On Fri, Jun 14, 2024 at 9:04 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 14.06.24 10:58, Barry Song wrote:
> > On Fri, Jun 14, 2024 at 8:56 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
> >>
> >> On Fri, Jun 14, 2024 at 8:51 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>
> >>> On 14.06.24 09:56, Barry Song wrote:
> >>>> On Thu, Jun 13, 2024 at 9:12 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>>>
> >>>>> On 13.06.24 11:06, Barry Song wrote:
> >>>>>> On Thu, Jun 13, 2024 at 8:49 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> On 13.06.24 10:46, Barry Song wrote:
> >>>>>>>> On Thu, Jun 13, 2024 at 12:08 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
> >>>>>>>>>
> >>>>>>>>> From: Barry Song <v-songbaohua@xxxxxxxx>
> >>>>>>>>>
> >>>>>>>>> The folio_test_anon(folio)==false case within do_swap_page() has been
> >>>>>>>>> relocated to folio_add_new_anon_rmap(). Additionally, two other callers
> >>>>>>>>> consistently pass anonymous folios.
> >>>>>>>>>
> >>>>>>>>> stack 1:
> >>>>>>>>> remove_migration_pmd
> >>>>>>>>> -> folio_add_anon_rmap_pmd
> >>>>>>>>> -> __folio_add_anon_rmap
> >>>>>>>>>
> >>>>>>>>> stack 2:
> >>>>>>>>> __split_huge_pmd_locked
> >>>>>>>>> -> folio_add_anon_rmap_ptes
> >>>>>>>>> -> __folio_add_anon_rmap
> >>>>>>>>>
> >>>>>>>>> __folio_add_anon_rmap() only needs to handle the cases
> >>>>>>>>> folio_test_anon(folio)==true now.
> >>>>>>>>
> >>>>>>>> My team reported a case where swapoff() is calling
> >>>>>>>> folio_add_anon_rmap_pte *not* folio_add_anon_rmap_ptes
> >>>>>>>> with one new anon (!folio_test_anon(folio)).
> >>>>>>>>
> >>>>>>>> I will double check all folio_add_anon_rmap_pte() cases.
> >>>>>>>
> >>>>>>> Right, swapoff() path is a bit special (e.g., we don't do any kind of
> >>>>>>> batching on the swapoff() path).
> >>>>>>>
> >>>>>>> But we should never get a new large anon folio there, or could that now
> >>>>>>> happen?
> >>>>>>
> >>>>>> My team encountered the following issue while testing this RFC
> >>>>>> series on real hardware. Let me take a look to identify the
> >>>>>> problem.
> >>>>>>
> >>>>>> [ 261.214195][T11285] page:000000004cdd779e refcount:4 mapcount:1
> >>>>>> mapping:00000000ba142c22 index:0x1 pfn:0x1b30a6
> >>>>>> [ 261.214213][T11285] memcg:ffffff8003678000
> >>>>>> [ 261.214217][T11285] aops:swap_aops
> >>>>>> [ 261.214233][T11285] flags:
> >>>>>> 0x2000000000081009(locked|uptodate|owner_priv_1|swapbacked|zone=1|kasantag=0x0)
> >>>>>> [ 261.214241][T11285] page_type: 0x0()
> >>>>>> [ 261.214246][T11285] raw: 2000000000081009 0000000000000000
> >>>>>> dead000000000122 0000000000000000
> >>>>>> [ 261.214251][T11285] raw: 0000000000000001 00000000000d84b3
> >>>>>> 0000000400000000 ffffff8003678000
> >>>>>> [ 261.214254][T11285] page dumped because:
> >>>>>> VM_WARN_ON_FOLIO(!folio_test_anon(folio))
> >>>>>> [ 261.214257][T11285] page_owner tracks the page as allocated
> >>>>>> [ 261.214260][T11285] page last allocated via order 0, migratetype
> >>>>>> Movable, gfp_mask 0x2140cca(GFP_HIGHUSER_MOVABLE|__GFP_COMP), pid
> >>>>>> 11285, tgid 11285 (swapoff), ts 261214177545, free_ts 261151875699
> >>>>>> [ 261.214268][T11285] post_alloc_hook+0x1b8/0x1c0
> >>>>>> [ 261.214284][T11285] prep_new_page+0x28/0x13c
> >>>>>> [ 261.214291][T11285] get_page_from_freelist+0x198c/0x1aa4
> >>>>>> [ 261.214298][T11285] __alloc_pages+0x15c/0x330
> >>>>>> [ 261.214304][T11285] __folio_alloc+0x1c/0x4c
> >>>>>> [ 261.214310][T11285] __read_swap_cache_async+0xd8/0x48c
> >>>>>> [ 261.214320][T11285] swap_cluster_readahead+0x158/0x324
> >>>>>> [ 261.214326][T11285] swapin_readahead+0x64/0x448
> >>>>>> [ 261.214331][T11285] __arm64_sys_swapoff+0x6ec/0x14b0
> >>>>>> [ 261.214337][T11285] invoke_syscall+0x58/0x114
> >>>>>> [ 261.214353][T11285] el0_svc_common+0xac/0xe0
> >>>>>> [ 261.214360][T11285] do_el0_svc+0x1c/0x28
> >>>>>> [ 261.214366][T11285] el0_svc+0x38/0x68
> >>>>>> [ 261.214372][T11285] el0t_64_sync_handler+0x68/0xbc
> >>>>>> [ 261.214376][T11285] el0t_64_sync+0x1a8/0x1ac
> >>>>>> [ 261.214381][T11285] page last free pid 90 tgid 90 stack trace:
> >>>>>> [ 261.214386][T11285] free_unref_page_prepare+0x338/0x374
> >>>>>> [ 261.214395][T11285] free_unref_page_list+0x84/0x378
> >>>>>> [ 261.214400][T11285] shrink_folio_list+0x1234/0x13e4
> >>>>>> [ 261.214409][T11285] evict_folios+0x1458/0x19b4
> >>>>>> [ 261.214417][T11285] try_to_shrink_lruvec+0x1c8/0x264
> >>>>>> [ 261.214422][T11285] shrink_one+0xa8/0x234
> >>>>>> [ 261.214427][T11285] shrink_node+0xb38/0xde0
> >>>>>> [ 261.214432][T11285] balance_pgdat+0x7a4/0xdb4
> >>>>>> [ 261.214437][T11285] kswapd+0x290/0x4e4
> >>>>>> [ 261.214442][T11285] kthread+0x114/0x1bc
> >>>>>> [ 261.214459][T11285] ret_from_fork+0x10/0x20
> >>>>>> [ 261.214477][T11285] ------------[ cut here ]------------
> >>>>>> [ 261.214480][T11285] WARNING: CPU: 3 PID: 11285 at mm/rmap.c:1305
> >>>>>> folio_add_anon_rmap_ptes+0x2b4/0x330
> >>>>>>
> >>>>>> [ 261.215403][T11285] pstate: 63400005 (nZCv daif +PAN -UAO +TCO +DIT
> >>>>>> -SSBS BTYPE=--)
> >>>>>> [ 261.215423][T11285] pc : folio_add_anon_rmap_ptes+0x2b4/0x330
> >>>>>> [ 261.215428][T11285] lr : folio_add_anon_rmap_ptes+0x2b4/0x330
> >>>>>> [ 261.215433][T11285] sp : ffffffc0a7dbbbf0
> >>>>>> [ 261.215437][T11285] x29: ffffffc0a7dbbbf0 x28: ffffff8040860a08
> >>>>>> x27: ffffff80db480000
> >>>>>> [ 261.215445][T11285] x26: fffffffe04cc2980 x25: ffffff808f77f120
> >>>>>> x24: 0000007b44941000
> >>>>>> [ 261.215452][T11285] x23: 0000000000000001 x22: 0000000000000001
> >>>>>> x21: fffffffe04cc2980
> >>>>>> [ 261.215459][T11285] x20: ffffff80db480000 x19: fffffffe04cc2980
> >>>>>> x18: ffffffed011dae80
> >>>>>> [ 261.215465][T11285] x17: 0000000000000001 x16: ffffffffffffffff
> >>>>>> x15: 0000000000000004
> >>>>>> [ 261.215471][T11285] x14: ffffff82fb73fac0 x13: 0000000000000003
> >>>>>> x12: 0000000000000003
> >>>>>> [ 261.215476][T11285] x11: 00000000fffeffff x10: c0000000fffeffff x9
> >>>>>> : 0d46c0889b468e00
> >>>>>> [ 261.215483][T11285] x8 : 0d46c0889b468e00 x7 : 545b5d3935343431 x6
> >>>>>> : 322e31363220205b
> >>>>>> [ 261.215489][T11285] x5 : ffffffed013de407 x4 : ffffffed00698967 x3
> >>>>>> : 0000000000000000
> >>>>>> [ 261.215495][T11285] x2 : 0000000000000000 x1 : ffffffc0a7dbb8c0 x0
> >>>>>> : ffffff8068c15c80
> >>>>>> [ 261.215501][T11285] Call trace:
> >>>>>> [ 261.215504][T11285] folio_add_anon_rmap_ptes+0x2b4/0x330
> >>>>>> [ 261.215509][T11285] __arm64_sys_swapoff+0x8cc/0x14b0
> >>>>>> [ 261.215516][T11285] invoke_syscall+0x58/0x114
> >>>>>> [ 261.215526][T11285] el0_svc_common+0xac/0xe0
> >>>>>> [ 261.215532][T11285] do_el0_svc+0x1c/0x28
> >>>>>> [ 261.215539][T11285] el0_svc+0x38/0x68
> >>>>>> [ 261.215544][T11285] el0t_64_sync_handler+0x68/0xbc
> >>>>>> [ 261.215548][T11285] el0t_64_sync+0x1a8/0x1ac
> >>>>>> [ 261.215552][T11285] ---[ end trace 0000000000000000 ]---
> >>>>>
> >>>>> Ah, yes. in unuse_pte(), you'll have to do the right thing if
> >>>>> !folio_test(anon) before doing the folio_add_anon_rmap_pte().
> >>>>>
> >>>>> You might have a fresh anon folio in the swapcache that was never mapped
> >>>>> (hopefully order-0, otherwise we'd likely be in trouble).
> >>>>
> >>>> Yes. It is order-0
> >>>>
> >>>> [ 261.214260][T11285] page last allocated via order 0, migratetype
> >>>>
> >>>> Otherwise, we would have encountered this VM_WARN_ON_FOLIO?
> >>>>
> >>>> __folio_add_anon_rmap()
> >>>> {
> >>>> ...
> >>>> VM_WARN_ON_FOLIO(folio_test_large(folio) &&
> >>>> level != RMAP_LEVEL_PMD, folio);
> >>>> ...
> >>>> }
> >>>>
> >>>> Given that nobody has ever reported this warning, I assume all callers
> >>>> using folio_add_anon_rmap_pte(s) are right now safe to move to ?
> >>>
> >>> Yes, and we should likely add a VM_WARN_ON_ONCE() here that we have a
> >>> small folio if !anon. If that ever changes, we can assess the situation.
> >>>
> >>
> >> this patch actually has a WARN_ON for all !anon, so it extends the WARN
> >> to small.
>
> Not what I mean, see below:
>
> >>
> >> - if (unlikely(!folio_test_anon(folio))) {
> >> - VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
> >> - /*
> >> - * For a PTE-mapped large folio, we only know that the single
> >> - * PTE is exclusive. Further, __folio_set_anon() might not get
> >> - * folio->index right when not given the address of the head
> >> - * page.
> >> - */
> >> - VM_WARN_ON_FOLIO(folio_test_large(folio) &&
> >> - level != RMAP_LEVEL_PMD, folio);
> >> - __folio_set_anon(folio, vma, address,
> >> - !!(flags & RMAP_EXCLUSIVE));
> >> - } else if (likely(!folio_test_ksm(folio))) {
> >> + VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
> >> +
> >>
> >>> Only swap created "new" anon folios without properly calling the right
> >>> function so far, all other code handles that correctly.
> >
> > as I assume patch2/3 should have moved all !anon to
> > folio_add_new_anon_rmap()
> >
> > diff --git a/mm/ksm.c b/mm/ksm.c
> > index d2641bc2efc9..c913ba8c37eb 100644
> > --- a/mm/ksm.c
> > +++ b/mm/ksm.c
> > @@ -1420,7 +1420,14 @@ static int replace_page(struct vm_area_struct
> > *vma, struct page *page,
> > */
> > if (!is_zero_pfn(page_to_pfn(kpage))) {
> > folio_get(kfolio);
> > - folio_add_anon_rmap_pte(kfolio, kpage, vma, addr, RMAP_NONE);
> > + /*
> > + * We currently ensure that new folios cannot be partially
> > + * exclusive: they are either fully exclusive or fully shared.
> > + */
> > + if (!folio_test_anon(kfolio))
> > + folio_add_new_anon_rmap(kfolio, vma, addr, RMAP_NONE);
> > + else
> > + folio_add_anon_rmap_pte(kfolio, kpage, vma, > addr, RMAP_NONE);
>
> I don't think that is required? We are only working with anon folios. Or
> were you able to trigger this? (which would be weird)
I didn't trigger this. but I am not sure if kfifo is always anon based on
the code context.
for page, it is 100% anon(otherwise "goto out"), but I am not quite
sure about kpage
by the code context.
static int try_to_merge_one_page(struct vm_area_struct *vma,
struct page *page, struct page *kpage)
{
pte_t orig_pte = __pte(0);
int err = -EFAULT;
if (page == kpage) /* ksm page forked */
return 0;
if (!PageAnon(page))
goto out;
....
}
Then I saw this
static int replace_page(struct vm_area_struct *vma, struct page *page,
struct page *kpage, pte_t orig_pte)
{
...
VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
VM_BUG_ON_FOLIO(folio_test_anon(kfolio) && PageAnonExclusive(kpage),
kfolio);
}
If kfolio is always anon, we should have used
VM_BUG_ON_FOLIO(PageAnonExclusive(kpage), folio)
just like
VM_BUG_ON_PAGE(PageAnonExclusive(page), page);
without "folio_test_anon(kfolio)".
So I lost my way.
>
> [...]
>
> > - folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
> > + /*
> > + * We currently ensure that new folios cannot be partially
> > + * exclusive: they are either fully exclusive or fully shared.
> > + */
> > + if (!folio_test_anon(folio))
>
> Here I suggest changing the comment to (and adding the VM_WARN_ON_ONCE):
>
> /*
> * We currently only expect small !anon folios, for which we now that
> * they are either fully exclusive or fully shared. If we ever get large
> * folios here, we have to be careful.
> */
> if (!folio_test_anon(folio) {
> VM_WARN_ON_ONCE(folio_test_large(folio));
> folio_add_new_anon_rmap(folio, vma, addr, rmap_flags);
> } else {
> ...
looks good to me.
>
>
> --
> Cheers,
>
> David / dhildenb
>
Thanks
Barry
