X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_inodebits.c;h=fad8fc63b0ccb59cb9da7c6e88ce7e5a19ef7524;hp=a0814dda94b08473613f71a434d0fd13619963b3;hb=da18ad5628556cb8eb0bcba786743b2d205a39d9;hpb=ccabce23bd9e366c345c852f565766a799f61238

diff --git a/lustre/ldlm/ldlm_inodebits.c b/lustre/ldlm/ldlm_inodebits.c
index a0814dd..fad8fc6 100644
--- a/lustre/ldlm/ldlm_inodebits.c
+++ b/lustre/ldlm/ldlm_inodebits.c
@@ -57,6 +57,89 @@
 #include "ldlm_internal.h"
 
 #ifdef HAVE_SERVER_SUPPORT
+
+/**
+ * It should iterate through all waiting locks on a given resource queue and
+ * attempt to grant them. An optimization is to check only heads waitintg
+ * locks for each inodebit type.
+ *
+ * Must be called with resource lock held.
+ */
+int ldlm_reprocess_inodebits_queue(struct ldlm_resource *res,
+				   struct list_head *queue,
+				   struct list_head *work_list,
+				   enum ldlm_process_intention intention,
+				   struct ldlm_lock *hint)
+{
+	__u64 flags;
+	int rc = LDLM_ITER_CONTINUE;
+	enum ldlm_error err;
+	LIST_HEAD(bl_ast_list);
+	struct ldlm_ibits_queues *queues = res->lr_ibits_queues;
+	int i;
+
+	ENTRY;
+
+	check_res_locked(res);
+
+	LASSERT(res->lr_type == LDLM_IBITS);
+	LASSERT(intention == LDLM_PROCESS_RESCAN ||
+		intention == LDLM_PROCESS_RECOVERY);
+
+	if (intention == LDLM_PROCESS_RECOVERY)
+		return ldlm_reprocess_queue(res, queue, work_list, intention,
+					    NULL);
+
+restart:
+	CDEBUG(D_DLMTRACE, "--- Reprocess resource "DLDLMRES" (%p)\n",
+	       PLDLMRES(res), res);
+
+	for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
+		LIST_HEAD(rpc_list);
+		struct list_head *head = &queues->liq_waiting[i];
+		struct ldlm_lock *pending;
+		struct ldlm_ibits_node *node;
+
+		if (list_empty(head))
+			continue;
+		if (hint && !(hint->l_policy_data.l_inodebits.bits & BIT(i)))
+			continue;
+
+		node = list_entry(head->next, struct ldlm_ibits_node,
+				  lin_link[i]);
+
+		pending = node->lock;
+		LDLM_DEBUG(pending, "Reprocessing lock from queue %d", i);
+
+		flags = 0;
+		rc = ldlm_process_inodebits_lock(pending, &flags, intention,
+						 &err, &rpc_list);
+		if (ldlm_is_granted(pending)) {
+			list_splice(&rpc_list, work_list);
+			/* Try to grant more locks from current queue */
+			i--;
+		} else {
+			list_splice(&rpc_list, &bl_ast_list);
+		}
+	}
+
+	if (!list_empty(&bl_ast_list)) {
+		unlock_res(res);
+
+		rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &bl_ast_list,
+				       LDLM_WORK_BL_AST);
+
+		lock_res(res);
+		if (rc == -ERESTART)
+			GOTO(restart, rc);
+	}
+
+	if (!list_empty(&bl_ast_list))
+		ldlm_discard_bl_list(&bl_ast_list);
+
+	RETURN(rc);
+}
+
 /**
  * Determine if the lock is compatible with all locks on the queue.
  *
@@ -74,8 +157,9 @@
  */
 static int
 ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
-			    struct list_head *work_list)
+			    __u64 *ldlm_flags, struct list_head *work_list)
 {
+	enum ldlm_mode req_mode = req->l_req_mode;
 	struct list_head *tmp;
 	struct ldlm_lock *lock;
 	__u64 req_bits = req->l_policy_data.l_inodebits.bits;
@@ -84,6 +168,8 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 
 	ENTRY;
 
+	lockmode_verify(req_mode);
+
 	/* There is no sense in lock with no bits set. Also such a lock
 	 * would be compatible with any other bit lock.
 	 * Meanwhile that can be true if there were just try_bits and all
@@ -92,6 +178,11 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 	if ((req_bits | *try_bits) == 0)
 		RETURN(0);
 
+	/* Group lock could be only DOM */
+	if (unlikely(req_mode == LCK_GROUP &&
+		     (req_bits | *try_bits) != MDS_INODELOCK_DOM))
+		RETURN(-EPROTO);
+
 	list_for_each(tmp, queue) {
 		struct list_head *mode_tail;
 
@@ -118,10 +209,41 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 			continue;
 		}
 
-		/* locks' mode are compatible, bits don't matter */
-		if (lockmode_compat(lock->l_req_mode, req->l_req_mode)) {
-			/* jump to last lock in mode group */
-			tmp = mode_tail;
+		if (lockmode_compat(lock->l_req_mode, req_mode)) {
+			/* non group locks are compatible, bits don't matter */
+			if (likely(req_mode != LCK_GROUP)) {
+				/* jump to last lock in mode group */
+				tmp = mode_tail;
+				continue;
+			}
+
+			if (req->l_policy_data.l_inodebits.li_gid ==
+			    lock->l_policy_data.l_inodebits.li_gid) {
+				if (ldlm_is_granted(lock))
+					RETURN(2);
+
+				if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
+					RETURN(-EWOULDBLOCK);
+
+				/* Place the same group together */
+				ldlm_resource_insert_lock_after(lock, req);
+				RETURN(0);
+			}
+		}
+
+		/* GROUP locks are placed to a head of the waiting list, but
+		 * grouped by gid. */
+		if (unlikely(req_mode == LCK_GROUP && !ldlm_is_granted(lock))) {
+			compat = 0;
+			if (lock->l_req_mode != LCK_GROUP) {
+				/* Already not a GROUP lock, insert before. */
+				ldlm_resource_insert_lock_before(lock, req);
+				break;
+			}
+			/* Still GROUP but a different gid(the same gid would
+			 * be handled above). Keep searching for the same gid */
+			LASSERT(req->l_policy_data.l_inodebits.li_gid !=
+				lock->l_policy_data.l_inodebits.li_gid);
 			continue;
 		}
 
@@ -158,14 +280,34 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 				    !ldlm_is_cos_incompat(req) &&
 				    ldlm_is_cos_enabled(req) &&
 				    lock->l_client_cookie == req->l_client_cookie)
-					goto not_conflicting;
+					goto skip_work_list;
+
+				compat = 0;
+
+				if (unlikely(lock->l_req_mode == LCK_GROUP)) {
+					LASSERT(ldlm_has_dom(lock));
+
+					if (*ldlm_flags & LDLM_FL_BLOCK_NOWAIT)
+						RETURN(-EWOULDBLOCK);
+
+					/* Local combined DOM lock came across
+					 * GROUP DOM lock, it makes the thread
+					 * to be blocked for a long time, not
+					 * allowed, the trybits to be used
+					 * instead.
+					 */
+					if (!req->l_export &&
+					    (req_bits & MDS_INODELOCK_DOM) &&
+					    (req_bits & ~MDS_INODELOCK_DOM))
+						LBUG();
+
+					goto skip_work_list;
+				}
 
 				/* Found a conflicting policy group. */
 				if (!work_list)
 					RETURN(0);
 
-				compat = 0;
-
 				/* Add locks of the policy group to @work_list
 				 * as blocking locks for @req */
 				if (lock->l_blocking_ast)
@@ -177,7 +319,7 @@ ldlm_inodebits_compat_queue(struct list_head *queue, struct ldlm_lock *req,
 						ldlm_add_ast_work_item(lock,
 								req, work_list);
 			}
-not_conflicting:
+skip_work_list:
 			if (tmp == mode_tail)
 				break;
 
@@ -197,31 +339,53 @@ not_conflicting:
  * waiting queues. The lock is granted if no conflicts are found in
  * either queue.
  */
-int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *flags,
+int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *ldlm_flags,
 				enum ldlm_process_intention intention,
 				enum ldlm_error *err,
 				struct list_head *work_list)
 {
 	struct ldlm_resource *res = lock->l_resource;
-	struct list_head rpc_list;
-	int rc;
-
+	struct list_head *grant_work = intention == LDLM_PROCESS_ENQUEUE ?
+							NULL : work_list;
+	int rc, rc2 = 0;
 	ENTRY;
 
-	LASSERT(lock->l_granted_mode != lock->l_req_mode);
-	LASSERT(list_empty(&res->lr_converting));
-	INIT_LIST_HEAD(&rpc_list);
+	*err = ELDLM_LOCK_ABORTED;
+	LASSERT(!ldlm_is_granted(lock));
 	check_res_locked(res);
 
 	if (intention == LDLM_PROCESS_RESCAN) {
-		*err = ELDLM_LOCK_ABORTED;
+		struct list_head *bl_list =
+			*ldlm_flags & LDLM_FL_BLOCK_NOWAIT ? NULL : work_list;
 
 		LASSERT(lock->l_policy_data.l_inodebits.bits != 0);
 
-		rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, NULL);
+		/* 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.
+		 */
+		*ldlm_flags = 0;
+		rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
+						 ldlm_flags, bl_list);
 		if (!rc)
 			RETURN(LDLM_ITER_STOP);
-		rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock, NULL);
+		rc = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
+						 ldlm_flags, NULL);
 		if (!rc)
 			RETURN(LDLM_ITER_STOP);
 
@@ -230,51 +394,54 @@ int ldlm_process_inodebits_lock(struct ldlm_lock *lock, __u64 *flags,
 			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_flags |= LDLM_FL_LOCK_CHANGED;
 		}
 		ldlm_resource_unlink_lock(lock);
-		ldlm_grant_lock(lock, work_list);
+		ldlm_grant_lock(lock, grant_work);
 
 		*err = ELDLM_OK;
 		RETURN(LDLM_ITER_CONTINUE);
 	}
 
-	LASSERT((intention == LDLM_PROCESS_ENQUEUE && work_list == NULL) ||
-		(intention == LDLM_PROCESS_RECOVERY && work_list != NULL));
-restart:
-	rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock, &rpc_list);
-	rc += ldlm_inodebits_compat_queue(&res->lr_waiting, lock, &rpc_list);
+	rc = ldlm_inodebits_compat_queue(&res->lr_granted, lock,
+					 ldlm_flags, work_list);
+	if (rc < 0)
+		GOTO(out, *err = rc);
 
 	if (rc != 2) {
+		rc2 = ldlm_inodebits_compat_queue(&res->lr_waiting, lock,
+						  ldlm_flags, work_list);
+		if (rc2 < 0)
+			GOTO(out, *err = rc = rc2);
+	}
+
+	if (rc + rc2 != 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) {
-			rc = ELDLM_LOCK_WOULDBLOCK;
-		} else {
-			rc = ldlm_handle_conflict_lock(lock, flags,
-						       &rpc_list, 0);
-			if (rc == -ERESTART)
-				GOTO(restart, rc);
+		     lock->l_policy_data.l_inodebits.try_bits)) {
+			/* There is no sense to set LDLM_FL_NO_TIMEOUT to
+			 * @ldlm_flags for DOM lock while they are enqueued
+			 * through intents, i.e. @lock here is local which does
+			 * not timeout. */
+			*err = ELDLM_OK;
 		}
-		*err = rc;
 	} 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;
+			*ldlm_flags |= LDLM_FL_LOCK_CHANGED;
 		}
 		LASSERT(lock->l_policy_data.l_inodebits.bits);
 		ldlm_resource_unlink_lock(lock);
-		ldlm_grant_lock(lock, work_list);
-		rc = 0;
+		ldlm_grant_lock(lock, grant_work);
+		*err = ELDLM_OK;
 	}
 
-	if (!list_empty(&rpc_list))
-		ldlm_discard_bl_list(&rpc_list);
-
-	RETURN(rc);
+	RETURN(LDLM_ITER_CONTINUE);
+out:
+	return rc;
 }
 #endif /* HAVE_SERVER_SUPPORT */
 
@@ -282,7 +449,10 @@ 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;
-	lpolicy->l_inodebits.try_bits = wpolicy->l_inodebits.try_bits;
+	/**
+	 * try_bits and li_gid 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,
@@ -291,9 +461,17 @@ void ldlm_ibits_policy_local_to_wire(const union ldlm_policy_data *lpolicy,
 	memset(wpolicy, 0, sizeof(*wpolicy));
 	wpolicy->l_inodebits.bits = lpolicy->l_inodebits.bits;
 	wpolicy->l_inodebits.try_bits = lpolicy->l_inodebits.try_bits;
+	wpolicy->l_inodebits.li_gid = lpolicy->l_inodebits.li_gid;
 }
 
-int ldlm_inodebits_drop(struct ldlm_lock *lock,  __u64 to_drop)
+/**
+ * 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;
 
@@ -301,7 +479,7 @@ int ldlm_inodebits_drop(struct ldlm_lock *lock,  __u64 to_drop)
 
 	/* 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\n",
+		LDLM_WARN(lock, "try to drop unset bits %#llx/%#llx",
 			  lock->l_policy_data.l_inodebits.bits, to_drop);
 		/* nothing to do */
 		RETURN(0);
@@ -315,3 +493,172 @@ int ldlm_inodebits_drop(struct ldlm_lock *lock,  __u64 to_drop)
 	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, bool tail)
+{
+	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 & BIT(i)))
+				continue;
+			if (tail)
+				list_add_tail(&lock->l_ibits_node->lin_link[i],
+					 &res->lr_ibits_queues->liq_waiting[i]);
+			else
+				list_add(&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;
+	} else if (head != &res->lr_granted) {
+		/* we are inserting in a middle of a list, after @head */
+		struct ldlm_lock *orig = list_entry(head, struct ldlm_lock,
+						    l_res_link);
+		LASSERT(orig->l_policy_data.l_inodebits.bits ==
+			lock->l_policy_data.l_inodebits.bits);
+		/* The is no a use case to insert before with exactly matched
+		 * set of bits */
+		LASSERT(tail == false);
+
+		for (i = 0; i < MDS_INODELOCK_NUMBITS; i++) {
+			if (!(lock->l_policy_data.l_inodebits.bits & (1 << i)))
+				continue;
+			list_add(&lock->l_ibits_node->lin_link[i],
+				 &orig->l_ibits_node->lin_link[i]);
+		}
+	}
+}
+
+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]);
+}