[PATCH v4 13/13] mm/gup: Handle hugetlb in the generic follow_page_mask code
David Hildenbrand
david at redhat.com
Tue Apr 2 09:39:31 PDT 2024
On 02.04.24 18:20, Peter Xu wrote:
> On Tue, Apr 02, 2024 at 05:26:28PM +0200, David Hildenbrand wrote:
>> On 02.04.24 16:48, Ryan Roberts wrote:
>>> Hi Peter,
>
> Hey, Ryan,
>
> Thanks for the report!
>
>>>
>>> On 27/03/2024 15:23, peterx at redhat.com wrote:
>>>> From: Peter Xu <peterx at redhat.com>
>>>>
>>>> Now follow_page() is ready to handle hugetlb pages in whatever form, and
>>>> over all architectures. Switch to the generic code path.
>>>>
>>>> Time to retire hugetlb_follow_page_mask(), following the previous
>>>> retirement of follow_hugetlb_page() in 4849807114b8.
>>>>
>>>> There may be a slight difference of how the loops run when processing slow
>>>> GUP over a large hugetlb range on cont_pte/cont_pmd supported archs: each
>>>> loop of __get_user_pages() will resolve one pgtable entry with the patch
>>>> applied, rather than relying on the size of hugetlb hstate, the latter may
>>>> cover multiple entries in one loop.
>>>>
>>>> A quick performance test on an aarch64 VM on M1 chip shows 15% degrade over
>>>> a tight loop of slow gup after the path switched. That shouldn't be a
>>>> problem because slow-gup should not be a hot path for GUP in general: when
>>>> page is commonly present, fast-gup will already succeed, while when the
>>>> page is indeed missing and require a follow up page fault, the slow gup
>>>> degrade will probably buried in the fault paths anyway. It also explains
>>>> why slow gup for THP used to be very slow before 57edfcfd3419 ("mm/gup:
>>>> accelerate thp gup even for "pages != NULL"") lands, the latter not part of
>>>> a performance analysis but a side benefit. If the performance will be a
>>>> concern, we can consider handle CONT_PTE in follow_page().
>>>>
>>>> Before that is justified to be necessary, keep everything clean and simple.
>>>>
>>>> Reviewed-by: Jason Gunthorpe <jgg at nvidia.com>
>>>> Signed-off-by: Peter Xu <peterx at redhat.com>
>>>
>>> Afraid I'm seeing an oops when running gup_longterm test on arm64 with current mm-unstable. Git bisect blames this patch. The oops reproduces for me every time on 2 different machines:
>>>
>>>
>>> [ 9.340416] kernel BUG at mm/gup.c:778!
>>> [ 9.340746] Internal error: Oops - BUG: 00000000f2000800 [#1] PREEMPT SMP
>>> [ 9.341199] Modules linked in:
>>> [ 9.341481] CPU: 1 PID: 1159 Comm: gup_longterm Not tainted 6.9.0-rc2-00210-g910ff1a347e4 #11
>>> [ 9.342232] Hardware name: linux,dummy-virt (DT)
>>> [ 9.342647] pstate: 60400005 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>> [ 9.343195] pc : follow_page_mask+0x4d4/0x880
>>> [ 9.343580] lr : follow_page_mask+0x4d4/0x880
>>> [ 9.344018] sp : ffff8000898b3aa0
>>> [ 9.344345] x29: ffff8000898b3aa0 x28: fffffdffc53973e8 x27: 00003c0005d08000
>>> [ 9.345028] x26: ffff00014e5cfd08 x25: ffffd3513a40c000 x24: fffffdffc5d08000
>>> [ 9.345682] x23: ffffc1ffc0000000 x22: 0000000000080101 x21: ffff8000898b3ba8
>>> [ 9.346337] x20: 0000fffff4200000 x19: ffff00014e52d508 x18: 0000000000000010
>>> [ 9.347005] x17: 5f656e6f7a5f7369 x16: 2120262620296567 x15: 6170286461654865
>>> [ 9.347713] x14: 6761502128454741 x13: 2929656761702865 x12: 6761705f65636976
>>> [ 9.348371] x11: 65645f656e6f7a5f x10: ffffd3513b31d6e0 x9 : ffffd3513852f090
>>> [ 9.349062] x8 : 00000000ffffefff x7 : ffffd3513b31d6e0 x6 : 0000000000000000
>>> [ 9.349753] x5 : ffff00017ff98cc8 x4 : 0000000000000fff x3 : 0000000000000000
>>> [ 9.350397] x2 : 0000000000000000 x1 : ffff000190e8b480 x0 : 0000000000000052
>>> [ 9.351097] Call trace:
>>> [ 9.351312] follow_page_mask+0x4d4/0x880
>>> [ 9.351700] __get_user_pages+0xf4/0x3e8
>>> [ 9.352089] __gup_longterm_locked+0x204/0xa70
>>> [ 9.352516] pin_user_pages+0x88/0xc0
>>> [ 9.352873] gup_test_ioctl+0x860/0xc40
>>> [ 9.353249] __arm64_sys_ioctl+0xb0/0x100
>>> [ 9.353648] invoke_syscall+0x50/0x128
>>> [ 9.354022] el0_svc_common.constprop.0+0x48/0xf8
>>> [ 9.354488] do_el0_svc+0x28/0x40
>>> [ 9.354822] el0_svc+0x34/0xe0
>>> [ 9.355128] el0t_64_sync_handler+0x13c/0x158
>>> [ 9.355489] el0t_64_sync+0x190/0x198
>>> [ 9.355793] Code: aa1803e0 d000d8e1 91220021 97fff560 (d4210000)
>>> [ 9.356280] ---[ end trace 0000000000000000 ]---
>>> [ 9.356651] note: gup_longterm[1159] exited with irqs disabled
>>> [ 9.357141] note: gup_longterm[1159] exited with preempt_count 2
>>> [ 9.358033] ------------[ cut here ]------------
>>> [ 9.358800] WARNING: CPU: 1 PID: 0 at kernel/context_tracking.c:128 ct_kernel_exit.constprop.0+0x108/0x120
>>> [ 9.360157] Modules linked in:
>>> [ 9.360541] CPU: 1 PID: 0 Comm: swapper/1 Tainted: G D 6.9.0-rc2-00210-g910ff1a347e4 #11
>>> [ 9.361626] Hardware name: linux,dummy-virt (DT)
>>> [ 9.362087] pstate: 204003c5 (nzCv DAIF +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>> [ 9.362758] pc : ct_kernel_exit.constprop.0+0x108/0x120
>>> [ 9.363306] lr : ct_idle_enter+0x10/0x20
>>> [ 9.363845] sp : ffff8000801abdc0
>>> [ 9.364222] x29: ffff8000801abdc0 x28: 0000000000000000 x27: 0000000000000000
>>> [ 9.364961] x26: 0000000000000000 x25: ffff00014149d780 x24: 0000000000000000
>>> [ 9.365557] x23: 0000000000000000 x22: ffffd3513b299d48 x21: ffffd3513a785730
>>> [ 9.366239] x20: ffffd3513b299c28 x19: ffff00017ffa7da0 x18: 0000fffff5ffffff
>>> [ 9.366869] x17: 0000000000000000 x16: 1fffe0002a21a8c1 x15: 0000000000000000
>>> [ 9.367524] x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000002
>>> [ 9.368207] x11: 0000000000000001 x10: 0000000000000ad0 x9 : ffffd35138589230
>>> [ 9.369123] x8 : ffff00014149e2b0 x7 : 0000000000000000 x6 : 000000000f8c0fb2
>>> [ 9.370403] x5 : 4000000000000002 x4 : ffff2cb045825000 x3 : ffff8000801abdc0
>>> [ 9.371170] x2 : ffffd3513a782da0 x1 : 4000000000000000 x0 : ffffd3513a782da0
>>> [ 9.372279] Call trace:
>>> [ 9.372519] ct_kernel_exit.constprop.0+0x108/0x120
>>> [ 9.373216] ct_idle_enter+0x10/0x20
>>> [ 9.373562] default_idle_call+0x3c/0x160
>>> [ 9.374055] do_idle+0x21c/0x280
>>> [ 9.374394] cpu_startup_entry+0x3c/0x50
>>> [ 9.374797] secondary_start_kernel+0x140/0x168
>>> [ 9.375220] __secondary_switched+0xb8/0xc0
>>> [ 9.375875] ---[ end trace 0000000000000000 ]---
>>>
>>>
>>> The oops trigger is at mm/gup.c:778:
>>> VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
>>>
>>>
>>> This is the output of gup_longterm (last output is just before oops):
>>>
>>> # [INFO] detected hugetlb page size: 2048 KiB
>>> # [INFO] detected hugetlb page size: 32768 KiB
>>> # [INFO] detected hugetlb page size: 64 KiB
>>> # [INFO] detected hugetlb page size: 1048576 KiB
>>> TAP version 13
>>> 1..70
>>> # [RUN] R/W longterm GUP pin in MAP_SHARED file mapping ... with memfd
>>> ok 1 Should have worked
>>> # [RUN] R/W longterm GUP pin in MAP_SHARED file mapping ... with tmpfile
>>> ok 2 Should have failed
>>> # [RUN] R/W longterm GUP pin in MAP_SHARED file mapping ... with local tmpfile
>>> ok 3 Should have failed
>>> # [RUN] R/W longterm GUP pin in MAP_SHARED file mapping ... with memfd hugetlb (2048 kB)
>>> ok 4 Should have worked
>>> # [RUN] R/W longterm GUP pin in MAP_SHARED file mapping ... with memfd hugetlb (32768 kB)
>>>
>>>
>>> So 2M passed ok, and its failing for 32M, which is cont-pmd. I'm guessing you're trying to iterate 2M into a cont-pmd folio and ending up with an unexpected tail page?
>>
>> I assume we find the expected tail page, it's just that the check
>>
>> VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
>>
>> Doesn't make sense with hugetlb folios. We might have a tail page mapped in
>> a cont-pmd entry. As soon as we call follow_huge_pmd() on "not the first
>> cont-pmd entry", we trigger this check.
>>
>> Likely this sanity check must also allow for hugetlb folios. Or we should
>> just remove it completely.
>
> Right, IMHO it'll be easier we remove it, actually I see there's one more
> at the end, so I think we need to remove both.
>
>>
>> In the past, we wanted to make sure that we never get tail pages of THP from
>> PMD entries, because something would currently be broken (we don't support
>> THP > PMD).
>
> There's probably one more thing we need to do, on allowing
> PageAnonExclusive() to work with hugetlb tails. Even if we remove the
> warnings and if I read the code right, we can BUG_ON again on checking tail
> pages over anon-exclusive for PageHuge.
>
> So I assume to fix it completely, we may need two changes: Patch 1 to
> prepare PageAnonExclusive() to work on hugetlb tails, then patch 2 to be
> squashed into the patch "mm/gup: handle huge pmd for follow_pmd_mask()".
> Note: not this patch to fixup, as this patch only does the "switchover" to
> the new path, the culprit should be the other patch..
>
> I have them attached below first, before I'll also go and see whether I can
> run some arm tests later today or tomorrow. David, any comments from
> anon-exclusive side?
I added the PageAnonExclusive checks for hugetlb back then, because
calling it on a tail page indicated real trouble for hugetlb.
Well, and I didn't want to have runtime-hugetlb checks in
PageAnonExclusive code called on certainly-not-hugetlb code paths.
Personally, I'd fixup the problematic callsite where we know nothing
nasty is happening (like we did for gup_must_unshare(), because we don't
expect hugetlb tail pages from arbitrary other code).
But as I'm getting closer to a folio_test_anon_exclusive()
implementation as we speak (closer, but not done :) ... ), where I'd
remove any such hugetlb special handling, I don't particularly care how
we handle GUP here in the meantime.
--
Cheers,
David / dhildenb
More information about the linux-arm-kernel
mailing list