From mboxrd@z Thu Jan 1 00:00:00 1970 From: Eric Dumazet Subject: Re: [PATCH] virtio-net: fix a race on 32bit arches Date: Wed, 06 Jun 2012 17:19:04 +0200 Message-ID: <1338995944.26966.6.camel@edumazet-glaptop> References: <1338971724.2760.3913.camel@edumazet-glaptop> <1338972341.2760.3944.camel@edumazet-glaptop> <20120606111357.GA15070@redhat.com> <1338988210.2760.4485.camel@edumazet-glaptop> <20120606144941.GA17092@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Cc: netdev@vger.kernel.org, linux-kernel@vger.kernel.org, virtualization@lists.linux-foundation.org, Stephen Hemminger To: "Michael S. Tsirkin" Return-path: In-Reply-To: <20120606144941.GA17092@redhat.com> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: virtualization-bounces@lists.linux-foundation.org Errors-To: virtualization-bounces@lists.linux-foundation.org List-Id: netdev.vger.kernel.org On Wed, 2012-06-06 at 17:49 +0300, Michael S. Tsirkin wrote: > On Wed, Jun 06, 2012 at 03:10:10PM +0200, Eric Dumazet wrote: > > On Wed, 2012-06-06 at 14:13 +0300, Michael S. Tsirkin wrote: > > > > > We currently do all stats either on napi callback or from > > > start_xmit callback. > > > This makes them safe, yes? > > > > Hmm, then _bh() variant is needed in virtnet_stats(), as explained in > > include/linux/u64_stats_sync.h section 6) > > > > * 6) If counter might be written by an interrupt, readers should block interrupts. > > * (On UP, there is no seqcount_t protection, a reader allowing interrupts could > > * read partial values) > > > > Yes, its tricky... > > Sounds good, but I have a question: this realies on counters > being atomic on 64 bit. > Would not it be better to always use a seqlock even on 64 bit? > This way counters would actually be correct and in sync. > As it is if we want e.g. average packet size, > we can not rely e.g. on it being bytes/packets. When this stuff was discussed, we chose to have a nop on 64bits. Your point has little to do with 64bit stats, it was already like that with 'long int' counters. Consider average driver doing : dev->stats.rx_bytes += skb->len; dev->stats.rx_packets++; A concurrent reader can read an updated rx_bytes and a 'previous' rx_packets one. 'fixing' this requires a lot of work and memory barriers (in all drivers), for a very litle gain (at most one packet error) u64_stats_sync was really meant to be 0-cost on 64bit arches.