Re: [PATCH v4 1/2] hung_task: Show the blocker task if the task is hung on mutex

[Lance Yang <lance.yang@linux.dev>]

On 2025/7/30 18:01, Masami Hiramatsu (Google) wrote:
> On Wed, 30 Jul 2025 17:36:04 +0800
> Lance Yang <lance.yang@linux.dev> wrote:
> 
>>
>>
>> On 2025/7/30 16:51, Masami Hiramatsu (Google) wrote:
>>> On Wed, 30 Jul 2025 16:59:22 +0900
>>> Sergey Senozhatsky <senozhatsky@chromium.org> wrote:
>>>
>>>> One thing that gives me a bit of "inconvenience" is that in certain
>>>> cases this significantly increases the amount of stack traces to go
>>>> through.  A distilled real life example:
>>>> - task T1 acquires lock L1, attempts to acquire L2
>>>> - task T2 acquires lock L2, attempts to acquire L3
>>>> - task T3 acquires lock L3, attempts to acquire L1
>>>>
>>>> So we'd now see:
>>>> - a backtrace of T1, followed by a backtrace of T2 (owner of L2)
>>>> - a backtrace of T2, followed by a backtrace of T3 (owner of L3)
>>>> - a backtrace of T3, followed by a backtrace of T1 (owner of L1)
>>>>
>>>> Notice how each task is backtraced twice.  I wonder if it's worth it
>>>> to de-dup the backtraces.  E.g. in
>>>>
>>>> 	task cat:115 is blocked on a mutex likely owned by task cat:114
>>>>
>>>> if we know that cat:114 is also blocked on a lock, then we probably
>>>> can just say "is blocked on a mutex likely owned by task cat:114" and
>>>> continue iterating through tasks.  That "cat:114" will be backtraced
>>>> individually later, as it's also blocked on a lock, owned by another
>>>> task.
>>>>
>>>> Does this make any sense?
>>>
>>> Hrm, OK. So what about dump the blocker task only if that task is
>>> NOT blocked? (because if the task is blocked, it should be dumped
>>> afterwards (or already))
>>
>> Hmm... I'm concerned about a potential side effect of that logic.
>>
>> Consider a simple, non-circular blocking chain like T1 -> T2 -> T3.
>>
>> In this scenario, T1, T2, and T3 would all be dumped as hung tasks.
>> However, with the proposed rule (dump only if NOT blocked), when the
>> detector processes T1, it would see that its blocker (T2) is also
>> blocked and would therefore skip printing any blocker information about
>> T2.
>>
>> The key issue is that we would lose the crucial T1 -> T2 relationship
>> information from the log.
> 
> I just skip printing T2's stack dump, but still show "T1 is blocked by T2"
> so the relationship is still clear.

Ah, I see! That approach makes sense to me ;)


Thanks,
Lance