[PATCH] fs/dcache: resched/chill only if we make no progress

[PATCH] fs/dcache: resched/chill only if we make no progress

[Sebastian Andrzej Siewior]

Upstream commit 47be61845c77 ("fs/dcache.c: avoid soft-lockup in
dput()") changed the condition _when_ cpu_relax() / cond_resched() was
invoked. This change was adapted in -RT into mostly the same thing
except that if cond_resched() did nothing we had to do cpu_chill() to
force the task off CPU for a tiny little bit in case the task had RT
priority and did not want to leave the CPU.
This change resulted in a performance regression (in my testcase the
build time on /dev/shm increased from 19min to 24min). The reason is
that with this change cpu_chill() was invoked even dput() made progress
(dentry_kill() returned a different dentry) instead only if we were
trying this operation on the same dentry over and over again.

This patch brings back to the old behavior back to cond_resched() &
chill if we make no progress. A little improvement is to invoke
cpu_chill() only if we are a RT task (and avoid the sleep otherwise).
Otherwise the scheduler should remove us from the CPU if we make no
progress.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 fs/dcache.c |   19 +++++++++++++------
 1 file changed, 13 insertions(+), 6 deletions(-)

--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -40,6 +40,8 @@
 #include <linux/ratelimit.h>
 #include <linux/list_lru.h>
 #include <linux/kasan.h>
+#include <linux/sched/rt.h>
+#include <linux/sched/deadline.h>
 
 #include "internal.h"
 #include "mount.h"
@@ -748,6 +750,8 @@ static inline bool fast_dput(struct dent
  */
 void dput(struct dentry *dentry)
 {
+	struct dentry *parent;
+
 	if (unlikely(!dentry))
 		return;
 
@@ -784,14 +788,17 @@ void dput(struct dentry *dentry)
 	return;
 
 kill_it:
-	dentry = dentry_kill(dentry);
-	if (dentry) {
+	parent = dentry_kill(dentry);
+	if (parent) {
 		int r;
 
-		/* the task with the highest priority won't schedule */
-		r = cond_resched();
-		if (!r)
-			cpu_chill();
+		if (parent == dentry) {
+			/* the task with the highest priority won't schedule */
+			r = cond_resched();
+			if (!r && (rt_task(current) || dl_task(current)))
+				cpu_chill();
+		} else
+			dentry = parent;
 		goto repeat;
 	}
 }

Re: [PATCH] fs/dcache: resched/chill only if we make no progress

[Thomas Gleixner]

On Thu, 8 Sep 2016, Sebastian Andrzej Siewior wrote:
> +		if (parent == dentry) {
> +			/* the task with the highest priority won't schedule */
> +			r = cond_resched();
> +			if (!r && (rt_task(current) || dl_task(current)))

That's wrong because a non RT task will just spin until its time slice is
up. Wasted cycles for nothing.

Ideally we make spinning conditional on 

    !on_cpu(dentry->lock->owner)

That would avoid going to sleep if the owner runs on another cpu. We have
no mechanism for that, but I think it'd be worth to think about one.

> +				cpu_chill();
> +		} else
> +			dentry = parent;

Missing braces

Thanks,

	tglx

Re: [PATCH] fs/dcache: resched/chill only if we make no progress

[Sebastian Andrzej Siewior]

On 2016-09-10 08:17:13 [+0200], Thomas Gleixner wrote:
> On Thu, 8 Sep 2016, Sebastian Andrzej Siewior wrote:
> > +		if (parent == dentry) {
> > +			/* the task with the highest priority won't schedule */
> > +			r = cond_resched();
> > +			if (!r && (rt_task(current) || dl_task(current)))
> 
> That's wrong because a non RT task will just spin until its time slice is
> up. Wasted cycles for nothing.
> 
> Ideally we make spinning conditional on 
> 
>     !on_cpu(dentry->lock->owner)
> 
> That would avoid going to sleep if the owner runs on another cpu. We have
> no mechanism for that, but I think it'd be worth to think about one.

something like

|#ifdef CONFIG_SMP
|                             task_running(task_rq(lock_owner), lock_owner))
|#else
|                             1)
|#endif
|                                 cpu_chill();

the lock_owner is tricky because we have two locks: inode->i_lock and
parent->d_lock. So we would have to do the same lookup, like
dentry_kill() is doing, and check both in worth case.

Sebastian