From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Michael S. Tsirkin" Subject: Re: [RFC PATCH net-next 0/6] Always use tx interrupt for virtio-net Date: Wed, 15 Oct 2014 14:58:32 +0300 Message-ID: <20141015115832.GE26918@redhat.com> References: <1413357930-45302-1-git-send-email-jasowang@redhat.com> <20141015102553.GF25776@redhat.com> <543E5706.10305@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Cc: eric.dumazet@gmail.com, netdev@vger.kernel.org, linux-kernel@vger.kernel.org, virtualization@lists.linux-foundation.org, davem@davemloft.net To: Jason Wang Return-path: Content-Disposition: inline In-Reply-To: <543E5706.10305@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, Oct 15, 2014 at 07:14:14PM +0800, Jason Wang wrote: > On 10/15/2014 06:25 PM, Michael S. Tsirkin wrote: > > On Wed, Oct 15, 2014 at 03:25:24PM +0800, Jason Wang wrote: > >> According to David, proper accounting and queueing (at all levels, not > >> just TCP sockets) is more important than trying to skim a bunch of > >> cycles by avoiding TX interrupts. > > He also mentioned we should find other ways to batch > > > > Right. > >> Having an event to free the SKB is > >> absolutely essential for the stack to operate correctly. > >> > >> This series tries to enable tx interrupt for virtio-net. The idea is > >> simple: enable tx interrupt and schedule a tx napi to free old xmit > >> skbs. > >> > >> Several notes: > >> - Tx interrupt storm avoidance when queue is about to be full is > >> kept. > > But queue is typically *not* full. More important to avoid interrupt > > storms in that case IMO. > > Yes. > >> Since we may enable callbacks in both ndo_start_xmit() and tx > >> napi, patch 1 adds a check to make sure used event never go > >> back. This will let the napi not enable the callbacks wrongly after > >> delayed callbacks was used. > > So why not just use delayed callbacks? > > This means the tx interrupt are coalesced in a somewhat adaptive way. > Need benchmark to see its effect. I think it's a minimal change, and does not need new APIs. If that's not optimal, we can do smarter things on top. > > > >> - For bulk dequeuing, there's no need to enable tx interrupt for each > >> packet. The last patch only enable tx interrupt for the final skb in > >> the chain through xmit_more and a new helper to publish current avail > >> idx as used event. > >> > >> This series fixes several issues of original rfc pointed out by Michael. > > Could you list the issues, for ease of review? > > Probably just one: > > - Move the virtqueue_disable_cb() from skb_xmit_done() into > virtnet_poll_tx() under tx lock. I think I did this already, I'll recheck.