When a fifo with no writers is opened nonblockingly for reading and the resulting file descriptor is added to a kqueue, kevent will immediately report an EOF condition. This is less than useful. The useful behaviour would be for kevent not to report EOF until at least one writer has come and gone. This matches how similar mechanisms in other operating systems work (e.g. epoll in Linux, /dev/poll in Solaris). Fix: Don't CANTRCVMORE the socketpair immediately after creating it. Add code to fifo_read_f to avoid calling soreceive blockingly when there are zero writers. Or just add a fi_seen_at_least_one_writer flag to struct fifoinfo... How-To-Repeat: #include <stdio.h> #include <sys/types.h> #include <sys/stat.h> #include <sys/event.h> #include <sys/time.h> #include <fcntl.h> #include <unistd.h> int main(void) { int rfd,wfd,qfd,n; static struct timespec zero; struct kevent kev; char fifopath[64]; sprintf(fifopath,"/tmp/less-than-useful.%ld",(long)getpid()); if (mkfifo(fifopath,S_IRUSR|S_IWUSR)==-1) { perror("mkfifo"); return 1; } rfd=open(fifopath,O_RDONLY|O_NONBLOCK); if (rfd==-1) { perror("open 1"); goto cleanup; } qfd=kqueue(); if (qfd==-1) { perror("kqueue"); goto cleanup; } EV_SET(&kev,rfd,EVFILT_READ,EV_ADD,0,0,NULL); n=kevent(qfd,&kev,1,NULL,0,&zero); if (n==-1) { perror("kevent 1"); goto cleanup; } n=kevent(qfd,NULL,0,&kev,1,&zero); if (n==-1) { perror("kevent 2"); goto cleanup; } if (n==1 && (kev.flags & EV_EOF)) { fputs("Immediate EOF (not useful)\n",stderr); } wfd=open(fifopath,O_WRONLY); if (wfd==-1) { perror("open 2"); goto cleanup; } n=kevent(qfd,NULL,0,&kev,1,&zero); if (n==-1) { perror("kevent 3"); goto cleanup; } if (n==1 && (kev.flags & EV_EOF)) { fputs("EOF while there's still a writer (BUG!)\n",stderr); } close(wfd); n=kevent(qfd,NULL,0,&kev,1,&zero); if (n==-1) { perror("kevent 4"); goto cleanup; } if (n==1 && (kev.flags & EV_EOF)) { fputs("EOF after the last writer went away (useful)\n",stderr); } close(qfd); close(rfd); unlink(fifopath); return 0; cleanup: unlink(fifopath); return 1; }
On Sun, 30 Sep 2007, Bo Lindbergh wrote: >> Description: > When a fifo with no writers is opened nonblockingly for reading and > the resulting file descriptor is added to a kqueue, kevent will > immediately report an EOF condition. This is less than useful. But it is an EOF condition: in this state, read() must return 0 to indicate EOF. select(), poll() and kqueue could have another mechanism for reporting this state, but don't in FreeBSD. Some other OS's have a specially broken select() and/or poll() for fifos but not for other file types so that the polling read condition doesn't actually report the read condition for fifos only. >> Fix: > Don't CANTRCVMORE the socketpair immediately after creating it. Add > code to fifo_read_f to avoid calling soreceive blockingly when there > are zero writers. That would break read(). > Or just add a fi_seen_at_least_one_writer flag to struct fifoinfo... Fixing this, or even implementing bug for bug compatibility with other OS's, is not easy. See PR's 34020, 53447, 76144, 76125, 76525, 94722 and the resulting commits for previous attempts to fix this. PR 94722 does something like this. It only tries to fix poll(). The behaviour of select() on a read descriptor cannot be changed, except to remove old buggy attempts to fix this problem, since select() on a read descriptor has no way to distinguish initial EOF from hangup. select() on an exceptional descriptor could consider hangup as an exception; there is no standard for this, but since exceptional descriptors are rarely used, changing the behaviour for them wouldn't break much. Kqueue has flags so it should be able to use the fix for poll() fairly easily. Bruce