Re: [PATCH v7 2/2] mm: support large folios swap-in for sync io devices

[Barry Song]

On Thu, Sep 5, 2024 at 11:23 AM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
>
>
>
> On 05/09/2024 00:10, Barry Song wrote:
> > On Thu, Sep 5, 2024 at 9:30 AM Usama Arif <usamaarif642@xxxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 03/09/2024 23:05, Yosry Ahmed wrote:
> >>> On Tue, Sep 3, 2024 at 2:36 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
> >>>>
> >>>> On Wed, Sep 4, 2024 at 8:08 AM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> >>>>>
> >>>>> On Tue, 3 Sep 2024 11:38:37 -0700 Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >>>>>
> >>>>>>> [   39.157954] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000007
> >>>>>>> [   39.158288] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000001
> >>>>>>> [   39.158634] R13: 0000000000002b9a R14: 0000000000000000 R15: 00007ffd619d5518
> >>>>>>> [   39.158998]  </TASK>
> >>>>>>> [   39.159226] ---[ end trace 0000000000000000 ]---
> >>>>>>>
> >>>>>>> After reverting this or Usama's "mm: store zero pages to be swapped
> >>>>>>> out in a bitmap", the problem is gone. I think these two patches may
> >>>>>>> have some conflict that needs to be resolved.
> >>>>>>
> >>>>>> Yup. I saw this conflict coming and specifically asked for this
> >>>>>> warning to be added in Usama's patch to catch it [1]. It served its
> >>>>>> purpose.
> >>>>>>
> >>>>>> Usama's patch does not handle large folio swapin, because at the time
> >>>>>> it was written we didn't have it. We expected Usama's series to land
> >>>>>> sooner than this one, so the warning was to make sure that this series
> >>>>>> handles large folio swapin in the zeromap code. Now that they are both
> >>>>>> in mm-unstable, we are gonna have to figure this out.
> >>>>>>
> >>>>>> I suspect Usama's patches are closer to land so it's better to handle
> >>>>>> this in this series, but I will leave it up to Usama and
> >>>>>> Chuanhua/Barry to figure this out :)
> >>>>
> >>>> I believe handling this in swap-in might violate layer separation.
> >>>> `swap_read_folio()` should be a reliable API to call, regardless of
> >>>> whether `zeromap` is present. Therefore, the fix should likely be
> >>>> within `zeromap` but not this `swap-in`. I’ll take a look at this with
> >>>> Usama :-)
> >>>
> >>> I meant handling it within this series to avoid blocking Usama
> >>> patches, not within this code. Thanks for taking a look, I am sure you
> >>> and Usama will figure out the best way forward :)
> >>
> >> Hi Barry and Yosry,
> >>
> >> Is the best (and quickest) way forward to have a v8 of this with
> >> https://lore.kernel.org/all/20240904055522.2376-1-21cnbao@xxxxxxxxx/
> >> as the first patch, and using swap_zeromap_entries_count in alloc_swap_folio
> >> in this support large folios swap-in patch?
> >
> > Yes, Usama. i can actually do a check:
> >
> > zeromap_cnt = swap_zeromap_entries_count(entry, nr);
> >
> > /* swap_read_folio() can handle inconsistent zeromap in multiple entries */
> > if (zeromap_cnt > 0 && zeromap_cnt < nr)
> >        try next order;
> >
> > On the other hand, if you read the code of zRAM, you will find zRAM has
> > exactly the same mechanism as zeromap but zRAM can even do more
> > by same_pages filled. since zRAM does the job in swapfile layer, there
> > is no this kind of consistency issue like zeromap.
> >
> > So I feel for zRAM case, we don't need zeromap at all as there are duplicated
> > efforts while I really appreciate your job which can benefit all swapfiles.
> > i mean, zRAM has the ability to check "zero"(and also non-zero but same
> > content). after zeromap checks zeromap, zRAM will check again:
> >
>
> Yes, so there is a reason for having the zeromap patches, which I have outlined
> in the coverletter.
>
> https://lore.kernel.org/all/20240627105730.3110705-1-usamaarif642@xxxxxxxxx/
>
> There are usecases where zswap/zram might not be used in production.
> We can reduce I/O and flash wear in those cases by a large amount.
>
> Also running in Meta production, we found that the number of non-zero filled
> complete pages were less than 1%, so essentially its only the zero-filled pages
> that matter.

I don't have data on Android phones, i'd like to see if phones have exactly
the same ratio that non-zero same page is rare.

>
> I believe after zeromap, it might be a good idea to remove the page_same_filled
> check from zram code? Its not really a problem if its kept as well as I dont
> believe any zero-filled pages should reach zram_write_page?
>
> > static int zram_write_page(struct zram *zram, struct page *page, u32 index)
> > {
> >        ...
> >
> >         if (page_same_filled(mem, &element)) {
> >                 kunmap_local(mem);
> >                 /* Free memory associated with this sector now. */
> >                 flags = ZRAM_SAME;
> >                 atomic64_inc(&zram->stats.same_pages);
> >                 goto out;
> >         }
> >         ...
> > }
> >
> > So it seems that zeromap might slightly impact my zRAM use case. I'm not
> > blaming you, just pointing out that there might be some overlap in effort
> > here :-)
> >
> >>
> >> Thanks,
> >> Usama
> >

Thanks
Barry