Re: [PATCH 2/2] Revert "mq-deadline: Fix request accounting"

[Bart Van Assche]

On 9/7/21 9:28 AM, Niklas Cassel wrote:
> On Tue, Sep 07, 2021 at 08:15:03AM -0700, Bart Van Assche wrote:
> > On 9/7/21 7:21 AM, Niklas Cassel wrote:
> > > blk-mq will no longer call the I/O scheduler .finish_request() callback
> > > for requests that were never inserted to the I/O scheduler.
> >
> > I do not agree. Even with patch 1/2 from this series applied, finish_request()
> > will still be called for requests inserted by blk_insert_cloned_request()
> > although these requests are never inserted to the I/O scheduler.
>
> Looking at blk_mq_free_request(),
> e->type->ops.finish_request() will only be called if RQF_ELVPRIV
> is set.
>
> blk_insert_cloned_request() doesn't seem to allocate a request
> itself, but instead takes an already cloned request.
>
> So I guess it depends on how the supplied request was cloned.
>
> I would assume if the original request doesn't have RQF_ELVPRIV set,
> then neither will the cloned request?
>
> I tried to look at blk_rq_prep_clone(), which seems to be a common
> cloning function, but I don't see req->rq_flags being copied
> (except for RQF_SPECIAL_PAYLOAD).
>
> Anyway, I don't see how .finish_request() will be called in relation
> to blk_insert_cloned_request(). Could you please help me out and
> give me an example of a call chain where this can happen?

Hi Niklas,

This is a bit outside my area of expertise. Anyway: map_request() calls
.clone_and_map_rq(). At least multipath_clone_and_map() calls
blk_get_request(). I think this shows that blk_insert_cloned_request()
may insert an entirely new request. Is my understanding correct that
blk_mq_rq_ctx_init() will set RQF_ELVPRIV for the cloned request if a
scheduler is associated with the request queue associated with the
cloned request?

Bart.