/*
* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011 Whamcloud, Inc.
+ *
*/
/*
* This file is part of Lustre, http://www.lustre.org/
static int cl_lock_invariant_trusted(const struct lu_env *env,
const struct cl_lock *lock)
{
- return
- cl_is_lock(lock) &&
- ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
+ return ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
cfs_atomic_read(&lock->cll_ref) >= lock->cll_holds &&
lock->cll_holds >= lock->cll_users &&
lock->cll_holds >= 0 &&
{
struct cl_object *obj = lock->cll_descr.cld_obj;
- LASSERT(cl_is_lock(lock));
LINVRNT(!cl_lock_is_mutexed(lock));
ENTRY;
{
struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
- LASSERT(cl_is_lock(lock));
CDEBUG(D_TRACE, "acquiring trusted reference: %d %p %lu\n",
cfs_atomic_read(&lock->cll_ref), lock, RETIP);
if (cfs_atomic_inc_return(&lock->cll_ref) == 1)
cfs_list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
int matched;
- LASSERT(cl_is_lock(lock));
matched = cl_lock_ext_match(&lock->cll_descr, need) &&
lock->cll_state < CLS_FREEING &&
lock->cll_error == 0 &&
struct cl_object_header *head;
struct cl_object *obj;
struct cl_lock *lock;
- int ok;
obj = need->cld_obj;
head = cl_object_header(obj);
cl_lock_mutex_get(env, lock);
if (lock->cll_state == CLS_INTRANSIT)
cl_lock_state_wait(env, lock); /* Don't care return value. */
- if (lock->cll_state == CLS_CACHED) {
- int result;
- result = cl_use_try(env, lock, 1);
- if (result < 0)
- cl_lock_error(env, lock, result);
- }
- ok = lock->cll_state == CLS_HELD;
- if (ok) {
- cl_lock_hold_add(env, lock, scope, source);
- cl_lock_user_add(env, lock);
+ cl_lock_hold_add(env, lock, scope, source);
+ cl_lock_user_add(env, lock);
+ if (lock->cll_state == CLS_CACHED)
+ cl_use_try(env, lock, 1);
+ if (lock->cll_state == CLS_HELD) {
+ cl_lock_mutex_put(env, lock);
+ cl_lock_lockdep_acquire(env, lock, 0);
cl_lock_put(env, lock);
- }
- cl_lock_mutex_put(env, lock);
- if (!ok) {
+ } else {
+ cl_unuse_try(env, lock);
+ cl_lock_unhold(env, lock, scope, source);
+ cl_lock_mutex_put(env, lock);
cl_lock_put(env, lock);
lock = NULL;
}
lu_ref_del(&lock->cll_holders, scope, source);
cl_lock_hold_mod(env, lock, -1);
if (lock->cll_holds == 0) {
+ CL_LOCK_ASSERT(lock->cll_state != CLS_HELD, env, lock);
if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
- lock->cll_descr.cld_mode == CLM_GROUP)
+ lock->cll_descr.cld_mode == CLM_GROUP ||
+ lock->cll_state != CLS_CACHED)
/*
* If lock is still phantom or grouplock when user is
* done with it---destroy the lock.
int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
{
cfs_waitlink_t waiter;
+ cfs_sigset_t blocked;
int result;
ENTRY;
cl_lock_trace(D_DLMTRACE, env, "state wait lock", lock);
result = lock->cll_error;
if (result == 0) {
+ /* To avoid being interrupted by the 'non-fatal' signals
+ * (SIGCHLD, for instance), we'd block them temporarily.
+ * LU-305 */
+ blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
+
cfs_waitlink_init(&waiter);
cfs_waitq_add(&lock->cll_wq, &waiter);
cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
cl_lock_mutex_put(env, lock);
LASSERT(cl_lock_nr_mutexed(env) == 0);
- cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
+
+ result = -EINTR;
+ if (likely(!OBD_FAIL_CHECK(OBD_FAIL_LOCK_STATE_WAIT_INTR))) {
+ cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
+ if (!cfs_signal_pending())
+ result = 0;
+ }
cl_lock_mutex_get(env, lock);
cfs_set_current_state(CFS_TASK_RUNNING);
cfs_waitq_del(&lock->cll_wq, &waiter);
- result = cfs_signal_pending() ? -EINTR : 0;
+
+ /* Restore old blocked signals */
+ cfs_restore_sigs(blocked);
}
RETURN(result);
}
ENTRY;
cl_lock_trace(D_DLMTRACE, env, "enqueue lock", lock);
do {
- result = 0;
-
LINVRNT(cl_lock_is_mutexed(lock));
- if (lock->cll_error != 0)
+ result = lock->cll_error;
+ if (result != 0)
break;
+
switch (lock->cll_state) {
case CLS_NEW:
cl_lock_state_set(env, lock, CLS_QUEUING);
LBUG();
}
} while (result == CLO_REPEAT);
- if (result < 0)
- cl_lock_error(env, lock, result);
- RETURN(result ?: lock->cll_error);
+ RETURN(result);
}
EXPORT_SYMBOL(cl_enqueue_try);
+/**
+ * Cancel the conflicting lock found during previous enqueue.
+ *
+ * \retval 0 conflicting lock has been canceled.
+ * \retval -ve error code.
+ */
+int cl_lock_enqueue_wait(const struct lu_env *env,
+ struct cl_lock *lock,
+ int keep_mutex)
+{
+ struct cl_lock *conflict;
+ int rc = 0;
+ ENTRY;
+
+ LASSERT(cl_lock_is_mutexed(lock));
+ LASSERT(lock->cll_state == CLS_QUEUING);
+ LASSERT(lock->cll_conflict != NULL);
+
+ conflict = lock->cll_conflict;
+ lock->cll_conflict = NULL;
+
+ cl_lock_mutex_put(env, lock);
+ LASSERT(cl_lock_nr_mutexed(env) == 0);
+
+ cl_lock_mutex_get(env, conflict);
+ cl_lock_trace(D_DLMTRACE, env, "enqueue wait", conflict);
+ cl_lock_cancel(env, conflict);
+ cl_lock_delete(env, conflict);
+
+ while (conflict->cll_state != CLS_FREEING) {
+ rc = cl_lock_state_wait(env, conflict);
+ if (rc != 0)
+ break;
+ }
+ cl_lock_mutex_put(env, conflict);
+ lu_ref_del(&conflict->cll_reference, "cancel-wait", lock);
+ cl_lock_put(env, conflict);
+
+ if (keep_mutex)
+ cl_lock_mutex_get(env, lock);
+
+ LASSERT(rc <= 0);
+ RETURN(rc);
+}
+EXPORT_SYMBOL(cl_lock_enqueue_wait);
+
static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
struct cl_io *io, __u32 enqflags)
{
do {
result = cl_enqueue_try(env, lock, io, enqflags);
if (result == CLO_WAIT) {
- result = cl_lock_state_wait(env, lock);
+ if (lock->cll_conflict != NULL)
+ result = cl_lock_enqueue_wait(env, lock, 1);
+ else
+ result = cl_lock_state_wait(env, lock);
if (result == 0)
continue;
}
break;
} while (1);
- if (result != 0) {
- cl_lock_user_del(env, lock);
- cl_lock_error(env, lock, result);
- }
+ if (result != 0)
+ cl_unuse_try(env, lock);
LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
lock->cll_state == CLS_HELD));
RETURN(result);
/**
* Tries to unlock a lock.
*
- * This function is called repeatedly by cl_unuse() until either lock is
- * unlocked, or error occurs.
- * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
- *
- * \pre lock->cll_state == CLS_HELD
+ * This function is called to release underlying resource:
+ * 1. for top lock, the resource is sublocks it held;
+ * 2. for sublock, the resource is the reference to dlmlock.
*
- * \post ergo(result == 0, lock->cll_state == CLS_CACHED)
+ * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
*
* \see cl_unuse() cl_lock_operations::clo_unuse()
* \see cl_lock_state::CLS_CACHED
ENTRY;
cl_lock_trace(D_DLMTRACE, env, "unuse lock", lock);
- LASSERT(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED);
if (lock->cll_users > 1) {
cl_lock_user_del(env, lock);
RETURN(0);
}
+ /* Only if the lock is in CLS_HELD or CLS_ENQUEUED state, it can hold
+ * underlying resources. */
+ if (!(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED)) {
+ cl_lock_user_del(env, lock);
+ RETURN(0);
+ }
+
/*
* New lock users (->cll_users) are not protecting unlocking
* from proceeding. From this point, lock eventually reaches
*/
result = 0;
} else {
- CERROR("result = %d, this is unlikely!\n", result);
- cl_lock_extransit(env, lock, state);
+ CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
+ /* Set the lock state to CLS_NEW so it will be destroyed.
+ * In lov_lock_unuse() it will release sublocks even if error
+ * occurs. */
+ cl_lock_extransit(env, lock, CLS_NEW);
}
-
- result = result ?: lock->cll_error;
- if (result < 0)
- cl_lock_error(env, lock, result);
- RETURN(result);
+ RETURN(result ?: lock->cll_error);
}
EXPORT_SYMBOL(cl_unuse_try);
-static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
-{
- int result;
- ENTRY;
-
- result = cl_unuse_try(env, lock);
- if (result)
- CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
-
- EXIT;
-}
-
/**
* Unlocks a lock.
*/
{
ENTRY;
cl_lock_mutex_get(env, lock);
- cl_unuse_locked(env, lock);
+ cl_unuse_try(env, lock);
cl_lock_mutex_put(env, lock);
cl_lock_lockdep_release(env, lock);
EXIT;
LASSERT(lock->cll_users > 0);
LASSERT(lock->cll_holds > 0);
- result = 0;
- if (lock->cll_error != 0)
+ result = lock->cll_error;
+ if (result != 0)
break;
if (cl_lock_is_intransit(lock)) {
cl_lock_state_set(env, lock, CLS_HELD);
}
} while (result == CLO_REPEAT);
- RETURN(result ?: lock->cll_error);
+ RETURN(result);
}
EXPORT_SYMBOL(cl_wait_try);
LASSERTF(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD,
"Wrong state %d \n", lock->cll_state);
LASSERT(lock->cll_holds > 0);
- cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
do {
result = cl_wait_try(env, lock);
break;
} while (1);
if (result < 0) {
- cl_lock_user_del(env, lock);
- cl_lock_error(env, lock, result);
+ cl_unuse_try(env, lock);
cl_lock_lockdep_release(env, lock);
}
+ cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
cl_lock_mutex_put(env, lock);
LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
RETURN(result);
LINVRNT(cl_lock_invariant(env, lock));
ENTRY;
- cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
if (lock->cll_error == 0 && error != 0) {
+ cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
lock->cll_error = error;
cl_lock_signal(env, lock);
cl_lock_cancel(env, lock);
* Finds an existing lock covering given page and optionally different from a
* given \a except lock.
*/
-struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
- struct cl_page *page, struct cl_lock *except,
- int pending, int canceld)
+struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env, struct cl_object *obj,
+ pgoff_t index, struct cl_lock *except,
+ int pending, int canceld)
{
struct cl_object_header *head;
struct cl_lock *scan;
need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
* not PHANTOM */
- need->cld_start = need->cld_end = page->cp_index;
+ need->cld_start = need->cld_end = index;
need->cld_enq_flags = 0;
cfs_spin_lock(&head->coh_lock_guard);
cfs_spin_unlock(&head->coh_lock_guard);
RETURN(lock);
}
-EXPORT_SYMBOL(cl_lock_at_page);
+EXPORT_SYMBOL(cl_lock_at_pgoff);
/**
- * Returns a list of pages protected (only) by a given lock.
- *
- * Scans an extent of page radix tree, corresponding to the \a lock and queues
- * all pages that are not protected by locks other than \a lock into \a queue.
+ * Calculate the page offset at the layer of @lock.
+ * At the time of this writing, @page is top page and @lock is sub lock.
*/
-void cl_lock_page_list_fixup(const struct lu_env *env,
- struct cl_io *io, struct cl_lock *lock,
- struct cl_page_list *queue)
+static pgoff_t pgoff_at_lock(struct cl_page *page, struct cl_lock *lock)
{
- struct cl_page *page;
- struct cl_page *temp;
- struct cl_page_list *plist = &cl_env_info(env)->clt_list;
+ struct lu_device_type *dtype;
+ const struct cl_page_slice *slice;
- LINVRNT(cl_lock_invariant(env, lock));
- ENTRY;
-
- /* Now, we have a list of cl_pages under the \a lock, we need
- * to check if some of pages are covered by other ldlm lock.
- * If this is the case, they aren't needed to be written out this time.
- *
- * For example, we have A:[0,200] & B:[100,300] PW locks on client, now
- * the latter is to be canceled, this means other client is
- * reading/writing [200,300] since A won't canceled. Actually
- * we just need to write the pages covered by [200,300]. This is safe,
- * since [100,200] is also protected lock A.
- */
+ dtype = lock->cll_descr.cld_obj->co_lu.lo_dev->ld_type;
+ slice = cl_page_at(page, dtype);
+ LASSERT(slice != NULL);
+ return slice->cpl_page->cp_index;
+}
- cl_page_list_init(plist);
- cl_page_list_for_each_safe(page, temp, queue) {
- pgoff_t idx = page->cp_index;
- struct cl_lock *found;
- struct cl_lock_descr *descr;
-
- /* The algorithm counts on the index-ascending page index. */
- LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
- page->cp_index < temp->cp_index));
-
- found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
- page, lock, 0, 0);
- if (found == NULL)
- continue;
-
- descr = &found->cll_descr;
- cfs_list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
- cp_batch) {
- idx = page->cp_index;
- if (descr->cld_start > idx || descr->cld_end < idx)
- break;
- cl_page_list_move(plist, queue, page);
+/**
+ * Check if page @page is covered by an extra lock or discard it.
+ */
+static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
+ struct cl_page *page, void *cbdata)
+{
+ struct cl_thread_info *info = cl_env_info(env);
+ struct cl_lock *lock = cbdata;
+ pgoff_t index = pgoff_at_lock(page, lock);
+
+ if (index >= info->clt_fn_index) {
+ struct cl_lock *tmp;
+
+ /* refresh non-overlapped index */
+ tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj, index, lock,
+ 1, 0);
+ if (tmp != NULL) {
+ /* Cache the first-non-overlapped index so as to skip
+ * all pages within [index, clt_fn_index). This
+ * is safe because if tmp lock is canceled, it will
+ * discard these pages. */
+ info->clt_fn_index = tmp->cll_descr.cld_end + 1;
+ if (tmp->cll_descr.cld_end == CL_PAGE_EOF)
+ info->clt_fn_index = CL_PAGE_EOF;
+ cl_lock_put(env, tmp);
+ } else if (cl_page_own(env, io, page) == 0) {
+ /* discard the page */
+ cl_page_unmap(env, io, page);
+ cl_page_discard(env, io, page);
+ cl_page_disown(env, io, page);
+ } else {
+ LASSERT(page->cp_state == CPS_FREEING);
}
- cl_lock_put(env, found);
}
- /* The pages in plist are covered by other locks, don't handle them
- * this time.
- */
- if (io != NULL)
- cl_page_list_disown(env, io, plist);
- cl_page_list_fini(env, plist);
- EXIT;
+ info->clt_next_index = index + 1;
+ return CLP_GANG_OKAY;
+}
+
+static int pageout_cb(const struct lu_env *env, struct cl_io *io,
+ struct cl_page *page, void *cbdata)
+{
+ struct cl_thread_info *info = cl_env_info(env);
+ struct cl_page_list *queue = &info->clt_queue.c2_qin;
+ struct cl_lock *lock = cbdata;
+ typeof(cl_page_own) *page_own;
+ int rc = CLP_GANG_OKAY;
+
+ page_own = queue->pl_nr ? cl_page_own_try : cl_page_own;
+ if (page_own(env, io, page) == 0) {
+ cl_page_list_add(queue, page);
+ info->clt_next_index = pgoff_at_lock(page, lock) + 1;
+ } else if (page->cp_state != CPS_FREEING) {
+ /* cl_page_own() won't fail unless
+ * the page is being freed. */
+ LASSERT(queue->pl_nr != 0);
+ rc = CLP_GANG_AGAIN;
+ }
+
+ return rc;
}
-EXPORT_SYMBOL(cl_lock_page_list_fixup);
/**
* Invalidate pages protected by the given lock, sending them out to the
struct cl_io *io = &info->clt_io;
struct cl_2queue *queue = &info->clt_queue;
struct cl_lock_descr *descr = &lock->cll_descr;
+ cl_page_gang_cb_t cb;
long page_count;
- int nonblock = 1, resched;
+ int res;
int result;
LINVRNT(cl_lock_invariant(env, lock));
if (result != 0)
GOTO(out, result);
+ cb = descr->cld_mode == CLM_READ ? check_and_discard_cb : pageout_cb;
+ info->clt_fn_index = info->clt_next_index = descr->cld_start;
do {
cl_2queue_init(queue);
- cl_page_gang_lookup(env, descr->cld_obj, io, descr->cld_start,
- descr->cld_end, &queue->c2_qin, nonblock,
- &resched);
+ res = cl_page_gang_lookup(env, descr->cld_obj, io,
+ info->clt_next_index, descr->cld_end,
+ cb, (void *)lock);
page_count = queue->c2_qin.pl_nr;
if (page_count > 0) {
+ /* must be writeback case */
+ LASSERTF(descr->cld_mode >= CLM_WRITE, "lock mode %s\n",
+ cl_lock_mode_name(descr->cld_mode));
+
result = cl_page_list_unmap(env, io, &queue->c2_qin);
if (!discard) {
long timeout = 600; /* 10 minutes. */
CWARN("Writing %lu pages error: %d\n",
page_count, result);
}
- cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
cl_2queue_discard(env, io, queue);
cl_2queue_disown(env, io, queue);
+ cl_2queue_fini(env, queue);
}
- cl_2queue_fini(env, queue);
- if (resched)
+ if (info->clt_next_index > descr->cld_end)
+ break;
+
+ if (res == CLP_GANG_RESCHED)
cfs_cond_resched();
- } while (resched || nonblock--);
+ } while (res != CLP_GANG_OKAY);
out:
cl_io_fini(env, io);
RETURN(result);
}
EXPORT_SYMBOL(cl_locks_prune);
-/**
- * Returns true if \a addr is an address of an allocated cl_lock. Used in
- * assertions. This check is optimistically imprecise, i.e., it occasionally
- * returns true for the incorrect addresses, but if it returns false, then the
- * address is guaranteed to be incorrect. (Should be named cl_lockp().)
- *
- * \see cl_is_page()
- */
-int cl_is_lock(const void *addr)
-{
- return cfs_mem_is_in_cache(addr, cl_lock_kmem);
-}
-EXPORT_SYMBOL(cl_is_lock);
-
static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
const struct cl_io *io,
const struct cl_lock_descr *need,
lock, enqflags);
break;
}
- cl_unuse_locked(env, lock);
+ cl_unuse_try(env, lock);
}
cl_lock_trace(D_DLMTRACE, env, "enqueue failed", lock);
cl_lock_hold_release(env, lock, scope, source);