From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752305AbcLUG5k (ORCPT ); Wed, 21 Dec 2016 01:57:40 -0500 Received: from mga01.intel.com ([192.55.52.88]:25045 "EHLO mga01.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750754AbcLUG5h (ORCPT ); Wed, 21 Dec 2016 01:57:37 -0500 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.33,382,1477983600"; d="scan'208";a="800753765" Subject: Re: [PATCH 2/2] usb: host: xhci: Handle the right timeout command To: Mathias Nyman , Baolin Wang References: <0c39bfed1cf6f7b747e702aa841f82c9d2140f27.1480922249.git.baolin.wang@linaro.org> <584EC7B9.6040100@linux.intel.com> <5857B794.2070100@linux.intel.com> <5857CEE4.6040007@intel.com> <5858B3AE.90705@linux.intel.com> <5858D231.7040407@linux.intel.com> <5858DB52.8060707@linux.intel.com> <58594A9A.10507@linux.intel.com> Cc: Mathias Nyman , Greg KH , USB , LKML , Mark Brown , "Lu, Baolu" From: Lu Baolu Message-ID: <585A27DF.4030003@linux.intel.com> Date: Wed, 21 Dec 2016 14:57:35 +0800 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.5.1 MIME-Version: 1.0 In-Reply-To: <58594A9A.10507@linux.intel.com> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi Mathias, I have some comments for the implementation of xhci_handle_command_timeout() as well. On 12/20/2016 11:13 PM, Mathias Nyman wrote: > On 20.12.2016 09:30, Baolin Wang wrote: > ... > > Alright, I gathered all current work related to xhci races and timeouts > and put them into a branch: > > git://git.kernel.org/pub/scm/linux/kernel/git/mnyman/xhci.git timeout_race_fixes > > Its based on 4.9 > It includes a few other patches just to avoid conflicts and make my life easier > > Interesting patches are: > > ee4eb91 xhci: remove unnecessary check for pending timer > 0cba67d xhci: detect stop endpoint race using pending timer instead of counter. > 4f2535f xhci: Handle command completion and timeout race > b9d00d7 usb: host: xhci: Fix possible wild pointer when handling abort command > 529a5a0 usb: xhci: fix possible wild pointer > 4766555 xhci: Fix race related to abort operation > de834a3 xhci: Use delayed_work instead of timer for command timeout > 69973b8 Linux 4.9 > > The fixes for command queue races will go to usb-linus and stable, the > reworks for stop ep watchdog timer will go to usb-next. > > Still completely untested, (well it compiles) > > Felipe gave instructions how to modify dwc3 driver to timeout on address > devicecommands to test these, I'll try to set that up. > > All additional testing is welcome, especially if you can trigger timeouts > and races > > -Mathias > > I post the code below and add my comments in line. 1276 void xhci_handle_command_timeout(struct work_struct *work) 1277 { 1278 struct xhci_hcd *xhci; 1279 int ret; 1280 unsigned long flags; 1281 u64 hw_ring_state; 1282 1283 xhci = container_of(to_delayed_work(work), struct xhci_hcd, cmd_timer); 1284 1285 spin_lock_irqsave(&xhci->lock, flags); 1286 1287 /* 1288 * If timeout work is pending, or current_cmd is NULL, it means we 1289 * raced with command completion. Command is handled so just return. 1290 */ 1291 if (!xhci->current_cmd || delayed_work_pending(&xhci->cmd_timer)) { 1292 spin_unlock_irqrestore(&xhci->lock, flags); 1293 return; 1294 } 1295 /* mark this command to be cancelled */ 1296 xhci->current_cmd->status = COMP_CMD_ABORT; 1297 1298 /* Make sure command ring is running before aborting it */ 1299 hw_ring_state = xhci_read_64(xhci, &xhci->op_regs->cmd_ring); 1300 if ((xhci->cmd_ring_state & CMD_RING_STATE_RUNNING) && 1301 (hw_ring_state & CMD_RING_RUNNING)) { 1302 /* Prevent new doorbell, and start command abort */ 1303 xhci->cmd_ring_state = CMD_RING_STATE_ABORTED; 1304 spin_unlock_irqrestore(&xhci->lock, flags); 1305 xhci_dbg(xhci, "Command timeout\n"); 1306 ret = xhci_abort_cmd_ring(xhci); 1307 if (unlikely(ret == -ESHUTDOWN)) { 1308 xhci_err(xhci, "Abort command ring failed\n"); 1309 xhci_cleanup_command_queue(xhci); 1310 usb_hc_died(xhci_to_hcd(xhci)->primary_hcd); 1311 xhci_dbg(xhci, "xHCI host controller is dead.\n"); 1312 } 1313 return; 1314 } 1315 1316 /* host removed. Bail out */ 1317 if (xhci->xhc_state & XHCI_STATE_REMOVING) { 1318 spin_unlock_irqrestore(&xhci->lock, flags); 1319 xhci_dbg(xhci, "host removed, ring start fail?\n"); 1320 xhci_cleanup_command_queue(xhci); 1321 return; 1322 } I think this part of code should be moved up to line 1295. 1323 1324 /* command timeout on stopped ring, ring can't be aborted */ 1325 xhci_dbg(xhci, "Command timeout on stopped ring\n"); 1326 xhci_handle_stopped_cmd_ring(xhci, xhci->current_cmd); 1327 spin_unlock_irqrestore(&xhci->lock, flags); This part of code is tricky. I have no idea about in which case should this code be executed? Anyway, we shouldn't call xhci_handle_stopped_cmd_ring() here, right? 1328 return; 1329 } Best regards, Lu Baolu