While adding patches to natively support FreeBSD in wayland, I noticed that epoll-shim (used to emulate epoll with kqueue) didn't report signals sent to self via kill() unless I added a non-zero timeout (even in a single-threaded test program). In https://github.com/jiixyj/epoll-shim/pull/32 Jan Kokemüller provided a reduced test case using only kqueue and kevent: ``` #ifdef NDEBUG #undef NDEBUG #endif #define _GNU_SOURCE #include <sys/types.h> #include <sys/event.h> #include <assert.h> #include <errno.h> #include <signal.h> #include <stdlib.h> #include <poll.h> #include <unistd.h> int main(void) { int rv; sigset_t set; rv = sigemptyset(&set); assert(rv == 0); rv = sigaddset(&set, SIGUSR1); assert(rv == 0); rv = sigprocmask(SIG_BLOCK, &set, NULL); assert(rv == 0); int skq = kqueue(); assert(skq >= 0); struct kevent kev; EV_SET(&kev, SIGUSR1, EVFILT_SIGNAL, EV_ADD, 0, 0, 0); rv = kevent(skq, &kev, 1, NULL, 0, NULL); assert(rv == 0); int kq = kqueue(); assert(kq >= 0); EV_SET(&kev, skq, EVFILT_READ, EV_ADD | EV_CLEAR, 0, 0, 0); rv = kevent(kq, &kev, 1, NULL, 0, NULL); assert(rv == 0); for (;;) { rv = kill(getpid(), SIGUSR1); assert(rv == 0); /* Turn this into `#if 1` to avoid the race. */ #if 0 rv = kevent(kq, NULL, 0, &kev, 1, NULL); #else rv = kevent(kq, NULL, 0, &kev, 1, &(struct timespec) { 0, 0 }); #endif assert(rv == 1); rv = kevent(kq, NULL, 0, &kev, 1, &(struct timespec) { 0, 0 }); assert(rv == 0); rv = kevent(skq, NULL, 0, &kev, 1, &(struct timespec) { 0, 0 }); assert(rv == 1); rv = kevent(skq, NULL, 0, &kev, 1, &(struct timespec) { 0, 0 }); assert(rv == 0); siginfo_t siginfo; rv = sigtimedwait(&set, &siginfo, &(struct timespec) { 0, 0 }); assert(rv == SIGUSR1); rv = sigtimedwait(&set, &siginfo, &(struct timespec) { 0, 0 }); assert(rv < 0); assert(errno == EAGAIN); } } ``` Is this behaviour expected or a bug?
So the problem is actually not that kevent does not get signal notification. It is that kqueue attached to kqueue with the signal event suddenly reports no events when you specify zero timeout, i.e. non-blocking peek. Am I right? This is because kqueue subsystem has to fight with the recursion and deadlock. If kqueue has another kqueue attached to it, then notification of the attached kqueue is delegated to a task. See knote_enqueue()->kqueue_wakeup()-> kqueue_schedtask(). Until task run, kq->kq_count == 0 and this together with timespec being 0 (AKA unset) causes kqueue_scan() to immediately return 0. I am not sure if this is worth complications to make the wakeup synchronous.
(In reply to Konstantin Belousov from comment #1) Yes, it's only the non-blocking peek that appears to be affected.
The crucial part is not the peek, but recursion. Anyway it is relatively simple to fix it seems https://reviews.freebsd.org/D31858
A commit in branch main references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=98168a6e6c12dab8f608f6b5f5b0b175d2b87ef0 commit 98168a6e6c12dab8f608f6b5f5b0b175d2b87ef0 Author: Konstantin Belousov <kib@FreeBSD.org> AuthorDate: 2021-09-06 11:43:06 +0000 Commit: Konstantin Belousov <kib@FreeBSD.org> CommitDate: 2021-09-06 23:43:34 +0000 kqueue: drain kqueue taskqueue if syscall tickled it Otherwise return from the syscall and next syscall, which could be kevent(2) on the kqueue that should be notified, races with the kqueue taskqueue thread, and potentially misses the wakeup. This is reliably visible when kevent(2) only peeks into events using zeroed timeout. PR: 258310 Reported by: arichardson, Jan Kokemüller <jan.kokemueller@gmail.com> Reviewed by: arichardson, markj Sponsored by: The FreeBSD Foundation MFC after: 1 week Differential revision: https://reviews.freebsd.org/D31858 sys/kern/kern_event.c | 11 +++++++++++ sys/kern/subr_trap.c | 7 ++++++- sys/sys/event.h | 1 + sys/sys/proc.h | 2 +- 4 files changed, 19 insertions(+), 2 deletions(-)
^Triage: assign to committer.
A commit in branch stable/13 references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=c43e14590772d8043048d7c9acf333764d21a1cf commit c43e14590772d8043048d7c9acf333764d21a1cf Author: Konstantin Belousov <kib@FreeBSD.org> AuthorDate: 2021-09-06 11:43:06 +0000 Commit: Konstantin Belousov <kib@FreeBSD.org> CommitDate: 2021-09-12 12:17:27 +0000 kqueue: drain kqueue taskqueue if syscall tickled it PR: 258310 (cherry picked from commit 98168a6e6c12dab8f608f6b5f5b0b175d2b87ef0) sys/kern/kern_event.c | 11 +++++++++++ sys/kern/subr_trap.c | 7 ++++++- sys/sys/event.h | 1 + sys/sys/proc.h | 2 +- 4 files changed, 19 insertions(+), 2 deletions(-)