* - the maximum extent
* - containing the requested extent
* - and not overlapping existing conflicting extents outside the requested one
- *
- * An alternative policy is to not shrink the new extent when conflicts exist */
+ */
static void
ldlm_extent_internal_policy(struct list_head *queue, struct ldlm_lock *req,
struct ldlm_extent *new_ex)
{
struct list_head *tmp;
ldlm_mode_t req_mode = req->l_req_mode;
- __u64 req_start = req->l_policy_data.l_extent.start;
- __u64 req_end = req->l_policy_data.l_extent.end;
+ __u64 req_start = req->l_req_extent.start;
+ __u64 req_end = req->l_req_extent.end;
ENTRY;
- if (new_ex->start == req_start && new_ex->end == req_end) {
- EXIT;
- return;
- }
+ lockmode_verify(req_mode);
list_for_each(tmp, queue) {
struct ldlm_lock *lock;
+ struct ldlm_extent *l_extent;
+
lock = list_entry(tmp, struct ldlm_lock, l_res_link);
+ l_extent = &lock->l_policy_data.l_extent;
- if (req == lock) {
+ if (new_ex->start == req_start && new_ex->end == req_end) {
EXIT;
return;
}
- /* if lock doesn't overlap new_ex, skip it. */
- if (lock->l_policy_data.l_extent.end < new_ex->start ||
- lock->l_policy_data.l_extent.start > new_ex->end)
+ /* Don't conflict with ourselves */
+ if (req == lock)
+ continue;
+
+ /* If lock doesn't overlap new_ex, skip it. */
+ if (l_extent->end < new_ex->start ||
+ l_extent->start > new_ex->end)
continue;
/* Locks are compatible, overlap doesn't matter */
if (lockmode_compat(lock->l_req_mode, req_mode))
continue;
- if (lock->l_policy_data.l_extent.start < req_start) {
- if (lock->l_policy_data.l_extent.end == ~0) {
+ /* Locks conflicting in requested extents and we can't satisfy
+ * both locks, so ignore it. Either we will ping-pong this
+ * extent (we would regardless of what extent we granted) or
+ * lock is unused and it shouldn't limit our extent growth. */
+ if (lock->l_req_extent.end >= req_start &&
+ lock->l_req_extent.start <= req_end)
+ continue;
+
+ /* We grow extents downwards only as far as they don't overlap
+ * with already-granted locks, on the assumtion that clients
+ * will be writing beyond the initial requested end and would
+ * then need to enqueue a new lock beyond the previous request.
+ * We don't grow downwards if there are lots of lockers. */
+ if (l_extent->start < req_start) {
+ if (atomic_read(&req->l_resource->lr_refcount) > 20)
new_ex->start = req_start;
- new_ex->end = req_end;
- EXIT;
- return;
- }
- new_ex->start = MIN(lock->l_policy_data.l_extent.end+1,
- req_start);
+ else
+ new_ex->start = min(l_extent->end+1, req_start);
}
- if (lock->l_policy_data.l_extent.end > req_end) {
- if (lock->l_policy_data.l_extent.start == 0) {
- new_ex->start = req_start;
- new_ex->end = req_end;
- EXIT;
- return;
- }
- new_ex->end = MAX(lock->l_policy_data.l_extent.start-1,
- req_end);
+ /* If we need to cancel this lock anyways because our request
+ * overlaps the granted lock, we grow up to its requested
+ * extent start instead of limiting this extent, assuming that
+ * clients are writing forwards and the lock had over grown
+ * its extent downwards before we enqueued our request. */
+ if (l_extent->end > req_end) {
+ if (l_extent->start <= req_end)
+ new_ex->end = max(lock->l_req_extent.start - 1,
+ req_end);
+ else
+ new_ex->end = max(l_extent->start - 1, req_end);
}
}
EXIT;
}
-/* Determine if the lock is compatible with all locks on the queue. */
+/* In order to determine the largest possible extent we can grant, we need
+ * to scan all of the queues. */
+static void ldlm_extent_policy(struct ldlm_resource *res,
+ struct ldlm_lock *lock, int *flags)
+{
+ struct ldlm_extent new_ex = { .start = 0, .end = ~0};
+
+ ldlm_extent_internal_policy(&res->lr_granted, lock, &new_ex);
+ ldlm_extent_internal_policy(&res->lr_waiting, lock, &new_ex);
+
+ if (new_ex.start != lock->l_policy_data.l_extent.start ||
+ new_ex.end != lock->l_policy_data.l_extent.end) {
+ *flags |= LDLM_FL_LOCK_CHANGED;
+ lock->l_policy_data.l_extent.start = new_ex.start;
+ lock->l_policy_data.l_extent.end = new_ex.end;
+ }
+}
+
+/* Determine if the lock is compatible with all locks on the queue.
+ * We stop walking the queue if we hit ourselves so we don't take
+ * conflicting locks enqueued after us into accound, or we'd wait forever. */
static int
ldlm_extent_compat_queue(struct list_head *queue, struct ldlm_lock *req,
int send_cbs)
struct list_head *tmp;
struct ldlm_lock *lock;
ldlm_mode_t req_mode = req->l_req_mode;
- __u64 req_start = req->l_policy_data.l_extent.start;
- __u64 req_end = req->l_policy_data.l_extent.end;
+ __u64 req_start = req->l_req_extent.start;
+ __u64 req_end = req->l_req_extent.end;
int compat = 1;
ENTRY;
+ lockmode_verify(req_mode);
+
list_for_each(tmp, queue) {
lock = list_entry(tmp, struct ldlm_lock, l_res_link);
ldlm_error_t *err)
{
struct ldlm_resource *res = lock->l_resource;
- struct ldlm_extent new_ex = {0, ~0};
struct list_head rpc_list = LIST_HEAD_INIT(rpc_list);
int rc;
ENTRY;
RETURN(LDLM_ITER_STOP);
ldlm_resource_unlink_lock(lock);
+
+ ldlm_extent_policy(res, lock, flags);
ldlm_grant_lock(lock, NULL, 0, 1);
RETURN(LDLM_ITER_CONTINUE);
}
- /* In order to determine the largest possible extent we can
- * grant, we need to scan all of the queues. */
- ldlm_extent_internal_policy(&res->lr_granted, lock, &new_ex);
- ldlm_extent_internal_policy(&res->lr_waiting, lock, &new_ex);
-
- if (new_ex.start != lock->l_policy_data.l_extent.start ||
- new_ex.end != lock->l_policy_data.l_extent.end) {
- *flags |= LDLM_FL_LOCK_CHANGED;
- lock->l_policy_data.l_extent.start = new_ex.start;
- lock->l_policy_data.l_extent.end = new_ex.end;
- }
-
restart:
LASSERT(res->lr_tmp == NULL);
res->lr_tmp = &rpc_list;
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. */
- ldlm_resource_unlink_lock(lock);
- ldlm_resource_add_lock(res, &res->lr_waiting, lock);
+ * 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);
GOTO(restart, -ERESTART);
*flags |= LDLM_FL_BLOCK_GRANTED;
} else {
+ ldlm_extent_policy(res, lock, flags);
ldlm_resource_unlink_lock(lock);
ldlm_grant_lock(lock, NULL, 0, 0);
}
RETURN(0);
}
+
+/* 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.
+ *
+ * NB: A lock on [x,y] protects a KMS of up to y + 1 bytes! */
+__u64 ldlm_extent_shift_kms(struct ldlm_lock *lock, __u64 old_kms)
+{
+ struct ldlm_resource *res = lock->l_resource;
+ struct list_head *tmp;
+ struct ldlm_lock *lck;
+ __u64 kms = 0;
+ ENTRY;
+
+ l_lock(&res->lr_namespace->ns_lock);
+ list_for_each(tmp, &res->lr_granted) {
+ lck = list_entry(tmp, struct ldlm_lock, l_res_link);
+
+ if (lock == lck)
+ continue;
+ if (lck->l_policy_data.l_extent.end >= old_kms)
+ GOTO(out, kms = old_kms);
+ kms = lck->l_policy_data.l_extent.end + 1;
+ }
+
+ GOTO(out, kms);
+ out:
+ l_unlock(&res->lr_namespace->ns_lock);
+ return kms;
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff