[PATCH] nvme-pci: cancel nvme device request before disabling

[Tong Zhang]

Hi Keith,

On Fri, Aug 14, 2020 at 11:04 AM Keith Busch <kbusch at kernel.org> wrote:
>
> On Fri, Aug 14, 2020 at 03:14:31AM -0400, Tong Zhang wrote:
> > diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> > index ba725ae47305..c4f1ce0ee1e3 100644
> > --- a/drivers/nvme/host/pci.c
> > +++ b/drivers/nvme/host/pci.c
> > @@ -1249,8 +1249,8 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
> >               dev_warn_ratelimited(dev->ctrl.device,
> >                        "I/O %d QID %d timeout, disable controller\n",
> >                        req->tag, nvmeq->qid);
> > -             nvme_dev_disable(dev, true);
> >               nvme_req(req)->flags |= NVME_REQ_CANCELLED;
> > +             nvme_dev_disable(dev, true);
> >               return BLK_EH_DONE;
>
> Shouldn't this flag have been set in nvme_cancel_request()?

nvme_cancel_request() is not setting this flag to cancelled and this is causing
__nvme_submit_sync_cmd() to return a positive number and eventually
this will allow nvme_init_identify() proceed with a timeout error where the
problematic nvme device has already been disabled by nvme_dev_disable(dev,true).
The reset of the logic in nvme_init_identify() is not going to check
those already
released resources. -- in fact this is UAF --

> like the timeout out command is the only command to have been cancelled
> by this action, nor is it guaranteed that getting here will mean the
> request was in fact cancelled. The controller could still provide a real
> completion.

what would you like to suggest here?