Message-ID: <CAGsJ_4zyhqwUS26bbx0a95rSO0jVANfC_RWdyTFbJQQG5KJ_cg@mail.gmail.com>
Date:   Thu, 30 Nov 2023 09:38:43 +1300
From:   Barry Song <21cnbao@...il.com>
To:     Ryan Roberts <ryan.roberts@....com>
Cc:     akpm@...ux-foundation.org, david@...hat.com,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org, mhocko@...e.com,
        shy828301@...il.com, wangkefeng.wang@...wei.com,
        willy@...radead.org, xiang@...nel.org, ying.huang@...el.com,
        yuzhao@...gle.com, hanchuanhua@...o.com
Subject: Re: [PATCH v3 0/4] Swap-out small-sized THP without splitting

On Thu, Nov 30, 2023 at 1:06 AM Ryan Roberts <ryan.roberts@....com> wrote:
>
> On 29/11/2023 07:47, Barry Song wrote:
> >> Hi All,
> >>
> >> This is v3 of a series to add support for swapping out small-sized THP without
> >> needing to first split the large folio via __split_huge_page(). It closely
> >> follows the approach already used by PMD-sized THP.
> >>
> >> "Small-sized THP" is an upcoming feature that enables performance improvements
> >> by allocating large folios for anonymous memory, where the large folio size is
> >> smaller than the traditional PMD-size. See [3].
> >>
> >> In some circumstances I've observed a performance regression (see patch 2 for
> >> details), and this series is an attempt to fix the regression in advance of
> >> merging small-sized THP support.
> >>
> >> I've done what I thought was the smallest change possible, and as a result, this
> >> approach is only employed when the swap is backed by a non-rotating block device
> >> (just as PMD-sized THP is supported today). Discussion against the RFC concluded
> >> that this is probably sufficient.
> >>
> >> The series applies against mm-unstable (1a3c85fa684a)
> >>
> >>
> >> Changes since v2 [2]
> >> ====================
> >>
> >>  - Reuse scan_swap_map_try_ssd_cluster() between order-0 and order > 0
> >>    allocation. This required some refactoring to make everything work nicely
> >>    (new patches 2 and 3).
> >>  - Fix bug where nr_swap_pages would say there are pages available but the
> >>    scanner would not be able to allocate them because they were reserved for the
> >>    per-cpu allocator. We now allow stealing of order-0 entries from the high
> >>    order per-cpu clusters (in addition to exisiting stealing from order-0
> >>    per-cpu clusters).
> >>
> >> Thanks to Huang, Ying for the review feedback and suggestions!
> >>
> >>
> >> Changes since v1 [1]
> >> ====================
> >>
> >>  - patch 1:
> >>     - Use cluster_set_count() instead of cluster_set_count_flag() in
> >>       swap_alloc_cluster() since we no longer have any flag to set. I was unable
> >>       to kill cluster_set_count_flag() as proposed against v1 as other call
> >>       sites depend explicitly setting flags to 0.
> >>  - patch 2:
> >>     - Moved large_next[] array into percpu_cluster to make it per-cpu
> >>       (recommended by Huang, Ying).
> >>     - large_next[] array is dynamically allocated because PMD_ORDER is not
> >>       compile-time constant for powerpc (fixes build error).
> >>
> >>
> >> Thanks,
> >> Ryan
> >
> >> P.S. I know we agreed this is not a prerequisite for merging small-sized THP,
> >> but given Huang Ying had provided some review feedback, I wanted to progress it.
> >> All the actual prerequisites are either complete or being worked on by others.
> >>
> >
> > Hi Ryan,
> >
> > this is quite important to a phone and a must-have component, so is large-folio
> > swapin, as i explained to you in another email.
>
> Yes understood; the "prerequisites" are just the things that must be merged
> *before* small-sized THP to ensure we don't regress existing behaviour or to
> ensure that small-size THP is correct/robust when enabled. Performance
> improvements can be merged after the initial small-sized series.

I completely agree. I didn't mean small-THP swap out as a whole  should be
a prerequisite for small-THP initial patchset, just describing how important
it is to a phone :-)

And actually we have done much further than this on phones by optimizing
zsmalloc/zram and allow a large folio compressed and decompressed
as a whole,  we have seen compressing/decompressing a whole large folio
can significantly improve compression ratio and decrease CPU consumption.

so that means large folios can not only save memory but also decrease
CPU consumption.

>
> > Luckily, we are having Chuanhua Han(Cc-ed) to prepare a patchset of largefolio
> > swapin on top of your this patchset, probably a port and cleanup of our
> > do_swap_page[1] againest yours.
>
> That's great to hear - welcome aboard, Chuanhua Han! Feel free to reach out if
> you have questions.
>
> I would guess that any large swap-in changes would be independent of this
> swap-out patch though? Wouldn't you just be looking for contiguous swap entries
> in the page table to determine a suitable folio order, then swap-in each of
> those entries into the folio? And if they happen to have contiguous swap offsets
> (enabled by this swap-out series) then you potentially get a batched disk access
> benefit.

I agree. Maybe we still need to check if the number of contiguous swap entries
is one of those supported large folio sizes?

>
> That's just a guess though, perhaps you can describe your proposed approach?

we have an ugly hack if we are swapping in from the dedicated zRAM for
large folios,
we assume we have a chance to swapin as a whole, but we do also handle corner
cases in which some entries might have been zap_pte_range()-ed.

My current proposal is as below,
A1. we get the number of contiguous swap entries with PTL and find it
is a valid large folio size
A2. we allocate large folio without PTL
A3. after getting PTL again, we re-check PTEs if the situation in A1
have been changed,
if no other threads change those PTEs, we set_ptes and finish the swap-in

but we have a chance to fail in A2, so in this case we still need to
fall back to basepage.

considering the MTE thread[1] I am handling, and MTE tag life cycle is
the same with swap
entry life cycle. it seems we will still need a page-level
arch_swap_restore even after
we support large folio swap-in for the below two reasons

1. contiguous PTEs might be partially dropped by madvise(DONTNEED) etc
2. we can still fall back to basepage for swap-in if we fail to get
large folio even PTEs are all
contiguous swap entries

Of course,  if we succeed in setting all PTEs for a large folio in A3,
we can have
a folio-level arch_swap_restore.

To me, an universal folio-level arch_swap_restore seems not sensible
to handle all kinds of
complex cases.

[1] [RFC V3 PATCH] arm64: mm: swap: save and restore mte tags for large folios
https://lore.kernel.org/linux-mm/20231114014313.67232-1-v-songbaohua@oppo.com/

>
> >
> > Another concern is that swapslots can be fragmented, if we place small/large folios
> > in a swap device, since large folios always require contiguous swapslot, we can
> > result in failure of getting slots even we still have many free slots which are not
> > contiguous.
>
> This series tries to mitigate that problem by reserving a swap cluster per
> order. That works well until we run out of swap clusters; a cluster can't be
> freed until all contained swap entries are swapped back in and deallocated.
>
> But I think we should start with the simple approach first, and only solve the
> problems as they arise through real testing.

I agree.

>
>  To avoid this, [2] dynamic hugepage solution have two swap devices,
> > one for basepage, the other one for CONTPTE. we have modified the priority-based
> > selection of swap devices to choose swap devices based on small/large folios.
> > i realize this approache is super ugly and might be very hard to find a way to
> > upstream though, it seems not universal especially if you are a linux server (-_-)
> >
> > two devices are not a nice approach though it works well for a real product,
> > we might still need some decent way to address this problem while the problem
> > is for sure not a stopper of your patchset.
>
> I guess that approach works for your case because A) you only have 2 sizes, and
> B) your swap device is zRAM, which dynamically allocate RAM as it needs it.
>
> The upstream small-sized THP solution can support multiple sizes, so you would
> need a swap device per size (I think 13 is the limit at the moment - PMD size
> for 64K base page). And if your swap device is a physical block device, you
> can't dynamically parition it the way you can with zRAM. Nether of those things
> scale particularly well IMHO.

right.

>
> >
> > [1] https://github.com/OnePlusOSS/android_kernel_oneplus_sm8550/blob/oneplus/sm8550_u_14.0.0_oneplus11/mm/memory.c#L4648
> > [2] https://github.com/OnePlusOSS/android_kernel_oneplus_sm8550/blob/oneplus/sm8550_u_14.0.0_oneplus11/mm/swapfile.c#L1129
> >
> >>
> >> [1] https://lore.kernel.org/linux-mm/20231010142111.3997780-1-ryan.roberts@arm.com/
> >> [2] https://lore.kernel.org/linux-mm/20231017161302.2518826-1-ryan.roberts@arm.com/
> >> [3] https://lore.kernel.org/linux-mm/15a52c3d-9584-449b-8228-1335e0753b04@arm.com/
> >>
> >>
> >> Ryan Roberts (4):
> >>   mm: swap: Remove CLUSTER_FLAG_HUGE from swap_cluster_info:flags
> >>   mm: swap: Remove struct percpu_cluster
> >>   mm: swap: Simplify ssd behavior when scanner steals entry
> >>   mm: swap: Swap-out small-sized THP without splitting
> >>
> >>  include/linux/swap.h |  31 +++---
> >>  mm/huge_memory.c     |   3 -
> >>  mm/swapfile.c        | 232 ++++++++++++++++++++++++-------------------
> >>  mm/vmscan.c          |  10 +-
> >>  4 files changed, 149 insertions(+), 127 deletions(-)
> >

Thanks
Barry

Powered by blists - more mailing lists