Avoid try_shrink_lock spinlock contention

Every lock_key_range call results in a spinlock on linfo->lock to do
housekeeping and reclaim idle locks. This housekeeping is non-critical
and can be deferred, but more importantly, while a reclaim pass is
ongoing, there's no need to attempt reclaim at the same time.

I ran into this while doing cross-node spam and it caused a near hard
lockup with my VM spinning on the lock for 7+ minutes, without making
any forward progress.

We can just trylock and give up. If the lock is held, another task
is already reclaiming and we're safe to skip. The force flag will
still cause a spin_lock to assure shutdown waits for other reclaims,
which needs to happen before shutdown can progress anyway.

Signed-off-by: Auke Kok <auke.kok@versity.com>

kmod/src/Makefile.kernelcompat

@@ -496,3 +496,12 @@ endif
 ifneq (,$(shell grep 'struct posix_acl.*get_inode_acl' include/linux/fs.h))
 ccflags-y += -DKC_GET_INODE_ACL
 endif
+
+#
+# v6.15-13744-g41cb08555c41
+#
+# from_timer renamed to timer_container_of.
+#
+ifneq (,$(shell grep 'define timer_container_of' include/linux/timer.h))
+ccflags-y += -DKC_TIMER_CONTAINER_OF
+endif

kmod/src/data.c

@@ -422,6 +422,8 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 
 	mutex_lock(&datinf->mutex);
 
+	scoutfs_inode_get_onoff(inode, &online, &offline);
+
 	/* default to single allocation at the written block */
 	start = iblock;
 	count = 1;
@@ -444,7 +446,6 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 		 * the preallocation size to the number of online
 		 * blocks.
 		 */
-		scoutfs_inode_get_onoff(inode, &online, &offline);
 		if (iblock > 1 && iblock == online) {
 			ret = scoutfs_ext_next(sb, &data_ext_ops, &args,
 					       iblock, 1, &found);
@@ -486,6 +487,13 @@ static int alloc_block(struct super_block *sb, struct inode *inode,
 		/* trim count by next extent after iblock */
 		if (found.len && found.start > start && found.start < start + count)
 			count = (found.start - start);
+
+		/*
+		 * Ramp the aligned region size up proportionally with
+		 * the file's online block count rather than jumping to
+		 * the full prealloc size.
+		 */
+		count = max_t(u64, 1, min(count, online));
 	}
 
 	/* overall prealloc limit */

kmod/src/fence.c

@@ -222,7 +222,7 @@ static struct attribute *fence_attrs[] = {
 
 static void fence_timeout(struct timer_list *timer)
 {
-	struct pending_fence *fence = from_timer(fence, timer, timer);
+	struct pending_fence *fence = timer_container_of(fence, timer, timer);
 	struct super_block *sb = fence->sb;
 	DECLARE_FENCE_INFO(sb, fi);
 

kmod/src/kernelcompat.h

@@ -489,4 +489,9 @@ static inline void stack_trace_print(unsigned long *entries, unsigned int nr_ent
 }
 #endif
 
+#ifndef KC_TIMER_CONTAINER_OF
+#define timer_container_of(var, callback_timer, timer_fieldname) \
+	from_timer(var, callback_timer, timer_fieldname)
+#endif
+
 #endif

kmod/src/lock.c

@@ -906,7 +906,11 @@ static bool try_shrink_lock(struct super_block *sb, struct lock_info *linfo, boo
 	               READ_ONCE(linfo->nr_locks) <= opts.lock_idle_count))
 		return false;
 
-	spin_lock(&linfo->lock);
+	/* Shrinking is best-effort housekeeping unless forced. */
+	if (force)
+		spin_lock(&linfo->lock);
+	else if (!spin_trylock(&linfo->lock))
+		return false;
 
 	lock = list_first_entry_or_null(&linfo->lru_list, struct scoutfs_lock, lru_head);
 	if (lock && (force || (linfo->nr_locks > opts.lock_idle_count))) {

kmod/src/recov.c

@@ -134,7 +134,7 @@ static int recov_finished(struct recov_info *recinf)
 
 static void timer_callback(struct timer_list *timer)
 {
-	struct recov_info *recinf = from_timer(recinf, timer, timer);
+	struct recov_info *recinf = timer_container_of(recinf, timer, timer);
 
 	recinf->timeout_fn(recinf->sb);
 }

tests/golden/data-prealloc

@@ -8,10 +8,10 @@
 /mnt/test/test/data-prealloc/file-1: extents: 32
 /mnt/test/test/data-prealloc/file-2: extents: 32
 == any writes to region prealloc get full extents
-/mnt/test/test/data-prealloc/file-1: extents: 4
-/mnt/test/test/data-prealloc/file-2: extents: 4
-/mnt/test/test/data-prealloc/file-1: extents: 4
-/mnt/test/test/data-prealloc/file-2: extents: 4
+/mnt/test/test/data-prealloc/file-1: extents: 8
+/mnt/test/test/data-prealloc/file-2: extents: 8
+/mnt/test/test/data-prealloc/file-1: extents: 8
+/mnt/test/test/data-prealloc/file-2: extents: 8
 == streaming offline writes get full extents either way
 /mnt/test/test/data-prealloc/file-1: extents: 4
 /mnt/test/test/data-prealloc/file-2: extents: 4
@@ -20,8 +20,8 @@
 == goofy preallocation amounts work
 /mnt/test/test/data-prealloc/file-1: extents: 6
 /mnt/test/test/data-prealloc/file-2: extents: 6
-/mnt/test/test/data-prealloc/file-1: extents: 6
-/mnt/test/test/data-prealloc/file-2: extents: 6
+/mnt/test/test/data-prealloc/file-1: extents: 10
+/mnt/test/test/data-prealloc/file-2: extents: 10
 /mnt/test/test/data-prealloc/file-1: extents: 3
 /mnt/test/test/data-prealloc/file-2: extents: 3
 == block writes into region allocs hole