[PATCH] nvme-pci: fix the timeout case when reset is ongoing

[PATCH] nvme-pci: fix the timeout case when reset is ongoing

[Jianchao Wang]

NVME_CTRL_RESETTING used to indicate the range of nvme initializing
strictly in fd634f41(nvme: merge probe_work and reset_work), but it
is not now. The NVME_CTRL_RESETTING is set before queue the
reset_work, there could be a big gap before the reset work handles
the outstanding requests. So when the NVME_CTRL_RESETTING is set,
nvme_timeout will not only meet the admin requests from the
initializing procedure, but also the IO and admin requests from
previous work before nvme_dev_disable is invoked.

To fix it, introduce a flag NVME_DEV_FLAG_INITIALIZING to mark the
range of initializing. When this flag is not set, handle the expried
requests as nvme_cancel_request. Otherwise, the requests should be
from the initializing procedure. Handle them as before. Because the
nvme_reset_work will see the error and disable the dev itself, so
discard the nvme_dev_disable here.

Signed-off-by: Jianchao Wang <jianchao.w.wang at oracle.com>
---
 drivers/nvme/host/pci.c | 26 +++++++++++++++++---------
 1 file changed, 17 insertions(+), 9 deletions(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index f5800c3..6e58de1 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -99,6 +99,9 @@ struct nvme_dev {
 	struct nvme_ctrl ctrl;
 	struct completion ioq_wait;
 
+	unsigned long flag;
+#define NVME_DEV_FLAG_INITIALIZING 0
+
 	/* shadow doorbell buffer support: */
 	u32 *dbbuf_dbs;
 	dma_addr_t dbbuf_dbs_dma_addr;
@@ -1207,17 +1210,19 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
 	}
 
 	/*
-	 * Shutdown immediately if controller times out while starting. The
-	 * reset work will see the pci device disabled when it gets the forced
-	 * cancellation error. All outstanding requests are completed on
-	 * shutdown, so we return BLK_EH_HANDLED.
+	 * There could be two kinds of expired reqs when reset is ongoing.
+	 * Outstanding IO or admin requests from previous work before the
+	 * nvme_reset_work invokes nvme_dev_disable. Handle them as the
+	 * nvme_cancel_request. Outstanding admin requests from the
+	 * initializing procedure. Set NVME_REQ_CANCELLED flag on them,
+	 * then nvme_reset_work will see the error, then disable the device
+	 * and remove the ctrl.
 	 */
 	if (dev->ctrl.state == NVME_CTRL_RESETTING) {
-		dev_warn(dev->ctrl.device,
-			 "I/O %d QID %d timeout, disable controller\n",
-			 req->tag, nvmeq->qid);
-		nvme_dev_disable(dev, false);
-		nvme_req(req)->flags |= NVME_REQ_CANCELLED;
+		if (test_bit(NVME_DEV_FLAG_INITIALIZING, &dev->flag))
+			nvme_req(req)->flags |= NVME_REQ_CANCELLED;
+		else
+			nvme_req(req)->status = NVME_SC_ABORT_REQ;
 		return BLK_EH_HANDLED;
 	}
 
@@ -2302,6 +2307,7 @@ static void nvme_reset_work(struct work_struct *work)
 	if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
 		nvme_dev_disable(dev, false);
 
+	set_bit(NVME_DEV_FLAG_INITIALIZING, &dev->flag);
 	result = nvme_pci_enable(dev);
 	if (result)
 		goto out;
@@ -2366,10 +2372,12 @@ static void nvme_reset_work(struct work_struct *work)
 		goto out;
 	}
 
+	clear_bit(NVME_DEV_FLAG_INITIALIZING, &dev->flag);
 	nvme_start_ctrl(&dev->ctrl);
 	return;
 
  out:
+	clear_bit(NVME_DEV_FLAG_INITIALIZING, &dev->flag);
 	nvme_remove_dead_ctrl(dev, result);
 }
 
-- 
2.7.4

[PATCH] nvme-pci: fix the timeout case when reset is ongoing

[Christoph Hellwig]

On Wed, Jan 03, 2018@06:31:44AM +0800, Jianchao Wang wrote:
> NVME_CTRL_RESETTING used to indicate the range of nvme initializing
> strictly in fd634f41(nvme: merge probe_work and reset_work), but it
> is not now. The NVME_CTRL_RESETTING is set before queue the
> reset_work, there could be a big gap before the reset work handles
> the outstanding requests. So when the NVME_CTRL_RESETTING is set,
> nvme_timeout will not only meet the admin requests from the
> initializing procedure, but also the IO and admin requests from
> previous work before nvme_dev_disable is invoked.
> 
> To fix it, introduce a flag NVME_DEV_FLAG_INITIALIZING to mark the
> range of initializing. When this flag is not set, handle the expried
> requests as nvme_cancel_request. Otherwise, the requests should be
> from the initializing procedure. Handle them as before. Because the
> nvme_reset_work will see the error and disable the dev itself, so
> discard the nvme_dev_disable here.

Instead of a parallel set of states we'll need to split
NVME_CTRL_RESET into NVME_CTRL_RESET_SCHEDULED and NVME_CTRL_RESETTING.

And if my memory doesn't fail me we were already considering that a while
ago.

[PATCH] nvme-pci: fix the timeout case when reset is ongoing

[jianchao.wang]

Hi Christoph

Many thanks for your kindly response.

On 01/04/2018 06:35 PM, Christoph Hellwig wrote:
> On Wed, Jan 03, 2018@06:31:44AM +0800, Jianchao Wang wrote:
>> NVME_CTRL_RESETTING used to indicate the range of nvme initializing
>> strictly in fd634f41(nvme: merge probe_work and reset_work), but it
>> is not now. The NVME_CTRL_RESETTING is set before queue the
>> reset_work, there could be a big gap before the reset work handles
>> the outstanding requests. So when the NVME_CTRL_RESETTING is set,
>> nvme_timeout will not only meet the admin requests from the
>> initializing procedure, but also the IO and admin requests from
>> previous work before nvme_dev_disable is invoked.
>>
>> To fix it, introduce a flag NVME_DEV_FLAG_INITIALIZING to mark the
>> range of initializing. When this flag is not set, handle the expried
>> requests as nvme_cancel_request. Otherwise, the requests should be
>> from the initializing procedure. Handle them as before. Because the
>> nvme_reset_work will see the error and disable the dev itself, so
>> discard the nvme_dev_disable here.
> 
> Instead of a parallel set of states we'll need to split
> NVME_CTRL_RESET into NVME_CTRL_RESET_SCHEDULED and NVME_CTRL_RESETTING.
> 
> And if my memory doesn't fail me we were already considering that a while
> ago.
> 
Yes, it is indeed more reasonable to split current NVME_CTRL_RESETTING into 
two states, but the nvme_dev_disable() in nvme_reset_work() should be the boundary.
After that, all the in-flight requests are requeued and request queue is quiesced,
the nvme driver is clear. So the new state maybe something like NEW_CTRL_RESET_PREPARE.:)

Thanks
Jianchao