[PATCH] nvme: fix NULL pointer dereference

[Keith Busch]

On Thu, Sep 17, 2020 at 11:32:12PM -0400, Tong Zhang wrote:
> Please correct me if I am wrong.
> After a bit more digging I found out that it is indeed command_id got
> corrupted is causing this problem. Although the tag and command_id
> range is checked like you said, the elements in rqs cannot be
> guaranteed to be not NULL. thus although the range check is passed,
> blk_mq_tag_to_rq() can still return NULL. 

I think your describing a sequence problem in initialization. We
shouldn't have interrupts wired up to uninitialized tagsets.

A more appropriate sequence would setup request_irq() after the tagset
is ready. It makes handling a failed irq setup a bit weird for io
queues, though.

> It is clear that the current sanitization is not enough and there's
> more implication about this -- when all rqs got populated, a corrupted
> command_id may silently corrupt other data not belonging to the
> current command.

The block layer doesn't do anything with requests that haven't been
started, so if your controller completes non-existent commands, then
nothing particular will happen with the rqs.

If the request had been started and the controller provides a corrupted
completion, then the fault lies entirely with the controller and you
should raise that issue with your vendor. There's no way the driver can
distinguish a genuine completion from a corrupted one.