Re: [cwilson@corp.sgi.com: Re: /dev/poll on irix...]

Some questions on /dev/poll:

Is the event cleared in the /dev/poll instance when delivered to the
application, forcing the application to re-register the FD to receive
further event, or is the registration persistent for the lifetime of
the filedescriptor or until manually unregistered?

If persistent, how to clear a event bit?

From this IRIX man page it seems the Solaris implementation is rather
braindead, but at the same time it is quite vague in itself so I do
not know.. smells like they have jumped onto something only because
Solaris has it.

Note on kernel design: The user->kernel->user copying of the pollfds
array is only one of the bottlenecks of poll(). Not always the most
significant... Another significant bottleneck is the redundant
processing of filedescriptors for which no event has occured only to
find that no event has occured.

Regards
Henrik

On Thursday 08 August 2002 03.41, Adrian Chadd wrote:
> Cute.
>
> Another platform that squid may run quite fast on in the future..
>
>
>
>
> Adrian
>
> ----- Forwarded message from Chan Wilson <cwilson@corp.sgi.com>
> -----
>
> Date: Tue, 06 Aug 2002 10:28:31 -0700 (PDT)
> To: adrian@squid-cache.org
> Subject: Re: /dev/poll on irix...
> From: Chan Wilson <cwilson@corp.sgi.com>
> X-Mailer: Mew version 2.2rc2 on Emacs 21.1 / Mule 5.0 (SAKAKI)
>
> 6.5.15m. Here's the man page:
>
>
> poll(7)
> poll(7)
>
> NAME
> poll - device for high speed poll on file descriptors
>
> SYNOPSIS
> #include <stropts.h>
> #include <poll.h>
> #include <sys/devpoll.h>
>
> DESCRIPTION
> /dev/poll is a device that allows users to efficiently monitor
> large numbers of file descriptors. /dev/poll works similarly to
> the poll system call and to the Sun Solaris version of the device
> except as noted below. File descriptors to be polled are
> registered and scanned with the write and ioctl system calls
> respectively.
>
> Registration of file descriptors is done by writing an array
> of pollfd structures to the /dev/poll device. These structures are
> handled the same as in poll and contain the following fields:
>
> int fd; /* file descriptor */
> short events; /* requested events */
> short revents; /* returned events */
>
> Each passed pollfd is cached within the /dev/poll device. If
> a previous pollfd is found referencing the same file descriptor,
> their events fields are combined. Users wishing to maintain
> separate sets of registered pollfd structures should open separate
> instances of the /dev/poll device. The write call returns the
> number of bytes written or -1 and an error code. Following a
> pollfd's registration, the device watches for the events specified
> in the structure's events field to occur to the file descriptor
> specified in the structure's fd field.
>
> Event notification is received through the ioctl system call
> using the DP_POLL command with a pointer to a dvpoll structure as
> its argument. The dvpoll structure has the following fields:
>
> struct pollfd * dp_fds; /* pollfd array */
> nfds_t dp_nfds; /* Number of pollfds at dp_fds */
> int dp_timeout; /* Time-out in milliseconds */
>
> dp_fds is an array to which up to dp_nfds pollfds with active
> events will be copied by the device. Each copied pollfd's revents
> field will include those events in the events field that are
> active. Each pollfd's fd field will be set to the file descriptor
> that it was monitoring, and its events field will be set to the
> events that it's watching for. If no active events are found, the
> call will block for the length of time in dp_timeout. If
> dp_timeout is 0 then the call will immediately return. If
> dp_timeout is -1 then the call will block until a registered event
> does become active. If dp_timeout is greater than 0 then it will
> block for that many milliseconds unless an event becomes active.
> The ioctl will return either the number of pollfds received, or -1
> if an error
>
>
> Page 1
>
> poll(7)
> poll(7)
>
> occurred.
>
> A user can unregister pollfds from the device's poll set by
> adding the POLLREMOVE flag to the events field of a pollfd and
> writing it to the device. When a monitored file descriptor is
> closed, all /dev/poll instances that had pollfds registered to
> monitor it will stop monitoring the file. The pollfds will remain
> cached in the /dev/poll device and can be removed with the
> POLLREMOVE flag as described above. If a new file or socket is
> opened with a file descriptor that is the same as a previously
> closed file or socket that had a registered pollfd that is still
> cached in a /dev/poll instance, the registered pollfd will not
> monitor the new file or socket. If a pollfd is written to one of
> these /dev/poll instances to monitor this new file or socket, it
> will overwrite the no longer used previous pollfd.
>
> Users may use the DP_REMOVE ioctl command on a /dev/poll
> instance to direct the device to try to remove cached pollfds from
> the device instance whenever the file descriptor that they are
> watching is closed. The device instance will only attempt this
> while doing other work and it may not always be successful.
>
> Users may use the DP_ISPOLLED ioctl command with a pollfd as
> the argument to check if the file descriptor specified in the
> pollfds fd field is part of the poll set for the /dev/poll instance
> queried. The ioctl returns 1 if it is and returns the currently
> polled events in the pollfd's events field. The call returns 0 if
> the file descriptor is not being polled and -1 if there was an
> error.
>
> When a program with an open instance of /dev/poll calls the
> fork system call to create a child process, any attempts by the
> child process to write or use ioctl on the instance will fail with
> the EACCES error code. It is suggested that child processes close
> the instance and open a new one. When the last open reference to
> an instance is closed, all cached pollfds will be removed as well.
>
> NOTES
> In the Sun Solaris version of /dev/poll, when a file
> descriptor is closed that has /dev/poll pollfds monitoring it, and
> a new file is opened with the same file descriptor, the pollfds
> will switch to monitor the new file. The IRIX version is different
> in that the pollfds will not switch to monitor the new file
> descriptor. They will remain dormant in their /dev/poll instances
> until they are overwritten with new pollfds referencing the file
> descriptor or they are removed with POLLREMOVE.
>
> HISTORY
> The /dev/poll device was added in IRIX 6.5.15.
>
> SEE ALSO
> fork(2), ioctl(2), poll(2), write(2).
>
>
> Page 2
>
>
>
> Adrian Chadd <adrian@squid-cache.org> spaketh thusly on Fri, 2 Aug
> 2002 21:15:01 -0600 about Re: /dev/poll on irix......
>
> > On Wed, Jul 31, 2002, Chan Wilson wrote:
> > > Adrian,
> > >
> > > How goes?
> > >
> > > Through happenstance, I have the attention of the engr who
> > > implemented /dev/poll in Irix. Got any questions?
> >
> > Ooer. When did it first find its way into the irix kernel?
> > I don't seem to remember finding manpages about it..
> >
> >
> >
> > Adrian
>
> ----- End forwarded message -----
Received on Sat Aug 10 2002 - 20:50:32 MDT