| Commit message (Collapse) | Author |
|---|
|
|
|
And that they're marked as finished. It immediately exposed some
concurrency bugs, so some locking has been fixed.
|
|
|
|
This is a major change, obviously; brought to me by Valgrind, which
noticed that we don't actually clean up after cimple-client threads.
For a more thorough explanation, please see the added comment in
tcp_server.c.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Is this an overkill? I don't know.
The thing is, correctly intercepting SIGTERM (also SIGINT, etc.) is
incredibly tricky. For example, before this commit, my I/O loops in
server.c and worker.c were inherently racy.
This was immediately obvious if you tried to run the tests. The tests
(especially the Valgrind flavour) would run a worker, wait until it
prints a "Waiting for a new command" line, and try to kill it using
SIGTERM. The problem is, the global_stop_flag check could have already
been executed by the worker, and it would hang forever in recv().
The solution seems to be to use signalfd and select()/poll(). I've never
used either before, but it seems to work well enough - at least the very
same tests pass and don't hang now.
|
|
|
|
|
|
|
|
|
|
|
|
OK, this is a major rework.
* tcp_server: connection threads are not detached anymore, the caller has
to clean them up. This was done so that the server can clean up the
threads cleanly.
* run_queue: simple refactoring, run_queue_entry is called just run now.
* server: worker threads are now killed when a run is assigned to a
worker.
* worker: the connection to server is no longer persistent. A worker
sends "new-worker", waits for a task, closes the connection, and when
it's done, sends the "complete" message and waits for a new task.
This is supposed to improve resilience, since the worker-server
connections don't have to be maintained while the worker is doing a CI
run.
|
|
|
|
|
|
Also, some minor refactoring.
|
|
|
|
* I don't really need to declare all variables at the top of the
function anymore.
* Default-initialize variables more.
* Don't set the output parameter until the object is completely
constructed.
|
|
|
|
|
|
|
|
Explicit is better than implicit.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
This allows free workers to pick up jobs after dead workers.
|
|
The problem is pthread_cond_destroy is unsafe to call if there're
threads waiting in pthread_cond_wait. I'm not sure this fix is enough:
what if the "broadcast" doesn't reach the threads until we call
pthread_cond_destroy? Does it even work that way? Idk
|
|
Previously, the client had no way to distinguish errors from succesful
calls.
|
|
|
|
|
|
Previously, I had a stupid system where I would create a thread after
every accept(), and put worker descriptors in a queue. A special
"scheduler" thread would then pick them out, and give out jobs to
complete.
The problem was, of course, I couldn't conveniently poll job status from
workers. I thought about using poll(), but that turned out to be a
horribly complicated API. How do I deal with partial reads, for example?
I don't honestly know.
Then it hit me that I could just use the threads that handle accept()ed
connections as "worker threads", which would synchronously schedule jobs
and wait for them to complete. This solves every problem and removes the
need for a lot of inter-thread synchronization magic. It even works now,
holy crap! You can launch and terminate workers at will, and they will
pick up new jobs automatically.
As a side not, msg_recv_and_handle turned out to be too limiting and
complicated for me, so I got rid of that, and do normal
msg_recv/msg_send calls.
|
|
|
|
pthread functions return positive error codes.
|
