[PATCH] NVMe: Call nvme_pci_disable on error path of nvme_probe_work

[Rashika Kheria]

Commit d5537e988eec ("NVMe: Don't unmap controller registers on reset"),
introduced a regression in which it did not replace nvme_dev_unmap()
with nvme_pci_disable() in the error path of nvme_probe_work().

This led to the following NVMe driver crash on systems where the devices
did not initialise in the first try.

BUG: unable to handle kernel paging request at ffffc90006da001c
IP: [<ffffffffa027b6bb>] nvme_dev_remove+0x5b/0xf0 [nvme]
RIP: e030:[<ffffffffa027b6bb>] [<ffffffffa027b6bb>]
nvme_dev_remove+0x5b/0xf0 [nvme]
RSP: e02b:ffff8806659c3cb8 EFLAGS: 00010286
RAX: ffffc90006da0000 RBX: ffff88067cbc3000 RCX: 0000000000000006
RDX: 0000000000000007 RSI: 0000000000000007 RDI: ffff8806864eda40
RBP: ffff8806659c3cd8 R08: 0000000000000006 R09: 000000000000fffe
R10: 0000000000000000 R11: 0000000000000000 R12: ffff88067e087000
R13: ffffffffa0281d20 R14: ffff88067e087098 R15: ffff8806799d8598
FS: 00007f880d5ba700(0000) GS:ffff8806864e0000(0000)
knlGS:0000000000000000
CS: e033 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: ffffc90006da001c CR3: 0000000676a97000 CR4: 0000000000042660
Call Trace:
[<ffffffffa027b7ea>] nvme_remove+0x9a/0x140 [nvme]
[<ffffffff813503ef>] pci_device_remove+0x3f/0xc0
[<ffffffff81449869>] ? __pm_runtime_idle+0x89/0x90
[<ffffffff8143ed4f>] __device_release_driver+0xaf/0x140
[<ffffffff8143eec8>] device_release_driver+0x28/0x40
[<ffffffff8143db66>] unbind_store+0x96/0xb0
[<ffffffff8143d027>] drv_attr_store+0x27/0x30
[<ffffffff8122e279>] sysfs_kf_write+0x39/0x40
[<ffffffff8122d9e4>] kernfs_fop_write+0xe4/0x160
[<ffffffff811b15df>] __vfs_write+0x2f/0x100
[<ffffffff81003640>] ? syscall_slow_exit_work+0x140/0x180
[<ffffffff81161db9>] ? vm_mmap_pgoff+0xb9/0xe0
[<ffffffff810af981>] ? percpu_down_read+0x11/0x60
[<ffffffff811b2bce>] vfs_write+0xbe/0x190
[<ffffffff811b2d81>] SyS_write+0x51/0xb0
[<ffffffff815b8aee>] entry_SYSCALL_64_fastpath+0x12/0x71

Cc: sta...@vger.kernel.org # 4.4.y
Cc: Jens Axboe <ax...@fb.com>
Cc: Keith Busch <keith...@intel.com>
Cc: Gabriel Krisman Bertazi <kri...@linux.vnet.ibm.com>
Cc: linux...@lists.infradead.org
Fixes: d5537e988eec ("NVMe: Don't unmap controller registers on reset")
Signed-off-by: Rashika Kheria <ras...@amazon.de>
---
drivers/nvme/host/pci.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
index c851bc5..f5d1579 100644
--- a/drivers/nvme/host/pci.c
+++ b/drivers/nvme/host/pci.c
@@ -3184,7 +3184,7 @@ static void nvme_probe_work(struct work_struct *work)
nvme_disable_queue(dev, 0);
nvme_dev_list_remove(dev);
unmap:
- nvme_dev_unmap(dev);
+ nvme_pci_disable(dev);
out:
if (!work_busy(&dev->reset_work))
nvme_dead_ctrl(dev);
--
2.10.2

[Kheria, Rashika]

________________________________________
From: Keith Busch <keith...@intel.com>
Sent: Monday, November 14, 2016 7:47 PM
To: Kheria, Rashika
Cc: Christoph Hellwig; Kheria, Rashika; linux...@lists.infradead.org
Subject: Re: [PATCH] NVMe: Call nvme_pci_disable on error path of nvme_probe_work

On Mon, Nov 14, 2016 at 09:57:27AM +0100, Rashika Kheria wrote:
> Hi everyone,
>
> Could you please review the following patch? This solves a regression in
> stable 4.4.y tree.

I missed the "Don't unmap" back-port to 4.4.y. I'm not sure, but I think
we may have addressed that differently with something less risky if we
needed that behaviour on 4.4-stable. That's okay, though, this new patch
looks correct. The original was part of a series that fixes this in its
following commit, but it should have looked like this from the beginning.

Acked-by: Keith Busch <keith...@intel.com>

> > ---end quoted text---

Amazon Development Center Germany GmbH
Berlin - Dresden - Aachen
main office: Krausenstr. 38, 10117 Berlin
Geschaeftsfuehrer: Dr. Ralf Herbrich, Christian Schlaeger
Ust-ID: DE289237879
Eingetragen am Amtsgericht Charlottenburg HRB 149173 B