Re: [PATCH v5 4/5] docs: counter: Document character device interface

[David Lechner]

On 10/13/20 1:58 PM, William Breathitt Gray wrote:
> On Mon, Oct 12, 2020 at 12:04:10PM -0500, David Lechner wrote:
> > On 10/8/20 7:28 AM, William Breathitt Gray wrote:
> > > On Thu, Oct 08, 2020 at 10:09:09AM +0200, Pavel Machek wrote:
> > > > Hi!
> > > >
> > > > > +        int main(void)
> > > > > +        {
> > > > > +                struct pollfd pfd = { .events = POLLIN };
> > > > > +                struct counter_event event_data[2];
> > > > > +
> > > > > +                pfd.fd = open("/dev/counter0", O_RDWR);
> > > > > +
> > > > > +                ioctl(pfd.fd, COUNTER_SET_WATCH_IOCTL, watches);
> > > > > +                ioctl(pfd.fd, COUNTER_SET_WATCH_IOCTL, watches + 1);
> > > > > +                ioctl(pfd.fd, COUNTER_LOAD_WATCHES_IOCTL);
> > > > > +
> > > > > +                for (;;) {
> > > > > +                        poll(&pfd, 1, -1);
> > > >
> > > > Why do poll, when you are doing blocking read?
> > > >
> > > > > +                        read(pfd.fd, event_data,  sizeof(event_data));
> > > >
> > > > Does your new chrdev always guarantee returning complete buffer?
> > > >
> > > > If so, should it behave like that?
> > > >
> > > > Best regards,
> > > > 									Pavel
> > > > --
> > > > (english) http://www.livejournal.com/~pavelmachek
> > > > (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
> > >
> > > I suppose you're right: a poll() should be redundant now with this
> > > version of the character device implementation because buffers will
> > > always return complete; so a blocking read() should achieve the same
> > > behavior that a poll() with read() would.
> > >
> > > I'll give some more time for additional feedback to come in for this
> > > version of the patchset, and then likely remove support for poll() in
> > > the v6 submission.
> > >
> > > William Breathitt Gray
> >
> > I hope that you mean that you will just remove it from the example
> > and not from the chardev. Otherwise it won't be possible to
> > integrate this with an event loop.
>
> Would you elaborate a bit further on this? My thought process is that
> because users must set the Counter Events they want to watch, and only
> those Counter Events show up in the character device node, a blocking
> read() would effectively behave the same as poll() with read(); if none
> of the Counter Events occur, the read() just blocks until one does, thus
> making the use of a poll() call redundant.
>
> William Breathitt Gray

If the counter device was the only file descriptor being read, then yes
it wouldn't matter. But if we are using this in combination with other
file descriptors, then it is common to poll all of the file descriptors
using a single syscall to see which one is ready to read rather than
doing a non-blocking read on all of the file descriptors, which would
result in many unnecessary syscalls.