From mboxrd@z Thu Jan 1 00:00:00 1970 From: Stephen Hemminger Subject: Re: [PATCH] sky2: Fix a race condition in sky2_poll Date: Mon, 6 Jul 2009 11:51:01 -0700 Message-ID: <20090706115101.68a8daaf@nehalam> References: <392fb48f0906200101y5c83e853w170829f2ce40b274@mail.gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: netdev@vger.kernel.org To: Mike McCormack Return-path: Received: from smtp1.linux-foundation.org ([140.211.169.13]:41620 "EHLO smtp1.linux-foundation.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753206AbZGFSvT (ORCPT ); Mon, 6 Jul 2009 14:51:19 -0400 In-Reply-To: <392fb48f0906200101y5c83e853w170829f2ce40b274@mail.gmail.com> Sender: netdev-owner@vger.kernel.org List-ID: On Sat, 20 Jun 2009 17:01:25 +0900 Mike McCormack wrote: > Clear interrupt only when the status buffer is fully drained, > Make sure to clear interrupt when work_done == work_limit > and the buffer is drained. > --- > drivers/net/sky2.c | 12 ++++++++---- > 1 files changed, 8 insertions(+), 4 deletions(-) > > diff --git a/drivers/net/sky2.c b/drivers/net/sky2.c > index 7681d28..ca1e9e5 100644 > --- a/drivers/net/sky2.c > +++ b/drivers/net/sky2.c > @@ -2524,9 +2524,6 @@ static int sky2_status_intr(struct sky2_hw *hw, > int to_do, u16 idx) > } > } while (hw->st_idx != idx); > > - /* Fully processed status ring so clear irq */ > - sky2_write32(hw, STAT_CTRL, SC_STAT_CLR_IRQ); > - > exit_loop: > sky2_rx_done(hw, 0, total_packets[0], total_bytes[0]); > sky2_rx_done(hw, 1, total_packets[1], total_bytes[1]); > @@ -2779,9 +2776,16 @@ static int sky2_poll(struct napi_struct *napi, > int work_limit) > if (status & Y2_IS_IRQ_PHY2) > sky2_phy_intr(hw, 1); > > - while ((idx = sky2_read16(hw, STAT_PUT_IDX)) != hw->st_idx) { > + idx = sky2_read16(hw, STAT_PUT_IDX); > + while (idx != hw->st_idx) { > work_done += sky2_status_intr(hw, work_limit - work_done, idx); > > + /* If we fully processed the status ring, clear the irq */ > + idx = sky2_read16(hw, STAT_PUT_IDX); > + if (idx == hw->st_idx) { > + sky2_write32(hw, STAT_CTRL, SC_STAT_CLR_IRQ); > + break; > + } > if (work_done >= work_limit) > goto done; > } Have you actually seen this race, or are you hypothesizing based on code review? I think the original works fine. There is a race where interrupt is cleared early, and the poll processing runs an extra time but that is harmless. But the patched code races the other way: > + while (idx != hw->st_idx) { > work_done += sky2_status_intr(hw, work_limit - work_done, idx); > > + /* If we fully processed the status ring, clear the irq */ > + idx = sky2_read16(hw, STAT_PUT_IDX); Packet arrives right here. The variable "idx" has old value, but chip register "put_idx" shows new packet. > + if (idx == hw->st_idx) { > + sky2_write32(hw, STAT_CTRL, SC_STAT_CLR_IRQ); This clears level triggered status interrupt. > + break; > + } > if (work_done >= work_limit) > goto done; > } Now the driver misses the status interrupt. --