On Mon, Jul 05, 2021 at 08:56:54PM +0100, Mark Rutland wrote:
> On Mon, Jul 05, 2021 at 06:16:49PM +0300, Anatoly Pugachev wrote:
> > Hello!
>
> Hi Anatoly,
>
> > latest sparc64 git kernel produces the following OOPS on running stress-ng as :
> >
> > $ stress-ng -v --mmap 1 -t 30s
> >
> > kernel OOPS (console logs):
> >
> > [ 27.276719] Unable to handle kernel NULL pointer dereference
> > [ 27.276782] tsk->{mm,active_mm}->context = 00000000000003cb
> > [ 27.276818] tsk->{mm,active_mm}->pgd = fff800003a2a0000
> > [ 27.276853] \|/ ____ \|/
> > [ 27.276853] "@'/ .. \`@"
> > [ 27.276853] /_| \__/ |_\
> > [ 27.276853] \__U_/
> > [ 27.276927] stress-ng(928): Oops [#1]
>
> I can reproduce this under QEMU; following your bisection (and working
> around the missing ifdeferry that breaks bisection), I can confirm that
> the first broken commit is:
>
> ff5b4f1ed580 ("locking/atomic: sparc: move to ARCH_ATOMIC")
>
> Sorry about this.
>
> > Can someone please look at this commit ids?
>
> From digging into this, I can't spot an obvious bug in the commit above.
Looking again with fresh eyes, there is a trivial bug after all.
Could you give the patch below a spin? It works for me locally under
QEMU.
Sorry again about this!
Thanks,
Mark
---->8----
>From afb683b2ce749dca426d27f05af3ea08455a52d7 Mon Sep 17 00:00:00 2001
From: Mark Rutland <mark.rutland@xxxxxxx>
Date: Tue, 6 Jul 2021 09:55:56 +0100
Subject: [PATCH] locking/atomic: sparc: fix arch_cmpxchg64_local()
Anatoly reports that since commit:
ff5b4f1ed580c59d ("locking/atomic: sparc: move to ARCH_ATOMIC")
... it's possible to reliably trigger an oops by running:
stress-ng -v --mmap 1 -t 30s
... which results in a NULL pointer dereference in
__split_huge_pmd_locked().
The underlying problem is that commit ff5b4f1ed580c59d left
arch_cmpxchg64_local() defined in terms of cmpxchg_local() rather than
arch_cmpxchg_local(). In <asm-generic/atomic-instrumented.h> we wrap
these with macros which use identically-named variables. When
cmpxchg_local() nests inside cmpxchg64_local(), this casues it to use an
unitialized variable as the pointer, which can be NULL.
This can also be seen in pmdp_establish(), where the compiler can
generate the pointer with a `clr` instruction:
0000000000000360 <pmdp_establish>:
360: 9d e3 bf 50 save %sp, -176, %sp
364: fa 5e 80 00 ldx [ %i2 ], %i5
368: 82 10 00 1b mov %i3, %g1
36c: 84 10 20 00 clr %g2
370: c3 f0 90 1d casx [ %g2 ], %i5, %g1
374: 80 a7 40 01 cmp %i5, %g1
378: 32 6f ff fc bne,a %xcc, 368 <pmdp_establish+0x8>
37c: fa 5e 80 00 ldx [ %i2 ], %i5
380: d0 5e 20 40 ldx [ %i0 + 0x40 ], %o0
384: 96 10 00 1b mov %i3, %o3
388: 94 10 00 1d mov %i5, %o2
38c: 92 10 00 19 mov %i1, %o1
390: 7f ff ff 84 call 1a0 <__set_pmd_acct>
394: b0 10 00 1d mov %i5, %i0
398: 81 cf e0 08 return %i7 + 8
39c: 01 00 00 00 nop
This patch fixes the problem by defining arch_cmpxchg64_local() in terms
of arch_cmpxchg_local(), avoiding potential shadowing, and resulting in
working cmpxchg64_local() and variants, e.g.
0000000000000360 <pmdp_establish>:
360: 9d e3 bf 50 save %sp, -176, %sp
364: fa 5e 80 00 ldx [ %i2 ], %i5
368: 82 10 00 1b mov %i3, %g1
36c: c3 f6 90 1d casx [ %i2 ], %i5, %g1
370: 80 a7 40 01 cmp %i5, %g1
374: 32 6f ff fd bne,a %xcc, 368 <pmdp_establish+0x8>
378: fa 5e 80 00 ldx [ %i2 ], %i5
37c: d0 5e 20 40 ldx [ %i0 + 0x40 ], %o0
380: 96 10 00 1b mov %i3, %o3
384: 94 10 00 1d mov %i5, %o2
388: 92 10 00 19 mov %i1, %o1
38c: 7f ff ff 85 call 1a0 <__set_pmd_acct>
390: b0 10 00 1d mov %i5, %i0
394: 81 cf e0 08 return %i7 + 8
398: 01 00 00 00 nop
39c: 01 00 00 00 nop
Fixes: ff5b4f1ed580c59d ("locking/atomic: sparc: move to ARCH_ATOMIC")
Signed-off-by: Mark Rutland <mark.rutland@xxxxxxx>
Reported-by: Anatoly Pugachev <matorola@xxxxxxxxx>
Cc: "David S. Miller" <davem@xxxxxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxxxxxxxx>
---
arch/sparc/include/asm/cmpxchg_64.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/sparc/include/asm/cmpxchg_64.h b/arch/sparc/include/asm/cmpxchg_64.h
index 8c39a9981187..12d00a42c0a3 100644
--- a/arch/sparc/include/asm/cmpxchg_64.h
+++ b/arch/sparc/include/asm/cmpxchg_64.h
@@ -201,7 +201,7 @@ static inline unsigned long __cmpxchg_local(volatile void *ptr,
#define arch_cmpxchg64_local(ptr, o, n) \
({ \
BUILD_BUG_ON(sizeof(*(ptr)) != 8); \
- cmpxchg_local((ptr), (o), (n)); \
+ arch_cmpxchg_local((ptr), (o), (n)); \
})
#define arch_cmpxchg64(ptr, o, n) arch_cmpxchg64_local((ptr), (o), (n))
--
2.11.0
