On Thu, Jun 29, 2023 at 9:32 PM John Hubbard <[email protected]> wrote:
>
> The following crash happens for me when running the -mm selftests
> (below). Specifically, it happens while running the uffd-stress
> subtests:
>
> kernel BUG at mm/hugetlb.c:7249!
> invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
> CPU: 0 PID: 3238 Comm: uffd-stress Not tainted 6.4.0-hubbard-github+ #109
> Hardware name: ASUS X299-A/PRIME X299-A, BIOS 1503 08/03/2018
> RIP: 0010:huge_pte_alloc+0x12c/0x1a0
> ...
> Call Trace:
> <TASK>
> ? __die_body+0x63/0xb0
> ? die+0x9f/0xc0
> ? do_trap+0xab/0x180
> ? huge_pte_alloc+0x12c/0x1a0
> ? do_error_trap+0xc6/0x110
> ? huge_pte_alloc+0x12c/0x1a0
> ? handle_invalid_op+0x2c/0x40
> ? huge_pte_alloc+0x12c/0x1a0
> ? exc_invalid_op+0x33/0x50
> ? asm_exc_invalid_op+0x16/0x20
> ? __pfx_put_prev_task_idle+0x10/0x10
> ? huge_pte_alloc+0x12c/0x1a0
> hugetlb_fault+0x1a3/0x1120
> ? finish_task_switch+0xb3/0x2a0
> ? lock_is_held_type+0xdb/0x150
> handle_mm_fault+0xb8a/0xd40
> ? find_vma+0x5d/0xa0
> do_user_addr_fault+0x257/0x5d0
> exc_page_fault+0x7b/0x1f0
> asm_exc_page_fault+0x22/0x30
>
> That happens because a BUG() statement in huge_pte_alloc() attempts to
> check that a pte, if present, is a hugetlb pte, but it does so in a
> non-lockless-safe manner that leads to a false BUG() report.
>
> We got here due to a couple of bugs, each of which by itself was not
> quite enough to cause a problem:
>
> First of all, before commit c33c794828f2("mm: ptep_get() conversion"),
> the BUG() statement in huge_pte_alloc() was itself fragile: it relied
> upon compiler behavior to only read the pte once, despite using it twice
> in the same conditional.
>
> Next, commit c33c794828f2 ("mm: ptep_get() conversion") broke that
> delicate situation, by causing all direct pte reads to be done via
> READ_ONCE(). And so READ_ONCE() got called twice within the same BUG()
> conditional, leading to comparing (potentially, occasionally) different
> versions of the pte, and thus to false BUG() reports.
>
> Fix this by taking a single snapshot of the pte before using it in the
> BUG conditional.
>
> Now, that commit is only partially to blame here but, people doing
> bisections will invariably land there, so this will help them find a fix
> for a real crash. And also, the previous behavior was unlikely to ever
> expose this bug--it was fragile, yet not actually broken.
>
> So that's why I chose this commit for the Fixes tag, rather than the
> commit that created the original BUG() statement.
>
> Fixes: c33c794828f2 ("mm: ptep_get() conversion")
Hi John,
Good catch, and thanks for the detailed explanation. It looks like
riscv and powerpc have equivalent problems in their huge_pte_alloc
implementations, perhaps it's worth taking a look at those. (riscv
looks like it has precisely the same problem except it's a WARN, but
powerpc looks more interesting.)
Either way,
Acked-by: James Houghton <[email protected]>
On 6/29/23 18:50, James Houghton wrote:
> Hi John,
>
> Good catch, and thanks for the detailed explanation. It looks like
> riscv and powerpc have equivalent problems in their huge_pte_alloc
> implementations, perhaps it's worth taking a look at those. (riscv
> looks like it has precisely the same problem except it's a WARN, but
> powerpc looks more interesting.)
>
> Either way,
>
> Acked-by: James Houghton <[email protected]>
Thanks for the ack, and yes, riscv and powerpc do look like the
same problem to me, too. I'm glad you thought to look there, as
I have fallen into tunnel vision while investigating this, heh.
OK, as long as I'm here, let me fix those up, too. I'll make those
a separate patch, because they are still in the "fragile" state,
rather than actually at the full crash state. :)
thanks,
--
John Hubbard
NVIDIA