Re: [patch] net, skbuff: do not prefer skb allocation fails early

[Eric Dumazet <eric.dumazet@gmail.com>]

On 01/02/2019 01:01 PM, David Rientjes wrote:
> Commit dcda9b04713c ("mm, tree wide: replace __GFP_REPEAT by 
> __GFP_RETRY_MAYFAIL with more useful semantic") replaced __GFP_REPEAT in
> alloc_skb_with_frags() with __GFP_RETRY_MAYFAIL when the allocation may
> directly reclaim.
> 
> The previous behavior would require reclaim up to 1 << order pages for
> skb aligned header_len of order > PAGE_ALLOC_COSTLY_ORDER before failing,
> otherwise the allocations in alloc_skb() would loop in the page allocator
> looking for memory.  __GFP_RETRY_MAYFAIL makes both allocations failable
> under memory pressure, including for the HEAD allocation.
> 
> This can cause, among many other things, write() to fail with ENOTCONN
> during RPC when under memory pressure.  
> 
> These allocations should succeed as they did previous to dcda9b04713c
> even if it requires calling the oom killer and additional looping in the
> page allocator to find memory.  There is no way to specify the previous
> behavior of __GFP_REPEAT, but it's unlikely to be necessary since the
> previous behavior only guaranteed that 1 << order pages would be reclaimed
> before failing for order > PAGE_ALLOC_COSTLY_ORDER.  That reclaim is not
> guaranteed to be contiguous memory, so repeating for such large orders is
> usually not beneficial.
> 
> Removing the setting of __GFP_RETRY_MAYFAIL to restore the previous
> behavior, specifically not allowing alloc_skb() to fail for small orders
> and oom kill if necessary rather than allowing RPCs to fail.
> 
> Fixes: dcda9b04713c ("mm, tree wide: replace __GFP_REPEAT by
> __GFP_RETRY_MAYFAIL with more useful semantic") 
> Signed-off-by: David Rientjes <rientjes@google.com>

Reviewed-by: Eric Dumazet <edumazet@google.com>

Thanks David.