = mode & ACCESSPERMS; out: mtx_unlock(&shm_timestamp_lock); return (error); } static int shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct shmfd *shmfd; int error; error = 0; shmfd = fp->f_data; mtx_lock(&shm_timestamp_lock); #ifdef MAC error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid); if (error != 0) goto out; #endif if (uid == (uid_t)-1) uid = shmfd->shm_uid; if (gid == (gid_t)-1) gid = shmfd->shm_gid; if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) || (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) && (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN))) goto out; shmfd->shm_uid = uid; shmfd->shm_gid = gid; out: mtx_unlock(&shm_timestamp_lock); return (error); } /* * Helper routines to allow the backing object of a shared memory file * descriptor to be mapped in the kernel. */ int shm_map(struct file *fp, size_t size, off_t offset, void **memp) { struct shmfd *shmfd; vm_offset_t kva, ofs; vm_object_t obj; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; obj = shmfd->shm_object; VM_OBJECT_WLOCK(obj); /* * XXXRW: This validation is probably insufficient, and subject to * sign errors. It should be fixed. */ if (offset >= shmfd->shm_size || offset + size > round_page(shmfd->shm_size)) { VM_OBJECT_WUNLOCK(obj); return (EINVAL); } shmfd->shm_kmappings++; vm_object_reference_locked(obj); VM_OBJECT_WUNLOCK(obj); /* Map the object into the kernel_map and wire it. */ kva = vm_map_min(kernel_map); ofs = offset & PAGE_MASK; offset = trunc_page(offset); size = round_page(size + ofs); rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0, VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE, VM_PROT_READ | VM_PROT_WRITE, 0); if (rv == KERN_SUCCESS) { rv = vm_map_wire(kernel_map, kva, kva + size, VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES); if (rv == KERN_SUCCESS) { *memp = (void *)(kva + ofs); return (0); } vm_map_remove(kernel_map, kva, kva + size); } else vm_object_deallocate(obj); /* On failure, drop our mapping reference. */ VM_OBJECT_WLOCK(obj); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (vm_mmap_to_errno(rv)); } /* * We require the caller to unmap the entire entry. This allows us to * safely decrement shm_kmappings when a mapping is removed. */ int shm_unmap(struct file *fp, void *mem, size_t size) { struct shmfd *shmfd; vm_map_entry_t entry; vm_offset_t kva, ofs; vm_object_t obj; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; vm_map_t map; int rv; if (fp->f_type != DTYPE_SHM) return (EINVAL); shmfd = fp->f_data; kva = (vm_offset_t)mem; ofs = kva & PAGE_MASK; kva = trunc_page(kva); size = round_page(size + ofs); map = kernel_map; rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry, &obj, &pindex, &prot, &wired); if (rv != KERN_SUCCESS) return (EINVAL); if (entry->start != kva || entry->end != kva + size) { vm_map_lookup_done(map, entry); return (EINVAL); } vm_map_lookup_done(map, entry); if (obj != shmfd->shm_object) return (EINVAL); vm_map_remove(map, kva, kva + size); VM_OBJECT_WLOCK(obj); KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped")); shmfd->shm_kmappings--; VM_OBJECT_WUNLOCK(obj); return (0); } static int shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list) { const char *path, *pr_path; size_t pr_pathlen; bool visible; sx_assert(&shm_dict_lock, SA_LOCKED); kif->kf_type = KF_TYPE_SHM; kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode; kif->kf_un.kf_file.kf_file_size = shmfd->shm_size; if (shmfd->shm_path != NULL) { path = shmfd->shm_path; pr_path = curthread->td_ucred->cr_prison->pr_path; if (strcmp(pr_path, "/") != 0) { /* Return the jail-rooted pathname. */ pr_pathlen = strlen(pr_path); visible = strncmp(path, pr_path, pr_pathlen) == 0 && path[pr_pathlen] == '/'; if (list && !visible) return (EPERM); if (visible) path += pr_pathlen; } strlcpy(kif->kf_path, path, sizeof(kif->kf_path)); } return (0); } static int shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp __unused) { int res; sx_slock(&shm_dict_lock); res = shm_fill_kinfo_locked(fp->f_data, kif, false); sx_sunlock(&shm_dict_lock); return (res); } static int shm_add_seals(struct file *fp, int seals) { struct shmfd *shmfd; void *rl_cookie; vm_ooffset_t writemappings; int error, nseals; error = 0; shmfd = fp->f_data; rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX); /* Even already-set seals should result in EPERM. */ if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) { error = EPERM; goto out; } nseals = seals & ~shmfd->shm_seals; if ((nseals & F_SEAL_WRITE) != 0) { if (shm_largepage(shmfd)) { error = ENOTSUP; goto out; } /* * The rangelock above prevents writable mappings from being * added after we've started applying seals. The RLOCK here * is to avoid torn reads on ILP32 arches as unmapping/reducing * writemappings will be done without a rangelock. */ VM_OBJECT_RLOCK(shmfd->shm_object); writemappings = shmfd->shm_object->un_pager.swp.writemappings; VM_OBJECT_RUNLOCK(shmfd->shm_object); /* kmappings are also writable */ if (writemappings > 0) { error = EBUSY; goto out; } } shmfd->shm_seals |= nseals; out: shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_get_seals(struct file *fp, int *seals) { struct shmfd *shmfd; shmfd = fp->f_data; *seals = shmfd->shm_seals; return (0); } static int shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags) { vm_object_t object; vm_pindex_t pistart, pi, piend; vm_ooffset_t off, len; int startofs, endofs, end; int error; off = *offset; len = *length; KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows")); if (off + len > shmfd->shm_size) len = shmfd->shm_size - off; object = shmfd->shm_object; startofs = off & PAGE_MASK; endofs = (off + len) & PAGE_MASK; pistart = OFF_TO_IDX(off); piend = OFF_TO_IDX(off + len); pi = OFF_TO_IDX(off + PAGE_MASK); error = 0; /* Handle the case when offset is on or beyond shm size. */ if ((off_t)len <= 0) { *length = 0; return (0); } VM_OBJECT_WLOCK(object); if (startofs != 0) { end = pistart != piend ? PAGE_SIZE : endofs; error = shm_partial_page_invalidate(object, pistart, startofs, end); if (error) goto out; off += end - startofs; len -= end - startofs; } if (pi < piend) { vm_object_page_remove(object, pi, piend, 0); off += IDX_TO_OFF(piend - pi); len -= IDX_TO_OFF(piend - pi); } if (endofs != 0 && pistart != piend) { error = shm_partial_page_invalidate(object, piend, 0, endofs); if (error) goto out; off += endofs; len -= endofs; } out: VM_OBJECT_WUNLOCK(shmfd->shm_object); *offset = off; *length = len; return (error); } static int shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, struct ucred *active_cred, struct thread *td) { void *rl_cookie; struct shmfd *shmfd; off_t off, len; int error; KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd")); KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, ("shm_fspacectl: non-zero flags")); KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, ("shm_fspacectl: offset/length overflow or underflow")); error = EINVAL; shmfd = fp->f_data; off = *offset; len = *length; rl_cookie = shm_rangelock_wlock(shmfd, off, off + len); switch (cmd) { case SPACECTL_DEALLOC: if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) { error = EPERM; break; } error = shm_deallocate(shmfd, &off, &len, flags); *offset = off; *length = len; break; default: __assert_unreachable(); } shm_rangelock_unlock(shmfd, rl_cookie); return (error); } static int shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) { void *rl_cookie; struct shmfd *shmfd; size_t size; int error; /* This assumes that the caller already checked for overflow. */ error = 0; shmfd = fp->f_data; size = offset + len; /* * Just grab the rangelock for the range that we may be attempting to * grow, rather than blocking read/write for regions we won't be * touching while this (potential) resize is in progress. Other * attempts to resize the shmfd will have to take a write lock from 0 to * OFF_MAX, so this being potentially beyond the current usable range of * the shmfd is not necessarily a concern. If other mechanisms are * added to grow a shmfd, this may need to be re-evaluated. */ rl_cookie = shm_rangelock_wlock(shmfd, offset, size); if (size > shmfd->shm_size) error = shm_dotruncate_cookie(shmfd, size, rl_cookie); shm_rangelock_unlock(shmfd, rl_cookie); /* Translate to posix_fallocate(2) return value as needed. */ if (error == ENOMEM) error = ENOSPC; return (error); } static int sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS) { struct shm_mapping *shmm; struct sbuf sb; struct kinfo_file kif; u_long i; int error, error2; sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req); sbuf_clear_flags(&sb, SBUF_INCLUDENUL); error = 0; sx_slock(&shm_dict_lock); for (i = 0; i < shm_hash + 1; i++) { LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) { error = shm_fill_kinfo_locked(shmm->sm_shmfd, &kif, true); if (error == EPERM) { error = 0; continue; } if (error != 0) break; pack_kinfo(&kif); error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ? 0 : ENOMEM; if (error != 0) break; } } sx_sunlock(&shm_dict_lock); error2 = sbuf_finish(&sb); sbuf_delete(&sb); return (error != 0 ? error : error2); } SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list, CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE, NULL, 0, sysctl_posix_shm_list, "", "POSIX SHM list"); int kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode, struct filecaps *caps) { return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL)); } /* * This version of the shm_open() interface leaves CLOEXEC behavior up to the * caller, and libc will enforce it for the traditional shm_open() call. This * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This * interface also includes a 'name' argument that is currently unused, but could * potentially be exported later via some interface for debugging purposes. * From the kernel's perspective, it is optional. Individual consumers like * memfd_create() may require it in order to be compatible with other systems * implementing the same function. */ int sys_shm_open2(struct thread *td, struct shm_open2_args *uap) { return (kern_shm_open2(td, uap->path, uap->flags, uap->mode, uap->shmflags, NULL, uap->name)); } diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c index 8012fab29081..f192c6798858 100644 --- a/sys/kern/vfs_subr.c +++ b/sys/kern/vfs_subr.c @@ -1,7395 +1,7393 @@ /*- * 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. */ /* * External virtual filesystem routines */ #include #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 #include #include #include #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS)) #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS #endif #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, bool isvnlru); static void v_init_counters(struct vnode *); static void vn_seqc_init(struct vnode *); static void vn_seqc_write_end_free(struct vnode *vp); static void vgonel(struct vnode *); static bool vhold_recycle_free(struct vnode *); static void vdropl_recycle(struct vnode *vp); static void vdrop_recycle(struct vnode *vp); 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); static SYSCTL_NODE(_vfs, OID_AUTO, vnode, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "vnode configuration and statistics"); static SYSCTL_NODE(_vfs_vnode, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "vnode configuration"); static SYSCTL_NODE(_vfs_vnode, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "vnode statistics"); static SYSCTL_NODE(_vfs_vnode, OID_AUTO, vnlru, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "vnode recycling"); /* * 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 (legacy)"); SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, count, 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 (legacy)"); SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, created, CTLFLAG_RD, &vnodes_created, "Number of vnodes created by getnewvnode"); /* * Conversion tables for conversion from vnode types to inode formats * and back. */ __enum_uint8(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; static long freevnodes_old; static u_long recycles_count; SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0, "Number of vnodes recycled to meet vnode cache targets (legacy)"); SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0, "Number of vnodes recycled to meet vnode cache targets"); static u_long recycles_free_count; SYSCTL_ULONG(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, &recycles_free_count, 0, "Number of free vnodes recycled to meet vnode cache targets (legacy)"); SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS, &recycles_free_count, 0, "Number of free vnodes recycled to meet vnode cache targets"); static counter_u64_t direct_recycles_free_count; SYSCTL_COUNTER_U64(_vfs_vnode_vnlru, OID_AUTO, direct_recycles_free, CTLFLAG_RD, &direct_recycles_free_count, "Number of free vnodes recycled by vn_alloc callers to meet vnode cache targets"); static counter_u64_t vnode_skipped_requeues; SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues, "Number of times LRU requeue was skipped due to lock contention"); static __read_mostly bool vnode_can_skip_requeue; SYSCTL_BOOL(_vfs_vnode_param, OID_AUTO, can_skip_requeue, CTLFLAG_RW, &vnode_can_skip_requeue, 0, "Is LRU requeue skippable"); static u_long deferred_inact; SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, 0, "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; 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 bool 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 (legacy)"); SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, "LU", "Target for maximum number of vnodes"); static int sysctl_freevnodes(SYSCTL_HANDLER_ARGS) { u_long rfreevnodes; rfreevnodes = vnlru_read_freevnodes(); return (sysctl_handle_long(oidp, &rfreevnodes, 0, req)); } SYSCTL_PROC(_vfs, OID_AUTO, freevnodes, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, "LU", "Number of \"free\" vnodes (legacy)"); SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes, "LU", "Number of \"free\" 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 (legacy)"); SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, "LU", "Target for minimum number of \"free\" vnodes"); static int vnlru_nowhere; SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS, &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; NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf); 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; } vgone(vp); putvnode: vput(vp); NDFREE_PNBUF(&nd); 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; 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. */ #define vnsz2log 8 #ifndef DEBUG_LOCKS _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log && sizeof(struct vnode) < 1UL << (vnsz2log + 1), "vnsz2log needs to be updated"); #endif /* * 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); } #ifdef KASAN static int vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused) { kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0); return (0); } static void vnode_dtor(void *mem, int size, void *arg __unused) { size_t end1, end2, off1, off2; _Static_assert(offsetof(struct vnode, v_vnodelist) < offsetof(struct vnode, v_dbatchcpu), "KASAN marks require updating"); off1 = offsetof(struct vnode, v_vnodelist); off2 = offsetof(struct vnode, v_dbatchcpu); end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist); end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu); /* * Access to the v_vnodelist and v_dbatchcpu fields are permitted even * after the vnode has been freed. Try to get some KASAN coverage by * marking everything except those two fields as invalid. Because * KASAN's tracking is not byte-granular, any preceding fields sharing * the same 8-byte aligned word must also be marked valid. */ /* Handle the area from the start until v_vnodelist... */ off1 = rounddown2(off1, KASAN_SHADOW_SCALE); kasan_mark(mem, off1, off1, KASAN_UMA_FREED); /* ... then the area between v_vnodelist and v_dbatchcpu ... */ off1 = roundup2(end1, KASAN_SHADOW_SCALE); off2 = rounddown2(off2, KASAN_SHADOW_SCALE); if (off2 > off1) kasan_mark((void *)((char *)mem + off1), off2 - off1, off2 - off1, KASAN_UMA_FREED); /* ... and finally the area from v_dbatchcpu to the end. */ off2 = roundup2(end2, KASAN_SHADOW_SCALE); kasan_mark((void *)((char *)mem + off2), size - off2, size - off2, KASAN_UMA_FREED); } #endif /* KASAN */ /* * 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; vp->v_state = VSTATE_DEAD; /* * 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)); kasan_mark(mem, size, size, 0); } /* * 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; uma_ctor ctor; uma_dtor dtor; int cpu, physvnodes, virtvnodes; /* * 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); #ifdef KASAN ctor = vnode_ctor; dtor = vnode_dtor; #else ctor = NULL; dtor = NULL; #endif vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor, vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN); 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); direct_recycles_free_count = counter_u64_alloc(M_WAITOK); vnode_skipped_requeues = 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"); 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: * 1. VOP_LOOKUP() obtains B while A is held * 2. vfs_busy() obtains a shared lock on F while A and B are held * 3. vput() releases lock on B * 4. vput() releases lock on A * 5. VFS_ROOT() obtains lock on D while shared lock on F is held * 6. vfs_unbusy() releases shared lock on F * 7. 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. Note that for stacked * filesystems, D and B in the example above may be the same lock, * which introdues potential lock order reversal deadlock between * dounmount() and step 5 above. These filesystems may avoid the LOR * by setting VV_CROSSLOCK on the covered vnode so that lock B will * remain held until after step 5. */ 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, 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(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 the mount it tried to busy is no longer valid. */ while (mp->mnt_kern_flag & MNTK_UNMOUNT) { KASSERT(TAILQ_EMPTY(&mp->mnt_uppers), ("%s: non-empty upper mount list with pending unmount", __func__)); 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, mpcpu)) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); 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 * for more free vnodes, not reduce their count. */ 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_recycle(vp); goto next_iter_unlocked; } if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { vdrop_recycle(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->handle == vp && vp->v_object->resident_page_count > trigger)) { VOP_UNLOCK(vp); vdropl_recycle(vp); vn_finished_write(mp); goto next_iter_unlocked; } recycles_count++; vgonel(vp); VOP_UNLOCK(vp); vdropl_recycle(vp); vn_finished_write(mp); done++; next_iter_unlocked: maybe_yield(); 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_free_per_call = 10000; SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0, "limit on vnode free requests per call to the vnlru_free routine (legacy)"); SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW, &max_free_per_call, 0, "limit on vnode free requests per call to the vnlru_free routine"); /* * Attempt to recycle requested amount of free vnodes. */ static int vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru) { struct vnode *vp; struct mount *mp; int ocount; bool retried; mtx_assert(&vnode_list_mtx, MA_OWNED); if (count > max_free_per_call) count = max_free_per_call; if (count == 0) { mtx_unlock(&vnode_list_mtx); return (0); } ocount = count; retried = false; vp = mvp; for (;;) { vp = TAILQ_NEXT(vp, v_vnodelist); if (__predict_false(vp == NULL)) { /* * The free vnode marker can be past eligible vnodes: * 1. if vdbatch_process trylock failed * 2. if vtryrecycle failed * * If so, start the scan from scratch. */ if (!retried && vnlru_read_freevnodes() > 0) { TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); vp = mvp; retried = true; continue; } /* * Give up */ TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); 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); /* * FIXME: ignores the return value, meaning it may be nothing * got recycled but it claims otherwise to the caller. * * Originally the value started being ignored in 2005 with * 114a1006a8204aa156e1f9ad6476cdff89cada7f . * * Respecting the value can run into significant stalls if most * vnodes belong to one file system and it has writes * suspended. In presence of many threads and millions of * vnodes they keep contending on the vnode_list_mtx lock only * to find vnodes they can't recycle. * * The solution would be to pre-check if the vnode is likely to * be recycle-able, but it needs to happen with the * vnode_list_mtx lock held. This runs into a problem where * VOP_GETWRITEMOUNT (currently needed to find out about if * writes are frozen) can take locks which LOR against it. * * Check nullfs for one example (null_getwritemount). */ vtryrecycle(vp, isvnlru); count--; if (count == 0) { break; } mtx_lock(&vnode_list_mtx); vp = mvp; } mtx_assert(&vnode_list_mtx, MA_NOTOWNED); return (ocount - count); } /* * XXX: returns without vnode_list_mtx locked! */ static int vnlru_free_locked_direct(int count) { int ret; mtx_assert(&vnode_list_mtx, MA_OWNED); ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false); mtx_assert(&vnode_list_mtx, MA_NOTOWNED); return (ret); } static int vnlru_free_locked_vnlru(int count) { int ret; mtx_assert(&vnode_list_mtx, MA_OWNED); ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true); mtx_assert(&vnode_list_mtx, MA_NOTOWNED); return (ret); } static int vnlru_free_vnlru(int count) { mtx_lock(&vnode_list_mtx); return (vnlru_free_locked_vnlru(count)); } void vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp) { MPASS(mnt_op != NULL); MPASS(mvp != NULL); VNPASS(mvp->v_type == VMARKER, mvp); mtx_lock(&vnode_list_mtx); vnlru_free_impl(count, mnt_op, mvp, true); mtx_assert(&vnode_list_mtx, MA_NOTOWNED); } struct vnode * vnlru_alloc_marker(void) { struct vnode *mvp; mvp = vn_alloc_marker(NULL); mtx_lock(&vnode_list_mtx); TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); return (mvp); } void vnlru_free_marker(struct vnode *mvp) { mtx_lock(&vnode_list_mtx); TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); vn_free_marker(mvp); } 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; static u_long vnlruproc_kicks; SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0, "Number of times vnlru awakened due to vnode shortage"); #define VNLRU_COUNT_SLOP 100 /* * The main freevnodes counter is only updated when a counter local to CPU * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally * walked to compute a more accurate total. * * 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 126 static void __noinline vfs_freevnodes_rollup(int8_t *lfreevnodes) { atomic_add_long(&freevnodes, *lfreevnodes); *lfreevnodes = 0; critical_exit(); } static __inline void vfs_freevnodes_inc(void) { int8_t *lfreevnodes; critical_enter(); lfreevnodes = PCPU_PTR(vfs_freevnodes); (*lfreevnodes)++; if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP)) vfs_freevnodes_rollup(lfreevnodes); else critical_exit(); } static __inline void vfs_freevnodes_dec(void) { int8_t *lfreevnodes; critical_enter(); lfreevnodes = PCPU_PTR(vfs_freevnodes); (*lfreevnodes)--; if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP)) vfs_freevnodes_rollup(lfreevnodes); else critical_exit(); } static u_long vnlru_read_freevnodes(void) { long slop, rfreevnodes, rfreevnodes_old; int cpu; rfreevnodes = atomic_load_long(&freevnodes); rfreevnodes_old = atomic_load_long(&freevnodes_old); if (rfreevnodes > rfreevnodes_old) slop = rfreevnodes - rfreevnodes_old; else slop = rfreevnodes_old - rfreevnodes; if (slop < VNLRU_FREEVNODES_SLOP) return (rfreevnodes >= 0 ? rfreevnodes : 0); CPU_FOREACH(cpu) { rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes; } atomic_store_long(&freevnodes_old, rfreevnodes); 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 void vnlru_kick_locked(void) { mtx_assert(&vnode_list_mtx, MA_OWNED); if (vnlruproc_sig == 0) { vnlruproc_sig = 1; vnlruproc_kicks++; wakeup(vnlruproc); } } static void vnlru_kick_cond(void) { if (vnlru_read_freevnodes() > wantfreevnodes) return; if (vnlruproc_sig) return; mtx_lock(&vnode_list_mtx); vnlru_kick_locked(); mtx_unlock(&vnode_list_mtx); } static void vnlru_proc_sleep(void) { if (vnlruproc_sig) { vnlruproc_sig = 0; wakeup(&vnlruproc_sig); } msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); } /* * A lighter version of the machinery below. * * Tries to reach goals only by recycling free vnodes and does not invoke * uma_reclaim(UMA_RECLAIM_DRAIN). * * This works around pathological behavior in vnlru in presence of tons of free * vnodes, but without having to rewrite the machinery at this time. Said * behavior boils down to continuously trying to reclaim all kinds of vnodes * (cycling through all levels of "force") when the count is transiently above * limit. This happens a lot when all vnodes are used up and vn_alloc * speculatively increments the counter. * * Sample testcase: vnode limit 8388608, 20 separate directory trees each with * 1 million files in total and 20 find(1) processes stating them in parallel * (one per each tree). * * On a kernel with only stock machinery this needs anywhere between 60 and 120 * seconds to execute (time varies *wildly* between runs). With the workaround * it consistently stays around 20 seconds [it got further down with later * changes]. * * That is to say the entire thing needs a fundamental redesign (most notably * to accommodate faster recycling), the above only tries to get it ouf the way. * * Return values are: * -1 -- fallback to regular vnlru loop * 0 -- do nothing, go to sleep * >0 -- recycle this many vnodes */ static long vnlru_proc_light_pick(void) { u_long rnumvnodes, rfreevnodes; if (vstir || vnlruproc_sig == 1) return (-1); rnumvnodes = atomic_load_long(&numvnodes); rfreevnodes = vnlru_read_freevnodes(); /* * vnode limit might have changed and now we may be at a significant * excess. Bail if we can't sort it out with free vnodes. * * Due to atomic updates the count can legitimately go above * the limit for a short period, don't bother doing anything in * that case. */ if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) { if (rnumvnodes - rfreevnodes >= desiredvnodes || rfreevnodes <= wantfreevnodes) { return (-1); } return (rnumvnodes - desiredvnodes); } /* * Don't try to reach wantfreevnodes target if there are too few vnodes * to begin with. */ if (rnumvnodes < wantfreevnodes) { return (0); } if (rfreevnodes < wantfreevnodes) { return (-1); } return (0); } static bool vnlru_proc_light(void) { long freecount; mtx_assert(&vnode_list_mtx, MA_NOTOWNED); freecount = vnlru_proc_light_pick(); if (freecount == -1) return (false); if (freecount != 0) { vnlru_free_vnlru(freecount); } mtx_lock(&vnode_list_mtx); vnlru_proc_sleep(); mtx_assert(&vnode_list_mtx, MA_NOTOWNED); return (true); } static u_long uma_reclaim_calls; SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS, &uma_reclaim_calls, 0, "Number of calls to uma_reclaim"); 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); if (force == 0 && vnlru_proc_light()) continue; 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 free vnodes. */ if (rnumvnodes > desiredvnodes + 10) { vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes); mtx_lock(&vnode_list_mtx); 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 free vnode count). Otherwise, try to reclaim * space for a 10% expansion. */ if (vstir && force == 0) { force = 1; vstir = false; } if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { vnlru_proc_sleep(); 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 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); /* * Total number of vnodes can transiently go slightly above the * limit (see vn_alloc_hard), no need to call uma_reclaim if * this happens. */ if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes && numvnodes <= desiredvnodes) { uma_reclaim_calls++; 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. */ static int vtryrecycle(struct vnode *vp, bool isvnlru) { struct mount *vnmp; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); VNPASS(vp->v_holdcnt > 0, 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_recycle(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_recycle(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_recycle(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)) { if (isvnlru) recycles_free_count++; else counter_u64_add(direct_recycles_free_count, 1); vgonel(vp); } VOP_UNLOCK(vp); vdropl_recycle(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 u_long vn_alloc_sleeps; SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0, "Number of times vnode allocation blocked waiting on vnlru"); static struct vnode * __noinline vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped) { u_long rfreevnodes; if (bumped) { if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) { atomic_subtract_long(&numvnodes, 1); bumped = false; } } mtx_lock(&vnode_list_mtx); if (vn_alloc_cyclecount != 0) { rnumvnodes = atomic_load_long(&numvnodes); if (rnumvnodes + 1 < desiredvnodes) { vn_alloc_cyclecount = 0; mtx_unlock(&vnode_list_mtx); goto alloc; } rfreevnodes = vnlru_read_freevnodes(); if (rfreevnodes < wantfreevnodes) { if (vn_alloc_cyclecount++ >= rfreevnodes) { vn_alloc_cyclecount = 0; vstir = true; } } else { vn_alloc_cyclecount = 0; } } /* * Grow the vnode cache if it will not be above its target max after * growing. Otherwise, if there is at least one free vnode, try to * reclaim 1 item from it before growing the cache (possibly above its * target max if the reclamation failed or is delayed). */ if (vnlru_free_locked_direct(1) > 0) goto alloc; mtx_assert(&vnode_list_mtx, MA_NOTOWNED); if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { /* * Wait for space for a new vnode. */ if (bumped) { atomic_subtract_long(&numvnodes, 1); bumped = false; } mtx_lock(&vnode_list_mtx); vnlru_kick_locked(); vn_alloc_sleeps++; msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && vnlru_read_freevnodes() > 1) vnlru_free_locked_direct(1); else mtx_unlock(&vnode_list_mtx); } alloc: mtx_assert(&vnode_list_mtx, MA_NOTOWNED); if (!bumped) atomic_add_long(&numvnodes, 1); vnlru_kick_cond(); 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, 0, false)); rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; if (__predict_false(vnlru_under(rnumvnodes, vlowat))) { return (vn_alloc_hard(mp, rnumvnodes, true)); } 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); } /* * Allocate a new vnode. */ 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)); cache_validate_vop_vector(mp, 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); vn_set_state(vp, VSTATE_UNINITIALIZED); /* * 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; vp->v_irflag = 0; v_init_counters(vp); vn_seqc_init(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; } /* * 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_free(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_rl.rl_waiters), vp, - ("Dangling rangelock waiters")); VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp, ("Leaked inactivation")); VI_UNLOCK(vp); cache_assert_no_entries(vp); #ifdef MAC mac_vnode_destroy(vp); #endif if (vp->v_pollinfo != NULL) { /* * Use LK_NOWAIT to shut up witness about the lock. We may get * here while having another vnode locked when trying to * satisfy a lookup and needing to recycle. */ VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT); destroy_vpollinfo(vp->v_pollinfo); VOP_UNLOCK(vp); vp->v_pollinfo = NULL; } vp->v_mountedhere = NULL; vp->v_unpcb = NULL; vp->v_rdev = NULL; vp->v_fifoinfo = NULL; 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; MNT_ILOCK(mp); VI_LOCK(vp); vp->v_mount = NULL; 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); /* * The caller expects the interlock to be still held. */ ASSERT_VI_LOCKED(vp, __func__); } static int insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr) { KASSERT(vp->v_mount == NULL, ("insmntque: vnode already on per mount vnode list")); VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) { ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); } else { KASSERT(!dtr, ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup", __func__)); } /* * 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) { vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } 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); } /* * Insert into list of vnodes for the new mount point, if available. * insmntque() reclaims the vnode on insertion failure, insmntque1() * leaves handling of the vnode to the caller. */ int insmntque(struct vnode *vp, struct mount *mp) { return (insmntque1_int(vp, mp, true)); } int insmntque1(struct vnode *vp, struct mount *mp) { return (insmntque1_int(vp, mp, false)); } /* * 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); do { error = BO_SYNC(bo, MNT_WAIT); } while (error == ERELOOKUP); if (error != 0) return (error); BO_LOCK(bo); if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) { BO_UNLOCK(bo); return (EBUSY); } } } /* * 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) { /* * Write out vnode metadata, e.g. indirect blocks. */ 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. Return zero on * success, EEXIST if a buffer with this identity already exists, or another * error on allocation failure. */ static inline int buf_vlist_find_or_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)) == xflags, ("buf_vlist_add: b_xflags %#x not set on bp %p", xflags, bp)); if (xflags & BX_VNDIRTY) bv = &bo->bo_dirty; else bv = &bo->bo_clean; error = BUF_PCTRIE_INSERT_LOOKUP_LE(&bv->bv_root, bp, &n); if (n == NULL) { KASSERT(error != EEXIST, ("buf_vlist_add: EEXIST but no existing buf found: bp %p", bp)); } else { KASSERT((uint64_t)n->b_lblkno <= (uint64_t)bp->b_lblkno, ("buf_vlist_add: out of order insert/lookup: bp %p n %p", bp, n)); KASSERT((n->b_lblkno == bp->b_lblkno) == (error == EEXIST), ("buf_vlist_add: inconsistent result for existing buf: " "error %d bp %p n %p", error, bp, n)); } if (error != 0) return (error); /* Keep the list ordered. */ if (n == NULL) { KASSERT(TAILQ_EMPTY(&bv->bv_hd) || (uint64_t)bp->b_lblkno < (uint64_t)TAILQ_FIRST(&bv->bv_hd)->b_lblkno, ("buf_vlist_add: queue order: " "%p should be before first %p", bp, TAILQ_FIRST(&bv->bv_hd))); TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); } else { KASSERT(TAILQ_NEXT(n, b_bobufs) == NULL || (uint64_t)bp->b_lblkno < (uint64_t)TAILQ_NEXT(n, b_bobufs)->b_lblkno, ("buf_vlist_add: queue order: " "%p should be before next %p", bp, TAILQ_NEXT(n, b_bobufs))); TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); } bv->bv_cnt++; return (0); } /* * 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) { int error; 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; error = buf_vlist_find_or_add(bp, bo, xflags); if (error) panic("buf_vlist_add: error=%d", error); } /* * 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. */ int bgetvp(struct vnode *vp, struct buf *bp) { struct bufobj *bo; int error; bo = &vp->v_bufobj; ASSERT_BO_UNLOCKED(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)); /* * Add the buf to the vnode's clean list unless we lost a race and find * an existing buf in either dirty or clean. */ bp->b_vp = vp; bp->b_bufobj = bo; bp->b_xflags |= BX_VNCLEAN; error = EEXIST; BO_LOCK(bo); if (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, bp->b_lblkno) == NULL) error = buf_vlist_find_or_add(bp, bo, BX_VNCLEAN); BO_UNLOCK(bo); if (__predict_true(error == 0)) { vhold(vp); return (0); } if (error != EEXIST) panic("bgetvp: buf_vlist_add error: %d", error); bp->b_vp = NULL; bp->b_bufobj = NULL; bp->b_xflags &= ~BX_VNCLEAN; return (error); } /* * 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)); } MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 || (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp, ("suspended mp syncing vp %p", vp)); 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); if ((bo->bo_flag & BO_NONSTERILE) == 0) { /* * Coordinate with getblk's unlocked lookup. Make * BO_NONSTERILE visible before the first reassignbuf produces * any side effect. This could be outside the bo lock if we * used a separate atomic flag field. */ bo->bo_flag |= BO_NONSTERILE; atomic_thread_fence_rel(); } 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); } /* * Get a usecount on a vnode. * * vget and vget_finish may fail to lock the vnode if they lose a race against * it being doomed. LK_RETRY can be passed in flags to lock it anyway. * * 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) { int old __diagused; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); old = refcount_acquire(&vp->v_usecount); VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); } 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); } static void vunlazy(struct vnode *vp) { struct mount *mp; ASSERT_VI_LOCKED(vp, __func__); VNPASS(!VN_IS_DOOMED(vp), vp); mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); /* * 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); } /* * 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__); VNPASS(vp->v_holdcnt > 0, vp); if (VN_IS_DOOMED(vp)) { vdropl(vp); return; } if (vp->v_iflag & VI_DEFINACT) { VNPASS(vp->v_holdcnt > 1, vp); 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); atomic_add_long(&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) { if (func == VUNREF) { VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp, ("recursive vunref")); vp->v_vflag |= VV_UNREF; } for (;;) { error = vinactive(vp); if (want_unlock) VOP_UNLOCK(vp); if (error != ERELOOKUP || !want_unlock) break; VOP_LOCK(vp, LK_EXCLUSIVE); } if (func == VUNREF) vp->v_vflag &= ~VV_UNREF; 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) vfs_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) vfs_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)) { vfs_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); /* * Attempt to requeue the passed batch, but give up easily. * * Despite batching the mechanism is prone to transient *significant* * lock contention, where vnode_list_mtx becomes the primary bottleneck * if multiple CPUs get here (one real-world example is highly parallel * do-nothing make , which will stat *tons* of vnodes). Since it is * quasi-LRU (read: not that great even if fully honoured) provide an * option to just dodge the problem. Parties which don't like it are * welcome to implement something better. */ if (vnode_can_skip_requeue) { if (!mtx_trylock(&vnode_list_mtx)) { counter_u64_add(vnode_skipped_requeues, 1); critical_enter(); for (i = 0; i < VDBATCH_SIZE; i++) { vp = vd->tab[i]; vd->tab[i] = NULL; MPASS(vp->v_dbatchcpu != NOCPU); vp->v_dbatchcpu = NOCPU; } vd->index = 0; critical_exit(); return; } /* fallthrough to locked processing */ } else { mtx_lock(&vnode_list_mtx); } mtx_assert(&vnode_list_mtx, MA_OWNED); critical_enter(); for (i = 0; i < VDBATCH_SIZE; i++) { vp = vd->tab[i]; vd->tab[i] = NULL; 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->index = 0; critical_exit(); } static void vdbatch_enqueue(struct vnode *vp) { struct vdbatch *vd; ASSERT_VI_LOCKED(vp, __func__); VNPASS(!VN_IS_DOOMED(vp), vp); 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; VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp); 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. * * It will only get freed if this is the last hold *and* it has been vgone'd. * * 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 __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))) { vfs_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); } static __always_inline void vdropl_impl(struct vnode *vp, bool enqueue) { ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (!refcount_release(&vp->v_holdcnt)) { VI_UNLOCK(vp); return; } VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp); VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); if (VN_IS_DOOMED(vp)) { vdropl_final(vp); return; } vfs_freevnodes_inc(); if (vp->v_mflag & VMP_LAZYLIST) { vunlazy(vp); } if (!enqueue) { VI_UNLOCK(vp); return; } /* * Also unlocks the interlock. We can't assert on it as we * released our hold and by now the vnode might have been * freed. */ vdbatch_enqueue(vp); } void vdropl(struct vnode *vp) { vdropl_impl(vp, true); } /* * vdrop a vnode when recycling * * This is a special case routine only to be used when recycling, differs from * regular vdrop by not requeieing the vnode on LRU. * * Consider a case where vtryrecycle continuously fails with all vnodes (due to * e.g., frozen writes on the filesystem), filling the batch and causing it to * be requeued. Then vnlru will end up revisiting the same vnodes. This is a * loop which can last for as long as writes are frozen. */ static void vdropl_recycle(struct vnode *vp) { vdropl_impl(vp, false); } static void vdrop_recycle(struct vnode *vp) { VI_LOCK(vp); vdropl_recycle(vp); } /* * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT * flags. DOINGINACT prevents us from recursing in calls to vinactive. */ static int vinactivef(struct vnode *vp) { int error; ASSERT_VOP_ELOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp); 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 ((vp->v_vflag & VV_NOSYNC) == 0) vnode_pager_clean_async(vp); error = VOP_INACTIVE(vp); VI_LOCK(vp); VNPASS(vp->v_iflag & VI_DOINGINACT, vp); vp->v_iflag &= ~VI_DOINGINACT; return (error); } int 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 (0); if (vp->v_iflag & VI_DOINGINACT) return (0); if (vp->v_usecount > 0) { vp->v_iflag &= ~VI_OWEINACT; return (0); } 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) { vnode_pager_clean_async(vp); do { error = VOP_FSYNC(vp, MNT_WAIT, td); } while (error == ERELOOKUP); 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. */ 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); } /* * Notify upper mounts about reclaimed or unlinked vnode. */ void vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event) { struct mount *mp; struct mount_upper_node *ump; mp = atomic_load_ptr(&vp->v_mount); if (mp == NULL) return; if (TAILQ_EMPTY(&mp->mnt_notify)) return; MNT_ILOCK(mp); mp->mnt_upper_pending++; KASSERT(mp->mnt_upper_pending > 0, ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending)); TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) { MNT_IUNLOCK(mp); switch (event) { case VFS_NOTIFY_UPPER_RECLAIM: VFS_RECLAIM_LOWERVP(ump->mp, vp); break; case VFS_NOTIFY_UPPER_UNLINK: VFS_UNLINK_LOWERVP(ump->mp, vp); break; } MNT_ILOCK(mp); } mp->mnt_upper_pending--; if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 && mp->mnt_upper_pending == 0) { mp->mnt_kern_flag &= ~MNTK_UPPER_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 (VN_IS_DOOMED(vp)) { VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \ vn_get_state(vp) == VSTATE_DEAD, vp); return; } /* * Paired with freevnode. */ vn_seqc_write_begin_locked(vp); vunlazy_gone(vp); vn_irflag_set_locked(vp, VIRF_DOOMED); vn_set_state(vp, VSTATE_DESTROYING); /* * 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 (!doinginact) { do { if (oweinact || active) { VI_LOCK(vp); vinactivef(vp); oweinact = (vp->v_iflag & VI_OWEINACT) != 0; VI_UNLOCK(vp); } } while (oweinact); } 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. */ if (vp->v_lockf != NULL) { (void)VOP_ADVLOCKPURGE(vp); vp->v_lockf = NULL; } /* * Delete from old mount point vnode list. */ if (vp->v_mount == NULL) { VI_LOCK(vp); } else { delmntque(vp); ASSERT_VI_LOCKED(vp, "vgonel 2"); } /* * Done with purge, reset to the standard lock and invalidate * the vnode. */ vp->v_vnlock = &vp->v_lock; vp->v_op = &dead_vnodeops; vp->v_type = VBAD; vn_set_state(vp, VSTATE_DEAD); } /* * Print out a description of a vnode. */ static const char *const vtypename[] = { [VNON] = "VNON", [VREG] = "VREG", [VDIR] = "VDIR", [VBLK] = "VBLK", [VCHR] = "VCHR", [VLNK] = "VLNK", [VSOCK] = "VSOCK", [VFIFO] = "VFIFO", [VBAD] = "VBAD", [VMARKER] = "VMARKER", }; _Static_assert(nitems(vtypename) == VLASTTYPE + 1, "vnode type name not added to vtypename"); static const char *const vstatename[] = { [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED", [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED", [VSTATE_DESTROYING] = "VSTATE_DESTROYING", [VSTATE_DEAD] = "VSTATE_DEAD", }; _Static_assert(nitems(vstatename) == VLASTSTATE + 1, "vnode state name not added to vstatename"); _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; short irflag; va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf("%p: ", (void *)vp); printf("type %s state %s op %p\n", vtypename[vp->v_type], vstatename[vp->v_state], vp->v_op); 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'; irflag = vn_irflag_read(vp); if (irflag & VIRF_DOOMED) strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); if (irflag & VIRF_PGREAD) strlcat(buf, "|VIRF_PGREAD", sizeof(buf)); if (irflag & VIRF_MOUNTPOINT) strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf)); if (irflag & VIRF_TEXT_REF) strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf)); flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF); 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_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_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); strlcat(buf, buf2, 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)); if (vp->v_iflag & VI_FOPENING) strlcat(buf, "|VI_FOPENING", sizeof(buf)); flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT | VI_OWEINACT | VI_DEFINACT | VI_FOPENING); 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)", buf + 1); if (mtx_owned(VI_MTX(vp))) printf(" VI_LOCKed"); printf("\n"); 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_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE) { 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_NOMSYNC); 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_RECURSE); MNT_KERN_FLAG(MNTK_UPPER_WAITER); MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE); MNT_KERN_FLAG(MNTK_USES_BCACHE); MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG); MNT_KERN_FLAG(MNTK_FPLOOKUP); MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER); 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_NULL_NOCACHE); MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); #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_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 */ 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__); VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp); 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; int lkflags; bool seen_defer; lkflags = LK_EXCLUSIVE | LK_INTERLOCK; if (flags != MNT_WAIT) lkflags |= LK_NOWAIT; 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) { if ((vp->v_vflag & VV_NOSYNC) == 0) { if (flags == MNT_WAIT) vnode_pager_clean_sync(vp); else vnode_pager_clean_async(vp); } 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, .vop_fsync = sync_fsync, .vop_getwritemount = vop_stdgetwritemount, .vop_inactive = sync_inactive, .vop_need_inactive = vop_stdneed_inactive, .vop_reclaim = sync_reclaim, .vop_lock1 = vop_stdlock, .vop_unlock = vop_stdunlock, .vop_islocked = vop_stdislocked, .vop_fplookup_vexec = VOP_EAGAIN, .vop_fplookup_symlink = VOP_EAGAIN, }; 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 = insmntque1(vp, mp); if (error != 0) panic("vfs_allocate_syncvnode: insmntque() failed"); vp->v_vflag &= ~VV_FORCEINSMQ; vn_set_state(vp, VSTATE_CONSTRUCTED); 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); VOP_UNLOCK(syncvp); 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_lock(syncvp, LK_EXCLUSIVE | LK_RETRY); 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; obj = vp->v_object; return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && vm_object_mightbedirty(obj)); } /* * 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_uint8(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 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) { if (KERNEL_PANICKED() || vp == NULL) return; #ifdef WITNESS if ((vp->v_irflag & VIRF_CROSSMP) == 0 && witness_is_owned(&vp->v_vnlock->lock_object) == -1) #else int locked = VOP_ISLOCKED(vp); if (locked == 0 || locked == LK_EXCLOTHER) #endif vfs_badlock("is not locked but should be", str, vp); } void assert_vop_unlocked(struct vnode *vp, const char *str) { if (KERNEL_PANICKED() || vp == NULL) return; #ifdef WITNESS if ((vp->v_irflag & VIRF_CROSSMP) == 0 && witness_is_owned(&vp->v_vnlock->lock_object) == 1) #else if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE) #endif vfs_badlock("is locked but should not be", str, vp); } void assert_vop_elocked(struct vnode *vp, const char *str) { if (KERNEL_PANICKED() || vp == NULL) return; if (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, int rc) { struct vop_fplookup_vexec_args *a; struct vnode *vp; a = ap; vp = a->a_vp; VFS_SMR_ASSERT_ENTERED(); if (rc == EOPNOTSUPP) VNPASS(VN_IS_DOOMED(vp), vp); } void vop_fplookup_symlink_debugpre(void *ap __unused) { VFS_SMR_ASSERT_ENTERED(); } void vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused) { VFS_SMR_ASSERT_ENTERED(); } static void vop_fsync_debugprepost(struct vnode *vp, const char *name) { if (vp->v_type == VCHR) ; /* * The shared vs. exclusive locking policy for fsync() * is actually determined by vp's write mount as indicated * by VOP_GETWRITEMOUNT(), which for stacked filesystems * may not be the same as vp->v_mount. However, if the * underlying filesystem which really handles the fsync() * supports shared locking, the stacked filesystem must also * be prepared for its VOP_FSYNC() operation to be called * with only a shared lock. On the other hand, if the * stacked filesystem claims support for shared write * locking but the underlying filesystem does not, and the * caller incorrectly uses a shared lock, this condition * should still be caught when the stacked filesystem * invokes VOP_FSYNC() on the underlying filesystem. */ else if (MNT_SHARED_WRITES(vp->v_mount)) ASSERT_VOP_LOCKED(vp, name); else ASSERT_VOP_ELOCKED(vp, name); } void vop_fsync_debugpre(void *a) { struct vop_fsync_args *ap; ap = a; vop_fsync_debugprepost(ap->a_vp, "fsync"); } void vop_fsync_debugpost(void *a, int rc __unused) { struct vop_fsync_args *ap; ap = a; vop_fsync_debugprepost(ap->a_vp, "fsync"); } void vop_fdatasync_debugpre(void *a) { struct vop_fdatasync_args *ap; ap = a; vop_fsync_debugprepost(ap->a_vp, "fsync"); } void vop_fdatasync_debugpost(void *a, int rc __unused) { struct vop_fdatasync_args *ap; ap = a; vop_fsync_debugprepost(ap->a_vp, "fsync"); } 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; struct vnode *vp = a->a_vp; VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp); ASSERT_VOP_LOCKED(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) { vp->v_vflag |= VV_UNLINKED; 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 |= kn->kn_sfflags & 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; KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ && kn->kn_filter != EVFILT_WRITE), ("READ/WRITE filter on a FIFO leaked through")); 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; off_t size; 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 (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0) return (0); VI_LOCK(vp); kn->kn_data = 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); } 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(uint64_t), 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, 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, mpcpu); return (vfs_cache_root_fallback(mp, flags, vpp)); } vrefact(vp); 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; maybe_yield(); 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) { maybe_yield(); 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, curthread)); } /* * 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_locked(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNPASS(vp->v_holdcnt > 0, vp); 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(struct vnode *vp) { VI_LOCK(vp); vn_seqc_write_begin_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); } /* * Special case handling for allocating and freeing vnodes. * * The counter remains unchanged on free so that a doomed vnode will * keep testing as in modify as long as it is accessible with SMR. */ static void vn_seqc_init(struct vnode *vp) { vp->v_seqc = 0; vp->v_seqc_users = 0; } static void vn_seqc_write_end_free(struct vnode *vp) { VNPASS(seqc_in_modify(vp->v_seqc), vp); VNPASS(vp->v_seqc_users == 1, vp); } void vn_irflag_set_locked(struct vnode *vp, short toset) { short flags; ASSERT_VI_LOCKED(vp, __func__); flags = vn_irflag_read(vp); VNASSERT((flags & toset) == 0, vp, ("%s: some of the passed flags already set (have %d, passed %d)\n", __func__, flags, toset)); atomic_store_short(&vp->v_irflag, flags | toset); } void vn_irflag_set(struct vnode *vp, short toset) { VI_LOCK(vp); vn_irflag_set_locked(vp, toset); VI_UNLOCK(vp); } void vn_irflag_set_cond_locked(struct vnode *vp, short toset) { short flags; ASSERT_VI_LOCKED(vp, __func__); flags = vn_irflag_read(vp); atomic_store_short(&vp->v_irflag, flags | toset); } void vn_irflag_set_cond(struct vnode *vp, short toset) { VI_LOCK(vp); vn_irflag_set_cond_locked(vp, toset); VI_UNLOCK(vp); } void vn_irflag_unset_locked(struct vnode *vp, short tounset) { short flags; ASSERT_VI_LOCKED(vp, __func__); flags = vn_irflag_read(vp); VNASSERT((flags & tounset) == tounset, vp, ("%s: some of the passed flags not set (have %d, passed %d)\n", __func__, flags, tounset)); atomic_store_short(&vp->v_irflag, flags & ~tounset); } void vn_irflag_unset(struct vnode *vp, short tounset) { VI_LOCK(vp); vn_irflag_unset_locked(vp, tounset); VI_UNLOCK(vp); } int vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred) { struct vattr vattr; int error; ASSERT_VOP_LOCKED(vp, __func__); error = VOP_GETATTR(vp, &vattr, cred); if (__predict_true(error == 0)) { if (vattr.va_size <= OFF_MAX) *size = vattr.va_size; else error = EFBIG; } return (error); } int vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred) { int error; VOP_LOCK(vp, LK_SHARED); error = vn_getsize_locked(vp, size, cred); VOP_UNLOCK(vp); return (error); } #ifdef INVARIANTS void vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state) { switch (vp->v_state) { case VSTATE_UNINITIALIZED: switch (state) { case VSTATE_CONSTRUCTED: case VSTATE_DESTROYING: return; default: break; } break; case VSTATE_CONSTRUCTED: ASSERT_VOP_ELOCKED(vp, __func__); switch (state) { case VSTATE_DESTROYING: return; default: break; } break; case VSTATE_DESTROYING: ASSERT_VOP_ELOCKED(vp, __func__); switch (state) { case VSTATE_DEAD: return; default: break; } break; case VSTATE_DEAD: switch (state) { case VSTATE_UNINITIALIZED: return; default: break; } break; } vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state); panic("invalid state transition %d -> %d\n", vp->v_state, state); } #endif diff --git a/sys/sys/rangelock.h b/sys/sys/rangelock.h index 62ccf77c03bc..310371bef879 100644 --- a/sys/sys/rangelock.h +++ b/sys/sys/rangelock.h @@ -1,109 +1,94 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2009 Konstantin Belousov * 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 unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 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. */ #ifndef _SYS_RANGELOCK_H #define _SYS_RANGELOCK_H -#include +#include +#ifndef _KERNEL +#include +#endif #define RL_LOCK_READ 0x0001 #define RL_LOCK_WRITE 0x0002 #define RL_LOCK_TYPE_MASK 0x0003 -#define RL_LOCK_GRANTED 0x0004 struct rl_q_entry; /* * The structure representing the range lock. Caller may request * read or write access to the range of bytes. Access is granted if * all existing lock owners are compatible with the request. Two lock * owners are compatible if their ranges do not overlap, or both * owners are for read. - * - * Access to the structure itself is synchronized with the externally - * supplied mutex. - * - * rl_waiters is the queue containing in order (a) granted write lock - * requests, (b) granted read lock requests, and (c) in order of arrival, - * lock requests which cannot be granted yet. - * - * rl_currdep is the first lock request that cannot be granted now due - * to the preceding requests conflicting with it (i.e., it points to - * position (c) in the list above). */ struct rangelock { - TAILQ_HEAD(, rl_q_entry) rl_waiters; - struct rl_q_entry *rl_currdep; + struct rl_q_entry *head; + bool sleepers; }; #ifdef _KERNEL -struct mtx; - void rangelock_init(struct rangelock *lock); void rangelock_destroy(struct rangelock *lock); -void rangelock_unlock(struct rangelock *lock, void *cookie, - struct mtx *ilk); -void *rangelock_unlock_range(struct rangelock *lock, void *cookie, - off_t start, off_t end, struct mtx *ilk); -void *rangelock_rlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk); -void *rangelock_tryrlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk); -void *rangelock_wlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk); -void *rangelock_trywlock(struct rangelock *lock, off_t start, off_t end, - struct mtx *ilk); -void rlqentry_free(struct rl_q_entry *rlqe); +void rangelock_unlock(struct rangelock *lock, void *cookie); +void *rangelock_rlock(struct rangelock *lock, vm_ooffset_t start, + vm_ooffset_t end); +void *rangelock_tryrlock(struct rangelock *lock, vm_ooffset_t start, + vm_ooffset_t end); +void *rangelock_wlock(struct rangelock *lock, vm_ooffset_t start, + vm_ooffset_t end); +void *rangelock_trywlock(struct rangelock *lock, vm_ooffset_t start, + vm_ooffset_t end); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void _rangelock_cookie_assert(void *cookie, int what, const char *file, int line); #endif #ifdef INVARIANTS #define rangelock_cookie_assert_(cookie, what, file, line) \ _rangelock_cookie_assert((cookie), (what), (file), (line)) #else #define rangelock_cookie_assert_(cookie, what, file, line) (void)0 #endif #define rangelock_cookie_assert(cookie, what) \ rangelock_cookie_assert_((cookie), (what), __FILE__, __LINE__) /* * Assertion flags. */ #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) #define RCA_LOCKED 0x0001 #define RCA_RLOCKED 0x0002 #define RCA_WLOCKED 0x0004 #endif #endif /* _KERNEL */ #endif /* _SYS_RANGELOCK_H */ diff --git a/sys/sys/vnode.h b/sys/sys/vnode.h index dc926d7a9c9e..0d0f228f7051 100644 --- a/sys/sys/vnode.h +++ b/sys/sys/vnode.h @@ -1,1210 +1,1207 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 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. */ #ifndef _SYS_VNODE_H_ #define _SYS_VNODE_H_ #include #include #include #include #include #include #include #include #include #include #include /* * The vnode is the focus of all file activity in UNIX. There is a * unique vnode allocated for each active file, each current directory, * each mounted-on file, text file, and the root. */ /* * Vnode types. VNON means no type. */ __enum_uint8_decl(vtype) { VNON, VREG, VDIR, VBLK, VCHR, VLNK, VSOCK, VFIFO, VBAD, VMARKER, VLASTTYPE = VMARKER, }; __enum_uint8_decl(vstate) { VSTATE_UNINITIALIZED, VSTATE_CONSTRUCTED, VSTATE_DESTROYING, VSTATE_DEAD, VLASTSTATE = VSTATE_DEAD, }; enum vgetstate { VGET_NONE, VGET_HOLDCNT, VGET_USECOUNT, }; /* * Each underlying filesystem allocates its own private area and hangs * it from v_data. If non-null, this area is freed in getnewvnode(). */ struct namecache; struct cache_fpl; struct vpollinfo { struct mtx vpi_lock; /* lock to protect below */ struct selinfo vpi_selinfo; /* identity of poller(s) */ short vpi_events; /* what they are looking for */ short vpi_revents; /* what has happened */ }; /* * Reading or writing any of these items requires holding the appropriate lock. * * Lock reference: * c - namecache mutex * i - interlock * l - mp mnt_listmtx or freelist mutex * I - updated with atomics, 0->1 and 1->0 transitions with interlock held * m - mount point interlock * p - pollinfo lock * u - Only a reference to the vnode is needed to read. * v - vnode lock * * Vnodes may be found on many lists. The general way to deal with operating * on a vnode that is on a list is: * 1) Lock the list and find the vnode. * 2) Lock interlock so that the vnode does not go away. * 3) Unlock the list to avoid lock order reversals. * 4) vget with LK_INTERLOCK and check for ENOENT, or * 5) Check for DOOMED if the vnode lock is not required. * 6) Perform your operation, then vput(). */ #if defined(_KERNEL) || defined(_KVM_VNODE) struct vnode { /* * Fields which define the identity of the vnode. These fields are * owned by the filesystem (XXX: and vgone() ?) */ __enum_uint8(vtype) v_type; /* u vnode type */ __enum_uint8(vstate) v_state; /* u vnode state */ short v_irflag; /* i frequently read flags */ seqc_t v_seqc; /* i modification count */ uint32_t v_nchash; /* u namecache hash */ u_int v_hash; struct vop_vector *v_op; /* u vnode operations vector */ void *v_data; /* u private data for fs */ /* * Filesystem instance stuff */ struct mount *v_mount; /* u ptr to vfs we are in */ TAILQ_ENTRY(vnode) v_nmntvnodes; /* m vnodes for mount point */ /* * Type specific fields, only one applies to any given vnode. */ union { struct mount *v_mountedhere; /* v ptr to mountpoint (VDIR) */ struct unpcb *v_unpcb; /* v unix domain net (VSOCK) */ struct cdev *v_rdev; /* v device (VCHR, VBLK) */ struct fifoinfo *v_fifoinfo; /* v fifo (VFIFO) */ }; /* * vfs_hash: (mount + inode) -> vnode hash. The hash value * itself is grouped with other int fields, to avoid padding. */ LIST_ENTRY(vnode) v_hashlist; /* * VFS_namecache stuff */ LIST_HEAD(, namecache) v_cache_src; /* c Cache entries from us */ TAILQ_HEAD(, namecache) v_cache_dst; /* c Cache entries to us */ struct namecache *v_cache_dd; /* c Cache entry for .. vnode */ /* * Locking */ struct lock v_lock; /* u (if fs don't have one) */ struct mtx v_interlock; /* lock for "i" things */ struct lock *v_vnlock; /* u pointer to vnode lock */ /* * The machinery of being a vnode */ TAILQ_ENTRY(vnode) v_vnodelist; /* l vnode lists */ TAILQ_ENTRY(vnode) v_lazylist; /* l vnode lazy list */ struct bufobj v_bufobj; /* * Buffer cache object */ /* * Hooks for various subsystems and features. */ struct vpollinfo *v_pollinfo; /* i Poll events, p for *v_pi */ struct label *v_label; /* MAC label for vnode */ struct lockf *v_lockf; /* Byte-level advisory lock list */ struct rangelock v_rl; /* Byte-range lock */ u_int v_holdcnt; /* I prevents recycling. */ u_int v_usecount; /* I ref count of users */ u_short v_iflag; /* i vnode flags (see below) */ u_short v_vflag; /* v vnode flags */ u_short v_mflag; /* l mnt-specific vnode flags */ short v_dbatchcpu; /* i LRU requeue deferral batch */ int v_writecount; /* I ref count of writers or (negative) text users */ int v_seqc_users; /* i modifications pending */ }; #ifndef DEBUG_LOCKS #ifdef _LP64 /* * Not crossing 448 bytes fits 9 vnodes per page. If you have to add fields * to the structure and there is nothing which can be done to prevent growth * then so be it. But don't grow it without a good reason. */ _Static_assert(sizeof(struct vnode) <= 448, "vnode size crosses 448 bytes"); #endif #endif #endif /* defined(_KERNEL) || defined(_KVM_VNODE) */ #define bo2vnode(bo) __containerof((bo), struct vnode, v_bufobj) /* XXX: These are temporary to avoid a source sweep at this time */ #define v_object v_bufobj.bo_object /* We don't need to lock the knlist */ #define VN_KNLIST_EMPTY(vp) ((vp)->v_pollinfo == NULL || \ KNLIST_EMPTY(&(vp)->v_pollinfo->vpi_selinfo.si_note)) #define VN_KNOTE(vp, b, a) \ do { \ if (!VN_KNLIST_EMPTY(vp)) \ KNOTE(&vp->v_pollinfo->vpi_selinfo.si_note, (b), \ (a) | KNF_NOKQLOCK); \ } while (0) #define VN_KNOTE_LOCKED(vp, b) VN_KNOTE(vp, b, KNF_LISTLOCKED) #define VN_KNOTE_UNLOCKED(vp, b) VN_KNOTE(vp, b, 0) /* * Vnode flags. * VI flags are protected by interlock and live in v_iflag * VV flags are protected by the vnode lock and live in v_vflag * * VIRF_DOOMED is doubly protected by the interlock and vnode lock. Both * are required for writing but the status may be checked with either. */ #define VHOLD_NO_SMR (1<<29) /* Disable vhold_smr */ #define VHOLD_ALL_FLAGS (VHOLD_NO_SMR) #define VIRF_DOOMED 0x0001 /* This vnode is being recycled */ #define VIRF_PGREAD 0x0002 /* Direct reads from the page cache are permitted, never cleared once set */ #define VIRF_MOUNTPOINT 0x0004 /* This vnode is mounted on */ #define VIRF_TEXT_REF 0x0008 /* Executable mappings ref the vnode */ #define VIRF_CROSSMP 0x0010 /* Cross-mp vnode, no locking */ #define VI_UNUSED0 0x0001 /* unused */ #define VI_MOUNT 0x0002 /* Mount in progress */ #define VI_DOINGINACT 0x0004 /* VOP_INACTIVE is in progress */ #define VI_OWEINACT 0x0008 /* Need to call inactive */ #define VI_DEFINACT 0x0010 /* deferred inactive */ #define VI_FOPENING 0x0020 /* In open, with opening process having the first right to advlock file */ #define VV_ROOT 0x0001 /* root of its filesystem */ #define VV_ISTTY 0x0002 /* vnode represents a tty */ #define VV_NOSYNC 0x0004 /* unlinked, stop syncing */ #define VV_ETERNALDEV 0x0008 /* device that is never destroyed */ #define VV_CACHEDLABEL 0x0010 /* Vnode has valid cached MAC label */ #define VV_VMSIZEVNLOCK 0x0020 /* object size check requires vnode lock */ #define VV_COPYONWRITE 0x0040 /* vnode is doing copy-on-write */ #define VV_SYSTEM 0x0080 /* vnode being used by kernel */ #define VV_PROCDEP 0x0100 /* vnode is process dependent */ #define VV_UNLINKED 0x0200 /* unlinked but stil open directory */ #define VV_DELETED 0x0400 /* should be removed */ #define VV_MD 0x0800 /* vnode backs the md device */ #define VV_FORCEINSMQ 0x1000 /* force the insmntque to succeed */ #define VV_READLINK 0x2000 /* fdescfs linux vnode */ #define VV_UNREF 0x4000 /* vunref, do not drop lock in inactive() */ #define VV_CROSSLOCK 0x8000 /* vnode lock is shared w/ root mounted here */ #define VMP_LAZYLIST 0x0001 /* Vnode is on mnt's lazy list */ /* * Vnode attributes. A field value of VNOVAL represents a field whose value * is unavailable (getattr) or which is not to be changed (setattr). */ struct vattr { __enum_uint8(vtype) va_type; /* vnode type (for create) */ u_short va_mode; /* files access mode and type */ u_short va_padding0; uid_t va_uid; /* owner user id */ gid_t va_gid; /* owner group id */ nlink_t va_nlink; /* number of references to file */ dev_t va_fsid; /* filesystem id */ ino_t va_fileid; /* file id */ u_quad_t va_size; /* file size in bytes */ long va_blocksize; /* blocksize preferred for i/o */ struct timespec va_atime; /* time of last access */ struct timespec va_mtime; /* time of last modification */ struct timespec va_ctime; /* time file changed */ struct timespec va_birthtime; /* time file created */ u_long va_gen; /* generation number of file */ u_long va_flags; /* flags defined for file */ dev_t va_rdev; /* device the special file represents */ u_quad_t va_bytes; /* bytes of disk space held by file */ u_quad_t va_filerev; /* file modification number */ u_int va_vaflags; /* operations flags, see below */ long va_spare; /* remain quad aligned */ }; /* * Flags for va_vaflags. */ #define VA_UTIMES_NULL 0x01 /* utimes argument was NULL */ #define VA_EXCLUSIVE 0x02 /* exclusive create request */ #define VA_SYNC 0x04 /* O_SYNC truncation */ /* * Flags for ioflag. (high 16 bits used to ask for read-ahead and * help with write clustering) * NB: IO_NDELAY and IO_DIRECT are linked to fcntl.h */ #define IO_UNIT 0x0001 /* do I/O as atomic unit */ #define IO_APPEND 0x0002 /* append write to end */ #define IO_NDELAY 0x0004 /* FNDELAY flag set in file table */ #define IO_NODELOCKED 0x0008 /* underlying node already locked */ #define IO_ASYNC 0x0010 /* bawrite rather then bdwrite */ #define IO_VMIO 0x0020 /* data already in VMIO space */ #define IO_INVAL 0x0040 /* invalidate after I/O */ #define IO_SYNC 0x0080 /* do I/O synchronously */ #define IO_DIRECT 0x0100 /* attempt to bypass buffer cache */ #define IO_NOREUSE 0x0200 /* VMIO data won't be reused */ #define IO_EXT 0x0400 /* operate on external attributes */ #define IO_NORMAL 0x0800 /* operate on regular data */ #define IO_NOMACCHECK 0x1000 /* MAC checks unnecessary */ #define IO_BUFLOCKED 0x2000 /* ffs flag; indir buf is locked */ #define IO_RANGELOCKED 0x4000 /* range locked */ #define IO_DATASYNC 0x8000 /* do only data I/O synchronously */ #define IO_SEQMAX 0x7F /* seq heuristic max value */ #define IO_SEQSHIFT 16 /* seq heuristic in upper 16 bits */ /* * Flags for accmode_t. */ #define VEXEC 000000000100 /* execute/search permission */ #define VWRITE 000000000200 /* write permission */ #define VREAD 000000000400 /* read permission */ #define VADMIN 000000010000 /* being the file owner */ #define VAPPEND 000000040000 /* permission to write/append */ /* * VEXPLICIT_DENY makes VOP_ACCESSX(9) return EPERM or EACCES only * if permission was denied explicitly, by a "deny" rule in NFSv4 ACL, * and 0 otherwise. This never happens with ordinary unix access rights * or POSIX.1e ACLs. Obviously, VEXPLICIT_DENY must be OR-ed with * some other V* constant. */ #define VEXPLICIT_DENY 000000100000 #define VREAD_NAMED_ATTRS 000000200000 /* not used */ #define VWRITE_NAMED_ATTRS 000000400000 /* not used */ #define VDELETE_CHILD 000001000000 #define VREAD_ATTRIBUTES 000002000000 /* permission to stat(2) */ #define VWRITE_ATTRIBUTES 000004000000 /* change {m,c,a}time */ #define VDELETE 000010000000 #define VREAD_ACL 000020000000 /* read ACL and file mode */ #define VWRITE_ACL 000040000000 /* change ACL and/or file mode */ #define VWRITE_OWNER 000100000000 /* change file owner */ #define VSYNCHRONIZE 000200000000 /* not used */ #define VCREAT 000400000000 /* creating new file */ #define VVERIFY 001000000000 /* verification required */ /* * Permissions that were traditionally granted only to the file owner. */ #define VADMIN_PERMS (VADMIN | VWRITE_ATTRIBUTES | VWRITE_ACL | \ VWRITE_OWNER) /* * Permissions that were traditionally granted to everyone. */ #define VSTAT_PERMS (VREAD_ATTRIBUTES | VREAD_ACL) /* * Permissions that allow to change the state of the file in any way. */ #define VMODIFY_PERMS (VWRITE | VAPPEND | VADMIN_PERMS | VDELETE_CHILD | \ VDELETE) /* * Token indicating no attribute value yet assigned. */ #define VNOVAL (-1) /* * LK_TIMELOCK timeout for vnode locks (used mainly by the pageout daemon) */ #define VLKTIMEOUT (hz / 20 + 1) #ifdef _KERNEL #ifdef MALLOC_DECLARE MALLOC_DECLARE(M_VNODE); #endif extern u_int ncsizefactor; extern const u_int io_hold_cnt; /* * Convert between vnode types and inode formats (since POSIX.1 * defines mode word of stat structure in terms of inode formats). */ extern __enum_uint8(vtype) iftovt_tab[]; extern int vttoif_tab[]; #define IFTOVT(mode) (iftovt_tab[((mode) & S_IFMT) >> 12]) #define VTTOIF(indx) (vttoif_tab[(int)(indx)]) #define MAKEIMODE(indx, mode) (int)(VTTOIF(indx) | (mode)) /* * Flags to various vnode functions. */ #define SKIPSYSTEM 0x0001 /* vflush: skip vnodes marked VSYSTEM */ #define FORCECLOSE 0x0002 /* vflush: force file closure */ #define WRITECLOSE 0x0004 /* vflush: only close writable files */ #define EARLYFLUSH 0x0008 /* vflush: early call for ffs_flushfiles */ #define V_SAVE 0x0001 /* vinvalbuf: sync file first */ #define V_ALT 0x0002 /* vinvalbuf: invalidate only alternate bufs */ #define V_NORMAL 0x0004 /* vinvalbuf: invalidate only regular bufs */ #define V_CLEANONLY 0x0008 /* vinvalbuf: invalidate only clean bufs */ #define V_VMIO 0x0010 /* vinvalbuf: called during pageout */ #define V_ALLOWCLEAN 0x0020 /* vinvalbuf: allow clean buffers after flush */ #define REVOKEALL 0x0001 /* vop_revoke: revoke all aliases */ #define V_WAIT 0x0001 /* vn_start_write: sleep for suspend */ #define V_NOWAIT 0x0002 /* vn_start_write: don't sleep for suspend */ #define V_XSLEEP 0x0004 /* vn_start_write: just return after sleep */ #define V_PCATCH 0x0008 /* vn_start_write: make the sleep interruptible */ #define V_VALID_FLAGS (V_WAIT | V_NOWAIT | V_XSLEEP | V_PCATCH) #define VR_START_WRITE 0x0001 /* vfs_write_resume: start write atomically */ #define VR_NO_SUSPCLR 0x0002 /* vfs_write_resume: do not clear suspension */ #define VS_SKIP_UNMOUNT 0x0001 /* vfs_write_suspend: fail if the filesystem is being unmounted */ #define VREF(vp) vref(vp) #ifdef DIAGNOSTIC #define VATTR_NULL(vap) vattr_null(vap) #else #define VATTR_NULL(vap) (*(vap) = va_null) /* initialize a vattr */ #endif /* DIAGNOSTIC */ #define NULLVP ((struct vnode *)NULL) /* * Global vnode data. */ extern struct vnode *rootvnode; /* root (i.e. "/") vnode */ extern struct mount *rootdevmp; /* "/dev" mount */ extern u_long desiredvnodes; /* number of vnodes desired */ extern struct uma_zone *namei_zone; extern struct vattr va_null; /* predefined null vattr structure */ extern u_int vn_lock_pair_pause_max; #define VI_LOCK(vp) mtx_lock(&(vp)->v_interlock) #define VI_LOCK_FLAGS(vp, flags) mtx_lock_flags(&(vp)->v_interlock, (flags)) #define VI_TRYLOCK(vp) mtx_trylock(&(vp)->v_interlock) #define VI_UNLOCK(vp) mtx_unlock(&(vp)->v_interlock) #define VI_MTX(vp) (&(vp)->v_interlock) #define VN_LOCK_AREC(vp) lockallowrecurse((vp)->v_vnlock) #define VN_LOCK_ASHARE(vp) lockallowshare((vp)->v_vnlock) #define VN_LOCK_DSHARE(vp) lockdisableshare((vp)->v_vnlock) #endif /* _KERNEL */ /* * Mods for extensibility. */ /* * Flags for vdesc_flags: */ #define VDESC_MAX_VPS 16 /* Low order 16 flag bits are reserved for willrele flags for vp arguments. */ #define VDESC_VP0_WILLRELE 0x0001 #define VDESC_VP1_WILLRELE 0x0002 #define VDESC_VP2_WILLRELE 0x0004 #define VDESC_VP3_WILLRELE 0x0008 /* * A generic structure. * This can be used by bypass routines to identify generic arguments. */ struct vop_generic_args { struct vnodeop_desc *a_desc; /* other random data follows, presumably */ }; typedef int vop_bypass_t(struct vop_generic_args *); /* * VDESC_NO_OFFSET is used to identify the end of the offset list * and in places where no such field exists. */ #define VDESC_NO_OFFSET -1 /* * This structure describes the vnode operation taking place. */ struct vnodeop_desc { char *vdesc_name; /* a readable name for debugging */ int vdesc_flags; /* VDESC_* flags */ int vdesc_vop_offset; vop_bypass_t *vdesc_call; /* Function to call */ /* * These ops are used by bypass routines to map and locate arguments. * Creds and procs are not needed in bypass routines, but sometimes * they are useful to (for example) transport layers. * Nameidata is useful because it has a cred in it. */ int *vdesc_vp_offsets; /* list ended by VDESC_NO_OFFSET */ int vdesc_vpp_offset; /* return vpp location */ int vdesc_cred_offset; /* cred location, if any */ int vdesc_thread_offset; /* thread location, if any */ int vdesc_componentname_offset; /* if any */ }; #ifdef _KERNEL /* * A list of all the operation descs. */ extern struct vnodeop_desc *vnodeop_descs[]; #define VOPARG_OFFSETOF(s_type, field) __offsetof(s_type, field) #define VOPARG_OFFSETTO(s_type, s_offset, struct_p) \ ((s_type)(((char*)(struct_p)) + (s_offset))) #ifdef DEBUG_VFS_LOCKS /* * Support code to aid in debugging VFS locking problems. Not totally * reliable since if the thread sleeps between changing the lock * state and checking it with the assert, some other thread could * change the state. They are good enough for debugging a single * filesystem using a single-threaded test. Note that the unreliability is * limited to false negatives; efforts were made to ensure that false * positives cannot occur. */ void assert_vi_locked(struct vnode *vp, const char *str); void assert_vi_unlocked(struct vnode *vp, const char *str); void assert_vop_elocked(struct vnode *vp, const char *str); void assert_vop_locked(struct vnode *vp, const char *str); void assert_vop_unlocked(struct vnode *vp, const char *str); #define ASSERT_VI_LOCKED(vp, str) assert_vi_locked((vp), (str)) #define ASSERT_VI_UNLOCKED(vp, str) assert_vi_unlocked((vp), (str)) #define ASSERT_VOP_ELOCKED(vp, str) assert_vop_elocked((vp), (str)) #define ASSERT_VOP_LOCKED(vp, str) assert_vop_locked((vp), (str)) #define ASSERT_VOP_UNLOCKED(vp, str) assert_vop_unlocked((vp), (str)) #define ASSERT_VOP_IN_SEQC(vp) do { \ struct vnode *_vp = (vp); \ \ VNPASS(seqc_in_modify(_vp->v_seqc), _vp); \ } while (0) #define ASSERT_VOP_NOT_IN_SEQC(vp) do { \ struct vnode *_vp = (vp); \ \ VNPASS(!seqc_in_modify(_vp->v_seqc), _vp); \ } while (0) #else /* !DEBUG_VFS_LOCKS */ #define ASSERT_VI_LOCKED(vp, str) ((void)0) #define ASSERT_VI_UNLOCKED(vp, str) ((void)0) #define ASSERT_VOP_ELOCKED(vp, str) ((void)0) #define ASSERT_VOP_LOCKED(vp, str) ((void)0) #define ASSERT_VOP_UNLOCKED(vp, str) ((void)0) #define ASSERT_VOP_IN_SEQC(vp) ((void)0) #define ASSERT_VOP_NOT_IN_SEQC(vp) ((void)0) #endif /* DEBUG_VFS_LOCKS */ /* * This call works for vnodes in the kernel. */ #define VCALL(c) ((c)->a_desc->vdesc_call(c)) #define DOINGASYNC(vp) \ (((vp)->v_mount->mnt_kern_flag & MNTK_ASYNC) != 0 && \ ((curthread->td_pflags & TDP_SYNCIO) == 0)) /* * VMIO support inline */ extern int vmiodirenable; static __inline int vn_canvmio(struct vnode *vp) { if (vp && (vp->v_type == VREG || (vmiodirenable && vp->v_type == VDIR))) return(TRUE); return(FALSE); } /* * Finally, include the default set of vnode operations. */ typedef void vop_getpages_iodone_t(void *, vm_page_t *, int, int); #include "vnode_if.h" /* vn_open_flags */ #define VN_OPEN_NOAUDIT 0x00000001 #define VN_OPEN_NOCAPCHECK 0x00000002 #define VN_OPEN_NAMECACHE 0x00000004 #define VN_OPEN_INVFS 0x00000008 #define VN_OPEN_WANTIOCTLCAPS 0x00000010 /* copy_file_range kernel flags */ #define COPY_FILE_RANGE_KFLAGS 0xff000000 #define COPY_FILE_RANGE_TIMEO1SEC 0x01000000 /* Return after 1sec. */ /* * Public vnode manipulation functions. */ struct componentname; struct file; struct mount; struct nameidata; struct ostat; struct freebsd11_stat; struct thread; struct proc; struct stat; struct nstat; struct ucred; struct uio; struct vattr; struct vfsops; struct vnode; typedef int (*vn_get_ino_t)(struct mount *, void *, int, struct vnode **); int bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn); /* cache_* may belong in namei.h. */ void cache_changesize(u_long newhashsize); #define VFS_CACHE_DROPOLD 0x1 void cache_enter_time_flags(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, struct timespec *tsp, struct timespec *dtsp, int flags); #define cache_enter(dvp, vp, cnp) \ cache_enter_time(dvp, vp, cnp, NULL, NULL) void cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, struct timespec *tsp, struct timespec *dtsp); int cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp); void cache_vnode_init(struct vnode *vp); void cache_purge(struct vnode *vp); void cache_purge_vgone(struct vnode *vp); void cache_purge_negative(struct vnode *vp); void cache_purgevfs(struct mount *mp); char *cache_symlink_alloc(size_t size, int flags); void cache_symlink_free(char *string, size_t size); int cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len); void cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp, struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp); void cache_vop_rmdir(struct vnode *dvp, struct vnode *vp); void cache_vop_vector_register(struct vop_vector *); #ifdef INVARIANTS void cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp); void cache_validate_vop_vector(struct mount *mp, struct vop_vector *vops); void cache_assert_no_entries(struct vnode *vp); #else static inline void cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { } static inline void cache_validate_vop_vector(struct mount *mp, struct vop_vector *vops) { } static inline void cache_assert_no_entries(struct vnode *vp) { } #endif void cache_fast_lookup_enabled_recalc(void); int change_dir(struct vnode *vp, struct thread *td); void cvtstat(struct stat *st, struct ostat *ost); int freebsd11_cvtnstat(struct stat *sb, struct nstat *nsb); int freebsd11_cvtstat(struct stat *st, struct freebsd11_stat *ost); int getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, struct vnode **vpp); void getnewvnode_reserve(void); void getnewvnode_drop_reserve(void); int insmntque(struct vnode *vp, struct mount *mp); int insmntque1(struct vnode *vp, struct mount *mp); u_quad_t init_va_filerev(void); int speedup_syncer(void); int vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen); int vn_getcwd(char *buf, char **retbuf, size_t *buflen); int vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf); int vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf); int vn_fullpath_hardlink(struct vnode *vp, struct vnode *dvp, const char *hdrl_name, size_t hrdl_name_length, char **retbuf, char **freebuf, size_t *buflen); struct vnode * vn_dir_dd_ino(struct vnode *vp); int vn_commname(struct vnode *vn, char *buf, u_int buflen); int vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, u_int pathlen); int vn_path_to_global_path_hardlink(struct thread *td, struct vnode *vp, struct vnode *dvp, char *path, u_int pathlen, const char *leaf_name, size_t leaf_length); int vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid, accmode_t accmode, struct ucred *cred); int vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred); int vaccess_acl_nfs4(__enum_uint8(vtype) type, uid_t file_uid, gid_t file_gid, struct acl *aclp, accmode_t accmode, struct ucred *cred); int vaccess_acl_posix1e(__enum_uint8(vtype) type, uid_t file_uid, gid_t file_gid, struct acl *acl, accmode_t accmode, struct ucred *cred); void vattr_null(struct vattr *vap); void vlazy(struct vnode *); void vdrop(struct vnode *); void vdropl(struct vnode *); int vflush(struct mount *mp, int rootrefs, int flags, struct thread *td); int vget(struct vnode *vp, int flags); enum vgetstate vget_prep_smr(struct vnode *vp); enum vgetstate vget_prep(struct vnode *vp); int vget_finish(struct vnode *vp, int flags, enum vgetstate vs); void vget_finish_ref(struct vnode *vp, enum vgetstate vs); void vget_abort(struct vnode *vp, enum vgetstate vs); void vgone(struct vnode *vp); void vhold(struct vnode *); void vholdnz(struct vnode *); bool vhold_smr(struct vnode *); int vinactive(struct vnode *vp); int vinvalbuf(struct vnode *vp, int save, int slpflag, int slptimeo); int vtruncbuf(struct vnode *vp, off_t length, int blksize); void v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, int blksize); void vunref(struct vnode *); void vn_printf(struct vnode *vp, const char *fmt, ...) __printflike(2,3); int vrecycle(struct vnode *vp); int vrecyclel(struct vnode *vp); int vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred); int vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred); int vn_close(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td); 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 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags, int ioflg, struct ucred *active_cred, struct ucred *file_cred); void vn_finished_write(struct mount *mp); void vn_finished_secondary_write(struct mount *mp); int vn_fsync_buf(struct vnode *vp, int waitfor); 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); int vn_need_pageq_flush(struct vnode *vp); bool vn_isdisk_error(struct vnode *vp, int *errp); bool vn_isdisk(struct vnode *vp); int _vn_lock(struct vnode *vp, int flags, const char *file, int line); #define vn_lock(vp, flags) _vn_lock(vp, flags, __FILE__, __LINE__) void vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1, struct vnode *vp2, bool vp2_locked, int lkflags2); int vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp); int vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, struct ucred *cred, struct file *fp); int vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, struct thread *td, struct file *fp); void vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end); void vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end); int vn_pollrecord(struct vnode *vp, struct thread *p, int events); 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); 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 vn_read_from_obj(struct vnode *vp, struct uio *uio); int vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, struct thread *td); int vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz, ssize_t *resid_adj, struct thread *td); void vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj); int vn_rlimit_trunc(u_quad_t size, struct thread *td); int vn_start_write(struct vnode *vp, struct mount **mpp, int flags); int vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags); int vn_truncate_locked(struct vnode *vp, off_t length, bool sync, struct ucred *cred); int vn_writechk(struct vnode *vp); int vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int *buflen, char *buf, struct thread *td); int vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int buflen, char *buf, struct thread *td); int vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, struct thread *td); int vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp); int vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, int lkflags, struct vnode **rvp); int vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *td); int vn_cmp(struct file *, struct file *, struct thread *td); int vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio); int vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, struct uio *uio); void vn_seqc_write_begin_locked(struct vnode *vp); void vn_seqc_write_begin(struct vnode *vp); void vn_seqc_write_end_locked(struct vnode *vp); void vn_seqc_write_end(struct vnode *vp); #define vn_seqc_read_any(vp) seqc_read_any(&(vp)->v_seqc) #define vn_seqc_read_notmodify(vp) seqc_read_notmodify(&(vp)->v_seqc) #define vn_seqc_consistent(vp, seq) seqc_consistent(&(vp)->v_seqc, seq) #define vn_rangelock_unlock(vp, cookie) \ - rangelock_unlock(&(vp)->v_rl, (cookie), VI_MTX(vp)) -#define vn_rangelock_unlock_range(vp, cookie, start, end) \ - rangelock_unlock_range(&(vp)->v_rl, (cookie), (start), (end), \ - VI_MTX(vp)) + rangelock_unlock(&(vp)->v_rl, (cookie)) #define vn_rangelock_rlock(vp, start, end) \ - rangelock_rlock(&(vp)->v_rl, (start), (end), VI_MTX(vp)) + rangelock_rlock(&(vp)->v_rl, (start), (end)) #define vn_rangelock_tryrlock(vp, start, end) \ - rangelock_tryrlock(&(vp)->v_rl, (start), (end), VI_MTX(vp)) + rangelock_tryrlock(&(vp)->v_rl, (start), (end)) #define vn_rangelock_wlock(vp, start, end) \ - rangelock_wlock(&(vp)->v_rl, (start), (end), VI_MTX(vp)) + rangelock_wlock(&(vp)->v_rl, (start), (end)) #define vn_rangelock_trywlock(vp, start, end) \ - rangelock_trywlock(&(vp)->v_rl, (start), (end), VI_MTX(vp)) + rangelock_trywlock(&(vp)->v_rl, (start), (end)) #define vn_irflag_read(vp) atomic_load_short(&(vp)->v_irflag) void vn_irflag_set_locked(struct vnode *vp, short toset); void vn_irflag_set(struct vnode *vp, short toset); void vn_irflag_set_cond_locked(struct vnode *vp, short toset); void vn_irflag_set_cond(struct vnode *vp, short toset); void vn_irflag_unset_locked(struct vnode *vp, short tounset); void vn_irflag_unset(struct vnode *vp, short tounset); int vfs_cache_lookup(struct vop_lookup_args *ap); int vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp); void vfs_timestamp(struct timespec *); void vfs_write_resume(struct mount *mp, int flags); int vfs_write_suspend(struct mount *mp, int flags); int vfs_write_suspend_umnt(struct mount *mp); struct vnode *vnlru_alloc_marker(void); void vnlru_free_marker(struct vnode *); void vnlru_free_vfsops(int, struct vfsops *, struct vnode *); int vop_stdbmap(struct vop_bmap_args *); int vop_stdfdatasync_buf(struct vop_fdatasync_args *); int vop_stdfsync(struct vop_fsync_args *); int vop_stdgetwritemount(struct vop_getwritemount_args *); int vop_stdgetpages(struct vop_getpages_args *); int vop_stdinactive(struct vop_inactive_args *); int vop_stdioctl(struct vop_ioctl_args *); int vop_stdneed_inactive(struct vop_need_inactive_args *); int vop_stdkqfilter(struct vop_kqfilter_args *); int vop_stdlock(struct vop_lock1_args *); int vop_stdunlock(struct vop_unlock_args *); int vop_stdislocked(struct vop_islocked_args *); int vop_lock(struct vop_lock1_args *); int vop_unlock(struct vop_unlock_args *); int vop_islocked(struct vop_islocked_args *); int vop_stdputpages(struct vop_putpages_args *); int vop_nopoll(struct vop_poll_args *); int vop_stdaccess(struct vop_access_args *ap); int vop_stdaccessx(struct vop_accessx_args *ap); int vop_stdadvise(struct vop_advise_args *ap); int vop_stdadvlock(struct vop_advlock_args *ap); int vop_stdadvlockasync(struct vop_advlockasync_args *ap); int vop_stdadvlockpurge(struct vop_advlockpurge_args *ap); int vop_stdallocate(struct vop_allocate_args *ap); int vop_stddeallocate(struct vop_deallocate_args *ap); int vop_stdset_text(struct vop_set_text_args *ap); int vop_stdpathconf(struct vop_pathconf_args *); int vop_stdpoll(struct vop_poll_args *); int vop_stdvptocnp(struct vop_vptocnp_args *ap); int vop_stdvptofh(struct vop_vptofh_args *ap); int vop_stdunp_bind(struct vop_unp_bind_args *ap); int vop_stdunp_connect(struct vop_unp_connect_args *ap); int vop_stdunp_detach(struct vop_unp_detach_args *ap); int vop_stdadd_writecount_nomsync(struct vop_add_writecount_args *ap); int vop_eopnotsupp(struct vop_generic_args *ap); int vop_ebadf(struct vop_generic_args *ap); int vop_einval(struct vop_generic_args *ap); int vop_enoent(struct vop_generic_args *ap); int vop_enotty(struct vop_generic_args *ap); int vop_eagain(struct vop_generic_args *ap); int vop_null(struct vop_generic_args *ap); int vop_panic(struct vop_generic_args *ap); int dead_poll(struct vop_poll_args *ap); int dead_read(struct vop_read_args *ap); int dead_write(struct vop_write_args *ap); /* These are called from within the actual VOPS. */ void vop_close_post(void *a, int rc); void vop_create_pre(void *a); void vop_create_post(void *a, int rc); void vop_whiteout_pre(void *a); void vop_whiteout_post(void *a, int rc); void vop_deleteextattr_pre(void *a); void vop_deleteextattr_post(void *a, int rc); void vop_link_pre(void *a); void vop_link_post(void *a, int rc); void vop_lookup_post(void *a, int rc); void vop_lookup_pre(void *a); void vop_mkdir_pre(void *a); void vop_mkdir_post(void *a, int rc); void vop_mknod_pre(void *a); void vop_mknod_post(void *a, int rc); void vop_open_post(void *a, int rc); void vop_read_post(void *a, int rc); void vop_read_pgcache_post(void *ap, int rc); void vop_readdir_post(void *a, int rc); void vop_reclaim_post(void *a, int rc); void vop_remove_pre(void *a); void vop_remove_post(void *a, int rc); void vop_rename_post(void *a, int rc); void vop_rename_pre(void *a); void vop_rmdir_pre(void *a); void vop_rmdir_post(void *a, int rc); void vop_setattr_pre(void *a); void vop_setattr_post(void *a, int rc); void vop_setacl_pre(void *a); void vop_setacl_post(void *a, int rc); void vop_setextattr_pre(void *a); void vop_setextattr_post(void *a, int rc); void vop_symlink_pre(void *a); void vop_symlink_post(void *a, int rc); int vop_sigdefer(struct vop_vector *vop, struct vop_generic_args *a); #ifdef DEBUG_VFS_LOCKS void vop_fdatasync_debugpre(void *a); void vop_fdatasync_debugpost(void *a, int rc); void vop_fplookup_vexec_debugpre(void *a); void vop_fplookup_vexec_debugpost(void *a, int rc); void vop_fplookup_symlink_debugpre(void *a); void vop_fplookup_symlink_debugpost(void *a, int rc); void vop_fsync_debugpre(void *a); void vop_fsync_debugpost(void *a, int rc); void vop_strategy_debugpre(void *a); void vop_lock_debugpre(void *a); void vop_lock_debugpost(void *a, int rc); void vop_unlock_debugpre(void *a); void vop_need_inactive_debugpre(void *a); void vop_need_inactive_debugpost(void *a, int rc); void vop_mkdir_debugpost(void *a, int rc); #else #define vop_fdatasync_debugpre(x) do { } while (0) #define vop_fdatasync_debugpost(x, y) do { } while (0) #define vop_fplookup_vexec_debugpre(x) do { } while (0) #define vop_fplookup_vexec_debugpost(x, y) do { } while (0) #define vop_fplookup_symlink_debugpre(x) do { } while (0) #define vop_fplookup_symlink_debugpost(x, y) do { } while (0) #define vop_fsync_debugpre(x) do { } while (0) #define vop_fsync_debugpost(x, y) do { } while (0) #define vop_strategy_debugpre(x) do { } while (0) #define vop_lock_debugpre(x) do { } while (0) #define vop_lock_debugpost(x, y) do { } while (0) #define vop_unlock_debugpre(x) do { } while (0) #define vop_need_inactive_debugpre(x) do { } while (0) #define vop_need_inactive_debugpost(x, y) do { } while (0) #define vop_mkdir_debugpost(x, y) do { } while (0) #endif void vop_rename_fail(struct vop_rename_args *ap); #define vop_stat_helper_pre(ap) ({ \ struct vop_stat_args *_ap = (ap); \ int _error; \ AUDIT_ARG_VNODE1(ap->a_vp); \ _error = mac_vnode_check_stat(_ap->a_active_cred, _ap->a_file_cred, _ap->a_vp);\ if (__predict_true(_error == 0)) { \ ap->a_sb->st_padding0 = 0; \ ap->a_sb->st_padding1 = 0; \ bzero(_ap->a_sb->st_spare, sizeof(_ap->a_sb->st_spare)); \ } \ _error; \ }) #define vop_stat_helper_post(ap, error) ({ \ struct vop_stat_args *_ap = (ap); \ int _error = (error); \ if (priv_check_cred_vfs_generation(_ap->a_active_cred)) \ _ap->a_sb->st_gen = 0; \ _error; \ }) #define VOP_WRITE_PRE(ap) \ struct vattr va; \ int error; \ off_t osize, ooffset, noffset; \ \ osize = ooffset = noffset = 0; \ if (!VN_KNLIST_EMPTY((ap)->a_vp)) { \ error = VOP_GETATTR((ap)->a_vp, &va, (ap)->a_cred); \ if (error) \ return (error); \ ooffset = (ap)->a_uio->uio_offset; \ osize = (off_t)va.va_size; \ } #define VOP_WRITE_POST(ap, ret) \ noffset = (ap)->a_uio->uio_offset; \ if (noffset > ooffset && !VN_KNLIST_EMPTY((ap)->a_vp)) { \ VFS_KNOTE_LOCKED((ap)->a_vp, NOTE_WRITE \ | (noffset > osize ? NOTE_EXTEND : 0)); \ } #define VOP_LOCK(vp, flags) VOP_LOCK1(vp, flags, __FILE__, __LINE__) #ifdef INVARIANTS #define VOP_ADD_WRITECOUNT_CHECKED(vp, cnt) \ do { \ int error_; \ \ error_ = VOP_ADD_WRITECOUNT((vp), (cnt)); \ VNASSERT(error_ == 0, (vp), ("VOP_ADD_WRITECOUNT returned %d", \ error_)); \ } while (0) #define VOP_SET_TEXT_CHECKED(vp) \ do { \ int error_; \ \ error_ = VOP_SET_TEXT((vp)); \ VNASSERT(error_ == 0, (vp), ("VOP_SET_TEXT returned %d", \ error_)); \ } while (0) #define VOP_UNSET_TEXT_CHECKED(vp) \ do { \ int error_; \ \ error_ = VOP_UNSET_TEXT((vp)); \ VNASSERT(error_ == 0, (vp), ("VOP_UNSET_TEXT returned %d", \ error_)); \ } while (0) #else #define VOP_ADD_WRITECOUNT_CHECKED(vp, cnt) VOP_ADD_WRITECOUNT((vp), (cnt)) #define VOP_SET_TEXT_CHECKED(vp) VOP_SET_TEXT((vp)) #define VOP_UNSET_TEXT_CHECKED(vp) VOP_UNSET_TEXT((vp)) #endif #define VN_IS_DOOMED(vp) __predict_false((vn_irflag_read(vp) & VIRF_DOOMED) != 0) void vput(struct vnode *vp); void vrele(struct vnode *vp); void vref(struct vnode *vp); void vrefact(struct vnode *vp); void v_addpollinfo(struct vnode *vp); static __inline int vrefcnt(struct vnode *vp) { return (vp->v_usecount); } #define vholdl(vp) do { \ ASSERT_VI_LOCKED(vp, __func__); \ vhold(vp); \ } while (0) #define vrefl(vp) do { \ ASSERT_VI_LOCKED(vp, __func__); \ vref(vp); \ } while (0) /* * The caller doesn't know the file size and vnode_create_vobject() should * determine the size on its own. */ #define VNODE_NO_SIZE ((off_t)-1) int vnode_create_vobject(struct vnode *vp, off_t size, struct thread *td); int vnode_create_disk_vobject(struct vnode *vp, off_t size, struct thread *td); void vnode_destroy_vobject(struct vnode *vp); extern struct vop_vector fifo_specops; extern struct vop_vector dead_vnodeops; extern struct vop_vector default_vnodeops; #define VOP_PANIC ((void*)(uintptr_t)vop_panic) #define VOP_NULL ((void*)(uintptr_t)vop_null) #define VOP_EBADF ((void*)(uintptr_t)vop_ebadf) #define VOP_ENOTTY ((void*)(uintptr_t)vop_enotty) #define VOP_EINVAL ((void*)(uintptr_t)vop_einval) #define VOP_ENOENT ((void*)(uintptr_t)vop_enoent) #define VOP_EOPNOTSUPP ((void*)(uintptr_t)vop_eopnotsupp) #define VOP_EAGAIN ((void*)(uintptr_t)vop_eagain) /* fifo_vnops.c */ int fifo_printinfo(struct vnode *); /* vfs_hash.c */ typedef int vfs_hash_cmp_t(struct vnode *vp, void *arg); void vfs_hash_changesize(u_long newhashsize); int vfs_hash_get(const struct mount *mp, u_int hash, int flags, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg); u_int vfs_hash_index(struct vnode *vp); int vfs_hash_insert(struct vnode *vp, u_int hash, int flags, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg); void vfs_hash_ref(const struct mount *mp, u_int hash, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg); void vfs_hash_rehash(struct vnode *vp, u_int hash); void vfs_hash_remove(struct vnode *vp); int vfs_kqfilter(struct vop_kqfilter_args *); struct dirent; int vn_dir_next_dirent(struct vnode *vp, struct thread *td, char *dirbuf, size_t dirbuflen, struct dirent **dpp, size_t *len, off_t *off, int *eofflag); int vn_dir_check_empty(struct vnode *vp); int vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off); int vfs_unixify_accmode(accmode_t *accmode); void vfs_unp_reclaim(struct vnode *vp); int setfmode(struct thread *td, struct ucred *cred, struct vnode *vp, int mode); int setfown(struct thread *td, struct ucred *cred, struct vnode *vp, uid_t uid, gid_t gid); int vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td); int vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td); int vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *active_cred); int vn_getsize(struct vnode *vp, off_t *size, struct ucred *active_cred); void vn_fsid(struct vnode *vp, struct vattr *va); int vn_dir_check_exec(struct vnode *vp, struct componentname *cnp); int vn_lktype_write(struct mount *mp, struct vnode *vp); #ifdef INVARIANTS void vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state); #endif static inline void vn_set_state(struct vnode *vp, __enum_uint8(vstate) state) { #ifdef INVARIANTS vn_set_state_validate(vp, state); #endif vp->v_state = state; } static inline __enum_uint8(vstate) vn_get_state(struct vnode *vp) { return (vp->v_state); } #define VOP_UNLOCK_FLAGS(vp, flags) ({ \ struct vnode *_vp = (vp); \ int _flags = (flags); \ int _error; \ \ if ((_flags & ~(LK_INTERLOCK | LK_RELEASE)) != 0) \ panic("%s: unsupported flags %x\n", __func__, flags); \ _error = VOP_UNLOCK(_vp); \ if (_flags & LK_INTERLOCK) \ VI_UNLOCK(_vp); \ _error; \ }) #include #define VFS_VOP_VECTOR_REGISTER(vnodeops) \ SYSINIT(vfs_vector_##vnodeops##_f, SI_SUB_VFS, SI_ORDER_ANY, \ vfs_vector_op_register, &vnodeops) #define VFS_SMR_DECLARE \ extern smr_t vfs_smr #define VFS_SMR() vfs_smr #define vfs_smr_enter() smr_enter(VFS_SMR()) #define vfs_smr_exit() smr_exit(VFS_SMR()) #define vfs_smr_synchronize() smr_synchronize(VFS_SMR()) #define vfs_smr_entered_load(ptr) smr_entered_load((ptr), VFS_SMR()) #define VFS_SMR_ASSERT_ENTERED() SMR_ASSERT_ENTERED(VFS_SMR()) #define VFS_SMR_ASSERT_NOT_ENTERED() SMR_ASSERT_NOT_ENTERED(VFS_SMR()) #define VFS_SMR_ZONE_SET(zone) uma_zone_set_smr((zone), VFS_SMR()) #define vn_load_v_data_smr(vp) ({ \ struct vnode *_vp = (vp); \ \ VFS_SMR_ASSERT_ENTERED(); \ atomic_load_consume_ptr(&(_vp)->v_data);\ }) #endif /* _KERNEL */ #endif /* !_SYS_VNODE_H_ */