changed net/connect to be non-blocking / asynchronous

[zevv]

This PR makes net/connect non-blocking, which I believe is (apart from name resolving) the last change needed to make the Janet network impl all nice and async.

I'd appreciate some feedback though:

• can anyone who knows the event loop code tells me if I did this right?
• closing the socket on failing connect is still missing; what's the right way to handle this?
• not sure what to do on getsockopt() failure, panic, assert?

[bakpakin]

With a cursory glance, everything looks correct. The intuition about calling janet_stream_close on a bad connect looks ok at first glance, any particular reason it is commented out?

not sure what to do on getsockopt() failure, panic, assert?

Don't assert, instead panic and clean up as if connect had failed. Generally, we don't assert anything about syscalls and try to convert errno -> janet panic.

[zevv]

With a cursory glance, everything looks correct. The intuition about calling janet_stream_close on a bad connect looks ok at first glance, any particular reason it is commented out?

Yes, SIGSEGV later because of a use-after-free, but I wasn't able yet to fully grasp the mechanics involved.

Don't assert, instead panic and clean up as if connect had failed. Generally, we don't assert anything about syscalls and try to convert errno -> janet panic.

Right, so that would be janet_cancel(s->fiber, janet_cstringv(strerror(r))); I guess, because the call fails "on behalf" of the fiber calling net/connect

[zevv]

@bakpakin When calling janet_stream_close():

connect
==894789== Invalid read of size 8
==894789==    at 0x132FCC: janet_loop1_impl (in /home/ico/external/janet/build/janet)
==894789==    by 0x1565B4: janet_loop1 (in /home/ico/external/janet/build/janet)
==894789==    by 0x15670C: janet_loop (in /home/ico/external/janet/build/janet)
==894789==    by 0x156CD2: janet_loop_fiber (in /home/ico/external/janet/build/janet)
==894789==    by 0x11BCAC: main (in /home/ico/external/janet/build/janet)
==894789==  Address 0x4cea900 is 0 bytes inside a block of size 88 free'd
==894789==    at 0x484317B: free (vg_replace_malloc.c:872)
==894789==    by 0x132A17: janet_stream_close (in /home/ico/external/janet/build/janet)
==894789==    by 0x132D68: ev_machine_write (in /home/ico/external/janet/build/janet)
==894789==    by 0x132F99: janet_loop1_impl (in /home/ico/external/janet/build/janet)
==894789==    by 0x1565B4: janet_loop1 (in /home/ico/external/janet/build/janet)
==894789==    by 0x15670C: janet_loop (in /home/ico/external/janet/build/janet)
==894789==    by 0x156CD2: janet_loop_fiber (in /home/ico/external/janet/build/janet)
==894789==    by 0x11BCAC: main (in /home/ico/external/janet/build/janet)
==894789==  Block was alloc'd at
==894789==    at 0x48407B4: malloc (vg_replace_malloc.c:381)
==894789==    by 0x133237: janet_listen (in /home/ico/external/janet/build/janet)
==894789==    by 0x13374B: janet_ev_connect (in /home/ico/external/janet/build/janet)
==894789==    by 0x145B42: cfun_net_connect (in /home/ico/external/janet/build/janet)
==894789==    by 0x15103C: run_vm (in /home/ico/external/janet/build/janet)
==894789==    by 0x1553EA: janet_continue_no_check (in /home/ico/external/janet/build/janet)
==894789==    by 0x15632D: janet_loop1 (in /home/ico/external/janet/build/janet)
==894789==    by 0x15670C: janet_loop (in /home/ico/external/janet/build/janet)
==894789==    by 0x156CD2: janet_loop_fiber (in /home/ico/external/janet/build/janet)
==894789==    by 0x11BCAC: main (in /home/ico/external/janet/build/janet)

[zevv]

So, on closer inspection the problem seems to be that closing the stream at that point is too early, because later in the loop implementation the stream and it states are still referenced and checked for other events.

I guess the right thing to do would be to flag the stream as 'to be closed' (add a stream flag?), and after the last loop iteration close the stream if this flag is set.

[bakpakin]

Is there any need to close the stream, or can you just drop it and let the GC take care of it? Might be easier that way - I haven't run with the code yet locally, but elsewhere in the event loop implementation we don't explicitly close anything until the user closes it.

[bakpakin]

On second thought, it is probably better to have to close rather than leaving file descriptors open for an undetermined amount of time.

[bakpakin]

LGTM, I will run some BSD testing and if that checks out I will merge.

[zevv]

Yeah, GC-ing file descriptors or sockets is kind of nasty, I agree we better clean them up.

[bakpakin]

There is a bug when JANET_NO_EPOLL is defined -> a JanetListenerState is used after being deallocated.

Here is a fix:

diff --git a/src/conf/janetconf.h b/src/conf/janetconf.h
index 73e39d55..82b173e4 100644
--- a/src/conf/janetconf.h
+++ b/src/conf/janetconf.h
@@ -49,7 +49,7 @@
 /* #define JANET_STACK_MAX 16384 */
 /* #define JANET_OS_NAME my-custom-os */
 /* #define JANET_ARCH_NAME pdp-8 */
-/* #define JANET_EV_NO_EPOLL */
+#define JANET_EV_NO_EPOLL
 /* #define JANET_EV_NO_KQUEUE */
 /* #define JANET_NO_INTERPRETER_INTERRUPT */
 
diff --git a/src/core/ev.c b/src/core/ev.c
index 31b0a2f6..fd3a7b9a 100644
--- a/src/core/ev.c
+++ b/src/core/ev.c
@@ -1969,6 +1969,7 @@ void janet_loop1_impl(int has_timeout, JanetTimestamp timeout) {
         JanetAsyncStatus status3 = JANET_ASYNC_STATUS_NOT_DONE;
         JanetAsyncStatus status4 = JANET_ASYNC_STATUS_NOT_DONE;
         state->event = pfd;
+        JanetStream *stream = state->stream;
         if (mask & POLLOUT)
             status1 = state->machine(state, JANET_ASYNC_EVENT_WRITE);
         if (mask & POLLIN)
@@ -1983,7 +1984,6 @@ void janet_loop1_impl(int has_timeout, JanetTimestamp timeout) {
                 status4 == JANET_ASYNC_STATUS_DONE)
             janet_unlisten(state, 0);
         /* Close the stream if requested and no more listeners are left */
-        JanetStream *stream = state->stream;
         if ((stream->flags & JANET_STREAM_TOCLOSE) && !stream->state) {
             janet_stream_close(stream);
         }

[zevv]

Applied, thanks.

Would it make sense to run CI with the JANET_EV_NO_EPOLL case as well?

[bakpakin]

We already do on sr.ht which is why I don't here, but I guess it would help for contributions.