Skip to contentSkip to conversation

5.9-rc7 null ptr deref in __i915_gem_userptr_get_pages_worker

PATCH In review Jason A. Donenfeld 4 participants 8 messages Sep 28, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Seeing a new crash in 5.9-rc7 I didn't have in 5.9-rc6:

[ 1311.596896] BUG: kernel NULL pointer dereference, address: 0000000000000064
[ 1311.596898] #PF: supervisor write access in kernel mode
[ 1311.596899] #PF: error_code(0x0002) - not-present page
[ 1311.596899] PGD 0 P4D 0
[ 1311.596901] Oops: 0002 [#1] SMP
[ 1311.596902] CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U
   O      5.9.0-rc7+ #140
[ 1311.596903] Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS
N2OET47W (1.34 ) 08/06/2020
[ 1311.596955] Workqueue: i915-userptr-acquire
__i915_gem_userptr_get_pages_worker [i915]
[ 1311.596959] RIP: 0010:__get_user_pages_remote+0xd7/0x310
[ 1311.596960] Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1
00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40
03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24
18 00
[ 1311.596961] RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
[ 1311.596962] RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
[ 1311.596962] RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
[ 1311.596963] RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
[ 1311.596963] R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
[ 1311.596963] R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
[ 1311.596964] FS:  0000000000000000(0000) GS:ffff888ffc480000(0000)
knlGS:0000000000000000
[ 1311.596965] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1311.596965] CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
[ 1311.596966] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 1311.596966] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 1311.596967] Call Trace:
[ 1311.596993]  __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
[ 1311.596996]  process_one_work+0x1ca/0x390
[ 1311.596997]  worker_thread+0x48/0x3c0
[ 1311.596998]  ? rescuer_thread+0x3d0/0x3d0
[ 1311.597000]  kthread+0x114/0x130
[ 1311.597001]  ? kthread_create_worker_on_cpu+0x40/0x40
[ 1311.597003]  ret_from_fork+0x1f/0x30
[ 1311.597031] CR2: 0000000000000064
[ 1311.597033] ---[ end trace e2b8ddde994a6f6d ]---
#1 lore raw

Increasing the CC list a bit, as i915 didn't really get much churn rc6->rc7, but mm/gup.c did, and mm has had a lot of recent changes.

Quoted 41 lines · Jason A. Donenfeld

Seeing a new crash in 5.9-rc7 I didn't have in 5.9-rc6:

[ 1311.596896] BUG: kernel NULL pointer dereference, address: 0000000000000064
[ 1311.596898] #PF: supervisor write access in kernel mode
[ 1311.596899] #PF: error_code(0x0002) - not-present page
[ 1311.596899] PGD 0 P4D 0
[ 1311.596901] Oops: 0002 [#1] SMP
[ 1311.596902] CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U
   O      5.9.0-rc7+ #140
[ 1311.596903] Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS
N2OET47W (1.34 ) 08/06/2020
[ 1311.596955] Workqueue: i915-userptr-acquire
__i915_gem_userptr_get_pages_worker [i915]
[ 1311.596959] RIP: 0010:__get_user_pages_remote+0xd7/0x310
[ 1311.596960] Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1
00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40
03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24
18 00
[ 1311.596961] RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
[ 1311.596962] RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
[ 1311.596962] RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
[ 1311.596963] RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
[ 1311.596963] R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
[ 1311.596963] R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
[ 1311.596964] FS:  0000000000000000(0000) GS:ffff888ffc480000(0000)
knlGS:0000000000000000
[ 1311.596965] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1311.596965] CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
[ 1311.596966] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 1311.596966] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 1311.596967] Call Trace:
[ 1311.596993]  __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
[ 1311.596996]  process_one_work+0x1ca/0x390
[ 1311.596997]  worker_thread+0x48/0x3c0
[ 1311.596998]  ? rescuer_thread+0x3d0/0x3d0
[ 1311.597000]  kthread+0x114/0x130
[ 1311.597001]  ? kthread_create_worker_on_cpu+0x40/0x40
[ 1311.597003]  ret_from_fork+0x1f/0x30
[ 1311.597031] CR2: 0000000000000064
[ 1311.597033] ---[ end trace e2b8ddde994a6f6d ]---
#2 lore raw

Alright, the failing code seems to be in mm:

if (flags & FOLL_PIN)
atomic_set(&current->mm->has_pinned, 1);

Apparently you can't rely on current->mm being valid in this context; it's null here, hence the +0x64 for has_pinned's offset.

This was added by 008cfe4418b3 ("mm: Introduce mm_struct.has_pinned"), which is new for rc7 indeed.

The crash goes away when changing that to:

        if ((flags & FOLL_PIN) && current->mm)
                atomic_set(&current->mm->has_pinned, 1);

But I haven't really evaluated whether or not that's racy or if I need to take locks to do such a thing.

#3 lore raw

Oh, this is just a copy and paste error, when the code was originally pasted from internal_get_user_pages_fast, which assumes a current. I'll fix this up and send a patch shortly.

#4 lore raw

PATCH 1/1 mm: do not rely on mm == current->mm in __get_user_pages_locked

It seems likely this block was pasted from internal_get_user_pages_fast, which is not passed an mm struct and therefore uses current's. But __get_user_pages_locked is passed an explicit mm, and current->mm is not always valid. This was hit when being called from i915, which uses:

pin_user_pages_remote->

    __get_user_pages_remote->
      __gup_longterm_locked->
        __get_user_pages_locked

Before, this would lead to an OOPS:

  BUG: kernel NULL pointer dereference, address: 0000000000000064
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP
  CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U     O      5.9.0-rc7+ #140
  Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS N2OET47W (1.34 ) 08/06/2020
  Workqueue: i915-userptr-acquire __i915_gem_userptr_get_pages_worker [i915]
  RIP: 0010:__get_user_pages_remote+0xd7/0x310
  Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1 00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40 03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24 18 00
  RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
  RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
  RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
  R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
  R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
  FS:  0000000000000000(0000) GS:ffff888ffc480000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
   process_one_work+0x1ca/0x390
   worker_thread+0x48/0x3c0
   ? rescuer_thread+0x3d0/0x3d0
   kthread+0x114/0x130
   ? kthread_create_worker_on_cpu+0x40/0x40
   ret_from_fork+0x1f/0x30
  CR2: 0000000000000064

This commit fixes the problem by using the mm pointer passed to the function rather than the bogus one in current.

Fixes 008cfe4418b3 ("mm: Introduce mm_struct.has_pinned")
Cc Jason Gunthorpe <[email protected]>
Cc Peter Xu <[email protected]>
---
 mm/gup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/gup.c +1 -1
index dfe781d2ad4c..e869c634cc9a 100644
@@ -1256,7 +1256,7 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
 	}
 
 	if (flags & FOLL_PIN)
-		atomic_set(&current->mm->has_pinned, 1);
+		atomic_set(&mm->has_pinned, 1);
 
 	/*
 	 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
Signature · 1 line
2.28.0
#5 lore raw

Re: [PATCH] mm: do not rely on mm == current->mm in __get_user_pages_locked

Quoting Jason A. Donenfeld (2020-09-28 11:35:07)

Quoted 59 lines

It seems likely this block was pasted from internal_get_user_pages_fast, which is not passed an mm struct and therefore uses current's. But __get_user_pages_locked is passed an explicit mm, and current->mm is not always valid. This was hit when being called from i915, which uses:

pin_user_pages_remote->

    __get_user_pages_remote->
      __gup_longterm_locked->
        __get_user_pages_locked

Before, this would lead to an OOPS:

  BUG: kernel NULL pointer dereference, address: 0000000000000064
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP
  CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U     O      5.9.0-rc7+ #140
  Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS N2OET47W (1.34 ) 08/06/2020
  Workqueue: i915-userptr-acquire __i915_gem_userptr_get_pages_worker [i915]
  RIP: 0010:__get_user_pages_remote+0xd7/0x310
  Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1 00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40 03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24 18 00
  RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
  RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
  RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
  R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
  R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
  FS:  0000000000000000(0000) GS:ffff888ffc480000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
   process_one_work+0x1ca/0x390
   worker_thread+0x48/0x3c0
   ? rescuer_thread+0x3d0/0x3d0
   kthread+0x114/0x130
   ? kthread_create_worker_on_cpu+0x40/0x40
   ret_from_fork+0x1f/0x30
  CR2: 0000000000000064

This commit fixes the problem by using the mm pointer passed to the function rather than the bogus one in current.

Fixes 008cfe4418b3 ("mm: Introduce mm_struct.has_pinned")
Cc Jason Gunthorpe <[email protected]>
Cc Peter Xu <[email protected]>
---
 mm/gup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/gup.c +0 -0
index dfe781d2ad4c..e869c634cc9a 100644

mm/gup.c static __always_inline long __get_user_pages_locked(struct mm_struct *mm, · line 1259

Quoted 5 lines
mm/gup.c +1 -1
@@ -1256,7 +1256,7 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
        }
 
        if (flags & FOLL_PIN)
-               atomic_set(&current->mm->has_pinned, 1);
+               atomic_set(&mm->has_pinned, 1);

That's literally the same diff as I was just testing :)

I can attest that it fixes the i915 issue, but since that's also your test case, I'm not adding much information.
-Chris

Reviewed-by on v1 1/1
#6 lore raw

Re: [PATCH] mm: do not rely on mm == current->mm in __get_user_pages_locked

In reply to Jason A. Donenfeld
Quoted 46 lines · Jason A. Donenfeld

It seems likely this block was pasted from internal_get_user_pages_fast, which is not passed an mm struct and therefore uses current's. But __get_user_pages_locked is passed an explicit mm, and current->mm is not always valid. This was hit when being called from i915, which uses:

pin_user_pages_remote->

    __get_user_pages_remote->
      __gup_longterm_locked->
        __get_user_pages_locked

Before, this would lead to an OOPS:

  BUG: kernel NULL pointer dereference, address: 0000000000000064
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP
  CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U     O      5.9.0-rc7+ #140
  Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS N2OET47W (1.34 ) 08/06/2020
  Workqueue: i915-userptr-acquire __i915_gem_userptr_get_pages_worker [i915]
  RIP: 0010:__get_user_pages_remote+0xd7/0x310
  Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1 00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40 03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24 18 00
  RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
  RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
  RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
  R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
  R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
  FS:  0000000000000000(0000) GS:ffff888ffc480000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
   process_one_work+0x1ca/0x390
   worker_thread+0x48/0x3c0
   ? rescuer_thread+0x3d0/0x3d0
   kthread+0x114/0x130
   ? kthread_create_worker_on_cpu+0x40/0x40
   ret_from_fork+0x1f/0x30
  CR2: 0000000000000064

This commit fixes the problem by using the mm pointer passed to the function rather than the bogus one in current.

Quoted 7 lines
Fixes 008cfe4418b3 ("mm: Introduce mm_struct.has_pinned")
Cc Jason Gunthorpe <[email protected]>
Cc Peter Xu <[email protected]>
---
 mm/gup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Yes this looks like the right fix

Thanks,
Jason

reviewer Reviewed-by on v1 1/1
#7 lore raw

Re: [PATCH] mm: do not rely on mm == current->mm in __get_user_pages_locked

In reply to Jason A. Donenfeld
Quoted 9 lines · Jason A. Donenfeld

It seems likely this block was pasted from internal_get_user_pages_fast, which is not passed an mm struct and therefore uses current's. But __get_user_pages_locked is passed an explicit mm, and current->mm is not always valid. This was hit when being called from i915, which uses:

pin_user_pages_remote->

    __get_user_pages_remote->
      __gup_longterm_locked->
        __get_user_pages_locked

Afaict it's not only an "current->mm can be NULL" issue - because this flag is used to mark "whether the mm pinned any page", so for remote pinning we definitely should mark the remote mm rather than the current mm, simply because it's the target mm page table that we'd want to stablize rather than the current->mm (even if current->mm always existed).

Quoted 61 lines

Before, this would lead to an OOPS:

  BUG: kernel NULL pointer dereference, address: 0000000000000064
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP
  CPU: 10 PID: 1431 Comm: kworker/u33:1 Tainted: P S   U     O      5.9.0-rc7+ #140
  Hardware name: LENOVO 20QTCTO1WW/20QTCTO1WW, BIOS N2OET47W (1.34 ) 08/06/2020
  Workqueue: i915-userptr-acquire __i915_gem_userptr_get_pages_worker [i915]
  RIP: 0010:__get_user_pages_remote+0xd7/0x310
  Code: f5 01 00 00 83 7d 00 01 0f 85 ed 01 00 00 f7 c1 00 00 04 00 0f 84 58 01 00 00 65 48 8b 04 25 00 6d 01 00 48 8b 80 40 03 00 00 <c7> 40 64 01 00 00 00 65 48 8b 04 25 00 6d 01 00 48 c7 44 24 18 00
  RSP: 0018:ffff888fdfe47de0 EFLAGS: 00010206
  RAX: 0000000000000000 RBX: 00007fe188531000 RCX: 0000000000040001
  RDX: 0000000000000001 RSI: 00007fe188531000 RDI: ffff888ff0748f00
  RBP: ffff888fdfe47e54 R08: ffff888fedc7d7c8 R09: 0000000000000000
  R10: 0000000000000018 R11: fefefefefefefeff R12: ffff888ff0748f00
  R13: ffff888fedc7d7c8 R14: ffff888f81fe3a40 R15: 0000000000042003
  FS:  0000000000000000(0000) GS:ffff888ffc480000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000064 CR3: 0000000002009003 CR4: 00000000003706e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   __i915_gem_userptr_get_pages_worker+0xc8/0x260 [i915]
   process_one_work+0x1ca/0x390
   worker_thread+0x48/0x3c0
   ? rescuer_thread+0x3d0/0x3d0
   kthread+0x114/0x130
   ? kthread_create_worker_on_cpu+0x40/0x40
   ret_from_fork+0x1f/0x30
  CR2: 0000000000000064

This commit fixes the problem by using the mm pointer passed to the function rather than the bogus one in current.

Fixes 008cfe4418b3 ("mm: Introduce mm_struct.has_pinned")
Cc Jason Gunthorpe <[email protected]>
Cc Peter Xu <[email protected]>
---
 mm/gup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/gup.c +1 -1
index dfe781d2ad4c..e869c634cc9a 100644
@@ -1256,7 +1256,7 @@ static __always_inline long __get_user_pages_locked(struct mm_struct *mm,
 	}
 
 	if (flags & FOLL_PIN)
-		atomic_set(&current->mm->has_pinned, 1);
+		atomic_set(&mm->has_pinned, 1);
 
 	/*
 	 * FOLL_PIN and FOLL_GET are mutually exclusive. Traditional behavior
Signature · 1 line
2.28.0

Thanks! And sorry for this silly mistake. I even didn't understand how it was written, because the normal gup change should have come earlier, anyway...

Signature · 6 lines
Peter Xu

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel