Date:   Sun, 11 Feb 2018 17:38:31 +0800
From:   Jianchao Wang <jianchao.w.wang@...cle.com>
To:     keith.busch@...el.com, axboe@...com, hch@....de, sagi@...mberg.me
Cc:     linux-nvme@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: [PATCH V3 0/6]nvme-pci: fixes on nvme_timeout and nvme_dev_disable

Hi Christoph, Keith and Sagi

Please consider and comment on the following patchset.
That's really appreciated.

There is a complicated relationship between nvme_timeout and nvme_dev_disable.
 - nvme_timeout has to invoke nvme_dev_disable to stop the
   controller doing DMA access before free the request.
 - nvme_dev_disable has to depend on nvme_timeout to complete
   adminq requests to set HMB or delete sq/cq when the controller
   has no response.
 - nvme_dev_disable will race with nvme_timeout when cancels the
   outstanding requests.
We have found some issues introduced by them, please refer the following link

http://lists.infradead.org/pipermail/linux-nvme/2018-January/015053.html 
http://lists.infradead.org/pipermail/linux-nvme/2018-January/015276.html
http://lists.infradead.org/pipermail/linux-nvme/2018-January/015328.html
Even we cannot ensure there is no other issue.

The best way to fix them is to break up the relationship between them.
With this patch, we could avoid nvme_dev_disable to be invoked
by nvme_timeout and eliminate the race between nvme_timeout and
nvme_dev_disable on outstanding requests.

Change V2->V3:
 - Keep queue frozen for reset case. If IO requests timeout during in RECONNECTING
   state, fail them and kill the controller. Really appreciate Keith's directive and advice.
 - Add 3 patches to fix some bug about namespaces_mutex and change it to rwsem. It could fix
   the deadlock risk introduced by namespace_mutex following.
 - other misc changes.

Changes V1->V2:
 - free and disable pci things in nvme_pci_disable_ctrl_directly
 - change comment and add reviewed-by in 1st patch
 - resort patches
 - other misc changes

There are 9 patches:

1st ~ 3th patches is to change the namespaces_mutex to rw_semphore lock.
4th, 6th, 7th is to do some preparations for the 8th patch.
5th fixes a bug found when test.
8th is to avoid nvme_dev_disable to be invoked by nvme_timeout, and implement
the synchronization between them. More details, please refer to the comment of
this patch.
9th fixes a bug after 8th patch is introduced. It let nvme_delete_io_queues can
only be wakeup by completion path.

This patchset was tested under debug patch for some days.
And some bugfix have been done.
The patches are available in following it branch:
https://github.com/jianchwa/linux-blcok.git nvme_fixes_V3_plus_rwsem

Jianchao Wang (9)
0001-nvme-fix-the-dangerous-reference-of-namespaces-list.patch
0002-nvme-fix-the-deadlock-in-nvme_update_formats.patch
0003-nvme-change-namespaces_mutext-to-namespaces_rwsem.patch
0004-nvme-pci-quiesce-IO-queues-prior-to-disabling-device.patch
0005-nvme-pci-suspend-queues-based-on-online_queues.patch
0006-nvme-pci-drain-the-entered-requests-after-ctrl-is-sh.patch
0007-blk-mq-make-blk_mq_rq_update_aborted_gstate-a-extern.patch
0008-nvme-pci-break-up-nvme_timeout-and-nvme_dev_disable.patch
0009-nvme-pci-discard-wait-timeout-when-delete-cq-sq.patch

diff stat following:
block/blk-mq.c                |   3 +-
drivers/nvme/host/core.c      |  88 +++++++++------
drivers/nvme/host/multipath.c |   4 +-
drivers/nvme/host/nvme.h      |   2 +-
drivers/nvme/host/pci.c       | 255 +++++++++++++++++++++++++++++++-----------
include/linux/blk-mq.h        |   1 +
6 files changed, 252 insertions(+), 101 deletions(-)

Thanks
Jianchao

Powered by blists - more mailing lists