[lkml]   [2021]   [May]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] pci: add NVMe FLR quirk to the SM951 SSD
On Fri Apr 30, 2021 at 3:51 PM CDT, Bjorn Helgaas wrote:

> Please make your subject line match ffb0863426eb ("PCI: Disable
> Samsung SM961/PM961 NVMe before FLR")

I had done this in a V2 of this patch, but after some additional
research I'm thinking the behavior of this quirk might be in-line
w/ the NVMe specification more generally, I'll elaborate more below.

> I don't see anything in the PCIe spec about software being required to
> do something special before initiating an FLR, so I assume this is a
> hardware defect in the Samsung 950 PRO? Has Samsung published an
> erratum or at least acknowledged it?
> There's always the possibility that we are doing something wrong in
> Linux *after* the FLR, e.g., not waiting long enough, not
> reinitializing something correctly, etc.

I did some dumping of registers both with and without this patch, and
determined the following to be true in my use-case:

1. My guest VM leaves the device in a state where SHN (shutdown
notification) is set to 0b01 (normal shutdown)

2. The guest also leaves CC.EN (controller enable) set to 0b1

3. vfio-pci attempts to issue an FLR while the device is in this state.

On page 40, sec 3.1.6 of the NVMe 1.1 spec, the documentation on SHST
states the following:

> To start executing commands on the controller after a shutdown
> operation (CSTS.SHST set to 10b), a reset (CC.EN cleared to ‘0’)
> is required. If host software submits commands to the controller
> without issuing a reset, the behavior is undefined.

In the case of the SM951/SM961 it appears the undefined behavior is that
they stop responding to attempts to change their configuration if you do
an FLR while the device is in this state. The reason this patch
resolved the issue I was seeing is because the toggle of the CC.EN flag
puts the drive in a known-good state after the guest's shutdown

Knowing this I would suspect we'd actually want to treat most NVMe
drives in this manner *if the kernel sees the SHN/SHST has been set
prior.* Perhaps other NVMe devices are more tolerant of not doing this?

~ robert straw

 \ /
  Last update: 2021-05-15 19:35    [W:0.128 / U:1.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site