Re: [PATCH 2/2] mm: do not loop over ALLOC_NO_WATERMARKS without triggering reclaim

[Tetsuo Handa]

Michal Hocko wrote:
> __alloc_pages_slowpath is looping over ALLOC_NO_WATERMARKS requests if
> __GFP_NOFAIL is requested. This is fragile because we are basically
> relying on somebody else to make the reclaim (be it the direct reclaim
> or OOM killer) for us. The caller might be holding resources (e.g.
> locks) which block other other reclaimers from making any progress for
> example. Remove the retry loop and rely on __alloc_pages_slowpath to
> invoke all allowed reclaim steps and retry logic.

This implies invoking OOM killer, doesn't it?

>   	/* Avoid recursion of direct reclaim */
> -	if (current->flags & PF_MEMALLOC)
> +	if (current->flags & PF_MEMALLOC) {
> +		/*
> +		 * __GFP_NOFAIL request from this context is rather bizarre
> +		 * because we cannot reclaim anything and only can loop waiting
> +		 * for somebody to do a work for us.
> +		 */
> +		if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
> +			cond_resched();
> +			goto retry;

I think that this "goto retry;" omits call to out_of_memory() which is allowed
for __GFP_NOFAIL allocations. Even if this is what you meant, current thread
can be a workqueue, which currently need a short sleep (as with
wait_iff_congested() changes), can't it?

> +		}
>   		goto nopage;
> +	}
>   
>   	/* Avoid allocations with no watermarks from looping endlessly */
>   	if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
> 

Well, is it cond_resched() which should include

  if (current->flags & PF_WQ_WORKER)
  	schedule_timeout(1);

than wait_iff_congested() because not all yield calls use wait_iff_congested()
and giving pending workqueue jobs a chance to be processed is anyway preferable?

  int __sched _cond_resched(void)
  {
  	if (should_resched(0)) {
  		if ((current->flags & PF_WQ_WORKER) && workqueue_has_pending_jobs())
  			schedule_timeout(1);
  		else
  			preempt_schedule_common();
  		return 1;
  	}
  	return 0;
  }

--
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/