Re: [libgpiod] [PATCH 11/19] API: add support for SET_CONFIG

[Kent Gibson <warthog618@xxxxxxxxx>]

On Thu, Nov 21, 2019 at 11:27:10AM +0100, Bartosz Golaszewski wrote:
> czw., 21 lis 2019 o 11:18 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> >
> > On Thu, Nov 21, 2019 at 11:03:32AM +0100, Bartosz Golaszewski wrote:
> > > czw., 21 lis 2019 o 10:30 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > >
> > > > On Thu, Nov 21, 2019 at 09:46:07AM +0100, Bartosz Golaszewski wrote:
> > > > > czw., 21 lis 2019 o 08:46 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > >
> > > > > > On Thu, Nov 21, 2019 at 08:13:42AM +0100, Bartosz Golaszewski wrote:
> > > > > > > czw., 21 lis 2019 o 01:34 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > > > >
> > > > > > > > On Wed, Nov 20, 2019 at 04:18:24PM +0100, Bartosz Golaszewski wrote:
> > > > > > > > > śr., 20 lis 2019 o 15:36 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > > > > > >
> > > > > > > > > > On Wed, Nov 20, 2019 at 03:18:36PM +0100, Bartosz Golaszewski wrote:
> > > > > > > > > > > śr., 20 lis 2019 o 15:13 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > > > > > > > >
> > > > > > > > > > > > On Wed, Nov 20, 2019 at 03:08:57PM +0100, Bartosz Golaszewski wrote:
> > > > > > > > > > > > > śr., 20 lis 2019 o 14:59 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > On Wed, Nov 20, 2019 at 12:00:45PM +0100, Bartosz Golaszewski wrote:
> > > > > > > > > > > > > > > wt., 19 lis 2019 o 16:53 Kent Gibson <warthog618@xxxxxxxxx> napisał(a):
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > On Mon, Nov 18, 2019 at 10:48:25PM +0800, Kent Gibson wrote:
> > > > > > > > > > > > > > > > > On Mon, Nov 18, 2019 at 02:52:04PM +0100, Bartosz Golaszewski
> > > > > > > > > > > > > > > > > > > +
> > > > > > > > > > > > > > > > > > > +int gpiod_line_set_flags_bulk(struct gpiod_line_bulk *bulk, int flags)
> > > > > > > > > > > > > > > > > > > +{
> > > > > > > > > > > > > > > > > > > +       struct gpiod_line *line;
> > > > > > > > > > > > > > > > > > > +       int values[GPIOD_LINE_BULK_MAX_LINES];
> > > > > > > > > > > > > > > > > > > +       unsigned int i;
> > > > > > > > > > > > > > > > > > > +       int direction;
> > > > > > > > > > > > > > > > > > > +
> > > > > > > > > > > > > > > > > > > +       line = gpiod_line_bulk_get_line(bulk, 0);
> > > > > > > > > > > > > > > > > > > +       if (line->as_is) {
> > > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > > Can you explain the purpose of this as_is field? I'm not sure this is
> > > > > > > > > > > > > > > > > > really needed.
> > > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > > It is there for gpiod_set_flags, which has to populate the direction
> > > > > > > > > > > > > > > > > flags in the SET_CONFIG ioctl. The existing line->direction is
> > > > > > > > > > > > > > > > > either input or output.  It is drawn from GPIOLINE_FLAG_IS_OUT, so
> > > > > > > > > > > > > > > > > as-is is gets mapped to input.
> > > > > > > > > > > > > > > > > I didn't want to change the existing line->direction, and adding the
> > > > > > > > > > > > > > > > > as-is seemed clearer than adding another flavour of direction that
> > > > > > > > > > > > > > > > > contained all three.
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Hmmm, I think I see what you were getting at - the line->direction is the
> > > > > > > > > > > > > > > > direction from the kernel, so it doesn't hurt to use that value to set the
> > > > > > > > > > > > > > > > corresponding request flags - even if the original request was as-is??
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > If that is the case then the line->as_is can be dropped throughout.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > Kent.
> > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > Yes, this is what I was thinking. Just need to make sure the value
> > > > > > > > > > > > > > > from the kernel is up-to-date.
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > So fail if needs_update?
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Kent.
> > > > > > > > > > > > >
> > > > > > > > > > > > > I'd say: do an implicit update before setting config.
> > > > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > So gpiod_line_update if needs_update, and fail if that fails?
> > > > > > > > > > > >
> > > > > > > > > > > > Kent.
> > > > > > > > > > >
> > > > > > > > > > > Without the if - needs_update is only set if an implicit update fails
> > > > > > > > > > > in line_maybe_update(). But in this case we need to be sure, so do it
> > > > > > > > > > > unconditionally.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > Given that line_maybe_update is called at the end of request creation, and
> > > > > > > > > > whenever set_config is called, how can line->direction be inconsistent
> > > > > > > > > > with the kernel state - as long as needs_update is false?
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > I don't think we should call line_maybe_update() on set_config() - in
> > > > > > > > > this case we should call gpiod_line_update() and fail in set_config()
> > > > > > > > > if it fails.
> > > > > > > > >
> > > > > > > > > I hope that's clearer.
> > > > > > > > >
> > > > > > > >
> > > > > > > > Not really.  I was already shaky on the needs_update and I'm getting more
> > > > > > > > confused about the overall needs_update handling policy by the minute.
> > > > > > > >
> > > > > > >
> > > > > > > Yeah it's not optimal. If you have better ideas on how to handle the
> > > > > > > fact that the kernel can't really notify us about the changes in
> > > > > > > line's flags introduced by other processes - I'll be more than glad to
> > > > > > > give them a try. At some point I was thinking about another ioctl()
> > > > > > > that - for a requested line - would return a file descriptor which
> > > > > > > would emit events when a line changes - for instance, it's requested
> > > > > > > by someone else or its direction is changed etc.
> > > > > > >
> > > > > >
> > > > > > I didn't realise it was possible for a requested line's flags to be
> > > > > > changed by other processes.  Quite the opposite - I thought that was one
> > > > > > of the reasons for GPIOD was to allow the userspace to prevent other from
> > > > > > processes messing with requested lines.
> > > > > >
> > > > >
> > > > > Ugh, sorry, was writing it before coffee. I was thinking about a
> > > > > non-requested line. Something like lineinfo ioctl() but returning an
> > > > > fd notifying about changes. Maybe we could even consider having
> > > > > lineinfo2 ioctl() which would be extended with this functionality -
> > > > > not only would it fill the relevant structure but also pass a new fd
> > > > > for notification about changes.
> > > > >
> > > >
> > > > Whew - that makes more sense. Had me worried there.
> > > >
> > > > Not sure how useful an async info ioctl would be.  Couldn't you build
> > > > something equivalent in userspace with the existing API - as long as you
> > > > don't mind the daemon holding the line, and so having to control the
> > > > line via the daemon.  You want to be able to monitor without requesting
> > > > the line?
> > > >
> > >
> > > I'm not sure if I was expressing myself clearly enough: a hypothetical
> > > daemon calls LINEINFO ioctl(). Now a different program or kernel
> > > driver requests this line. The daemon is not up-to-date on its state
> > > unless it polls the line all the time. If a user now asks the daemon
> > > about this line's state - it will be given outdated info. Listening on
> > > this fd would allow us to be informed about such changes immediately.
> > >
> >
> > I think I understand you - but you might not be getting my meaning...
> > I was thinking the daemon would request the lines it wanted to monitor
> > - which is why you would then have to control the line via the daemon.
> 
> No, I don't think requesting the line should be obligatory. In my WiP
> dbus daemon, I expose line info for all lines in the system by reading
> LINEINFO for each one. Then - for unrequested lines - every time the
> client asks for any line info again - I call gpiod_line_update()
> before responding. This could be optimized by this lineinfo fd
> feature.
> 
> I don't want to force the user-space to choose between using a single
> central daemon or dealing with lines separately.
> 
> > The daemon then always knows the state of the line.
> > That obviously isn't the case if you want to monitor a line without
> > requesting it, hence the "You want to be able to monitor without requesting
> > the line?" question.
> >
> 
> In other words: yes.
> 
> >
> > > > I'm still puzzled as to when the existing info ioctl could fail on a
> > > > requested line - which is when needs_update gets set in
> > > > line_maybe_update().  Hardware being unplugged?
> > > >
> > >
> > > If the ioctl() can fail, then we're obligated to check the return
> > > value. As you say: unplugging the device is a good example - it may be
> > > a GPIO expander on an HID device (e.g. Silicon Labs CP2112) that can
> > > be easily disconnected from USB.
> > >
> >
> > Fair enough. But for failures of that scale shouldn't the line request
> > fail - rather than just setting needs_update?  Or are there less
> > catastrohpic failure modes?
> >
> 
> What if the disconnect happens after the request but before the
> update? It's super unlikely, but again: the lineinfo ioctl() can fail,
> so we need to check the return value. We also can't update line info
> before requesting the line as it's racy - someone can change the state
> between the update and the request.
> 
> (I hope I'm getting this right :))
> 

I understand that the disconnect can occur between the request ioctl and
the info ioctl, but both of those are called within
line_request_values(), which implements the core of
gpiod_line_request_bulk(), so the opportunity exists to propagate the info
failure back as part of the request, but instead the error is absorbed
and needs_update is set.  This puts the onus on the caller to always
check gpiod_line_needs_update() between requesting a line and calling
any of the state accessors - else they may be returning garbage.

Similarly the event case in line_request_event_single().

I was wondering what the reasoning was for this approach?

Kent.