Skip to contentSkip to conversation

xfs: xfs: non-blocking inodegc pushes

PATCH In review Dave Chinner 3 participants 13 messages Jun 15, 2022 – Jun 25, 2022 XFS filesystem

lore
Share & tools
lore
#0 lore raw

v2 cover xfs: xfs: non-blocking inodegc pushes

Hi folks,

These patches introduce non-blocking inodegc pushes to fix long hold-offs in statfs() operations when inodegc is performing long running inode inactivation operations.

The first patch introduces the bound maximum work start time for the inodegc queues - it's short, only 1 jiffie (10ms) - but we don't want to delay inodegc for an arbitrarily long period of time. However, it means that work always starts quickly and so that reduces the need for statfs() to have to wait for background inodegc to start and complete to catch space "freed" by recent unlinks.

The second patch converts statfs to use a "push" rather than a "flush". The push simply schedules any pending work that hasn't yet timed out to run immediately and returns. It does not wait for the inodegc work to complete - that's what a flush does, and that's what caused all the problems for statfs(). Hence statfs() is converted to push semantics at the same time, thereby removing the blocking behaviour it currently has.

This should prevent a large amount of the issues that have been seeing with lots of processes stuck in statfs() - that will no long happen. The only time user processes should get stuck now is when the inodegc throttle kicks in (unlinks only at this point) or if we are waiting for a lock a long running inodegc operation holds to be released. We had those specific problems before background inodegc - they manifested as unkillable unlink operations that had every backed up on them instead of background inodegc that has everything backed up on them.

This patch has been running in my test environment for nearly a month now without regressions occurring. While there are likely still going to be inodegc flush related hold-offs in certain circumstances, nothing appears to be impacting the correctness of fstests tests or creating new issues. The 0-day kernel testing bot also indicates that certain benchmarks (such as aim7 and stress-ng.rename) run significantly faster with bound maximum delays and non-blocking statfs operations.

Comments, thoughts and testing appreciated.

-Dave.

Version 2:
- Also convert quota reportting inodegc flushes to a push.

#1 lore raw

PATCH v2 1/2 xfs: bound maximum wait time for inodegc work

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +22 -14
index 374b3bafaeb0..46b30ecf498c 100644
@@ -442,7 +442,7 @@ xfs_inodegc_queue_all(
 	for_each_online_cpu(cpu) {
 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
 		if (!llist_empty(&gc->list))
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 	}
 }
 
@@ -1844,8 +1844,8 @@ void
 xfs_inodegc_worker(
 	struct work_struct	*work)
 {
-	struct xfs_inodegc	*gc = container_of(work, struct xfs_inodegc,
-							work);
+	struct xfs_inodegc	*gc = container_of(to_delayed_work(work),
+						struct xfs_inodegc, work);
 	struct llist_node	*node = llist_del_all(&gc->list);
 	struct xfs_inode	*ip, *n;
 
@@ -2017,6 +2017,7 @@ xfs_inodegc_queue(
 	struct xfs_inodegc	*gc;
 	int			items;
 	unsigned int		shrinker_hits;
+	unsigned long		queue_delay = 1;
 
 	trace_xfs_inode_set_need_inactive(ip);
 	spin_lock(&ip->i_flags_lock);
@@ -2028,19 +2029,26 @@ xfs_inodegc_queue(
 	items = READ_ONCE(gc->items);
 	WRITE_ONCE(gc->items, items + 1);
 	shrinker_hits = READ_ONCE(gc->shrinker_hits);
-	put_cpu_ptr(gc);
 
-	if (!xfs_is_inodegc_enabled(mp))
+	/*
+	 * We queue the work while holding the current CPU so that the work
+	 * is scheduled to run on this CPU.
+	 */
+	if (!xfs_is_inodegc_enabled(mp)) {
+		put_cpu_ptr(gc);
 		return;
-
-	if (xfs_inodegc_want_queue_work(ip, items)) {
-		trace_xfs_inodegc_queue(mp, __return_address);
-		queue_work(mp->m_inodegc_wq, &gc->work);
 	}
 
+	if (xfs_inodegc_want_queue_work(ip, items))
+		queue_delay = 0;
+
+	trace_xfs_inodegc_queue(mp, __return_address);
+	mod_delayed_work(mp->m_inodegc_wq, &gc->work, queue_delay);
+	put_cpu_ptr(gc);
+
 	if (xfs_inodegc_want_flush_work(ip, items, shrinker_hits)) {
 		trace_xfs_inodegc_throttle(mp, __return_address);
-		flush_work(&gc->work);
+		flush_delayed_work(&gc->work);
 	}
 }
 
@@ -2057,7 +2065,7 @@ xfs_inodegc_cpu_dead(
 	unsigned int		count = 0;
 
 	dead_gc = per_cpu_ptr(mp->m_inodegc, dead_cpu);
-	cancel_work_sync(&dead_gc->work);
+	cancel_delayed_work_sync(&dead_gc->work);
 
 	if (llist_empty(&dead_gc->list))
 		return;
@@ -2076,12 +2084,12 @@ xfs_inodegc_cpu_dead(
 	llist_add_batch(first, last, &gc->list);
 	count += READ_ONCE(gc->items);
 	WRITE_ONCE(gc->items, count);
-	put_cpu_ptr(gc);
 
 	if (xfs_is_inodegc_enabled(mp)) {
 		trace_xfs_inodegc_queue(mp, __return_address);
-		queue_work(mp->m_inodegc_wq, &gc->work);
+		mod_delayed_work(mp->m_inodegc_wq, &gc->work, 0);
 	}
+	put_cpu_ptr(gc);
 }
 
 /*
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}
 	}
fs/xfs/xfs_mount.h +1 -1
index ba5d42abf66e..d2eaebd85abf 100644
@@ -61,7 +61,7 @@ struct xfs_error_cfg {
  */
 struct xfs_inodegc {
 	struct llist_head	list;
-	struct work_struct	work;
+	struct delayed_work	work;
 
 	/* approximate count of inodes in the list */
 	unsigned int		items;
fs/xfs/xfs_super.c +1 -1
index 48a7239ed1b2..651ae75a7e23 100644
@@ -1075,7 +1075,7 @@ xfs_inodegc_init_percpu(
 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
 		init_llist_head(&gc->list);
 		gc->items = 0;
-		INIT_WORK(&gc->work, xfs_inodegc_worker);
+		INIT_DELAYED_WORK(&gc->work, xfs_inodegc_worker);
 	}
 	return 0;
 }
Signature · 1 line
2.35.1
#2 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 8 lines
fs/xfs/xfs_icache.c +1 -1
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..? If so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Brian

Quoted 30 lines
fs/xfs/xfs_icache.c +0 -0
 	}
fs/xfs/xfs_mount.h +1 -1
index ba5d42abf66e..d2eaebd85abf 100644
@@ -61,7 +61,7 @@ struct xfs_error_cfg {
  */
 struct xfs_inodegc {
 	struct llist_head	list;
-	struct work_struct	work;
+	struct delayed_work	work;
 
 	/* approximate count of inodes in the list */
 	unsigned int		items;
fs/xfs/xfs_super.c +1 -1
index 48a7239ed1b2..651ae75a7e23 100644
@@ -1075,7 +1075,7 @@ xfs_inodegc_init_percpu(
 		gc = per_cpu_ptr(mp->m_inodegc, cpu);
 		init_llist_head(&gc->list);
 		gc->items = 0;
-		INIT_WORK(&gc->work, xfs_inodegc_worker);
+		INIT_DELAYED_WORK(&gc->work, xfs_inodegc_worker);
 	}
 	return 0;
 }
Signature · 1 line
2.35.1
#3 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 4 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
Reviewed-by
#4 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 8 lines
Quoted 4 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#5 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

Quoted 9 lines
Quoted 8 lines
Quoted 4 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#6 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 63 lines · Brian Foster
Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 6 lines

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

It can still be useful for certain g*dawful scenarios --

Let's say you have a horribly misconfigured cloudy system with a tiny log, hundreds of CPUs, a memory hogging process, another process with many hundreds of threads that are performing small appending synchronous writes to a large number of files, and some other process repeatedly opens and closes files. Background writeback completion will create enough workers to tie up the log such that writeback and inodegc contend for log grant space and make slow progress. If memory is also tight, we want to slow down the file scanning process so that it doesn't shove /more/ inodes into the cache and push the system towards OOM behavior.

Back in the old days when inodegc was a radix tree tag it was fairly easy to get OOMs when the delay interval was long (5 seconds). The OOM probability went down pretty sharply as the interval approached zero, but even at 1 jiffy I could still occasionally trip it, whereas the pre-deferred-inactivation kernels would never OOM.

I haven't tested it all that rigorously with Dave's fancy new per-cpu list design, but I did throw on my silly test setup (see below) and still got it to OOM once in 3 runs with the shrinker bits turned off.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

<nod> That comment /does/ need to be updated to note the subtlety that a lot of shrinker activity can slow down close()ing a file by making user tasks wait for the inodegc workers to clear the backlog.

Quoted 8 lines
Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Yeah. I snapshot /home (which has ~20 million inodes) then race fsstress and xfs_scrub -n in a VM with 560MB of RAM.

--D

Quoted 14 lines

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#7 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 73 lines · Darrick J. Wong
Quoted 63 lines · Brian Foster
Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 6 lines

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

It can still be useful for certain g*dawful scenarios --

Quoted 10 lines

Let's say you have a horribly misconfigured cloudy system with a tiny log, hundreds of CPUs, a memory hogging process, another process with many hundreds of threads that are performing small appending synchronous writes to a large number of files, and some other process repeatedly opens and closes files. Background writeback completion will create enough workers to tie up the log such that writeback and inodegc contend for log grant space and make slow progress. If memory is also tight, we want to slow down the file scanning process so that it doesn't shove /more/ inodes into the cache and push the system towards OOM behavior.

Yeah, I get the general idea/purpose of the throttling. What I'm probing at here is whether a case like this is still handled effectively with such a short scheduling delay. Presumably there is some window before list size based throttling triggers for which the shrinker is expected to cover, so that implies the shrinker historically is able to detect and push populated queues and trigger throttling from the point it is invoked (whether directly via repeated shrinker invocations or indirectly via causing larger queue sizes is not clear to me).

The thing that stands out to me as a question wrt to this change is that the trigger for shrinker induced throttling is the list size at the time of the callback(s), and that goes from having a lifecycle associated with the size-oriented scheduling algorithm to a time-based scheduling lifecycle of one jiffy (also noting that the inodegc worker resets shrinker_hits before it begins to process inodes). So with that in mind, how reliable is this lowmem signal based on the list size back to the tasks creating more work and memory pressure? Once a shrinker invocation occurs, what are the odds that the callback is able to detect a populated list and act accordingly?

These questions are somewhat rhetorical because this all seems rather unpredictable when we consider varying resource availability. The relevant question for this patch is probably just that somebody has tested and confirmed that the shrinker hasn't been subtly or indirectly broken in cases like the one you describe above (where perhaps we might not have many shrinker callback opportunities to act on before OOM).

Quoted 10 lines

Back in the old days when inodegc was a radix tree tag it was fairly easy to get OOMs when the delay interval was long (5 seconds). The OOM probability went down pretty sharply as the interval approached zero, but even at 1 jiffy I could still occasionally trip it, whereas the pre-deferred-inactivation kernels would never OOM.

I haven't tested it all that rigorously with Dave's fancy new per-cpu list design, but I did throw on my silly test setup (see below) and still got it to OOM once in 3 runs with the shrinker bits turned off.

Ok.. so that implies we still need throttling, but I'm not sure what "fancy percpu list design" refers to. If you have a good test case, I think the interesting immediate question is: are those OOMs avoided with this patch but the shrinker infrastructure still in place?

If not, then I wonder if something is going wonky there. If so, I'm still a bit curious what the behavior looks like and whether it can be simplified in light of this change, but that's certainly beyond the scope of this patch.

Quoted 22 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

<nod> That comment /does/ need to be updated to note the subtlety that a lot of shrinker activity can slow down close()ing a file by making user tasks wait for the inodegc workers to clear the backlog.

Quoted 4 lines
Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Quoted 8 lines
Quoted 4 lines

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Yeah. I snapshot /home (which has ~20 million inodes) then race fsstress and xfs_scrub -n in a VM with 560MB of RAM.

Yeah small delay or yeah large delay?

Brian

Quoted 17 lines

--D

Quoted 14 lines

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#8 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 85 lines · Brian Foster
Quoted 73 lines · Darrick J. Wong
Quoted 63 lines · Brian Foster
Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 6 lines

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

It can still be useful for certain g*dawful scenarios --

Quoted 10 lines

Let's say you have a horribly misconfigured cloudy system with a tiny log, hundreds of CPUs, a memory hogging process, another process with many hundreds of threads that are performing small appending synchronous writes to a large number of files, and some other process repeatedly opens and closes files. Background writeback completion will create enough workers to tie up the log such that writeback and inodegc contend for log grant space and make slow progress. If memory is also tight, we want to slow down the file scanning process so that it doesn't shove /more/ inodes into the cache and push the system towards OOM behavior.

Quoted 3 lines

Yeah, I get the general idea/purpose of the throttling. What I'm probing at here is whether a case like this is still handled effectively with such a short scheduling delay.

Given the trace_xfs_inodegc_shrinker_* data I've collected on my simulator, I think the 1 jiffy delay is handling this well enough that we only trip that tracepoint about a half dozen times in 2h of trying to exercising the inode caches. That said, I haven't really had time to try this on (say) a 20CPU machine with ridiculously small memory to see if I get different results.

(Note: I'm still trying to figure out why generic/522 reports corruption, and hoping that willy's folios pull today just fixes it magically...)

Quoted 6 lines

Presumably there is some window before
list size based throttling triggers for which the shrinker is expected to cover, so that implies the shrinker historically is able to detect and push populated queues and trigger throttling from the point it is invoked (whether directly via repeated shrinker invocations or indirectly via causing larger queue sizes is not clear to me).

Back when I merged the inodegc series, it was fairly easy to get the shrinker to trip dozens of times during a stress run, even when the only throttling criteria was the queue size. I think the reflink/rmap transaction reservation optimizations have made the log grant bottlenecks much less severe, since I now see a lot less contention on the grant heads.

Quoted 10 lines

The thing that stands out to me as a question wrt to this change is that the trigger for shrinker induced throttling is the list size at the time of the callback(s), and that goes from having a lifecycle associated with the size-oriented scheduling algorithm to a time-based scheduling lifecycle of one jiffy (also noting that the inodegc worker resets shrinker_hits before it begins to process inodes). So with that in mind, how reliable is this lowmem signal based on the list size back to the tasks creating more work and memory pressure? Once a shrinker invocation occurs, what are the odds that the callback is able to detect a populated list and act accordingly?

Hrmmm. Decent, I think? If the list is populated but the inactivation worker is not running, then we queue the worker and clear the list; if the worker is already running when the shrinker gets called, we might end up requeuing it unnecessarily, in which case it'll just clear more.

Quoted 2 lines

These questions are somewhat rhetorical because this all seems rather unpredictable when we consider varying resource availability.

Not to mention the shrinker itself undergoing various behavioral changes over the years... :(

Quoted 5 lines

The
relevant question for this patch is probably just that somebody has tested and confirmed that the shrinker hasn't been subtly or indirectly broken in cases like the one you describe above (where perhaps we might not have many shrinker callback opportunities to act on before OOM).

FWIW, the only time I /ever/ saw OOMs (either now with 5.19 or ages ago with 5.9) was with extreme lowmem testing. On VMs with more than a gigabyte or so of memory, I notice that we usually hit the time/length thresholds without shrinkers getting involved. Granted, I did try to pick the shrinker values so that we only get called on the *second* level of shrinking, which is after we've freed some but not enough memory.

If my responses seem a bit like handwaving, they are, because (at the moment) this is well off in the weeds.

Quoted 6 lines
Quoted 6 lines

Back in the old days when inodegc was a radix tree tag it was fairly easy to get OOMs when the delay interval was long (5 seconds). The OOM probability went down pretty sharply as the interval approached zero, but even at 1 jiffy I could still occasionally trip it, whereas the pre-deferred-inactivation kernels would never OOM.

Quoted 7 lines
Quoted 4 lines

I haven't tested it all that rigorously with Dave's fancy new per-cpu list design, but I did throw on my silly test setup (see below) and still got it to OOM once in 3 runs with the shrinker bits turned off.

Ok.. so that implies we still need throttling, but I'm not sure what "fancy percpu list design" refers to. If you have a good test case, I

You might recall that the original versions of deferred inode inactivation would set radix tree tags, and a per-AG workqueue function would scan that AG's radix tree for tagged inodes and inactivate them. Dave observed that letting the cpu caches grow cold led to performance problems, and replaced the tagging mechanism with per-cpu lists, which is what we have now.

Quoted 2 lines

think the interesting immediate question is: are those OOMs avoided with this patch but the shrinker infrastructure still in place?

That's the $64000 question that I don't know definitively yet.

Quoted 4 lines

If not, then I wonder if something is going wonky there. If so, I'm still a bit curious what the behavior looks like and whether it can be simplified in light of this change, but that's certainly beyond the scope of this patch.

<nod> For now I'd like to get this going for 5.19 to fix the original complaint about statfs blocking in D state while waiting for inodegc after deleting a ~100m file extents[1], and defer the decision of whether or not we /really/ need the shrinker to a later time (like 5.20).

[1] https://lore.kernel.org/linux-xfs/[email protected]/

Quoted 28 lines
Quoted 28 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

<nod> That comment /does/ need to be updated to note the subtlety that a lot of shrinker activity can slow down close()ing a file by making user tasks wait for the inodegc workers to clear the backlog.

Quoted 8 lines
Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Quoted 5 lines
Quoted 3 lines

Yeah. I snapshot /home (which has ~20 million inodes) then race fsstress and xfs_scrub -n in a VM with 560MB of RAM.

Yeah small delay or yeah large delay?

Both -- with large inactivation delays and no shrinker, OOMs happen fairly frequently; with a short delay and no shrinker, they're harder (but still not impossible) to trigger.

Granted ... even the overall frequency of OOMs with large inactivation delays seems to have gone down a bit from when I was more actively testing in the 5.9 era.

--D

Quoted 20 lines

Brian

Quoted 17 lines

--D

Quoted 14 lines

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#9 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 126 lines · Darrick J. Wong
Quoted 85 lines · Brian Foster
Quoted 73 lines · Darrick J. Wong
Quoted 63 lines · Brian Foster
Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 6 lines

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

It can still be useful for certain g*dawful scenarios --

Quoted 10 lines

Let's say you have a horribly misconfigured cloudy system with a tiny log, hundreds of CPUs, a memory hogging process, another process with many hundreds of threads that are performing small appending synchronous writes to a large number of files, and some other process repeatedly opens and closes files. Background writeback completion will create enough workers to tie up the log such that writeback and inodegc contend for log grant space and make slow progress. If memory is also tight, we want to slow down the file scanning process so that it doesn't shove /more/ inodes into the cache and push the system towards OOM behavior.

Quoted 3 lines

Yeah, I get the general idea/purpose of the throttling. What I'm probing at here is whether a case like this is still handled effectively with such a short scheduling delay.

Given the trace_xfs_inodegc_shrinker_* data I've collected on my simulator, I think the 1 jiffy delay is handling this well enough that we only trip that tracepoint about a half dozen times in 2h of trying to exercising the inode caches. That said, I haven't really had time to try this on (say) a 20CPU machine with ridiculously small memory to see if I get different results.

(Note: I'm still trying to figure out why generic/522 reports corruption, and hoping that willy's folios pull today just fixes it magically...)

Quoted 6 lines

Presumably there is some window before
list size based throttling triggers for which the shrinker is expected to cover, so that implies the shrinker historically is able to detect and push populated queues and trigger throttling from the point it is invoked (whether directly via repeated shrinker invocations or indirectly via causing larger queue sizes is not clear to me).

Back when I merged the inodegc series, it was fairly easy to get the shrinker to trip dozens of times during a stress run, even when the only throttling criteria was the queue size. I think the reflink/rmap transaction reservation optimizations have made the log grant bottlenecks much less severe, since I now see a lot less contention on the grant heads.

Quoted 10 lines

The thing that stands out to me as a question wrt to this change is that the trigger for shrinker induced throttling is the list size at the time of the callback(s), and that goes from having a lifecycle associated with the size-oriented scheduling algorithm to a time-based scheduling lifecycle of one jiffy (also noting that the inodegc worker resets shrinker_hits before it begins to process inodes). So with that in mind, how reliable is this lowmem signal based on the list size back to the tasks creating more work and memory pressure? Once a shrinker invocation occurs, what are the odds that the callback is able to detect a populated list and act accordingly?

Quoted 5 lines

Hrmmm. Decent, I think? If the list is populated but the inactivation worker is not running, then we queue the worker and clear the list; if the worker is already running when the shrinker gets called, we might end up requeuing it unnecessarily, in which case it'll just clear more.

Sure, that sounds harmless, but I was thinking more about the likelihood of missing throttles than spurious work cycles and such..

Generally with this change, the worker is pretty much always scheduled once an inode is queued. It executes once a tick passes where another destroy hasn't otherwise come along to kick it out by another tick. So if I run a sustained removal, I see similar batching/scheduling as since before this change. If I run something with a mix of removes and $something_else, then it doesn't seem to take much to skip batching and run a worker cycle per inode.

So what happens if the shrinker scan sees one or more workers that happen to execute before the scan callback is able to actually trigger throttling? Is that critical, or is this enough of a fallback mechanism that it's unlikely to be important? Similarly if one is running a workload that results in behavior as described above, is this sort of "oneshot shrinker to worker throttle" actually still going to slow down the worker tasks? Is shrinker callback frequency high enough to actually make an impact? Etc.

These are questions I don't really have enough context to answer atm, nor can say whether or not may result in problems...

Quoted 44 lines
Quoted 2 lines

These questions are somewhat rhetorical because this all seems rather unpredictable when we consider varying resource availability.

Not to mention the shrinker itself undergoing various behavioral changes over the years... :(

Quoted 5 lines

The
relevant question for this patch is probably just that somebody has tested and confirmed that the shrinker hasn't been subtly or indirectly broken in cases like the one you describe above (where perhaps we might not have many shrinker callback opportunities to act on before OOM).

FWIW, the only time I /ever/ saw OOMs (either now with 5.19 or ages ago with 5.9) was with extreme lowmem testing. On VMs with more than a gigabyte or so of memory, I notice that we usually hit the time/length thresholds without shrinkers getting involved. Granted, I did try to pick the shrinker values so that we only get called on the *second* level of shrinking, which is after we've freed some but not enough memory.

If my responses seem a bit like handwaving, they are, because (at the moment) this is well off in the weeds.

Quoted 6 lines
Quoted 6 lines

Back in the old days when inodegc was a radix tree tag it was fairly easy to get OOMs when the delay interval was long (5 seconds). The OOM probability went down pretty sharply as the interval approached zero, but even at 1 jiffy I could still occasionally trip it, whereas the pre-deferred-inactivation kernels would never OOM.

Quoted 7 lines
Quoted 4 lines

I haven't tested it all that rigorously with Dave's fancy new per-cpu list design, but I did throw on my silly test setup (see below) and still got it to OOM once in 3 runs with the shrinker bits turned off.

Ok.. so that implies we still need throttling, but I'm not sure what "fancy percpu list design" refers to. If you have a good test case, I

You might recall that the original versions of deferred inode inactivation would set radix tree tags, and a per-AG workqueue function would scan that AG's radix tree for tagged inodes and inactivate them. Dave observed that letting the cpu caches grow cold led to performance problems, and replaced the tagging mechanism with per-cpu lists, which is what we have now.

Quoted 5 lines
Quoted 2 lines

think the interesting immediate question is: are those OOMs avoided with this patch but the shrinker infrastructure still in place?

That's the $64000 question that I don't know definitively yet.

Ok. If the answer to the above questions is "don't really know, it might be broken, but historical testing shows the shrinker is much less important with smaller queuing delays such that it's unlikely to cause any new problems on anything but the most pathological workloads," then that sounds reasonable to me.

I don't mean to harp on it. I just want to make sure the question/concern gets across clearly enough that somebody who knows this system better than I can grok it and make a more intelligent assessment.

Brian

Quoted 78 lines
Quoted 4 lines

If not, then I wonder if something is going wonky there. If so, I'm still a bit curious what the behavior looks like and whether it can be simplified in light of this change, but that's certainly beyond the scope of this patch.

<nod> For now I'd like to get this going for 5.19 to fix the original complaint about statfs blocking in D state while waiting for inodegc after deleting a ~100m file extents[1], and defer the decision of whether or not we /really/ need the shrinker to a later time (like 5.20).

[1] https://lore.kernel.org/linux-xfs/[email protected]/

Quoted 28 lines
Quoted 28 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

<nod> That comment /does/ need to be updated to note the subtlety that a lot of shrinker activity can slow down close()ing a file by making user tasks wait for the inodegc workers to clear the backlog.

Quoted 8 lines
Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Quoted 5 lines
Quoted 3 lines

Yeah. I snapshot /home (which has ~20 million inodes) then race fsstress and xfs_scrub -n in a VM with 560MB of RAM.

Yeah small delay or yeah large delay?

Both -- with large inactivation delays and no shrinker, OOMs happen fairly frequently; with a short delay and no shrinker, they're harder (but still not impossible) to trigger.

Granted ... even the overall frequency of OOMs with large inactivation delays seems to have gone down a bit from when I was more actively testing in the 5.9 era.

--D

Quoted 20 lines

Brian

Quoted 17 lines

--D

Quoted 14 lines

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#10 lore raw

Re: [PATCH 1/2] xfs: bound maximum wait time for inodegc work

Quoted 144 lines · Brian Foster
Quoted 126 lines · Darrick J. Wong
Quoted 85 lines · Brian Foster
Quoted 73 lines · Darrick J. Wong
Quoted 63 lines · Brian Foster
Quoted 52 lines · Darrick J. Wong
Quoted 46 lines · Dave Chinner
Quoted 30 lines · Brian Foster
Quoted 25 lines · Dave Chinner

From: Dave Chinner <[email protected]>

Currently inodegc work can sit queued on the per-cpu queue until the workqueue is either flushed of the queue reaches a depth that triggers work queuing (and later throttling). This means that we could queue work that waits for a long time for some other event to trigger flushing.

Hence instead of just queueing work at a specific depth, use a delayed work that queues the work at a bound time. We can still schedule the work immediately at a given depth, but we no long need to worry about leaving a number of items on the list that won't get processed until external events prevail.

---
 fs/xfs/xfs_icache.c | 36 ++++++++++++++++++++++--------------
 fs/xfs/xfs_mount.h  |  2 +-
 fs/xfs/xfs_super.c  |  2 +-
 3 files changed, 24 insertions(+), 16 deletions(-)
fs/xfs/xfs_icache.c +0 -0
index 374b3bafaeb0..46b30ecf498c 100644

...

Quoted 3 lines
fs/xfs/xfs_icache.c +0 -0
@@ -2176,7 +2184,7 @@ xfs_inodegc_shrinker_scan(
 			unsigned int	h = READ_ONCE(gc->shrinker_hits);
 
Quoted 9 lines

fs/xfs/xfs_icache.c xfs_inodegc_shrinker_scan( · line 2189

Quoted 5 lines
fs/xfs/xfs_icache.c +1 -1
 			WRITE_ONCE(gc->shrinker_hits, h + 1);
-			queue_work_on(cpu, mp->m_inodegc_wq, &gc->work);
+			mod_delayed_work_on(cpu, mp->m_inodegc_wq, &gc->work, 0);
 			no_items = false;
 		}

This all seems reasonable to me, but is there much practical benefit to shrinker infra/feedback just to expedite a delayed work item by one jiffy? Maybe there's a use case to continue to trigger throttling..?

I haven't really considered doing anything other than fixing the reported bug. That just requires an API conversion for the existing "queue immediately" semantics and is the safest minimum change to fix the issue at hand.

Quoted 4 lines

So, yes, the shrinker code may (or may not) be superfluous now, but I haven't looked at it and done analysis of the behaviour without the shrinkers enabled. I'll do that in a completely separate patchset if it turns out that it is not needed now.

Quoted 9 lines

I think the shrinker part is still necessary -- bulkstat and xfs_scrub on a very low memory machine (~560M RAM) opening and closing tens of millions of files can still OOM the machine if one doesn't have a means to slow down ->destroy_inode (and hence the next open()) when reclaim really starts to dig in. Without the shrinker bits, it's even easier to trigger OOM storms when xfs has timer-delayed inactivation... which is something that Brian pointed out a year ago when we were reviewing the initial inodegc patchset.

Quoted 6 lines

It wouldn't surprise me if the infrastructure is still necessary for the throttling use case. In that case, I'm more curious about things like whether it's still as effective as intended with such a small scheduling delay, or whether it still might be worth simplifying in various ways (i.e., does the scheduling delay actually make a difference? do we still need a per cpu granular throttle? etc.).

It can still be useful for certain g*dawful scenarios --

Quoted 10 lines

Let's say you have a horribly misconfigured cloudy system with a tiny log, hundreds of CPUs, a memory hogging process, another process with many hundreds of threads that are performing small appending synchronous writes to a large number of files, and some other process repeatedly opens and closes files. Background writeback completion will create enough workers to tie up the log such that writeback and inodegc contend for log grant space and make slow progress. If memory is also tight, we want to slow down the file scanning process so that it doesn't shove /more/ inodes into the cache and push the system towards OOM behavior.

Quoted 3 lines

Yeah, I get the general idea/purpose of the throttling. What I'm probing at here is whether a case like this is still handled effectively with such a short scheduling delay.

Given the trace_xfs_inodegc_shrinker_* data I've collected on my simulator, I think the 1 jiffy delay is handling this well enough that we only trip that tracepoint about a half dozen times in 2h of trying to exercising the inode caches. That said, I haven't really had time to try this on (say) a 20CPU machine with ridiculously small memory to see if I get different results.

(Note: I'm still trying to figure out why generic/522 reports corruption, and hoping that willy's folios pull today just fixes it magically...)

Quoted 6 lines

Presumably there is some window before
list size based throttling triggers for which the shrinker is expected to cover, so that implies the shrinker historically is able to detect and push populated queues and trigger throttling from the point it is invoked (whether directly via repeated shrinker invocations or indirectly via causing larger queue sizes is not clear to me).

Back when I merged the inodegc series, it was fairly easy to get the shrinker to trip dozens of times during a stress run, even when the only throttling criteria was the queue size. I think the reflink/rmap transaction reservation optimizations have made the log grant bottlenecks much less severe, since I now see a lot less contention on the grant heads.

Quoted 10 lines

The thing that stands out to me as a question wrt to this change is that the trigger for shrinker induced throttling is the list size at the time of the callback(s), and that goes from having a lifecycle associated with the size-oriented scheduling algorithm to a time-based scheduling lifecycle of one jiffy (also noting that the inodegc worker resets shrinker_hits before it begins to process inodes). So with that in mind, how reliable is this lowmem signal based on the list size back to the tasks creating more work and memory pressure? Once a shrinker invocation occurs, what are the odds that the callback is able to detect a populated list and act accordingly?

Quoted 5 lines

Hrmmm. Decent, I think? If the list is populated but the inactivation worker is not running, then we queue the worker and clear the list; if the worker is already running when the shrinker gets called, we might end up requeuing it unnecessarily, in which case it'll just clear more.

Sure, that sounds harmless, but I was thinking more about the likelihood of missing throttles than spurious work cycles and such..

Generally with this change, the worker is pretty much always scheduled once an inode is queued. It executes once a tick passes where another destroy hasn't otherwise come along to kick it out by another tick. So if I run a sustained removal, I see similar batching/scheduling as since before this change. If I run something with a mix of removes and $something_else, then it doesn't seem to take much to skip batching and run a worker cycle per inode.

Quoted 1 line

So what happens if the shrinker scan sees one or more workers that

Er... by "shrinker scan", do you mean xfs_inodegc_shrinker_count? And by "scan callback", do you mean xfs_inodegc_shrinker_scan?

Quoted 3 lines

happen to execute before the scan callback is able to actually trigger throttling? Is that critical, or is this enough of a fallback mechanism that it's unlikely to be important?

I /think/ it's possible to miss throttling opportunities, but I also suspect it's relatively unimportant because of the low maximum queue size.

Let's say _count runs, discovers non-empty lists, and triggers the callback. If the workers run after _count returns but before _scan gets a chance to run, then they'll have processed the queued inodes and _scan won't set shrinker_hits anywhere. Yes, a _destroy_inode won't get throttled, but the system is at least making some forward progress clearing the backlog and (hopefully) freeing memory.

If there are a /lot/ of destroy_inode calls, they'll eventually hit the queue size limit (32) and throttle on the workers. If there aren't any further destroy_inode calls, there's nothing to throttle. If however there's some destroy_inode calls but not enough to hit the queue limit, then I suppose we're reliant on the shrinker to get called again. This is of course trivially satisfied if there are other processes that also enter the shrinker code and trigger more scans.

However, I suppose there /is/ a small hole here -- if there was only that one process in the shrinker and _scan doesn't find anything, the shrinker will not call us back. In that case, the only way we trigger throttling is if there's enough destroy_inode activity to hit a queue limit.

Dave and I have talked about closing this hole by queuing the worker immediately if the inode being destroyed has more than a certain number of extents or something that acts as a proxy for memory consumption, but so far as we can tell, shrinker invocations are rare and missed throttling hasn't shown up in a noticeable way.

Quoted 5 lines

Similarly if one is running a
workload that results in behavior as described above, is this sort of "oneshot shrinker to worker throttle" actually still going to slow down the worker tasks? Is shrinker callback frequency high enough to actually make an impact? Etc.

It seems to, but with a smallish sample size on a normal system, which means I've only seen the shrinker throttle show up in the trace output if I use the insane lowmem scenario or artificially crank up the delay to 30 seconds. The good news is that in those scenarios, it /does/ slow things down enough that we mostly don't OOM.

(Though at this point it's really hard to go back for an apples-apples comparison with the pre-inodegc code.)

Quoted 50 lines

These are questions I don't really have enough context to answer atm, nor can say whether or not may result in problems...

Quoted 47 lines
Quoted 2 lines

These questions are somewhat rhetorical because this all seems rather unpredictable when we consider varying resource availability.

Not to mention the shrinker itself undergoing various behavioral changes over the years... :(

Quoted 5 lines

The
relevant question for this patch is probably just that somebody has tested and confirmed that the shrinker hasn't been subtly or indirectly broken in cases like the one you describe above (where perhaps we might not have many shrinker callback opportunities to act on before OOM).

FWIW, the only time I /ever/ saw OOMs (either now with 5.19 or ages ago with 5.9) was with extreme lowmem testing. On VMs with more than a gigabyte or so of memory, I notice that we usually hit the time/length thresholds without shrinkers getting involved. Granted, I did try to pick the shrinker values so that we only get called on the *second* level of shrinking, which is after we've freed some but not enough memory.

If my responses seem a bit like handwaving, they are, because (at the moment) this is well off in the weeds.

Quoted 6 lines
Quoted 6 lines

Back in the old days when inodegc was a radix tree tag it was fairly easy to get OOMs when the delay interval was long (5 seconds). The OOM probability went down pretty sharply as the interval approached zero, but even at 1 jiffy I could still occasionally trip it, whereas the pre-deferred-inactivation kernels would never OOM.

Quoted 7 lines
Quoted 4 lines

I haven't tested it all that rigorously with Dave's fancy new per-cpu list design, but I did throw on my silly test setup (see below) and still got it to OOM once in 3 runs with the shrinker bits turned off.

Ok.. so that implies we still need throttling, but I'm not sure what "fancy percpu list design" refers to. If you have a good test case, I

You might recall that the original versions of deferred inode inactivation would set radix tree tags, and a per-AG workqueue function would scan that AG's radix tree for tagged inodes and inactivate them. Dave observed that letting the cpu caches grow cold led to performance problems, and replaced the tagging mechanism with per-cpu lists, which is what we have now.

Quoted 2 lines

think the interesting immediate question is: are those OOMs avoided with this patch but the shrinker infrastructure still in place?

Quoted 8 lines
Quoted 2 lines

That's the $64000 question that I don't know definitively yet.

Ok. If the answer to the above questions is "don't really know, it might be broken, but historical testing shows the shrinker is much less important with smaller queuing delays such that it's unlikely to cause any new problems on anything but the most pathological workloads," then that sounds reasonable to me.

Yep, pretty much. :/

Quoted 3 lines

I don't mean to harp on it. I just want to make sure the question/concern gets across clearly enough that somebody who knows this system better than I can grok it and make a more intelligent assessment.

Oh, I'm glad you're harping on it -- anything involving shrinkers and deferred work ought to get a closer look.

--D

Quoted 81 lines

Brian

Quoted 78 lines
Quoted 4 lines

If not, then I wonder if something is going wonky there. If so, I'm still a bit curious what the behavior looks like and whether it can be simplified in light of this change, but that's certainly beyond the scope of this patch.

<nod> For now I'd like to get this going for 5.19 to fix the original complaint about statfs blocking in D state while waiting for inodegc after deleting a ~100m file extents[1], and defer the decision of whether or not we /really/ need the shrinker to a later time (like 5.20).

[1] https://lore.kernel.org/linux-xfs/[email protected]/

Quoted 28 lines
Quoted 28 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

If
so, it looks like decent enough overhead to cycle through every cpu in both callbacks that it might be worth spelling out more clearly in the top-level comment.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

I'm not sure what you are asking here - mod_delayed_work_on() has pretty much the same overhead and behaviour as queue_work() in this case, so... ?

I'm just pointing out that the comment around the shrinker infrastructure isn't very informative if the shrinker turns out to still be necessary for reasons other than making the workers run sooner.

<nod> That comment /does/ need to be updated to note the subtlety that a lot of shrinker activity can slow down close()ing a file by making user tasks wait for the inodegc workers to clear the backlog.

Quoted 8 lines
Quoted 3 lines

<shrug> Looks ok to me, since djwong-dev has had some variant of timer delayed inactivation in it longer than it hasn't:

Was that with a correspondingly small delay or something larger (on the order of seconds or so)? Either way, it sounds like you have a predictable enough workload that can actually test this continues to work as expected..?

Quoted 5 lines
Quoted 3 lines

Yeah. I snapshot /home (which has ~20 million inodes) then race fsstress and xfs_scrub -n in a VM with 560MB of RAM.

Yeah small delay or yeah large delay?

Both -- with large inactivation delays and no shrinker, OOMs happen fairly frequently; with a short delay and no shrinker, they're harder (but still not impossible) to trigger.

Granted ... even the overall frequency of OOMs with large inactivation delays seems to have gone down a bit from when I was more actively testing in the 5.9 era.

--D

Quoted 20 lines

Brian

Quoted 17 lines

--D

Quoted 14 lines

Brian

Quoted 11 lines

--D

Quoted 6 lines

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#11 lore raw

PATCH v2 2/2 xfs: introduce xfs_inodegc_push()

In reply to Dave Chinner

From: Dave Chinner <[email protected]>

The current blocking mechanism for pushing the inodegc queue out to disk can result in systems becoming unusable when there is a long running inodegc operation. This is because the statfs() implementation currently issues a blocking flush of the inodegc queue and a significant number of common system utilities will call statfs() to discover something about the underlying filesystem.

This can result in userspace operations getting stuck on inodegc progress, and when trying to remove a heavily reflinked file on slow storage with a full journal, this can result in delays measuring in hours.

Avoid this problem by adding "push" function that expedites the flushing of the inodegc queue, but doesn't wait for it to complete.

Convert xfs_fs_statfs() and xfs_qm_scall_getquota() to use this mechanism so they don't block but still ensure that queued operations are expedited.

Fixes ab23a7768739 ("xfs: per-cpu deferred inode inactivation queues")
---
 fs/xfs/xfs_icache.c      | 20 +++++++++++++++-----
 fs/xfs/xfs_icache.h      |  1 +
 fs/xfs/xfs_qm_syscalls.c |  7 +++++--
 fs/xfs/xfs_super.c       |  7 +++++--
 fs/xfs/xfs_trace.h       |  1 +
 5 files changed, 27 insertions(+), 9 deletions(-)
fs/xfs/xfs_icache.c +15 -5
index 46b30ecf498c..aef4097ffd3e 100644
@@ -1865,19 +1865,29 @@ xfs_inodegc_worker(
 }
 
 /*
- * Force all currently queued inode inactivation work to run immediately and
- * wait for the work to finish.
+ * Expedite all pending inodegc work to run immediately. This does not wait for
+ * completion of the work.
  */
 void
-xfs_inodegc_flush(
+xfs_inodegc_push(
 	struct xfs_mount	*mp)
 {
 	if (!xfs_is_inodegc_enabled(mp))
 		return;
+	trace_xfs_inodegc_push(mp, __return_address);
+	xfs_inodegc_queue_all(mp);
+}
 
+/*
+ * Force all currently queued inode inactivation work to run immediately and
+ * wait for the work to finish.
+ */
+void
+xfs_inodegc_flush(
+	struct xfs_mount	*mp)
+{
+	xfs_inodegc_push(mp);
 	trace_xfs_inodegc_flush(mp, __return_address);
-
-	xfs_inodegc_queue_all(mp);
 	flush_workqueue(mp->m_inodegc_wq);
 }
 
fs/xfs/xfs_icache.h +1 -0
index 2e4cfddf8b8e..6cd180721659 100644
@@ -76,6 +76,7 @@ void xfs_blockgc_stop(struct xfs_mount *mp);
 void xfs_blockgc_start(struct xfs_mount *mp);
 
 void xfs_inodegc_worker(struct work_struct *work);
+void xfs_inodegc_push(struct xfs_mount *mp);
 void xfs_inodegc_flush(struct xfs_mount *mp);
 void xfs_inodegc_stop(struct xfs_mount *mp);
 void xfs_inodegc_start(struct xfs_mount *mp);
fs/xfs/xfs_qm_syscalls.c +5 -2
index 74ac9ca9e119..a30f4d067746 100644
@@ -454,9 +454,12 @@ xfs_qm_scall_getquota(
 	struct xfs_dquot	*dqp;
 	int			error;
 
-	/* Flush inodegc work at the start of a quota reporting scan. */
+	/*
+	 * Expedite pending inodegc work at the start of a quota reporting
+	 * scan but don't block waiting for it to complete.
+	 */
 	if (id == 0)
-		xfs_inodegc_flush(mp);
+		xfs_inodegc_push(mp);
 
 	/*
 	 * Try to get the dquot. We don't want it allocated on disk, so don't
fs/xfs/xfs_super.c +5 -2
index 651ae75a7e23..4edee1d3784a 100644
@@ -798,8 +798,11 @@ xfs_fs_statfs(
 	xfs_extlen_t		lsize;
 	int64_t			ffree;
 
-	/* Wait for whatever inactivations are in progress. */
-	xfs_inodegc_flush(mp);
+	/*
+	 * Expedite background inodegc but don't wait. We do not want to block
+	 * here waiting hours for a billion extent file to be truncated.
+	 */
+	xfs_inodegc_push(mp);
 
 	statp->f_type = XFS_SUPER_MAGIC;
 	statp->f_namelen = MAXNAMELEN - 1;
fs/xfs/xfs_trace.h +1 -0
index d32026585c1b..0fa1b7a2918c 100644
@@ -240,6 +240,7 @@ DEFINE_EVENT(xfs_fs_class, name,					\
 	TP_PROTO(struct xfs_mount *mp, void *caller_ip), \
 	TP_ARGS(mp, caller_ip))
 DEFINE_FS_EVENT(xfs_inodegc_flush);
+DEFINE_FS_EVENT(xfs_inodegc_push);
 DEFINE_FS_EVENT(xfs_inodegc_start);
 DEFINE_FS_EVENT(xfs_inodegc_stop);
 DEFINE_FS_EVENT(xfs_inodegc_queue);
Signature · 1 line
2.35.1
Reviewed-by on v2 2/2
#12 lore raw

Re: [PATCH 2/2] xfs: introduce xfs_inodegc_push()

Quoted 17 lines · Dave Chinner

From: Dave Chinner <[email protected]>

The current blocking mechanism for pushing the inodegc queue out to disk can result in systems becoming unusable when there is a long running inodegc operation. This is because the statfs() implementation currently issues a blocking flush of the inodegc queue and a significant number of common system utilities will call statfs() to discover something about the underlying filesystem.

This can result in userspace operations getting stuck on inodegc progress, and when trying to remove a heavily reflinked file on slow storage with a full journal, this can result in delays measuring in hours.

Avoid this problem by adding "push" function that expedites the flushing of the inodegc queue, but doesn't wait for it to complete.

Quoted 7 lines

Convert xfs_fs_statfs() and xfs_qm_scall_getquota() to use this mechanism so they don't block but still ensure that queued operations are expedited.

Fixes ab23a7768739 ("xfs: per-cpu deferred inode inactivation queues")

Looks good,

--D

Quoted 111 lines
---
 fs/xfs/xfs_icache.c      | 20 +++++++++++++++-----
 fs/xfs/xfs_icache.h      |  1 +
 fs/xfs/xfs_qm_syscalls.c |  7 +++++--
 fs/xfs/xfs_super.c       |  7 +++++--
 fs/xfs/xfs_trace.h       |  1 +
 5 files changed, 27 insertions(+), 9 deletions(-)
fs/xfs/xfs_icache.c +15 -5
index 46b30ecf498c..aef4097ffd3e 100644
@@ -1865,19 +1865,29 @@ xfs_inodegc_worker(
 }
 
 /*
- * Force all currently queued inode inactivation work to run immediately and
- * wait for the work to finish.
+ * Expedite all pending inodegc work to run immediately. This does not wait for
+ * completion of the work.
  */
 void
-xfs_inodegc_flush(
+xfs_inodegc_push(
 	struct xfs_mount	*mp)
 {
 	if (!xfs_is_inodegc_enabled(mp))
 		return;
+	trace_xfs_inodegc_push(mp, __return_address);
+	xfs_inodegc_queue_all(mp);
+}
 
+/*
+ * Force all currently queued inode inactivation work to run immediately and
+ * wait for the work to finish.
+ */
+void
+xfs_inodegc_flush(
+	struct xfs_mount	*mp)
+{
+	xfs_inodegc_push(mp);
 	trace_xfs_inodegc_flush(mp, __return_address);
-
-	xfs_inodegc_queue_all(mp);
 	flush_workqueue(mp->m_inodegc_wq);
 }
 
fs/xfs/xfs_icache.h +1 -0
index 2e4cfddf8b8e..6cd180721659 100644
@@ -76,6 +76,7 @@ void xfs_blockgc_stop(struct xfs_mount *mp);
 void xfs_blockgc_start(struct xfs_mount *mp);
 
 void xfs_inodegc_worker(struct work_struct *work);
+void xfs_inodegc_push(struct xfs_mount *mp);
 void xfs_inodegc_flush(struct xfs_mount *mp);
 void xfs_inodegc_stop(struct xfs_mount *mp);
 void xfs_inodegc_start(struct xfs_mount *mp);
fs/xfs/xfs_qm_syscalls.c +5 -2
index 74ac9ca9e119..a30f4d067746 100644
@@ -454,9 +454,12 @@ xfs_qm_scall_getquota(
 	struct xfs_dquot	*dqp;
 	int			error;
 
-	/* Flush inodegc work at the start of a quota reporting scan. */
+	/*
+	 * Expedite pending inodegc work at the start of a quota reporting
+	 * scan but don't block waiting for it to complete.
+	 */
 	if (id == 0)
-		xfs_inodegc_flush(mp);
+		xfs_inodegc_push(mp);
 
 	/*
 	 * Try to get the dquot. We don't want it allocated on disk, so don't
fs/xfs/xfs_super.c +5 -2
index 651ae75a7e23..4edee1d3784a 100644
@@ -798,8 +798,11 @@ xfs_fs_statfs(
 	xfs_extlen_t		lsize;
 	int64_t			ffree;
 
-	/* Wait for whatever inactivations are in progress. */
-	xfs_inodegc_flush(mp);
+	/*
+	 * Expedite background inodegc but don't wait. We do not want to block
+	 * here waiting hours for a billion extent file to be truncated.
+	 */
+	xfs_inodegc_push(mp);
 
 	statp->f_type = XFS_SUPER_MAGIC;
 	statp->f_namelen = MAXNAMELEN - 1;
fs/xfs/xfs_trace.h +1 -0
index d32026585c1b..0fa1b7a2918c 100644
@@ -240,6 +240,7 @@ DEFINE_EVENT(xfs_fs_class, name,					\
 	TP_PROTO(struct xfs_mount *mp, void *caller_ip), \
 	TP_ARGS(mp, caller_ip))
 DEFINE_FS_EVENT(xfs_inodegc_flush);
+DEFINE_FS_EVENT(xfs_inodegc_push);
 DEFINE_FS_EVENT(xfs_inodegc_start);
 DEFINE_FS_EVENT(xfs_inodegc_stop);
 DEFINE_FS_EVENT(xfs_inodegc_queue);
Signature · 1 line
2.35.1