Motivation & Approach
=====================

While testing workloads with high sustained memory pressure on large machines
in the Meta fleet (1Tb memory, 316 CPUs), we saw an unexpectedly high number
of softlockups. Further investigation showed that the zone lock in
free_pcppages_bulk was being held for a long time, and was called to free
2k+ pages over 100 times just during boot.

This causes starvation in other processes for the zone lock, which can lead
to the system stalling as multiple threads cannot make progress without the
locks. We can see these issues manifesting as warnings:

[ 4512.591979] rcu: INFO: rcu_sched self-detected stall on CPU
[ 4512.604370] rcu:     20-....: (9312 ticks this GP) idle=a654/1/0x4000000000000000 softirq=309340/309344 fqs=5426
[ 4512.626401] rcu:              hardirqs   softirqs   csw/system
[ 4512.638793] rcu:      number:        0        145            0
[ 4512.651177] rcu:     cputime:       30      10410          174   ==> 10558(ms)
[ 4512.666657] rcu:     (t=21077 jiffies g=783665 q=1242213 ncpus=316)

While these warnings are benign, they do point to the underlying issue of
lock contention. To prevent starvation in both locks, batch the freeing of
pages using pcp->batch.

Because free_pcppages_bulk is called with the pcp lock and acquires the zone
lock, relinquishing and reacquiring the locks are only effective when both of
them are broken together (unless the system was built with queued spinlocks).
Thus, instead of modifying free_pcppages_bulk to break both locks, batch the
freeing from its callers instead.

A similar fix has been implemented in the Meta fleet, and we have seen
significantly less softlockups.

Testing
=======
The following are a few synthetic benchmarks, made on three machines. The
first is a large machine with 754GiB memory and 316 processors.
The second is a relatively smaller machine with 251GiB memory and 176
processors. The third and final is the smallest of the three, which has 62GiB
memory and 36 processors.

On all machines, I kick off a kernel build with -j$(nproc).
Negative delta is better (faster compilation).

Large machine (754GiB memory, 316 processors)
make -j$(nproc)
+------------+---------------+-----------+
| Metric (s) | Variation (%) | Delta(%)  |
+------------+---------------+-----------+
| real       |        0.8070 |  - 1.4865 |
| user       |        0.2823 |  + 0.4081 |
| sys        |        5.0267 |  -11.8737 |
+------------+---------------+-----------+

Medium machine (251GiB memory, 176 processors)
make -j$(nproc)
+------------+---------------+----------+
| Metric (s) | Variation (%) | Delta(%) |
+------------+---------------+----------+
| real       |        0.2806 |  +0.0351 |
| user       |        0.0994 |  +0.3170 |
| sys        |        0.6229 |  -0.6277 |
+------------+---------------+----------+

Small machine (62GiB memory, 36 processors)
make -j$(nproc)
+------------+---------------+----------+
| Metric (s) | Variation (%) | Delta(%) |
+------------+---------------+----------+
| real       |        0.1503 |  -2.6585 |
| user       |        0.0431 |  -2.2984 |
| sys        |        0.1870 |  -3.2013 |
+------------+---------------+----------+

Here, variation is the coefficient of variation, i.e. standard deviation / mean.

Based on these results, it seems like there are varying degrees to how much
lock contention this reduces. For the largest and smallest machines that I ran
the tests on, it seems like there is quite some significant reduction. There
is also some performance increases visible from userspace.

Interestingly, the performance gains don't scale with the size of the machine,
but rather there seems to be a dip in the gain there is for the medium-sized
machine.

Changelog
=========
v3 --> v4:
- Patches 1/3 and 2/3 were left untouched, other than adding review tags
  and a small clairification in 2/3 to note impact on the zone lock.
- Patch 3/3 now uses a while loop, instead of a confusing goto statement.
- Patch 3/3 now checks ZONE_BELOW_HIGH once at the end of the function, and
  high is calculated just once as well, before the while loop. Both suggestions
  were made by Vlastimil Babka, to improve readability and to stick more closely
  to the original scope of the function.
- It turns out that omitting the repeated zone flag check and high calculation
  leads to a performance increase for all machine types.
  The cover letter includes the most recent test results.
- I've also included the test results in patch 3/3, so that the numbers are
  there and can be referenced in the commit log in the future as well.

v2 --> v3:
- Refactored on top of mm-new
- Wordsmithing the cover letter & commit messages to clarify which lock
  is contended, as suggested by Hillf Danton.
- Ran new tests for the cover letter, instead of running stress-ng, I decided
  to compile the kernel which I think will be more reflective of the "default"
  workload that might be run. Also ran on a smaller machines to show the
  expected behavior of this patchset when there is lock contention vs.
  lower lock contention.
- Removed patch 2/4, which would have batched page freeing for
  drain_pages_zone. It is not a good candidate for this series since it is
  called on each CPU in __drain_all_pages.
- Small change in 1/4 to initialize todo, as suggested by Christoph Lameter
- Small change in 1/4 to avoid bit manipulation, as suggested by SeongJae Park.
- Change in 4/4 to handle the case when the thread gets migrated to a different
  CPU during the window between unlocking & reacquiring the pcp lock, as
  suggested by Vlastimil Babka.
- Small change in 4/4 to handle the case when pcp lock could not be acquired
  within the loop in free_unref_folios.

v1 --> v2:
- Reworded cover letter to be more explicit about what kinds of issues
  running processes might face as a result of the existing lock starvation
- Reworded cover letter to be in sections to make it easier to read
- Fixed patch 4/4 to properly store & restore UP flags.
- Re-ran tests, updated the testing results and interpretation

Joshua Hahn (3):
  mm/page_alloc/vmstat: Simplify refresh_cpu_vm_stats change detection
  mm/page_alloc: Batch page freeing in decay_pcp_high
  mm/page_alloc: Batch page freeing in free_frozen_page_commit

 include/linux/gfp.h |  2 +-
 mm/page_alloc.c     | 83 ++++++++++++++++++++++++++++++++++++---------
 mm/vmstat.c         | 28 ++++++++-------
 3 files changed, 83 insertions(+), 30 deletions(-)


base-commit: 53e573001f2b5168f9b65d2b79e9563a3b479c17
-- 
2.47.3

On Mon, 13 Oct 2025 12:08:08 -0700 Joshua Hahn wrote:
> Motivation & Approach
> =====================
> 
> While testing workloads with high sustained memory pressure on large machines
> in the Meta fleet (1Tb memory, 316 CPUs), we saw an unexpectedly high number
> of softlockups. Further investigation showed that the zone lock in
> free_pcppages_bulk was being held for a long time, and was called to free
> 2k+ pages over 100 times just during boot.
> 
> This causes starvation in other processes for the zone lock, which can lead
> to the system stalling as multiple threads cannot make progress without the
> locks. We can see these issues manifesting as warnings:
> 
> [ 4512.591979] rcu: INFO: rcu_sched self-detected stall on CPU
> [ 4512.604370] rcu:     20-....: (9312 ticks this GP) idle=a654/1/0x4000000000000000 softirq=309340/309344 fqs=5426
> [ 4512.626401] rcu:              hardirqs   softirqs   csw/system
> [ 4512.638793] rcu:      number:        0        145            0
> [ 4512.651177] rcu:     cputime:       30      10410          174   ==> 10558(ms)
> [ 4512.666657] rcu:     (t=21077 jiffies g=783665 q=1242213 ncpus=316)
> 
> While these warnings are benign, they do point to the underlying issue of

No fix is needed if it is benign.

> lock contention. To prevent starvation in both locks, batch the freeing of
> pages using pcp->batch.
> 
> Because free_pcppages_bulk is called with the pcp lock and acquires the zone
> lock, relinquishing and reacquiring the locks are only effective when both of
> them are broken together (unless the system was built with queued spinlocks).
> Thus, instead of modifying free_pcppages_bulk to break both locks, batch the
> freeing from its callers instead.
> 
> A similar fix has been implemented in the Meta fleet, and we have seen
> significantly less softlockups.
> 
Fine, softlockup is not cured.

> Testing
> =======
> The following are a few synthetic benchmarks, made on three machines. The
> first is a large machine with 754GiB memory and 316 processors.
> The second is a relatively smaller machine with 251GiB memory and 176
> processors. The third and final is the smallest of the three, which has 62GiB
> memory and 36 processors.
> 
> On all machines, I kick off a kernel build with -j$(nproc).
> Negative delta is better (faster compilation).
> 
> Large machine (754GiB memory, 316 processors)
> make -j$(nproc)
> +------------+---------------+-----------+
> | Metric (s) | Variation (%) | Delta(%)  |
> +------------+---------------+-----------+
> | real       |        0.8070 |  - 1.4865 |
> | user       |        0.2823 |  + 0.4081 |
> | sys        |        5.0267 |  -11.8737 |
> +------------+---------------+-----------+
> 
> Medium machine (251GiB memory, 176 processors)
> make -j$(nproc)
> +------------+---------------+----------+
> | Metric (s) | Variation (%) | Delta(%) |
> +------------+---------------+----------+
> | real       |        0.2806 |  +0.0351 |
> | user       |        0.0994 |  +0.3170 |
> | sys        |        0.6229 |  -0.6277 |
> +------------+---------------+----------+
> 
> Small machine (62GiB memory, 36 processors)
> make -j$(nproc)
> +------------+---------------+----------+
> | Metric (s) | Variation (%) | Delta(%) |
> +------------+---------------+----------+
> | real       |        0.1503 |  -2.6585 |
> | user       |        0.0431 |  -2.2984 |
> | sys        |        0.1870 |  -3.2013 |
> +------------+---------------+----------+
> 
> Here, variation is the coefficient of variation, i.e. standard deviation / mean.
> 
> Based on these results, it seems like there are varying degrees to how much
> lock contention this reduces. For the largest and smallest machines that I ran
> the tests on, it seems like there is quite some significant reduction. There
> is also some performance increases visible from userspace.
> 
> Interestingly, the performance gains don't scale with the size of the machine,
> but rather there seems to be a dip in the gain there is for the medium-sized
> machine.
>
Explaining the dip helps land this work in the next tree.

On Tue, 14 Oct 2025 19:29:45 +0800 Hillf Danton <hdanton@sina.com> wrote:

> On Mon, 13 Oct 2025 12:08:08 -0700 Joshua Hahn wrote:
> > Motivation & Approach
> > =====================
> > 
> > While testing workloads with high sustained memory pressure on large machines
> > in the Meta fleet (1Tb memory, 316 CPUs), we saw an unexpectedly high number
> > of softlockups. Further investigation showed that the zone lock in
> > free_pcppages_bulk was being held for a long time, and was called to free
> > 2k+ pages over 100 times just during boot.
> > 
> > This causes starvation in other processes for the zone lock, which can lead
> > to the system stalling as multiple threads cannot make progress without the
> > locks. We can see these issues manifesting as warnings:
> > 
> > [ 4512.591979] rcu: INFO: rcu_sched self-detected stall on CPU
> > [ 4512.604370] rcu:     20-....: (9312 ticks this GP) idle=a654/1/0x4000000000000000 softirq=309340/309344 fqs=5426
> > [ 4512.626401] rcu:              hardirqs   softirqs   csw/system
> > [ 4512.638793] rcu:      number:        0        145            0
> > [ 4512.651177] rcu:     cputime:       30      10410          174   ==> 10558(ms)
> > [ 4512.666657] rcu:     (t=21077 jiffies g=783665 q=1242213 ncpus=316)

Hello Hillf, thank you for your review.

> > While these warnings are benign, they do point to the underlying issue of
> 
> No fix is needed if it is benign.

Maybe this is poor wording on my part. What I mean to say is that these
warning messages can help us understand that the system is trending negatively,
even though the warning messages themselves may not indiate that something
has crashed or broken completely.

> > lock contention. To prevent starvation in both locks, batch the freeing of
> > pages using pcp->batch.
> > 
> > Because free_pcppages_bulk is called with the pcp lock and acquires the zone
> > lock, relinquishing and reacquiring the locks are only effective when both of
> > them are broken together (unless the system was built with queued spinlocks).
> > Thus, instead of modifying free_pcppages_bulk to break both locks, batch the
> > freeing from its callers instead.
> > 
> > A similar fix has been implemented in the Meta fleet, and we have seen
> > significantly less softlockups.
> > 
> Fine, softlockup is not cured.
> 
> > Testing
> > =======
> > The following are a few synthetic benchmarks, made on three machines. The
> > first is a large machine with 754GiB memory and 316 processors.
> > The second is a relatively smaller machine with 251GiB memory and 176
> > processors. The third and final is the smallest of the three, which has 62GiB
> > memory and 36 processors.
> > 
> > On all machines, I kick off a kernel build with -j$(nproc).
> > Negative delta is better (faster compilation).
> > 
> > Large machine (754GiB memory, 316 processors)
> > make -j$(nproc)
> > +------------+---------------+-----------+
> > | Metric (s) | Variation (%) | Delta(%)  |
> > +------------+---------------+-----------+
> > | real       |        0.8070 |  - 1.4865 |
> > | user       |        0.2823 |  + 0.4081 |
> > | sys        |        5.0267 |  -11.8737 |
> > +------------+---------------+-----------+
> > 
> > Medium machine (251GiB memory, 176 processors)
> > make -j$(nproc)
> > +------------+---------------+----------+
> > | Metric (s) | Variation (%) | Delta(%) |
> > +------------+---------------+----------+
> > | real       |        0.2806 |  +0.0351 |
> > | user       |        0.0994 |  +0.3170 |
> > | sys        |        0.6229 |  -0.6277 |
> > +------------+---------------+----------+
> > 
> > Small machine (62GiB memory, 36 processors)
> > make -j$(nproc)
> > +------------+---------------+----------+
> > | Metric (s) | Variation (%) | Delta(%) |
> > +------------+---------------+----------+
> > | real       |        0.1503 |  -2.6585 |
> > | user       |        0.0431 |  -2.2984 |
> > | sys        |        0.1870 |  -3.2013 |
> > +------------+---------------+----------+
> > 
> > Here, variation is the coefficient of variation, i.e. standard deviation / mean.
> > 
> > Based on these results, it seems like there are varying degrees to how much
> > lock contention this reduces. For the largest and smallest machines that I ran
> > the tests on, it seems like there is quite some significant reduction. There
> > is also some performance increases visible from userspace.
> > 
> > Interestingly, the performance gains don't scale with the size of the machine,
> > but rather there seems to be a dip in the gain there is for the medium-sized
> > machine.
> >
> Explaining the dip helps land this work in the next tree.

I do agree that I left this on a bit of a cliffhanger here. I'm a bit confused
as to why there is this kind of behavior as well, although I have a theory
as to why this behavior is seen. Going back to why we see zone lock contention
in the first place, I think it might have to do with the memory vs. processors
ratio that leads to such contention issues.

The lower the memory:processor ratio is, it seems like there is already
less zone lock contention. If we rank these machines by their mem:proc ratio:

Large machine : 2.38
Small machine : 1.72
Medium machine: 1.42

It seems like this is the order in which we see the gains as well. I this
explanation also kind of makes sense -- the more memory we have, the more
memory each pcp will have, and the longer free_pcppages_bulk would have taken
before (and vice versa). This is the case, at least for my setup, where each
machine is onlined in one node (zone) and so the pcp watermarks really
do scale with the size of the system.

I didn't want to include this in the cover letter, because this was purely an
untested conjecture.

I hope this helps!
Joshua