- struct ldlm_resource *res = lock->l_resource;
- struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
- int rc;
- ENTRY;
-
- LASSERT(list_empty(&res->lr_converting));
-
- if (!first_enq) {
- LASSERT(res->lr_tmp != NULL);
- rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, 0);
- if (!rc)
- RETURN(LDLM_ITER_STOP);
- rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, 0);
- if (!rc)
- RETURN(LDLM_ITER_STOP);
-
- ldlm_resource_unlink_lock(lock);
- ldlm_grant_lock(lock, NULL, 0, 1);
- RETURN(LDLM_ITER_CONTINUE);
- }
-
- restart:
- LASSERT(res->lr_tmp == NULL);
- res->lr_tmp = &rpc_list;
- rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, 1);
- rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, 1);
- res->lr_tmp = NULL;
-
- if (rc != 2) {
- /* If either of the compat_queue()s returned 0, 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);
- l_unlock(&res->lr_namespace->ns_lock);
- rc = ldlm_run_ast_work(res->lr_namespace, &rpc_list);
- l_lock(&res->lr_namespace->ns_lock);
- if (rc == -ERESTART)
- GOTO(restart, -ERESTART);
- *flags |= LDLM_FL_BLOCK_GRANTED;
- } else {
- ldlm_resource_unlink_lock(lock);
- ldlm_grant_lock(lock, NULL, 0, 0);
- }
- RETURN(0);
+ struct ldlm_resource *res = lock->l_resource;
+ struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
+ NULL : work_list;
+ int rc;
+
+ ENTRY;
+
+ LASSERT(!ldlm_is_granted(lock));
+ check_res_locked(res);
+
+ if (intention == LDLM_PROCESS_RESCAN) {
+ struct list_head *bl_list;
+
+ if (*flags & LDLM_FL_BLOCK_NOWAIT) {
+ bl_list = NULL;
+ *err = ELDLM_LOCK_WOULDBLOCK;
+ } else {
+ bl_list = work_list;
+ *err = ELDLM_LOCK_ABORTED;
+ }
+
+ LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
+
+ /* It is possible that some of granted locks was not canceled
+ * but converted and is kept in granted queue. So there is
+ * a window where lock with 'ast_sent' might become granted
+ * again. Meanwhile a new lock may appear in that window and
+ * conflicts with the converted lock so the following scenario
+ * is possible:
+ *
+ * 1) lock1 conflicts with lock2
+ * 2) bl_ast was sent for lock2
+ * 3) lock3 comes and conflicts with lock2 too
+ * 4) no bl_ast sent because lock2->l_bl_ast_sent is 1
+ * 5) lock2 was converted for lock1 but not for lock3
+ * 6) lock1 granted, lock3 still is waiting for lock2, but
+ * there will never be another bl_ast for that
+ *
+ * To avoid this scenario the work_list is used below to collect
+ * any blocked locks from granted queue during every reprocess
+ * and bl_ast will be sent if needed.
+ */
+ rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
+ bl_list);
+ if (!rc)
+ RETURN(LDLM_ITER_STOP);
+ rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL);
+ if (!rc)
+ RETURN(LDLM_ITER_STOP);
+
+ /* grant also try_bits if any */
+ if (lock->l_policy_data.l_inodebits.try_bits != 0) {
+ lock->l_policy_data.l_inodebits.bits |=
+ lock->l_policy_data.l_inodebits.try_bits;
+ lock->l_policy_data.l_inodebits.try_bits = 0;
+ *flags |= LDLM_FL_LOCK_CHANGED;
+ }
+ ldlm_resource_unlink_lock(lock);
+ ldlm_grant_lock(lock, grant_work);
+
+ *err = ELDLM_OK;
+ RETURN(LDLM_ITER_CONTINUE);
+ }
+
+ rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, work_list);
+ rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, work_list);
+
+ if (rc != 2) {
+ /* if there were only bits to try and all are conflicting */
+ if ((lock->l_policy_data.l_inodebits.bits |
+ lock->l_policy_data.l_inodebits.try_bits) == 0) {
+ *err = ELDLM_LOCK_WOULDBLOCK;
+ } else {
+ *err = ELDLM_OK;
+ }
+ } else {
+ /* grant also all remaining try_bits */
+ if (lock->l_policy_data.l_inodebits.try_bits != 0) {
+ lock->l_policy_data.l_inodebits.bits |=
+ lock->l_policy_data.l_inodebits.try_bits;
+ lock->l_policy_data.l_inodebits.try_bits = 0;
+ *flags |= LDLM_FL_LOCK_CHANGED;
+ }
+ LASSERT(lock->l_policy_data.l_inodebits.bits);
+ ldlm_resource_unlink_lock(lock);
+ ldlm_grant_lock(lock, grant_work);
+ *err = ELDLM_OK;
+ }
+
+ RETURN(LDLM_ITER_CONTINUE);
+}
+#endif /* HAVE_SERVER_SUPPORT */
+
+void ldlm_ibits_policy_wire_to_local(const union ldlm_wire_policy_data *wpolicy,
+ union ldlm_policy_data *lpolicy)
+{
+ lpolicy->l_inodebits.bits = wpolicy->l_inodebits.bits;
+ /**
+ * try_bits are to be handled outside of generic write_to_local due
+ * to different behavior on a server and client.
+ */
+}
+
+void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
+ union ldlm_wire_policy_data *wpolicy)
+{
+ memset(wpolicy, 0, sizeof(*wpolicy));
+ wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
+ wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
+}
+
+/**
+ * Attempt to convert already granted IBITS lock with several bits set to
+ * a lock with less bits (downgrade).
+ *
+ * Such lock conversion is used to keep lock with non-blocking bits instead of
+ * cancelling it, introduced for better support of DoM files.
+ */
+int ldlm_inodebits_drop(struct ldlm_lock *lock, __u64 to_drop)
+{
+ ENTRY;
+
+ check_res_locked(lock->l_resource);
+
+ /* Just return if there are no conflicting bits */
+ if ((lock->l_policy_data.l_inodebits.bits & to_drop) == 0) {
+ LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
+ lock->l_policy_data.l_inodebits.bits, to_drop);
+ /* nothing to do */
+ RETURN(0);
+ }
+
+ /* remove lock from a skiplist and put in the new place
+ * according with new inodebits */
+ ldlm_resource_unlink_lock(lock);
+ lock->l_policy_data.l_inodebits.bits &= ~to_drop;
+ ldlm_grant_lock_with_skiplist(lock);
+ RETURN(0);
+}
+EXPORT_SYMBOL(ldlm_inodebits_drop);
+
+/* convert single lock */
+int ldlm_cli_inodebits_convert(struct ldlm_lock *lock,
+ enum ldlm_cancel_flags cancel_flags)
+{
+ struct ldlm_namespace *ns = ldlm_lock_to_ns(lock);
+ struct ldlm_lock_desc ld = { { 0 } };
+ __u64 drop_bits, new_bits;
+ __u32 flags = 0;
+ int rc;
+
+ ENTRY;
+
+ check_res_locked(lock->l_resource);
+
+ /* Lock is being converted already */
+ if (ldlm_is_converting(lock)) {
+ if (!(cancel_flags & LCF_ASYNC)) {
+ unlock_res_and_lock(lock);
+ wait_event_idle(lock->l_waitq,
+ is_lock_converted(lock));
+ lock_res_and_lock(lock);
+ }
+ RETURN(0);
+ }
+
+ /* lru_cancel may happen in parallel and call ldlm_cli_cancel_list()
+ * independently.
+ */
+ if (ldlm_is_canceling(lock))
+ RETURN(-EINVAL);
+
+ /* no need in only local convert */
+ if (lock->l_flags & (LDLM_FL_LOCAL_ONLY | LDLM_FL_CANCEL_ON_BLOCK))
+ RETURN(-EINVAL);
+
+ drop_bits = lock->l_policy_data.l_inodebits.cancel_bits;
+ /* no cancel bits - means that caller needs full cancel */
+ if (drop_bits == 0)
+ RETURN(-EINVAL);
+
+ new_bits = lock->l_policy_data.l_inodebits.bits & ~drop_bits;
+ /* check if all lock bits are dropped, proceed with cancel */
+ if (!new_bits)
+ RETURN(-EINVAL);
+
+ /* check if no dropped bits, consider this as successful convert */
+ if (lock->l_policy_data.l_inodebits.bits == new_bits)
+ RETURN(0);
+
+ ldlm_set_converting(lock);
+ /* Finally call cancel callback for remaining bits only.
+ * It is important to have converting flag during that
+ * so blocking_ast callback can distinguish convert from
+ * cancels.
+ */
+ ld.l_policy_data.l_inodebits.cancel_bits = drop_bits;
+ unlock_res_and_lock(lock);
+ lock->l_blocking_ast(lock, &ld, lock->l_ast_data, LDLM_CB_CANCELING);
+ /* now notify server about convert */
+ rc = ldlm_cli_convert_req(lock, &flags, new_bits);
+ lock_res_and_lock(lock);
+ if (rc)
+ GOTO(full_cancel, rc);
+
+ /* Finally clear these bits in lock ibits */
+ ldlm_inodebits_drop(lock, drop_bits);
+
+ /* Being locked again check if lock was canceled, it is important
+ * to do and don't drop cbpending below
+ */
+ if (ldlm_is_canceling(lock))
+ GOTO(full_cancel, rc = -EINVAL);
+
+ /* also check again if more bits to be cancelled appeared */
+ if (drop_bits != lock->l_policy_data.l_inodebits.cancel_bits)
+ GOTO(clear_converting, rc = -EAGAIN);
+
+ /* clear cbpending flag early, it is safe to match lock right after
+ * client convert because it is downgrade always.
+ */
+ ldlm_clear_cbpending(lock);
+ ldlm_clear_bl_ast(lock);
+ spin_lock(&ns->ns_lock);
+ if (list_empty(&lock->l_lru))
+ ldlm_lock_add_to_lru_nolock(lock);
+ spin_unlock(&ns->ns_lock);
+
+ /* the job is done, zero the cancel_bits. If more conflicts appear,
+ * it will result in another cycle of ldlm_cli_inodebits_convert().
+ */
+full_cancel:
+ lock->l_policy_data.l_inodebits.cancel_bits = 0;
+clear_converting:
+ ldlm_clear_converting(lock);
+ RETURN(rc);
+}
+
+int ldlm_inodebits_alloc_lock(struct ldlm_lock *lock)
+{
+ if (ldlm_is_ns_srv(lock)) {
+ int i;
+
+ OBD_SLAB_ALLOC_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
+ if (lock->l_ibits_node == NULL)
+ return -ENOMEM;
+ for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
+ INIT_LIST_HEAD(&lock->l_ibits_node->lin_link[i]);
+ lock->l_ibits_node->lock = lock;
+ } else {
+ lock->l_ibits_node = NULL;
+ }
+ return 0;
+}
+
+void ldlm_inodebits_add_lock(struct ldlm_resource *res, struct list_head *head,
+ struct ldlm_lock *lock)
+{
+ int i;
+
+ if (!ldlm_is_ns_srv(lock))
+ return;
+
+ if (head == &res->lr_waiting) {
+ for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
+ if (lock->l_policy_data.l_inodebits.bits & (1 << i))
+ list_add_tail(&lock->l_ibits_node->lin_link[i],
+ &res->lr_ibits_queues->liq_waiting[i]);
+ }
+ } else if (head == &res->lr_granted && lock->l_ibits_node != NULL) {
+ for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
+ LASSERT(list_empty(&lock->l_ibits_node->lin_link[i]));
+ OBD_SLAB_FREE_PTR(lock->l_ibits_node, ldlm_inodebits_slab);
+ lock->l_ibits_node = NULL;
+ }
+}
+
+void ldlm_inodebits_unlink_lock(struct ldlm_lock *lock)
+{
+ int i;
+
+ ldlm_unlink_lock_skiplist(lock);
+ if (!ldlm_is_ns_srv(lock))
+ return;
+
+ for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
+ list_del_init(&lock->l_ibits_node->lin_link[i]);