Re: [patch] don't allocate ratnodes under PF_MEMALLOC

[Christoph Hellwig]

>  
>  /*
> + * On the swap_out path, the radix-tree node allocations are performing
> + * GFP_ATOMIC allocations under PF_MEMALLOC.  They can completely
> + * exhaust the page allocator.  This is bad; some pages should be left
> + * available for the I/O system to start sending the swapcache contents
> + * to disk.
> + *
> + * So PF_MEMALLOC is dropped here.  This causes the slab allocations to fail
> + * earlier, so radix-tree nodes will then be allocated from the mempool
> + * reserves.
> + */
> +static inline int
> +swap_out_add_to_swap_cache(struct page *page, swp_entry_t entry)
> +{
> +	int flags = current->flags;
> +	int ret;
> +
> +	current->flags &= ~PF_MEMALLOC;
> +	ret = add_to_swap_cache(page, entry);
> +	current->flags = flags;
> +	return ret;
> +}

I don't like this soloution very - I think porting th add_to_swap() logic
from -rmap and implementing the flags fiddling in that function makes
more sense.  I will do so once a -pre4 is out to resync.

	Christoph