Re: [RFC]extension of the anchor API

[Oliver Neukum <oneukum@suse.com>]

Am Mittwoch, den 14.04.2021, 10:56 -0400 schrieb Alan Stern:
> On Wed, Apr 14, 2021 at 10:12:01AM +0200, Oliver Neukum wrote:
> > Am Montag, den 12.04.2021, 11:06 -0400 schrieb Alan Stern:
> > > On Mon, Apr 12, 2021 at 11:58:16AM +0200, Oliver Neukum wrote:
> > > > That presumes that the URBs will finish in order. I don't think such
> > > > an assumption can be made.
> > > 
> > > I don't understand -- I can't detect any such presumption.
> > 
> > OK, this shows that I am bad at explaining.
> > > As far as I can tell, the only reason for maintaining the URBs in any 
> > > particular order on the anchor list is so that usb_kill_anchored_urbs 
> > > and usb_poison_anchored_urbs can kill them in reverse order of 
> > > submission.  THat's why the current code moves completed URBs to the end 
> > > of the list.
> > 
> > No longer strictly true, as the API has a call to submit everything
> > on an anchor, but I think it boils down to the same thing.
> > 
> > > If you keep a pointer to the most recently submitted URB, killing them 
> > > easy enough to do.  Start with that URB, then go backward through the 
> > > list (wrapping to the end when you reach the beginning of the list).
> > 
> > Yes, but that supposes that the next on the list has not been
> > resubmitted _before_ the one after it.
> > 
> > If you do not keep the list ordered, but in the initial order,
> > we can have the situation that A (happens most recently submitted)
> > is followed by B and C, but C was submitted before B.
> 
> I think the only reasonable alternative is to move an URB to the end of 
> the list when it is submitted, rather than when it completes.  Have you 
> considered doing it that way?

No, that did not occur to me. Back to the drawing board.
Still I have to put it somewhere when I anchor an URB. Head or tail?

> The real problem with usb_submit_anchored_urbs is that the core can't 
> know in what order the caller wants the URBs to be submitted.  If the 

I think the reasonable assumption is that they need to be submitted
in the order they were anchored.

> In the kerneldoc you can explain that if the anchor has not been used 
> since its URBs were added then the URBs will be submitted in the order 
> they were added to the anchor, but otherwise they will be submitted in 
> an unspecified order, which may not be suitable.

Yes.

> > > The order in which the URBs complete doesn't matter, because trying to 
> > > unlink a completed URB won't cause any harm.
> > 
> > As long as it stays completed.
> 
> Rather, as long as they complete in order of submission.
> 
> > >   The only assumption here 
> > > is that URBs get submitted in the list's order (possibly circularly) -- 
> > > this should always be true.
> > 
> > I am afraid we cannot guarantee that. It might intuitively seem so,
> > but nothing guarantees that all URBs are going to the same endpoint.
> 
> I hadn't thought of that.  Do anchors get used that way anywhere?

I haven't found an example, but I thought it could not be ruled out.
So you think that that case should be discouraged in documentation
and henceforth ignored?

So we do agree that we need the following:

a - submit in the order you
anchored
b - kill or poison in the reverse order
c - unpoison does not really matter but better do it in the submit
order?

Does that mean that the list needs to be kept ordered by sequence
of submission? I think so.

	Regards
		Oliver