+ * 0 if the lock is not compatible
+ * 1 if the lock is compatible
+ * 2 if this group lock is compatible and requires no further checking
+ * negative error, such as EWOULDBLOCK for group locks
+ */
+static int
+ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
+ int *flags, ldlm_error_t *err,
+ struct list_head *work_list, int *contended_locks)
+{
+ struct list_head *tmp;
+ struct ldlm_lock *lock;
+ struct ldlm_resource *res = req->l_resource;
+ ldlm_mode_t req_mode = req->l_req_mode;
+ __u64 req_start = req->l_req_extent.start;
+ __u64 req_end = req->l_req_extent.end;
+ int compat = 1;
+ int scan = 0;
+ int check_contention;
+ ENTRY;
+
+ lockmode_verify(req_mode);
+
+ /* Using interval tree for granted lock */
+ if (queue == &res->lr_granted) {
+ struct ldlm_interval_tree *tree;
+ struct ldlm_extent_compat_args data = {.work_list = work_list,
+ .lock = req,
+ .locks = contended_locks,
+ .compat = &compat };
+ struct interval_node_extent ex = { .start = req_start,
+ .end = req_end };
+ int idx, rc;
+
+ for (idx = 0; idx < LCK_MODE_NUM; idx++) {
+ tree = &res->lr_itree[idx];
+ if (tree->lit_root == NULL) /* empty tree, skipped */
+ continue;
+
+ data.mode = tree->lit_mode;
+ if (lockmode_compat(req_mode, tree->lit_mode)) {
+ struct ldlm_interval *node;
+ struct ldlm_extent *extent;
+
+ if (req_mode != LCK_GROUP)
+ continue;
+
+ /* group lock, grant it immediately if
+ * compatible */
+ node = to_ldlm_interval(tree->lit_root);
+ extent = ldlm_interval_extent(node);
+ if (req->l_policy_data.l_extent.gid ==
+ extent->gid)
+ RETURN(2);
+ }
+
+ if (tree->lit_mode == LCK_GROUP) {
+ if (*flags & LDLM_FL_BLOCK_NOWAIT) {
+ compat = -EWOULDBLOCK;
+ goto destroylock;
+ }
+
+ *flags |= LDLM_FL_NO_TIMEOUT;
+ if (!work_list)
+ RETURN(0);
+
+ /* if work list is not NULL,add all
+ locks in the tree to work list */
+ compat = 0;
+ interval_iterate(tree->lit_root,
+ ldlm_extent_compat_cb, &data);
+ continue;
+ }
+
+ if (!work_list) {
+ rc = interval_is_overlapped(tree->lit_root,&ex);
+ if (rc)
+ RETURN(0);
+ } else {
+ interval_search(tree->lit_root, &ex,
+ ldlm_extent_compat_cb, &data);
+ if (!list_empty(work_list) && compat)
+ compat = 0;
+ }
+ }
+ } else { /* for waiting queue */
+ list_for_each(tmp, queue) {
+ check_contention = 1;
+
+ lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+
+ if (req == lock)
+ break;
+
+ if (unlikely(scan)) {
+ /* We only get here if we are queuing GROUP lock
+ and met some incompatible one. The main idea of this
+ code is to insert GROUP lock past compatible GROUP
+ lock in the waiting queue or if there is not any,
+ then in front of first non-GROUP lock */
+ if (lock->l_req_mode != LCK_GROUP) {
+ /* Ok, we hit non-GROUP lock, there should
+ * be no more GROUP locks later on, queue in
+ * front of first non-GROUP lock */
+
+ ldlm_resource_insert_lock_after(lock, req);
+ list_del_init(&lock->l_res_link);
+ ldlm_resource_insert_lock_after(req, lock);
+ compat = 0;
+ break;
+ }
+ if (req->l_policy_data.l_extent.gid ==
+ lock->l_policy_data.l_extent.gid) {
+ /* found it */
+ ldlm_resource_insert_lock_after(lock, req);
+ compat = 0;
+ break;
+ }
+ continue;
+ }
+
+ /* locks are compatible, overlap doesn't matter */
+ if (lockmode_compat(lock->l_req_mode, req_mode)) {
+ if (req_mode == LCK_PR &&
+ ((lock->l_policy_data.l_extent.start <=
+ req->l_policy_data.l_extent.start) &&
+ (lock->l_policy_data.l_extent.end >=
+ req->l_policy_data.l_extent.end))) {
+ /* If we met a PR lock just like us or wider,
+ and nobody down the list conflicted with
+ it, that means we can skip processing of
+ the rest of the list and safely place
+ ourselves at the end of the list, or grant
+ (dependent if we met an conflicting locks
+ before in the list).
+ In case of 1st enqueue only we continue
+ traversing if there is something conflicting
+ down the list because we need to make sure
+ that something is marked as AST_SENT as well,
+ in cse of empy worklist we would exit on
+ first conflict met. */
+ /* There IS a case where such flag is
+ not set for a lock, yet it blocks
+ something. Luckily for us this is
+ only during destroy, so lock is
+ exclusive. So here we are safe */
+ if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
+ RETURN(compat);
+ }
+ }
+
+ /* non-group locks are compatible, overlap doesn't
+ matter */
+ if (likely(req_mode != LCK_GROUP))
+ continue;
+
+ /* If we are trying to get a GROUP lock and there is
+ another one of this kind, we need to compare gid */
+ if (req->l_policy_data.l_extent.gid ==
+ lock->l_policy_data.l_extent.gid) {
+ /* If existing lock with matched gid is granted,
+ we grant new one too. */
+ if (lock->l_req_mode == lock->l_granted_mode)
+ RETURN(2);
+
+ /* Otherwise we are scanning queue of waiting
+ * locks and it means current request would
+ * block along with existing lock (that is
+ * already blocked.
+ * If we are in nonblocking mode - return
+ * immediately */
+ if (*flags & LDLM_FL_BLOCK_NOWAIT) {
+ compat = -EWOULDBLOCK;
+ goto destroylock;
+ }
+ /* If this group lock is compatible with another
+ * group lock on the waiting list, they must be
+ * together in the list, so they can be granted
+ * at the same time. Otherwise the later lock
+ * can get stuck behind another, incompatible,
+ * lock. */
+ ldlm_resource_insert_lock_after(lock, req);
+ /* Because 'lock' is not granted, we can stop
+ * processing this queue and return immediately.
+ * There is no need to check the rest of the
+ * list. */
+ RETURN(0);
+ }
+ }
+
+ if (unlikely(req_mode == LCK_GROUP &&
+ (lock->l_req_mode != lock->l_granted_mode))) {
+ scan = 1;
+ compat = 0;
+ if (lock->l_req_mode != LCK_GROUP) {
+ /* Ok, we hit non-GROUP lock, there should be no
+ more GROUP locks later on, queue in front of
+ first non-GROUP lock */
+
+ ldlm_resource_insert_lock_after(lock, req);
+ list_del_init(&lock->l_res_link);
+ ldlm_resource_insert_lock_after(req, lock);
+ break;
+ }
+ if (req->l_policy_data.l_extent.gid ==
+ lock->l_policy_data.l_extent.gid) {
+ /* found it */
+ ldlm_resource_insert_lock_after(lock, req);
+ break;
+ }
+ continue;
+ }
+
+ if (unlikely(lock->l_req_mode == LCK_GROUP)) {
+ /* If compared lock is GROUP, then requested is PR/PW/
+ * so this is not compatible; extent range does not
+ * matter */
+ if (*flags & LDLM_FL_BLOCK_NOWAIT) {
+ compat = -EWOULDBLOCK;
+ goto destroylock;
+ } else {
+ *flags |= LDLM_FL_NO_TIMEOUT;
+ }
+ } else if (lock->l_policy_data.l_extent.end < req_start ||
+ lock->l_policy_data.l_extent.start > req_end) {
+ /* if a non group lock doesn't overlap skip it */
+ continue;
+ } else if (lock->l_req_extent.end < req_start ||
+ lock->l_req_extent.start > req_end) {
+ /* false contention, the requests doesn't really overlap */
+ check_contention = 0;
+ }
+
+ if (!work_list)
+ RETURN(0);
+
+ /* don't count conflicting glimpse locks */
+ if (lock->l_req_mode == LCK_PR &&
+ lock->l_policy_data.l_extent.start == 0 &&
+ lock->l_policy_data.l_extent.end == OBD_OBJECT_EOF)
+ check_contention = 0;
+
+ *contended_locks += check_contention;
+
+ compat = 0;
+ if (lock->l_blocking_ast)
+ ldlm_add_ast_work_item(lock, req, work_list);
+ }
+ }
+
+ if (ldlm_check_contention(req, *contended_locks) &&
+ compat == 0 &&
+ (*flags & LDLM_FL_DENY_ON_CONTENTION) &&
+ req->l_req_mode != LCK_GROUP &&
+ req_end - req_start <=
+ req->l_resource->lr_namespace->ns_max_nolock_size)
+ GOTO(destroylock, compat = -EUSERS);
+
+ RETURN(compat);
+destroylock:
+ list_del_init(&req->l_res_link);
+ ldlm_lock_destroy_nolock(req);
+ *err = compat;
+ RETURN(compat);
+}
+
+static void discard_bl_list(struct list_head *bl_list)
+{
+ struct list_head *tmp, *pos;
+ ENTRY;
+
+ list_for_each_safe(pos, tmp, bl_list) {
+ struct ldlm_lock *lock =
+ list_entry(pos, struct ldlm_lock, l_bl_ast);
+
+ list_del_init(&lock->l_bl_ast);
+ LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
+ lock->l_flags &= ~LDLM_FL_AST_SENT;
+ LASSERT(lock->l_bl_ast_run == 0);
+ LASSERT(lock->l_blocking_lock);
+ LDLM_LOCK_RELEASE(lock->l_blocking_lock);
+ lock->l_blocking_lock = NULL;
+ LDLM_LOCK_RELEASE(lock);
+ }
+ EXIT;
+}
+
+/* If first_enq is 0 (ie, called from ldlm_reprocess_queue):
+ * - blocking ASTs have already been sent
+ * - must call this function with the ns lock held
+ *
+ * If first_enq is 1 (ie, called from ldlm_lock_enqueue):
+ * - blocking ASTs have not been sent
+ * - must call this function with the ns lock held once */
+int ldlm_process_extent_lock(struct ldlm_lock *lock, int *flags, int first_enq,
+ ldlm_error_t *err, struct list_head *work_list)
+{
+ struct ldlm_resource *res = lock->l_resource;
+ CFS_LIST_HEAD(rpc_list);
+ int rc, rc2;
+ int contended_locks = 0;
+ ENTRY;
+
+ LASSERT(list_empty(&res->lr_converting));
+ LASSERT(!(*flags & LDLM_FL_DENY_ON_CONTENTION) ||
+ !(lock->l_flags & LDLM_AST_DISCARD_DATA));
+ check_res_locked(res);
+ *err = ELDLM_OK;
+
+ if (!first_enq) {
+ /* Careful observers will note that we don't handle -EWOULDBLOCK
+ * here, but it's ok for a non-obvious reason -- compat_queue
+ * can only return -EWOULDBLOCK if (flags & BLOCK_NOWAIT).
+ * flags should always be zero here, and if that ever stops
+ * being true, we want to find out. */
+ LASSERT(*flags == 0);
+ rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags,
+ err, NULL, &contended_locks);
+ if (rc == 1) {
+ rc = ldlm_extent_compat_queue(&res->lr_waiting, lock,
+ flags, err, NULL,
+ &contended_locks);
+ }
+ if (rc == 0)
+ RETURN(LDLM_ITER_STOP);
+
+ ldlm_resource_unlink_lock(lock);
+
+ if (!OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_EVICT_RACE))
+ ldlm_extent_policy(res, lock, flags);
+ ldlm_grant_lock(lock, work_list);
+ RETURN(LDLM_ITER_CONTINUE);
+ }
+
+ restart:
+ contended_locks = 0;
+ rc = ldlm_extent_compat_queue(&res->lr_granted, lock, flags, err,
+ &rpc_list, &contended_locks);
+ if (rc < 0)
+ GOTO(out, rc); /* lock was destroyed */
+ if (rc == 2)
+ goto grant;
+
+ rc2 = ldlm_extent_compat_queue(&res->lr_waiting, lock, flags, err,
+ &rpc_list, &contended_locks);
+ if (rc2 < 0)
+ GOTO(out, rc = rc2); /* lock was destroyed */
+
+ if (rc + rc2 == 2) {
+ grant:
+ ldlm_extent_policy(res, lock, flags);
+ ldlm_resource_unlink_lock(lock);
+ ldlm_grant_lock(lock, NULL);
+ } else {
+ /* If either of the compat_queue()s returned failure, then we
+ * have ASTs to send and must go onto the waiting list.
+ *
+ * bug 2322: we used to unlink and re-add here, which was a
+ * terrible folly -- if we goto restart, we could get
+ * re-ordered! Causes deadlock, because ASTs aren't sent! */
+ if (list_empty(&lock->l_res_link))
+ ldlm_resource_add_lock(res, &res->lr_waiting, lock);
+ unlock_res(res);
+ rc = ldlm_run_ast_work(&rpc_list, LDLM_WORK_BL_AST);
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
+ !ns_is_client(res->lr_namespace))
+ class_fail_export(lock->l_export);
+
+ lock_res(res);
+ if (rc == -ERESTART) {
+
+ /* 15715: The lock was granted and destroyed after
+ * resource lock was dropped. Interval node was freed
+ * in ldlm_lock_destroy. Anyway, this always happens
+ * when a client is being evicted. So it would be
+ * ok to return an error. -jay */
+ if (lock->l_destroyed) {
+ *err = -EAGAIN;
+ GOTO(out, rc = -EAGAIN);
+ }
+
+ /* lock was granted while resource was unlocked. */
+ if (lock->l_granted_mode == lock->l_req_mode) {
+ /* bug 11300: if the lock has been granted,
+ * break earlier because otherwise, we will go
+ * to restart and ldlm_resource_unlink will be
+ * called and it causes the interval node to be
+ * freed. Then we will fail at
+ * ldlm_extent_add_lock() */
+ *flags &= ~(LDLM_FL_BLOCK_GRANTED | LDLM_FL_BLOCK_CONV |
+ LDLM_FL_BLOCK_WAIT);
+ GOTO(out, rc = 0);
+ }
+
+ GOTO(restart, -ERESTART);
+ }
+
+ *flags |= LDLM_FL_BLOCK_GRANTED;
+ /* this way we force client to wait for the lock
+ * endlessly once the lock is enqueued -bzzz */
+ *flags |= LDLM_FL_NO_TIMEOUT;
+
+ }
+ RETURN(0);
+out:
+ if (!list_empty(&rpc_list)) {
+ LASSERT(!(lock->l_flags & LDLM_AST_DISCARD_DATA));
+ discard_bl_list(&rpc_list);
+ }
+ RETURN(rc);
+}
+
+/* When a lock is cancelled by a client, the KMS may undergo change if this
+ * is the "highest lock". This function returns the new KMS value.
+ * Caller must hold ns_lock already.