On Thu, Feb 11, 2021 at 5:10 AM Vlastimil Babka [off-list ref] wrote:

On 2/9/21 6:46 PM, Yang Shi wrote:

quoted

The number of deferred objects might get windup to an absurd number, and it
results in clamp of slab objects.  It is undesirable for sustaining workingset.

So shrink deferred objects proportional to priority and cap nr_deferred to twice
of cache items.

Makes sense to me, minimally it's simpler than the old code and avoiding absurd
growth of nr_deferred should be a good thing, as well as the "proportional to
priority" part.

Thanks.

I just suspect there's a bit of unnecessary bias in the implementation, as
explained below:

quoted

The idea is borrowed from Dave Chinner's patch:
https://lore.kernel.org/linux-xfs/20191031234618.15403-13-david@fromorbit.com/ (local)

Tested with kernel build and vfs metadata heavy workload in our production
environment, no regression is spotted so far.

Signed-off-by: Yang Shi <redacted>
---
 mm/vmscan.c | 40 +++++-----------------------------------
 1 file changed, 5 insertions(+), 35 deletions(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 66163082cc6f..d670b119d6bd 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c

@@ -654,7 +654,6 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
       */
      nr = count_nr_deferred(shrinker, shrinkctl);

-     total_scan = nr;
      if (shrinker->seeks) {
              delta = freeable >> priority;
              delta *= 4;

@@ -668,37 +667,9 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
              delta = freeable / 2;
      }

+     total_scan = nr >> priority;
      total_scan += delta;

So, our scan goal consists of the part based on freeable objects (delta), plus a
part of the defferred objects (nr >> priority). Fine.

quoted

-     if (total_scan < 0) {
-             pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n",
-                    shrinker->scan_objects, total_scan);
-             total_scan = freeable;
-             next_deferred = nr;
-     } else
-             next_deferred = total_scan;
-
-     /*
-      * We need to avoid excessive windup on filesystem shrinkers
-      * due to large numbers of GFP_NOFS allocations causing the
-      * shrinkers to return -1 all the time. This results in a large
-      * nr being built up so when a shrink that can do some work
-      * comes along it empties the entire cache due to nr >>>
-      * freeable. This is bad for sustaining a working set in
-      * memory.
-      *
-      * Hence only allow the shrinker to scan the entire cache when
-      * a large delta change is calculated directly.
-      */
-     if (delta < freeable / 4)
-             total_scan = min(total_scan, freeable / 2);
-
-     /*
-      * Avoid risking looping forever due to too large nr value:
-      * never try to free more than twice the estimate number of
-      * freeable entries.
-      */
-     if (total_scan > freeable * 2)
-             total_scan = freeable * 2;
+     total_scan = min(total_scan, (2 * freeable));

Probably unnecessary as we cap next_deferred below anyway? So total_scan cannot
grow without limits anymore. But can't hurt.

quoted

      trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
                                 freeable, delta, total_scan, priority);

@@ -737,10 +708,9 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
              cond_resched();
      }

-     if (next_deferred >= scanned)
-             next_deferred -= scanned;
-     else
-             next_deferred = 0;
+     next_deferred = max_t(long, (nr - scanned), 0) + total_scan;

And here's the bias I think. Suppose we scanned 0 due to e.g. GFP_NOFS. We count
as newly deferred both the "delta" part of total_scan, which is fine, but also
the "nr >> priority" part, where we failed to our share of the "reduce
nr_deferred" work, but I don't think it means we should also increase
nr_deferred by that amount of failed work.

Here "nr" is the saved deferred work since the last scan, "scanned" is
the scanned work in this round, total_scan is the *unscanned" work
which is actually "total_scan - scanned" (total_scan is decreased by
scanned in each loop). So, the logic is "decrease any scanned work
from deferred then add newly unscanned work to deferred". IIUC this is
what "deferred" means even before this patch.

OTOH if we succeed and scan exactly the whole goal, we are subtracting from
nr_deferred both the "nr >> priority" part, which is correct, but also delta,
which was new work, not deferred one, so that's incorrect IMHO as well.

I don't think so. The deferred comes from new work, why not dec new
work from deferred?

And, the old code did:

if (next_deferred >= scanned)
                next_deferred -= scanned;
        else
                next_deferred = 0;

IIUC, it also decreases the new work (the scanned includes both last
deferred and new delata).

So the calculation should probably be something like this?

        next_deferred = max_t(long, nr + delta - scanned, 0);

Thanks,
Vlastimil

quoted

+     next_deferred = min(next_deferred, (2 * freeable));
+
      /*
       * move the unused scan count back into the shrinker in a
       * manner that handles concurrent updates.