Skip to contentSkip to conversation

Edge triggered epoll with pts devices acts as level triggered

DISCUSSION Adam Langley 2 participants 3 messages Aug 12, 2005 – Aug 13, 2005

lore
Share & tools
lore
#0 lore raw

(please cc me on replies)

Waiting for edge triggered events (with EPOLLET) on pseudo terminal devices appears to act as if it were level triggered; when data is ready the fd is always returned by epoll_wait.

You can test this with the code below. Compile, run and press return. If edge triggering is working correctly a single event will be generated, otherwise a never ending stream will start.

This works *correctly* at a real terminal, but fails for pseudo terminals (specifically an xterm). As far as I can test with other terminals and ssh this is a general problem with pseudo terminals.

% uname -a
Linux ice 2.6.12 #4 SMP Sun Jul 31 11:42:15 BST 2005 i686 Pentium II (Deschutes) GenuineIntel GNU/Linux
(vanilla kernel sources)

% cat et2.c
#include <sys/epoll.h>
#include <unistd.h>
#include <string.h>

int

main(int argc, char **argv) {
        const int epoll_fd = epoll_create(4);
        struct epoll_event ev;
        struct epoll_event events[4];

        memset(&ev, 0, sizeof(ev));
        ev.events = EPOLLIN | EPOLLET;

        epoll_ctl(epoll_fd, EPOLL_CTL_ADD, 0, &ev);

        for (;;) {
                epoll_wait(epoll_fd, events, 4, -1);
                write(1, ".", 1);
        }
}

Thanks

AGL

Signature · 3 lines
Adam Langley                                      [email protected]
http://www.imperialviolet.org                       (+44) (0)7906 332512
PGP: 9113   256A   CC0F   71A6   4C84   5087   CDA5   52DF   2CB6   3D60
#1 lore raw
Quoted 3 lines · Adam Langley

Waiting for edge triggered events (with EPOLLET) on pseudo terminal devices appears to act as if it were level triggered; when data is ready the fd is always returned by epoll_wait.

This occurs because writing to the terminal happens to cause a read event to occur for pseudo terminals, but not for real terminals. This is much less of a problem than the orginal message would suggest.

AGL

Signature · 3 lines
Adam Langley                                      [email protected]
http://www.imperialviolet.org                       (+44) (0)7906 332512
PGP: 9113   256A   CC0F   71A6   4C84   5087   CDA5   52DF   2CB6   3D60
#2 lore raw
In reply to Adam Langley
Quoted 2 lines · Adam Langley
               epoll_wait(epoll_fd, events, 4, -1);
               write(1, ".", 1);

The test case is faulty. write will trigger the event again infinitely. fds 0 and 1 are the same in many cases. Try this:

$ ./epolltest > foo

Now you will only get 1 trigger for each input.

Signature · 3 lines
Heikki Orsila			Barbie's law:
[email protected]		"Math is hard, let's go shopping!"
http://www.iki.fi/shd