Re: [Patch] sched/fair: Avoid throttle_list starvation with low cfs quota

Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx> · Sun, 11 Nov 2018 12:20:27 +0300

On 10.10.2018 21:37, Phil Auld wrote:
On Wed, Oct 10, 2018 at 10:49:25AM -0700 bsegall@xxxxxxxxxx wrote:
Ingo Molnar <mingo@xxxxxxxxxx> writes:

I've Cc:-ed a handful of gents who worked on CFS bandwidth details to widen the discussion.
Patch quoted below.

Looks like a real bug that needs to be fixed - and at first sight the quota of 1000 looks very
low - could we improve the arithmetics perhaps?

A low quota of 1000 is used because there's many VMs or containers provisioned on the system
that is triggering the bug, right?

Thanks,

	Ingo

* Phil Auld <pauld@xxxxxxxxxx> wrote:

From: "Phil Auld" <pauld@xxxxxxxxxx>

sched/fair: Avoid throttle_list starvation with low cfs quota

With a very low cpu.cfs_quota_us setting, such as the minimum of 1000,
distribute_cfs_runtime may not empty the throttled_list before it runs
out of runtime to distribute. In that case, due to the change from
c06f04c7048 to put throttled entries at the head of the list, later entries
on the list will starve.  Essentially, the same X processes will get pulled
off the list, given CPU time and then, when expired, get put back on the
head of the list where distribute_cfs_runtime will give runtime to the same
set of processes leaving the rest.

Fix the issue by setting a bit in struct cfs_bandwidth when
distribute_cfs_runtime is running, so that the code in throttle_cfs_rq can
decide to put the throttled entry on the tail or the head of the list.  The
bit is set/cleared by the callers of distribute_cfs_runtime while they hold
cfs_bandwidth->lock.

Signed-off-by: Phil Auld <pauld@xxxxxxxxxx>
Fixes: c06f04c70489 ("sched: Fix potential near-infinite distribute_cfs_runtime() loop")
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx

Reviewed-by: Ben Segall <bsegall@xxxxxxxxxx>


In theory this does mean the unfairness could still happen if distribute is still
running, but while a tiny quota makes it more likely, the fact that
we're not getting through much of the list makes it not really a worry.
If you wanted to be even more careful there could be some generation
counter or something, but it doesn't seem necessary.


Thanks for the review.  Yeah, I thought about a few other approaches, not explicitly
what you suggested, but they all complicated things. This one seemed the closest to
"obviously correct".

I've sent patch about same problem couple years ago:
https://lore.kernel.org/patchwork/patch/750523/

I think my approach is closer to FIFO and more fair.



Cheers,
Phil



---

This is easy to reproduce with a handful of cpu consumers. I use crash on
the live system. In some cases you can simply look at the throttled list and
see the later entries are not changing:

crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1"  "$4}' | pr -t -n3
   1     ffff90b56cb2d200  -976050
   2     ffff90b56cb2cc00  -484925
   3     ffff90b56cb2bc00  -658814
   4     ffff90b56cb2ba00  -275365
   5     ffff90b166a45600  -135138
   6     ffff90b56cb2da00  -282505
   7     ffff90b56cb2e000  -148065
   8     ffff90b56cb2fa00  -872591
   9     ffff90b56cb2c000  -84687
  10     ffff90b56cb2f000  -87237
  11     ffff90b166a40a00  -164582
crash> list cfs_rq.throttled_list -H 0xffff90b54f6ade40 -s cfs_rq.runtime_remaining | paste - - | awk '{print $1"  "$4}' | pr -t -n3
   1     ffff90b56cb2d200  -994147
   2     ffff90b56cb2cc00  -306051
   3     ffff90b56cb2bc00  -961321
   4     ffff90b56cb2ba00  -24490
   5     ffff90b166a45600  -135138
   6     ffff90b56cb2da00  -282505
   7     ffff90b56cb2e000  -148065
   8     ffff90b56cb2fa00  -872591
   9     ffff90b56cb2c000  -84687
  10     ffff90b56cb2f000  -87237
  11     ffff90b166a40a00  -164582

Sometimes it is easier to see by finding a process getting starved and looking
at the sched_info:

crash> task ffff8eb765994500 sched_info
PID: 7800   TASK: ffff8eb765994500  CPU: 16  COMMAND: "cputest"
   sched_info = {
     pcount = 8,
     run_delay = 697094208,
     last_arrival = 240260125039,
     last_queued = 240260327513
   },
crash> task ffff8eb765994500 sched_info
PID: 7800   TASK: ffff8eb765994500  CPU: 16  COMMAND: "cputest"
   sched_info = {
     pcount = 8,
     run_delay = 697094208,
     last_arrival = 240260125039,
     last_queued = 240260327513
   },


  fair.c  |   22 +++++++++++++++++++---
  sched.h |    2 ++
  2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 7fc4a371bdd2..f88e00705b55 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4476,9 +4476,13 @@ static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
  
  	/*
  	 * Add to the _head_ of the list, so that an already-started
-	 * distribute_cfs_runtime will not see us
+	 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
+	 * not running add to the tail so that later runqueues don't get starved.
  	 */
-	list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
+	if (cfs_b->distribute_running)
+		list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
+	else
+		list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);

  	/*
  	 * If we're the first throttled task, make sure the bandwidth
@@ -4622,14 +4626,16 @@ static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
  	 * in us over-using our runtime if it is all used during this loop, but
  	 * only by limited amounts in that extreme case.
  	 */
-	while (throttled && cfs_b->runtime > 0) {
+	while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
  		runtime = cfs_b->runtime;
+		cfs_b->distribute_running = 1;
  		raw_spin_unlock(&cfs_b->lock);
  		/* we can't nest cfs_b->lock while distributing bandwidth */
  		runtime = distribute_cfs_runtime(cfs_b, runtime,
  						 runtime_expires);
  		raw_spin_lock(&cfs_b->lock);
  
+		cfs_b->distribute_running = 0;
  		throttled = !list_empty(&cfs_b->throttled_cfs_rq);
  
  		cfs_b->runtime -= min(runtime, cfs_b->runtime);
@@ -4740,6 +4746,11 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
  
  	/* confirm we're still not at a refresh boundary */
  	raw_spin_lock(&cfs_b->lock);
+	if (cfs_b->distribute_running) {
+		raw_spin_unlock(&cfs_b->lock);
+		return;
+	}
+
  	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
  		raw_spin_unlock(&cfs_b->lock);
  		return;
@@ -4749,6 +4760,9 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
  		runtime = cfs_b->runtime;
  
  	expires = cfs_b->runtime_expires;
+	if (runtime)
+		cfs_b->distribute_running = 1;
+
  	raw_spin_unlock(&cfs_b->lock);
  
  	if (!runtime)
@@ -4759,6 +4773,7 @@ static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
  	raw_spin_lock(&cfs_b->lock);
  	if (expires == cfs_b->runtime_expires)
  		cfs_b->runtime -= min(runtime, cfs_b->runtime);
+	cfs_b->distribute_running = 0;
  	raw_spin_unlock(&cfs_b->lock);
  }
  
@@ -4867,6 +4882,7 @@ void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
  	cfs_b->period_timer.function = sched_cfs_period_timer;
  	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
  	cfs_b->slack_timer.function = sched_cfs_slack_timer;
+	cfs_b->distribute_running = 0;
  }
  
  static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 455fa330de04..9683f458aec7 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -346,6 +346,8 @@ struct cfs_bandwidth {
  	int			nr_periods;
  	int			nr_throttled;
  	u64			throttled_time;
+
+	bool                    distribute_running;
  #endif
  };
  


--








