Re: [PATCH] stop swapoff 3/3 OOMkill

[Andrew Morton]

Hugh Dickins <[e-mail]> wrote:
>
> Current behaviour is that once swapoff has filled memory, other tasks
> get OOMkilled one by one until it completes, or more likely hangs.
> Better that swapoff be the first choice for OOMkill.

This calls for __GFP_NORETRY.  It will disable the oom-kill and the infinite
retry in the page allocator.  So we will have:

__GFP_REPEAT:

	retry the allocation, but the caller can handle a failure.
	eg: pte_alloc_one().

	__GFP_REPEAT _may_ end up returning NULL.  It depends on the VM
	implemention - eg it would in -aa kernels.  

__GFP_NOFAIL:

	retry the allocation inifinitely, regardless of the VM implementation. 
	For jbd_kmalloc() and others.

__GFP_NORETRY:

	Don't oom-kill and don't retry.   For swapoff.

I've implemented a __GFP_REPEAT, and don't like it, because it blurs the
__GFP_REPEAT and __GFP_NOFAIL requirements.  I'll add __GFP_NORETRY and we
can then pass that into read_swap_cache_async() and handle the error.

Sound good?