Re: [PATCH v2] mm: khugepaged: Fix kernel BUG in hpage_collapse_scan_file

[Zach O'Keefe]

On Mar 30 19:53, Ivan Orlov wrote:
> Syzkaller reported the following issue:
> 
> kernel BUG at mm/khugepaged.c:1823!
> invalid opcode: 0000 [#1] PREEMPT SMP KASAN
> CPU: 1 PID: 5097 Comm: syz-executor220 Not tainted 6.2.0-syzkaller-13154-g857f1268a591 #0
> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/16/2023
> RIP: 0010:collapse_file mm/khugepaged.c:1823 [inline]
> RIP: 0010:hpage_collapse_scan_file+0x67c8/0x7580 mm/khugepaged.c:2233
> Code: 00 00 89 de e8 c9 66 a3 ff 31 ff 89 de e8 c0 66 a3 ff 45 84 f6 0f 85 28 0d 00 00 e8 22 64 a3 ff e9 dc f7 ff ff e8 18 64 a3 ff <0f> 0b f3 0f 1e fa e8 0d 64 a3 ff e9 93 f6 ff ff f3 0f 1e fa 4c 89
> RSP: 0018:ffffc90003dff4e0 EFLAGS: 00010093
> RAX: ffffffff81e95988 RBX: 00000000000001c1 RCX: ffff8880205b3a80
> RDX: 0000000000000000 RSI: 00000000000001c0 RDI: 00000000000001c1
> RBP: ffffc90003dff830 R08: ffffffff81e90e67 R09: fffffbfff1a433c3
> R10: 0000000000000000 R11: dffffc0000000001 R12: 0000000000000000
> R13: ffffc90003dff6c0 R14: 00000000000001c0 R15: 0000000000000000
> FS:  00007fdbae5ee700(0000) GS:ffff8880b9900000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00007fdbae6901e0 CR3: 000000007b2dd000 CR4: 00000000003506e0
> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> Call Trace:
>  <TASK>
>  madvise_collapse+0x721/0xf50 mm/khugepaged.c:2693
>  madvise_vma_behavior mm/madvise.c:1086 [inline]
>  madvise_walk_vmas mm/madvise.c:1260 [inline]
>  do_madvise+0x9e5/0x4680 mm/madvise.c:1439
>  __do_sys_madvise mm/madvise.c:1452 [inline]
>  __se_sys_madvise mm/madvise.c:1450 [inline]
>  __x64_sys_madvise+0xa5/0xb0 mm/madvise.c:1450
>  do_syscall_x64 arch/x86/entry/common.c:50 [inline]
>  do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80
>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
> 

Thanks, Ivan.

In the process of reviewing this, I starting thinking if the !shmem case was
also susceptible to a similar race, and I *think* it might be. Unfortunately, my
time has ran out, and I haven't been able to validate ; I'm less familiar with
the file-side of things.

The underlying problem is race with truncation/hole-punch  under OOM condition.
The nice do-while loop near the top of collapse_file() attempts to avoid this
scenario by making sure enough slots are available. However, when we drop xarray
lock, we open ourselves up to concurrent removal + slot deletion. Those slots
then need to be allocated again -- which under OOM condition is failable.

The syzbot reproducer picks on shmem, but I think this can occur for file as
well. If we find a hole, we unlock the xarray and call
page_cache_sync_readahead(), which if it succeeds, IIUC, will have allocated a
new slot in our mapping pointing to the new page. We *then* locks the page. Only
after the page is locked are we protected from concurrent removal (Note: this is
what provides us protection in many of the xas_store() cases ; we've held the
slot's contained page-lock since verifying the slot exists, protecting us from
removal / reallocation races).

Maybe I'm just low on caffeine at the end of the day, and am missing something,
but if I had more time, I'd be looking into the file-side some more to verify.
Apologies that hasn't occurred to me until now ; I was looking at one of your
comments and double-checked why I *thought* we were safe.

Anyways, irrespective of that looming issues, some more notes to follow:

> The 'xas_store' call during page cache scanning can potentially
> translate 'xas' into the error state (with the reproducer provided
> by the syzkaller the error code is -ENOMEM). However, there are no
> further checks after the 'xas_store', and the next call of 'xas_next'
> at the start of the scanning cycle doesn't increase the xa_index,
> and the issue occurs.
> 
> This patch will add the xarray state error checking after the
> 'xas_store' and the corresponding result error code. It will
> also add xarray state error checking via WARN_ON_ONCE macros,
> to be sure that ENOMEM or other possible errors don't occur
> at the places they shouldn't.

Thanks for the additions here. I think it's worthwhile providing even more
details about the specifics of the race we are fixing and/or guarding against to
help ppl understand how that -ENOMEM comes about if the do-while loop has
"Ensured" we have slots available (additionally, I think that comment can be
augmented).

> Tested via syzbot.
> 
> Reported-by: syzbot+9578faa5475acb35fa50@xxxxxxxxxxxxxxxxxxxxxxxxx
> Link: https://syzkaller.appspot.com/bug?id=7d6bb3760e026ece7524500fe44fb024a0e959fc
> Signed-off-by: Ivan Orlov <ivan.orlov0322@xxxxxxxxx>
> ---
> V1 -> V2: Add WARN_ON_ONCE error checking and comments
> 
>  mm/khugepaged.c | 20 ++++++++++++++++++++
>  1 file changed, 20 insertions(+)
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 92e6f56a932d..8b6580b13339 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -55,6 +55,7 @@ enum scan_result {
>  	SCAN_CGROUP_CHARGE_FAIL,
>  	SCAN_TRUNCATED,
>  	SCAN_PAGE_HAS_PRIVATE,
> +	SCAN_STORE_FAILED,
>  };

I'm still reluctant to add a new error code for this as this seems like quite a
rare race that requires OOM to trigger. I'd be happier just reusing SCAN_FAIL,
or, something we might get some millage out of later: SCAN_OOM.

Also, a reminder to update include/trace/events/huge_memory.h, if you go that
route.

>  
>  #define CREATE_TRACE_POINTS
> @@ -1840,6 +1841,15 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
>  					goto xa_locked;
>  				}
>  				xas_store(&xas, hpage);
> +				if (xas_error(&xas)) {
> +					/* revert shmem_charge performed
> +					 * in the previous condition
> +					 */

Nit: Here, and following, I think standard convention for multiline comment is
to have an empty first and last line, eg:

 +					/*
 +					 * revert shmem_charge performed
 +					 * in the previous condition
 +					 */

Though, checkpatch.pl --strict didn't seem to care.

> +					mapping->nrpages--;
> +					shmem_uncharge(mapping->host, 1);
> +					result = SCAN_STORE_FAILED;
> +					goto xa_locked;
> +				}
>  				nr_none++;
>  				continue;
>  			}
> @@ -1992,6 +2002,11 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
>  
>  		/* Finally, replace with the new page. */
>  		xas_store(&xas, hpage);
> +		/* We can't get an ENOMEM here (because the allocation happened before)
> +		 * but let's check for errors (XArray implementation can be
> +		 * changed in the future)
> +		 */
> +		WARN_ON_ONCE(xas_error(&xas));

Nit: it's not just that allocation happened before -- need some guarantee we've
been protected from concurrent removal. This is what made me look at the file
side.

>  		continue;
>  out_unlock:
>  		unlock_page(page);
> @@ -2029,6 +2044,11 @@ static int collapse_file(struct mm_struct *mm, unsigned long addr,
>  	/* Join all the small entries into a single multi-index entry */
>  	xas_set_order(&xas, start, HPAGE_PMD_ORDER);
>  	xas_store(&xas, hpage);
> +	/* Here we can't get an ENOMEM (because entries were
> +	 * previously allocated) But let's check for errors
> +	 * (XArray implementation can be changed in the future)
> +	 */
> +	WARN_ON_ONCE(xas_error(&xas));

Ditto.

Apologies I won't be around to see this change through -- I'm just out of time,
and will be shutting my computer down tomorrow for 3 months.  Sorry for the poor
timing, for raising issues, then disappearing. Hopefully I'm wrong and the
file-side isn't a concern.

Best,
Zach

>  xa_locked:
>  	xas_unlock_irq(&xas);
>  xa_unlocked:
> -- 
> 2.34.1
>