Re: mm: slub: gpf in deactivate_slab

[Michal Hocko]

On Tue 25-03-14 11:54:43, Sasha Levin wrote:
> On 03/12/2014 12:25 PM, Sasha Levin wrote:
> >Hi all,
> >
> >While fuzzing with trinity inside a KVM tools guest running latest -next kernel I've stumbled
> >on the following spew:
> >
> >[  241.916559] BUG: unable to handle kernel paging request at ffff880029aa5e58
> >[  241.917961] IP: [<ffffffff812c5fa3>] deactivate_slab+0x103/0x560
> >[  241.919439] PGD 88f9067 PUD 88fa067 PMD 102fd35067 PTE 8000000029aa5060
> >[  241.920339] Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
> >[  241.920339] Dumping ftrace buffer:
> >[  241.920339]    (ftrace buffer empty)
> >[  241.920339] Modules linked in:
> >[  241.920339] CPU: 17 PID: 9910 Comm: trinity-c183 Tainted: G        W    3.14.0-rc6-next-20140311-sasha-00009-g6c028cd-dirty #146
> >[  241.920339] task: ffff88090eb68000 ti: ffff88090eb70000 task.ti: ffff88090eb70000
> >[  241.920339] RIP: 0010:[<ffffffff812c5fa3>]  [<ffffffff812c5fa3>] deactivate_slab+0x103/0x560
> >[  241.920339] RSP: 0018:ffff88090eb71c18  EFLAGS: 00010082
> >[  241.920339] RAX: 0000000000000418 RBX: ffff880229ae2010 RCX: 0000000180170017
> >[  241.920339] RDX: 0000000000000000 RSI: ffffffff812c5f52 RDI: ffffffff812c5e29
> >[  241.920339] RBP: ffff88090eb71d28 R08: ffff880229ae2010 R09: 0000000000000080
> >[  241.920339] R10: ffff880229ae2ed0 R11: 0000000000000000 R12: ffffea0008a6b800
> >[  241.920339] R13: ffff88012b4da580 R14: ffff880029aa5a40 R15: ffff880229ae2010
> >[  241.920339] FS:  00007fb615415700(0000) GS:ffff88022ba00000(0000) knlGS:0000000000000000
> >[  241.920339] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
> >[  241.920339] CR2: ffff880029aa5e58 CR3: 000000090eb4a000 CR4: 00000000000006a0
> >[  241.920339] DR0: 0000000000005bf2 DR1: 0000000000000000 DR2: 0000000000000000
> >[  241.920339] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000070602
> >[  241.920339] Stack:
> >[  241.920339]  ffffffff856e37ab ffffffff84486272 0000000000000000 000000002b405600
> >[  241.920339]  ffff88090eb71c58 ffff88090eb70000 0000000000000011 ffff88022ba03fc0
> >[  241.920339]  ffff88090eb71cb8 000000000000000f ffff88022b405600 ffff880029aa5a40
> >[  241.920339] Call Trace:
> >[  241.920339]  [<ffffffff84486272>] ? preempt_count_sub+0xe2/0x120
> >[  241.920339]  [<ffffffff8130f151>] ? alloc_inode+0x41/0xa0
> >[  241.954886]  [<ffffffff8130f151>] ? alloc_inode+0x41/0xa0
> >[  241.954886]  [<ffffffff8130f151>] ? alloc_inode+0x41/0xa0
> >[  241.954886]  [<ffffffff812c36eb>] ? set_track+0xab/0x100
> >[  241.954886]  [<ffffffff812c731f>] __slab_alloc+0x42f/0x4d0
> >[  241.954886]  [<ffffffff81073e6d>] ? sched_clock+0x1d/0x30
> >[  241.954886]  [<ffffffff8130f151>] ? alloc_inode+0x41/0xa0
> >[  241.961624]  [<ffffffff812c870f>] kmem_cache_alloc+0x12f/0x2e0
> >[  241.961624]  [<ffffffff8130f151>] ? alloc_inode+0x41/0xa0
> >[  241.961624]  [<ffffffff8130f151>] alloc_inode+0x41/0xa0
> >[  241.961624]  [<ffffffff8130f1cb>] new_inode_pseudo+0x1b/0x70
> >[  241.961624]  [<ffffffff812f985c>] get_pipe_inode+0x1c/0xf0
> >[  241.961624]  [<ffffffff812f995c>] create_pipe_files+0x2c/0x170
> >[  241.961624]  [<ffffffff812f9ae1>] __do_pipe_flags+0x41/0xf0
> >[  241.961624]  [<ffffffff812f9bbb>] SyS_pipe2+0x2b/0xb0
> >[  241.961624]  [<ffffffff8448b3b1>] ? tracesys+0x7e/0xe2
> >[  241.961624]  [<ffffffff8448b410>] tracesys+0xdd/0xe2
> >[  241.972975] Code: 4d 85 f6 75 8b eb 45 90 4d 85 f6 75 13 49 8b 5c 24 10 45 31 ff 0f 1f 00 eb 32 66 0f 1f 44 00 00 4c 89 b5 48 ff ff ff 49 63 45 20 <49> 8b 0c 06 48 85 c9 74 10 4d 89 f7 49 8b 54 24 10 49 89 ce e9
> >[  241.972975] RIP  [<ffffffff812c5fa3>] deactivate_slab+0x103/0x560
> >[  241.972975]  RSP <ffff88090eb71c18>
> >[  241.972975] CR2: ffff880029aa5e58
> 
> I have a lead on this. Consider the following:
> 
>   kmem_cache_alloc
> 	__slab_alloc
> 		local_irq_save()
> 		deactivate_slab
> 			__cmpxchg_double_slab
> 				slab_unlock
> 					__bit_spin_unlock
> 						preempt_enable
> 		[ Page Fault ]
> 
> With this trace, it manifests as a "BUG: sleeping function called from invalid
> context at arch/x86/mm/fault.c" on a might_sleep() in the page fault handler
> (which is an issue on it's own), but I suspect it's also the cause of the trace
> above - preemption enabled and a race that removed the page.
> 
> Could someone confirm please?

You are right. The function even does VM_BUG_ON(!irqs_disabled())...
Unfortunatelly we do not seem to have an _irq alternative of the bit
spinlock.
Not sure what to do about it. Christoph? 

Btw. it seems to go way back to 3.1 (1d07171c5e58e).
-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/