On 20-06-16 15:54, Keith Busch wrote:

On Mon, Jun 20, 2016@11:21:09AM -0700, Jethro Beekman wrote:

quoted

On 20-06-16 08:26, Keith Busch wrote:

Would this just be a matter of setting req->retries and checking for it in
nvme_req_needs_retry? How does one keep track of the number of tries so far?

I just sent a patch out earlier today to use req->retries to track the
retry count, and nvme module parameter to set the max retries. I think
that would fix the long delays you're seeing, assuming the patch is okay.

Your patch "nvme: Limit command retries" works for me and obviates the need for
this patch.

quoted

You're right, I assumed that admin commands can't have namespace ids, but
looking at the spec, that's not the case. Turns out there's a problem with the
driver then: nvme_ioctl never includes the ns for NVME_IOCTL_ADMIN_CMD.

The NVME_IOCTL_ADMIN_CMD already takes any namespace identifier the user
put in that field.

I see, the ns argument is just to specify the queue. I assume userspace is
supposed to obtain the ns using NVME_IOCTL_ID? This seems broken, if I have an
open block device handle I can send commands to any nvme namespace as well as
the controller? I think on the block devices you should only be able to send
commands with your nsid. There was some discussion on the security implications
of this about a year ago [1], and it was decided to fix this, but it doesn't
look like this was actually merged?

[1] http://lists.infradead.org/pipermail/linux-nvme/2015-January/001446.html

Jethro