Re: [PATCH] mm: slub: Panic if the object corruption is checked.

[Vlastimil Babka <vbabka@suse.cz>]

On 1/21/25 1:40 AM, Hyesoo Yu wrote:
> On Mon, Jan 20, 2025 at 03:36:08PM +0000, Matthew Wilcox wrote:
>> On Mon, Jan 20, 2025 at 05:28:21PM +0900, Hyesoo Yu wrote:
>>> If a slab object is corrupted or an error occurs in its internal
>>> value, continuing after restoration may cause other side effects.
>>> At this point, it is difficult to debug because the problem occurred
>>> in the past. A flag has been added that can cause a panic when there
>>> is a problem with the object.
>>>
>>> Signed-off-by: Hyesoo Yu <hyesoo.yu@samsung.com>
>>> Change-Id: I4e7e5e0ec3421a7f6c84d591db052f79d3775493
>>
>> Linux does not use Change IDs.  Please omit these from future patches.
>>
>> Panicing is a very unfriendly approach.  I think a better approach would
>> be to freeze the slab where corruption is detected.  That is, no future
>> objects are allocated from that slab, and attempts to free objects from
>> that slab become no-ops.  I don't think that should be hard to implement.

Freezing of slab is already done in some cases when corruption is
detected - all objects are marked as used, and further freeing attempts
on the slab are discarded. Perhaps not all cases, which could be improved.

> Thanks you for your responce. That is my mistake. I will remove the change ID.
> 
> I agree that freezing is better than recovery or panic for the system's stability.
> However what I want from the patch is not just to make the system run stably.
> I need to immediately trigger a panic to investigate the slub.

IMHO it's a valid goal to panic more quickly when debugging, and
enabling slub_debug means debugging is in progress (as opposed to normal
production when we try to avoid panic).
But making it possible to reuse the general panic_on_warn mechanism
(which can be also expected to be enabled when debugging) is indeed
preferable to introducing a new slab-specific flag.

> I would like to analyze the corrupted data at that moment to check issues
> like cache problem, user errors, system clock frequency and similar problems,
> not just passing by without any issues.
> 
> However I agree that panic is not a friendly approach.
> I will modify it to notify the problem using warn() and then use
> panic_on_warn to trigger panic.
> 
> Thanks,
> Regards.
> 
>