bug#64297: [Cuirass] Remote server not picking up job, losing workers

["Ludovic Courtès" <ludo@gnu.org>]

Ludovic Courtès <ludovic.courtes@inria.fr> skribis:

> As of cuirass@1.1.0-16.b825967, ‘cuirass remote-server’ appears to not
> pick jobs as quickly as it should and to lose sight of workers (you can
> see them come and go on <https://ci.guix.gnu.org/workers>).
>
> /var/log/cuirass-remote-worker.log shows that it does build things, but
> only sporadically.  Then there are things like:
>
>   2023-06-26 10:07:58 warning: Poll loop busy during 3404 seconds.
>
> This is presumably related to Cuirass commit
> c4743b54720e86b0e0b0295fb6d33977e4293644 (previously ‘remote-worker’
> would have a database worker thread; now it opens a new connection every
> time—a stopgap before it’s fiberized, but apparently not a good one).

Apparently this has to do with squee calling ‘current-read-waiter’
(i.e., poll(2)) while waiting for its response and passing checking the
wrong FD for some reason, as in this case:

--8<---------------cut here---------------start------------->8---
18484 15:01:00 connect(55, {sa_family=AF_UNIX, sun_path="/tmp/ephemeralpg.58xnKh/.s.PGSQL.5432"}, 110) = 0 <0.000019>
18484 15:01:00 getsockopt(55, SOL_SOCKET, SO_ERROR, [0], [4]) = 0 <0.000014>
18484 15:01:00 getsockname(55, {sa_family=AF_UNIX}, [128 => 2]) = 0 <0.000010>
18484 15:01:00 poll([{fd=55, events=POLLOUT|POLLERR}], 1, -1) = 1 ([{fd=55, revents=POLLOUT}]) <0.000011>
18484 15:01:00 sendto(55, "\0\0\0!\0\3\0\0user\0ludo\0database\0test\0\0", 33, MSG_NOSIGNAL, NULL, 0) = 33 <0.000014>
18484 15:01:00 poll([{fd=55, events=POLLIN|POLLERR}], 1, -1 <unfinished ...>
18484 15:01:00 <... poll resumed>)      = 1 ([{fd=55, revents=POLLIN}]) <0.001786>
18484 15:01:00 recvfrom(55, "R\0\0\0\10\0\0\0\0S\0\0\0\26application_name\0\0S\0\0\0\31client_encoding\0UTF8\0S\0\0\0\27DateStyle\0ISO, MD"..., 16384, 0, NULL, NULL) = 376 <0.000016>
18484 15:01:00 sendto(55, "P\0\0\0\366\0\nUPDATE Builds SET status = -2, worker = null FROM\n(SELECT id FROM Workers"..., 291, MSG_NOSIGNAL, NULL, 0 <unfinished ...>
18484 15:01:00 <... sendto resumed>)    = 291 <0.000589>
18484 15:01:00 poll([{fd=54, events=POLLIN}], 1, -1 <unfinished ...>
18484 15:03:14 <... poll resumed>)      = 1 ([{fd=54, revents=POLLIN}]) <134.319198>
18484 15:03:14 recvfrom(55,  <unfinished ...>
18484 15:03:14 <... recvfrom resumed>"1\0\0\0\0042\0\0\0\4n\0\0\0\4C\0\0\0\rUPDATE 0\0Z\0\0\0\5I", 16384, 0, NULL, NULL) = 35 <0.000026>
18484 15:03:14 sendto(55, "P\0\0\0V\0DELETE FROM Workers WHERE\n(extract(epoch from now())::int - last_seen) >  "..., 131, MSG_NOSIGNAL, NULL, 0 <unfinished ...>
18484 15:03:14 <... sendto resumed>)    = 131 <0.000084>
18484 15:03:14 poll([{fd=54, events=POLLIN}], 1, -1) = 1 ([{fd=54, revents=POLLNVAL}]) <0.000013>
18484 15:03:14 recvfrom(55, "1\0\0\0\0042\0\0\0\4n\0\0\0\4C\0\0\0\rDELETE 4\0Z\0\0\0\5I", 16384, 0, NULL, NULL) = 35 <0.000016>
18484 15:03:14 sendto(55, "X\0\0\0\4", 5, MSG_NOSIGNAL, NULL, 0) = 5 <0.000025>
18484 15:03:14 close(55)                = 0 <0.000018>
18484 15:03:14 openat(AT_FDCWD, "/etc/localtime", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory) <0.000019>
18484 15:03:14 write(2, "2023-06-30T15:03:14 warning: Poll loop busy during 134 seconds.\n", 64) = 64 <0.000024>
--8<---------------cut here---------------end--------------->8---

In this case FD 54 is a connection with a worker process; terminating
that process led poll(2) to return, thus unblocking the “poll loop”.

The problem is most likely with the connection-to-port caching in
squee’s ‘connection-socket-port’, as can be seen in this other trace
where I added ‘pk’ calls in ‘connection-socket-port’:

--8<---------------cut here---------------start------------->8---
19468 15:37:43 connect(55, {sa_family=AF_UNIX, sun_path="/tmp/ephemeralpg.58xnKh/.s.PGSQL.5432"}, 110) = 0 <0.000018>
19468 15:37:43 getsockopt(55, SOL_SOCKET, SO_ERROR, [0], [4]) = 0 <0.000010>
19468 15:37:43 getsockname(55, {sa_family=AF_UNIX}, [128 => 2]) = 0 <0.000009>
19468 15:37:43 poll([{fd=55, events=POLLOUT|POLLERR}], 1, -1) = 1 ([{fd=55, revents=POLLOUT}]) <0.000009>
19468 15:37:43 sendto(55, "\0\0\0!\0\3\0\0user\0ludo\0database\0test\0\0", 33, MSG_NOSIGNAL, NULL, 0) = 33 <0.000012>
19468 15:37:43 poll([{fd=55, events=POLLIN|POLLERR}], 1, -1) = 1 ([{fd=55, revents=POLLIN}]) <0.002109>
19468 15:37:43 recvfrom(55, "R\0\0\0\10\0\0\0\0S\0\0\0\26application_name\0\0S\0\0\0\31client_encoding\0UTF8\0S\0\0\0\27DateStyle\0ISO, MD"..., 16384, 0, NULL, NULL) = 376 <0.000009>
19468 15:37:43 sendto(55, "P\0\0\0\366\0\nUPDATE Builds SET status = -2, worker = null FROM\n(SELECT id FROM Workers"..., 291, MSG_NOSIGNAL, NULL, 0) = 291 <0.000012>
19468 15:37:43 write(1, "\n", 1)        = 1 <0.000015>
19468 15:37:43 ioctl(54, TCGETS, 0x7ffd8fde6660) = -1 ENOTTY (Inappropriate ioctl for device) <0.000012>
19468 15:37:43 write(1, ";;; (cached #<pg-conn cdac20 (connected)> #<input-output: file 54>)\n", 68) = 68 <0.000012>
--8<---------------cut here---------------end--------------->8---

Above we open a fresh connection on FD 55, but ‘connection-socket-port’
determines that the connection object is cached and associated with a
port for FD 54.

To be continued…

Ludo’.