On Mon, Oct 28, 2024 at 8:07 PM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
>
>
>
> On 27/10/2024 01:14, Barry Song wrote:
> > From: Barry Song <v-songbaohua@xxxxxxxx>
> >
> > In a memcg where mTHP is always utilized, even at full capacity, it
> > might not be the best option. Consider a system that uses only small
> > folios: after each reclamation, a process has at least SWAP_CLUSTER_MAX
> > of buffer space before it can initiate the next reclamation. However,
> > large folios can quickly fill this space, rapidly bringing the memcg
> > back to full capacity, even though some portions of the large folios
> > may not be immediately needed and used by the process.
> >
> > Usama and Kanchana identified a regression when building the kernel in
> > a memcg with memory.max set to a small value while enabling large
> > folio swap-in support on zswap[1].
> >
> > The issue arises from an edge case where the memory cgroup remains
> > nearly full most of the time. Consequently, bringing in mTHP can
> > quickly cause a memcg overflow, triggering a swap-out. The subsequent
> > swap-in then recreates the overflow, resulting in a repetitive cycle.
> >
> > We need a mechanism to stop the cup from overflowing continuously.
> > One potential solution is to slow the filling process when we identify
> > that the cup is nearly full.
> >
> > Usama reported an improvement when we mitigate mTHP swap-in as the
> > memcg approaches full capacity[2]:
> >
> > int mem_cgroup_swapin_charge_folio(...)
> > {
> > ...
> > if (folio_test_large(folio) &&
> > mem_cgroup_margin(memcg) < max(MEMCG_CHARGE_BATCH, folio_nr_pages(folio)))
> > ret = -ENOMEM;
> > else
> > ret = charge_memcg(folio, memcg, gfp);
> > ...
> > }
> >
> > AMD 16K+32K THP=always
> > metric mm-unstable mm-unstable + large folio zswapin series mm-unstable + large folio zswapin + no swap thrashing fix
> > real 1m23.038s 1m23.050s 1m22.704s
> > user 53m57.210s 53m53.437s 53m52.577s
> > sys 7m24.592s 7m48.843s 7m22.519s
> > zswpin 612070 999244 815934
> > zswpout 2226403 2347979 2054980
> > pgfault 20667366 20481728 20478690
> > pgmajfault 385887 269117 309702
> >
> > AMD 16K+32K+64K THP=always
> > metric mm-unstable mm-unstable + large folio zswapin series mm-unstable + large folio zswapin + no swap thrashing fix
> > real 1m22.975s 1m23.266s 1m22.549s
> > user 53m51.302s 53m51.069s 53m46.471s
> > sys 7m40.168s 7m57.104s 7m25.012s
> > zswpin 676492 1258573 1225703
> > zswpout 2449839 2714767 2899178
> > pgfault 17540746 17296555 17234663
> > pgmajfault 429629 307495 287859
> >
> > I wonder if we can extend the mitigation to do_anonymous_page() as
> > well. Without hardware like AMD and ARM with hardware TLB coalescing
> > or CONT-PTE, I conducted a quick test on my Intel i9 workstation with
> > 10 cores and 2 threads. I enabled one 12 GiB zRAM while running kernel
> > builds in a memcg with memory.max set to 1 GiB.
> >
> > $ echo always > /sys/kernel/mm/transparent_hugepage/hugepages-64kB/enabled
> > $ echo always > /sys/kernel/mm/transparent_hugepage/hugepages-32kB/enabled
> > $ echo always > /sys/kernel/mm/transparent_hugepage/hugepages-16kB/enabled
> > $ echo never > /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled
> >
> > $ time systemd-run --scope -p MemoryMax=1G make ARCH=arm64 \
> > CROSS_COMPILE=aarch64-linux-gnu- Image -10 1>/dev/null 2>/dev/null
> >
> > disable-mTHP-swapin mm-unstable with-this-patch
> > Real: 6m54.595s 7m4.832s 6m45.811s
> > User: 66m42.795s 66m59.984s 67m21.150s
> > Sys: 12m7.092s 15m18.153s 12m52.644s
> > pswpin: 4262327 11723248 5918690
> > pswpout: 14883774 19574347 14026942
> > 64k-swpout: 624447 889384 480039
> > 32k-swpout: 115473 242288 73874
> > 16k-swpout: 158203 294672 109142
> > 64k-swpin: 0 495869 159061
> > 32k-swpin: 0 219977 56158
> > 16k-swpin: 0 223501 81445
> >
>
Hi Usama,
> hmm, both the user and sys time are worse with the patch compared to
> disable-mTHP-swapin. I wonder if the real time is an anomaly and if you
> repeat the experiment the real time might be worse as well?
Well, I've improved my script to include a loop:
echo always > /sys/kernel/mm/transparent_hugepage/hugepages-64kB/enabled
echo always > /sys/kernel/mm/transparent_hugepage/hugepages-32kB/enabled
echo always > /sys/kernel/mm/transparent_hugepage/hugepages-16kB/enabled
echo never > /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled
for ((i=1; i<=100; i++))
do
echo "Executing round $i"
make ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- clean 1>/dev/null 2>/dev/null
echo 3 > /proc/sys/vm/drop_caches
time systemd-run --scope -p MemoryMax=1G make ARCH=arm64 \
CROSS_COMPILE=aarch64-linux-gnu- vmlinux -j15 1>/dev/null 2>/dev/null
cat /proc/vmstat | grep pswp
echo -n 64k-swpout: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-64kB/stats/swpout
echo -n 32k-swpout: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-32kB/stats/swpout
echo -n 16k-swpout: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-16kB/stats/swpout
echo -n 64k-swpin: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-64kB/stats/swpin
echo -n 32k-swpin: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-32kB/stats/swpin
echo -n 16k-swpin: ; cat
/sys/kernel/mm/transparent_hugepage/hugepages-16kB/stats/swpin
done
I've noticed that the user/sys/real time on my i9 machine fluctuates
constantly, could be things
like:
real 6m52.087s
user 67m12.463s
sys 13m8.281s
...
real 7m42.937s
user 66m55.250s
sys 12m56.330s
...
real 6m49.374s
user 66m37.040s
sys 12m44.542s
...
real 6m54.205s
user 65m49.732s
sys 11m33.078s
...
likely due to unstable temperatures and I/O latency. As a result, my
data doesn’t seem
reference-worthy.
As a phone engineer, we never use phones to run kernel builds. I'm also
quite certain that phones won't provide stable and reliable data for this
type of workload. Without access to a Linux server to conduct the test,
I really need your help.
I used to work on optimizing the ARM server scheduler and memory
management, and I really miss that machine I had until three years ago :-)
>
> > I need Usama's assistance to identify a suitable patch, as I lack
> > access to hardware such as AMD machines and ARM servers with TLB
> > optimization.
> >
> > [1] https://lore.kernel.org/all/b1c17b5e-acd9-4bef-820e-699768f1426d@xxxxxxxxx/
> > [2] https://lore.kernel.org/all/7a14c332-3001-4b9a-ada3-f4d6799be555@xxxxxxxxx/
> >
> > Cc: Kanchana P Sridhar <kanchana.p.sridhar@xxxxxxxxx>
> > Cc: Usama Arif <usamaarif642@xxxxxxxxx>
> > Cc: David Hildenbrand <david@xxxxxxxxxx>
> > Cc: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> > Cc: Chris Li <chrisl@xxxxxxxxxx>
> > Cc: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
> > Cc: "Huang, Ying" <ying.huang@xxxxxxxxx>
> > Cc: Kairui Song <kasong@xxxxxxxxxxx>
> > Cc: Ryan Roberts <ryan.roberts@xxxxxxx>
> > Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
> > Cc: Michal Hocko <mhocko@xxxxxxxxxx>
> > Cc: Roman Gushchin <roman.gushchin@xxxxxxxxx>
> > Cc: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> > Cc: Muchun Song <muchun.song@xxxxxxxxx>
> > Signed-off-by: Barry Song <v-songbaohua@xxxxxxxx>
> > ---
> > include/linux/memcontrol.h | 9 ++++++++
> > mm/memcontrol.c | 45 ++++++++++++++++++++++++++++++++++++++
> > mm/memory.c | 17 ++++++++++++++
> > 3 files changed, 71 insertions(+)
> >
> > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > index 524006313b0d..8bcc8f4af39f 100644
> > --- a/include/linux/memcontrol.h
> > +++ b/include/linux/memcontrol.h
> > @@ -697,6 +697,9 @@ static inline int mem_cgroup_charge(struct folio *folio, struct mm_struct *mm,
> > int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg, gfp_t gfp,
> > long nr_pages);
> >
> > +int mem_cgroup_precharge_large_folio(struct mm_struct *mm,
> > + swp_entry_t *entry);
> > +
> > int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
> > gfp_t gfp, swp_entry_t entry);
> >
> > @@ -1201,6 +1204,12 @@ static inline int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg,
> > return 0;
> > }
> >
> > +static inline int mem_cgroup_precharge_large_folio(struct mm_struct *mm,
> > + swp_entry_t *entry)
> > +{
> > + return 0;
> > +}
> > +
> > static inline int mem_cgroup_swapin_charge_folio(struct folio *folio,
> > struct mm_struct *mm, gfp_t gfp, swp_entry_t entry)
> > {
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 17af08367c68..f3d92b93ea6d 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -4530,6 +4530,51 @@ int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg, gfp_t gfp,
> > return 0;
> > }
> >
> > +static inline bool mem_cgroup_has_margin(struct mem_cgroup *memcg)
> > +{
> > + for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
> > + if (mem_cgroup_margin(memcg) < HPAGE_PMD_NR)
>
> There might be 3 issues with the approach:
>
> Its a very big margin, lets say you have ARM64_64K_PAGES, and you have
> 256K THP set to always. As HPAGE_PMD is 512M for 64K page, you are
> basically saying you need 512M free memory to swapin just 256K?
Right, sorry for the noisy code. I was just thinking about 4KB pages
and wondering
if we could simplify the code.
>
> Its an uneven margin for different folio sizes.
> For 16K folio swapin, you are checking if there is margin for 128 folios,
> but for 1M folio swapin, you are checking there is margin for just 2 folios.
>
> Maybe it might be better to make this dependent on some factor of folio_nr_pages?
Agreed. This is similar to what we discussed regarding your zswap mTHP
swap-in series:
int mem_cgroup_swapin_charge_folio(...)
{
...
if (folio_test_large(folio) &&
mem_cgroup_margin(memcg) < max(MEMCG_CHARGE_BATCH,
folio_nr_pages(folio)))
ret = -ENOMEM;
else
ret = charge_memcg(folio, memcg, gfp);
...
}
As someone focused on phones, my challenge is the absence of stable platforms to
benchmark this type of workload. If possible, Usama, I would greatly
appreciate it if
you could take the lead on the patch.
>
> As Johannes pointed out, the charging code already does the margin check.
> So for 4K, the check just checks if there is 4K available, but for 16K it checks
> if a lot more than 16K is available. Maybe there should be a similar policy for
> all? I guess this is similar to my 2nd point, but just considers 4K folios as
> well.
I don't think the charging code performs a margin check. It simply
tries to charge
the specified nr_pages (whether 1 or more). If nr_pages are available,
the charge
proceeds; otherwise, if GFP allows blocking, it triggers memory reclamation to
reclaim max(SWAP_CLUSTER_MAX, nr_pages) base pages.
If, after reclamation, we have exactly SWAP_CLUSTER_MAX pages available, a
large folio with nr_pages == SWAP_CLUSTER_MAX will successfully charge,
immediately filling the memcg.
Shortly after, smaller folios—typically with blockable GFP—will quickly trigger
additional reclamation. While nr_pages - 1 subpages of the large folio may not
be immediately needed, they still occupy enough space to fill the memcg to
capacity.
My second point about the mitigation is as follows: For a system (or
memcg) under severe memory pressure, especially one without hardware TLB
optimization, is enabling mTHP always the right choice? Since mTHP operates at
a larger granularity, some internal fragmentation is unavoidable, regardless
of optimization. Could the mitigation code help in automatically tuning
this fragmentation?
>
> Thanks,
> Usama
>
>
> > + return false;
> > + }
> > +
> > + return true;
> > +}
> > +
> > +/**
> > + * mem_cgroup_swapin_precharge_large_folio: Precharge large folios.
> > + *
> > + * @mm: mm context of the victim
> > + * @entry: swap entry for which the folio will be allocated
> > + *
> > + * If we are arriving the edge of an almost full memcg, return error so that
> > + * swap-in and anon faults can quickly fall back to small folios to avoid swap
> > + * thrashing.
> > + *
> > + * Returns 0 on success, an error code on failure.
> > + */
> > +int mem_cgroup_precharge_large_folio(struct mm_struct *mm, swp_entry_t *entry)
> > +{
> > + struct mem_cgroup *memcg = NULL;
> > + unsigned short id;
> > + bool has_margin;
> > +
> > + if (mem_cgroup_disabled())
> > + return 0;
> > +
> > + rcu_read_lock();
> > + if (entry) {
> > + id = lookup_swap_cgroup_id(*entry);
> > + memcg = mem_cgroup_from_id(id);
> > + }
> > + if (!memcg || !css_tryget_online(&memcg->css))
> > + memcg = get_mem_cgroup_from_mm(mm);
> > + has_margin = mem_cgroup_has_margin(memcg);
> > + rcu_read_unlock();
> > +
> > + css_put(&memcg->css);
> > + return has_margin ? 0 : -ENOMEM;
> > +}
> > +
> > /**
> > * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
> > * @folio: folio to charge.
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 0f614523b9f4..96368ba0e8a6 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -4168,6 +4168,16 @@ static struct folio *alloc_swap_folio(struct vm_fault *vmf)
> >
> > pte_unmap_unlock(pte, ptl);
> >
> > + if (!orders)
> > + goto fallback;
> > +
> > + /*
> > + * Avoid swapping in large folios when memcg is nearly full, as it
> > + * may quickly trigger additional swap-out and swap-in cycles.
> > + */
> > + if (mem_cgroup_precharge_large_folio(vma->vm_mm, &entry))
> > + goto fallback;
> > +
> > /* Try allocating the highest of the remaining orders. */
> > gfp = vma_thp_gfp_mask(vma);
> > while (orders) {
> > @@ -4707,6 +4717,13 @@ static struct folio *alloc_anon_folio(struct vm_fault *vmf)
> > if (!orders)
> > goto fallback;
> >
> > + /*
> > + * When memcg is nearly full, large folios can rapidly fill
> > + * the margin and trigger new reclamation
> > + */
> > + if (mem_cgroup_precharge_large_folio(vma->vm_mm, NULL))
> > + goto fallback;
> > +
> > /* Try allocating the highest of the remaining orders. */
> > gfp = vma_thp_gfp_mask(vma);
> > while (orders) {
>
Thanks
Barry
