From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754516Ab2DCM3W (ORCPT ); Tue, 3 Apr 2012 08:29:22 -0400 Received: from mail-ey0-f174.google.com ([209.85.215.174]:41208 "EHLO mail-ey0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752644Ab2DCM3V (ORCPT ); Tue, 3 Apr 2012 08:29:21 -0400 Subject: Re: [PATCH v6 0/2] Ethernet drivers for WIZnet chips From: Eric Dumazet To: Mike Sinkovsky Cc: netdev@vger.kernel.org, linux-kernel@vger.kernel.org In-Reply-To: <4F7AE9D3.80204@permonline.ru> References: <1332752876-1650-1-git-send-email-msink@permonline.ru> <1333450726-24455-1-git-send-email-msink@permonline.ru> <1333453170.18626.65.camel@edumazet-glaptop> <4F7AE9D3.80204@permonline.ru> Content-Type: text/plain; charset="UTF-8" Date: Tue, 03 Apr 2012 14:29:16 +0200 Message-ID: <1333456156.18626.93.camel@edumazet-glaptop> Mime-Version: 1.0 X-Mailer: Evolution 2.28.3 Content-Transfer-Encoding: 8bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 2012-04-03 at 18:15 +0600, Mike Sinkovsky wrote: > 03.04.2012 17:39, Eric Dumazet написал: > > On Tue, 2012-04-03 at 16:58 +0600, Mike Sinkovsky wrote: > >> Based on original driver from chip manufacturer, but nearly full rewite. > >> Tested and used in production with Blackfin BF531 embedded processor. > >> > >> Signed-off-by: Mike Sinkovsky > >> --- > >> v6: > >> - remove (S0_TX_FSR< skb->len) check in TX handler, it doesn't work > >> anyway. Chip can transmit only one frame with MTU 1500 at a time, > >> and tx buffer size is bigger. > > > > So what happens if XXX frames are given to start_xmit() in a flood ? > > > > You removed any flow control, how can this work ? > > > > Device has an infinite queue ? > > As I understand from datasheet, device doesn't have tx queue at all. > It have tx buffer, processor must save transmitted frame to it, and then > save SEND command to command register. When transmission completed, > SENDOK bit in interrupt register will be set to '1', and interrupt > handler will be called. > > So, according to datasheet, driver must stop queue in start_xmit() > routine, and wake in interrupt handler of SENDOK bit. > I tried this, and it basically works, but SOMETIMES, very rare, I see > tx_timeout(). > > And without any flow control - driver works perfectly. > Weird, don't know why. Really this should be fixed, since your driver makes qdisc flow control impossible (unless adding a rate limiter like HTB/CBQ) You probably had a race on your xmit routine, and interrupt routine. You need proper synchronization : A spinlock to make this easy, or a smart barrier game.