On Mon, May 08, 2017 at 08:46:39PM +0800, Ming Lei wrote:

On Mon, May 08, 2017 at 07:24:57PM +0800, Ming Lei wrote:

quoted

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index c8541c3dcd19..ebe13e157c00 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c

@@ -2185,8 +2185,8 @@ static void nvme_remove(struct pci_dev *pdev)
 	}
 
 	flush_work(&dev->reset_work);
-	nvme_uninit_ctrl(&dev->ctrl);
 	nvme_dev_disable(dev, true);
+	nvme_uninit_ctrl(&dev->ctrl);
 	nvme_dev_remove_admin(dev);
 	nvme_free_queues(dev, 0);
 	nvme_release_cmb(dev);

This patch should be wrong, and looks the correct fix should be
flushing 'dev->remove_work' before calling nvme_uninit_ctrl().

Yeah, disabling the device before calling "nvme_uninit_ctrl" shouldn't
be required. If you disable the device first, del_gendisk can't flush
dirty data on an orderly removal request.

But it might cause deadloack by calling flush_work(&dev->remove_work)
here simply.

I'm almost certain the remove_work shouldn't even be running in this
case. If the reset work can't transition the controller state correctly,
it should assume something is handling the controller.

---

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index 26a5fd0..d81104d 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c

@@ -1792,7 +1797,7 @@ static void nvme_reset_work(struct work_struct *work)
 		nvme_dev_disable(dev, false);
 
 	if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
-		goto out;
+		return;
 
 	result = nvme_pci_enable(dev);
 	if (result)
--