From mboxrd@z Thu Jan  1 00:00:00 1970
From: Joe Perches <joe@perches.com>
Subject: Re: [PATCH] net: use __GFP_NORETRY for high order allocations
Date: Thu, 06 Feb 2014 13:34:04 -0800
Message-ID: <1391722444.15777.28.camel@joe-AO722>
References: <1391712162.10160.8.camel@edumazet-glaptop2.roam.corp.google.com>
	 <1391718270.15777.20.camel@joe-AO722>
	 <alpine.DEB.2.02.1402061302020.9567@chino.kir.corp.google.com>
Mime-Version: 1.0
Content-Type: text/plain; charset="ISO-8859-1"
Content-Transfer-Encoding: 7bit
Cc: Eric Dumazet <eric.dumazet@gmail.com>,
	David Miller <davem@davemloft.net>,
	netdev <netdev@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
To: David Rientjes <rientjes@google.com>
Return-path: <linux-kernel-owner@vger.kernel.org>
In-Reply-To: <alpine.DEB.2.02.1402061302020.9567@chino.kir.corp.google.com>
Sender: linux-kernel-owner@vger.kernel.org
List-Id: netdev.vger.kernel.org

On Thu, 2014-02-06 at 13:03 -0800, David Rientjes wrote:
> On Thu, 6 Feb 2014, Joe Perches wrote:
> 
> > On Thu, 2014-02-06 at 10:42 -0800, Eric Dumazet wrote:
> > > sock_alloc_send_pskb() & sk_page_frag_refill()
> > > have a loop trying high order allocations to prepare
> > > skb with low number of fragments as this increases performance.
> > > 
> > > Problem is that under memory pressure/fragmentation, this can
> > > trigger OOM while the intent was only to try the high order
> > > allocations, then fallback to order-0 allocations.
> > []
> > > Call Trace:
> > >  [<ffffffff8043766c>] dump_header+0xe1/0x23e
> > >  [<ffffffff80437a02>] oom_kill_process+0x6a/0x323
> > >  [<ffffffff80438443>] out_of_memory+0x4b3/0x50d
> > >  [<ffffffff8043a4a6>] __alloc_pages_may_oom+0xa2/0xc7
> > >  [<ffffffff80236f42>] __alloc_pages_nodemask+0x1002/0x17f0
> > >  [<ffffffff8024bd23>] alloc_pages_current+0x103/0x2b0
> > >  [<ffffffff8028567f>] sk_page_frag_refill+0x8f/0x160
> > []
> > > diff --git a/net/core/sock.c b/net/core/sock.c
> > []
> > > @@ -1775,7 +1775,9 @@ struct sk_buff *sock_alloc_send_pskb(struct sock *sk, unsigned long header_len,
> > >  			while (order) {
> > >  				if (npages >= 1 << order) {
> > >  					page = alloc_pages(sk->sk_allocation |
> > > -							   __GFP_COMP | __GFP_NOWARN,
> > > +							   __GFP_COMP |
> > > +							   __GFP_NOWARN |
> > > +							   __GFP_NORETRY,
> > >  							   order);
> > >  					if (page)
> > >  						goto fill_page;
> > > @@ -1845,7 +1847,7 @@ bool skb_page_frag_refill(unsigned int sz, struct page_frag *pfrag, gfp_t prio)
> > >  		gfp_t gfp = prio;
> > >  
> > >  		if (order)
> > > -			gfp |= __GFP_COMP | __GFP_NOWARN;
> > > +			gfp |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY;
> > 
> > Perhaps add __GFP_THISNODE too ?
> > 
> 
> How does __GFP_THISNODE have anything to do with avoiding oom killing due 
> to high-order fragmentation?

I don't think it does.

> If they absolutely require local memory to 
> currnet's cpu node then that would make sense,

I presumed THISNODE would be used only with NORETRY

> but the fallback still 
> allocates order-0 memory remotely and with __GFP_THISNODE on this attempt 
> we wouldn't even attempt remote reclaim.
any other alloc attempt could work on other cpus.

It was just a thought, ignore it if it's a dumb thought.