Václav Haisman wrote:

Václav Haisman wrote:

...

The attached patch should fix the problem with dying imap on "Unrecognized event". The problem is that when we register a handle for IO_ERROR only, we still can get readable/writable event without EV_EOF being set. This case was not handled.

-- Vaclav Haisman The fix uncovered another related glitch. If we are waiting just on IO_ERROR then we get readable/writable event all the time, because kqueue is level triggered by default and the codition is always present, and the result is busy loop. Adding EV_CLEAR bit in this case should fix it.

-- Vaclav Haisman

Bah, I was too fast to post. Attached is a corrected patch. The condition I have added in the previous should be different.

-- Vaclav Haisman

*** ioloop-kqueue.c.~1.4.~ Tue May 2 14:36:56 2006 --- ioloop-kqueue.c Wed May 10 12:02:31 2006

*** 74,80 **** { struct ioloop_handler_context *ctx = ioloop->handler_context; const int fd = io->fd; ! struct kevent ev = { fd, 0, EV_ADD | EV_EOF, 0, 0, NULL }; enum io_condition condition = io->condition & MASK;

i_assert(io->callback != NULL);

--- 74,80 ---- { struct ioloop_handler_context *ctx = ioloop->handler_context; const int fd = io->fd; ! struct kevent ev = { fd, 0, EV_ADD, 0, 0, NULL }; enum io_condition condition = io->condition & MASK;

i_assert(io->callback != NULL);

*** 101,106 **** --- 101,108 ---- if (condition & (IO_READ | IO_ERROR)) { ctx->fds[fd].mode |= condition; ev.filter = EVFILT_READ;

•   if (!(condition & ~IO_ERROR))

•   	ev.flags |= EV_CLEAR;
    if (kevent(ctx->kq, &ev, 1, NULL, 0, NULL) < 0) {
    	i_error("kevent(%d) in io_loop_handle_add() failed: %m",
    		fd);

*** 109,114 **** --- 111,118 ---- if (condition & (IO_WRITE | IO_ERROR)) { ctx->fds[fd].mode |= condition; ev.filter = EVFILT_WRITE;

•   if (!(condition & ~IO_ERROR))

•   	ev.flags |= EV_CLEAR;
    if (kevent(ctx->kq, &ev, 1, NULL, 0, NULL) < 0) {
    	i_error("kevent(%d) in io_loop_handle_add() failed: %m",
    		fd);

*** 184,190 ****

	i_assert(ctx->evbuf[i].ident < ctx->fds_size);
	if ((ctx->fds[ctx->evbuf[i].ident].mode & IO_ERROR) &&

! (ctx->evbuf[i].flags & EV_EOF)) { struct io *errio = ctx->fds[ctx->evbuf[i].ident].errio;

		t_id = t_push();

--- 188,194 ----

	i_assert(ctx->evbuf[i].ident < ctx->fds_size);
	if ((ctx->fds[ctx->evbuf[i].ident].mode & IO_ERROR) &&

! (ctx->evbuf[i].flags & (EV_EOF | EV_ERROR))) { struct io *errio = ctx->fds[ctx->evbuf[i].ident].errio;

		t_id = t_push();

*** 203,216 **** " in I/O handler %p", (void *)io->callback); } } else i_panic("Unrecognized event: kevent {.ident = %u," " .filter = 0x%04x," " .flags = 0x%04x," " .fflags = 0x%08x," ! " .data = 0x%08x}", ctx->evbuf[i].ident, ctx->evbuf[i].filter, ctx->evbuf[i].flags, ! ctx->evbuf[i].fflags, ctx->evbuf[i].data); } }

--- 207,228 ---- " in I/O handler %p", (void *)io->callback); }

•   } else if (ctx->fds[ctx->evbuf[i].ident].mode & IO_ERROR) {

•   	/* 

•   	   NO-OP. If the handle is registered only for 

•   	   IO_ERROR, then we can get readable/writable event

•   	   but no IO_READ | IO_WRITE set.

•   	 */
    } else
    	i_panic("Unrecognized event: kevent {.ident =  %u,"
                              " .filter = 0x%04x,"
                              " .flags = 0x%04x,"
                              " .fflags = 0x%08x,"

  ! " .data = 0x%08x}\n" ! "mode: 0x%x04x", ctx->evbuf[i].ident, ctx->evbuf[i].filter, ctx->evbuf[i].flags, ! ctx->evbuf[i].fflags, ctx->evbuf[i].data, ! ctx->fds[ctx->evbuf[i].ident].mode); } }

yes, the maillog should be like this:

Unrecognized event: kevent {.ident = 0, .filter = 0xffffffff, .flags = 0x8001, .fflags = 0x00000000, .data = 0x00000000} mode: 0x004x killed with signal 6

thank you

在 2006/5/11 的來信中，"Václav Haisman" <V.Haisman@sh.cvut.cz> 提及：

John Wong wrote:

...

i use openbsd/i386 + dovecot-cvs(10-may) + this patch, my maillog still have kqueue error message like this:

Unrecognized event: kevent {.ident = 0, .filter = 0xffffffff, .flags = 0x8001, .fflags = 0x00000000, .data = 0x00000000} There should be one more line of the log following this one. Could you please post it too?

...

thank you

-- Vaclav Haisman