On Tue, 16 Dec 2025 at 02:31, Johannes Weiner <hannes@cmpxchg.org> wrote:
>
> On Mon, Dec 15, 2025 at 02:54:19PM +0000, Dipendra Khadka wrote:
> > In try_charge_memcg(), reorder the retry logic checks to follow the
> > early-exit pattern by testing for dying task before decrementing the
> > retry counter:
> >
> > Before:
> >     if (nr_retries--)
> >         goto retry;
> >    
> >     if (passed_oom && task_is_dying())
> >         goto nomem;
> >
> > After:
> >     if (passed_oom && task_is_dying())
> >         goto nomem;
> >    
> >     if (nr_retries--)
> >         goto retry;
> >
> > This makes the control flow more obvious: check exit conditions first,
> > then decide whether to retry. When current task is dying (e.g., has
> > received SIGKILL or is exiting), we should exit immediately rather than
> > consuming a retry count first.
> >
> > No functional change for the common case where task is not dying.
>
> It's definitely a functional change, not just code clarification.
>
> The oom kill resets nr_retries. This means that currently, even an OOM
> victim is going to retry a full set of reclaims, even if they are
> hopeless. After your patch, it'll retry for other reasons but can bail
> much earlier as well. Check the other conditions.
>
> The dying task / OOM victim allocation path is tricky and it tends to
> fail us in the rarest and most difficult to debug scenarios. There
> should be a good reason to change it.

The task_is_dying() check in try_charge_memcg() identifies when the
CURRENT task (the caller) is the OOM victim - not when some other 

process was killed.

Two scenarios:

1. Normal allocator triggers OOM:
  - Process A allocates → triggers OOM
  - Process B is killed (victim)
  - Process A continues with reset retries - task_is_dying() = false for A
  → Unchanged by my patch

2. Victim tries to allocate:
 - Process B (victim, TIF_MEMDIE set) tries to allocate
  - task_is_dying() = true
  - Current code: wastes retries on hopeless reclaims
  - My patch: exits immediately
  → Optimization for this case

The victim has three safety mechanisms that make the retries unnecessary:
1. oom_reaper proactively frees its memory
2. __alloc_pages_slowpath() grants reserves via oom_reserves_allowed()
3. Critical allocations with __GFP_NOFAIL still reach force: label

The retry loop for a dying victim is futile because:
- Reclaim won't help (victim is being killed to free memory!)
- Victim will exit regardless
- Just wastes CPU cycles

Would you like me to provide evidence showing the unnecessary retries,