Hang waiting for create_server

[jimfulton]

• uvloop 0.4.33:
• python 3.5.1, 3.5.2:
• platform Mac OS X, Ubuntu 16.04/4.4.12-boot2docker:

I need some suggestions for debugging server startup hangs.

Startup code looks like:

 cr = loop.create_server(self.factory, addr[0], addr[1],
                                reuse_address=True, ssl=ssl)
f = asyncio.async(cr, loop=loop)
server = loop.run_until_complete(f)

What's sometimes happening, in the context of the ZEO tests, is that the run_loop_until_complete call above never returns.

This is in a port of ZEO, github.com/zopefoundation/ZEO, to asyncio. The ZEO tests are exhaustive, and I suspect in uvloop's case, exhausting. :) The tests start and stop servers 10s, maybe 100s of times in a test runs that typically lasts a few minutes. If I run a smaller subset of the tests I don't get a hang.

Most tests run the server as a subprocess using multiprocessing. Some tests run the server using threading. It appears that the server isn't exiting, but merely hanging.

Do you have any suggestions for how to debug this? I've tried setting PYTHONASYNCIODEBUG=1 and enabling debug logging, but that isn't yielding any additional information..

[1st1]

The general advice is to create a minimal testcase and then debug the code with gdb (or lldb) and print statements ;)

In this particular case, could you please tell me the values of addr and ssl? Does the loop have any other scheduled tasks? Does it run in the main thread?

[1st1]

I'd also suggest to add prints before each await in Loop.create_server. That should give us better understanding where exactly the loop hangs.

[1st1]

Hm, actually, there is only one await in Loop.create_server, and it waits on getaddrinfo. So it might be a bug in the DNS resolver. What's the value of addr?

[1st1]

And there's also an option to make a debug build with make debug. When built in debug mode, loop has a print_debug_info() method which outputs a bunch of useful debug info.

[jimfulton]

On Thu, Jul 7, 2016 at 5:39 PM, Yury Selivanov notifications@github.com
wrote:

The general advice is to create a minimal testcase and then debug the code
with gdb (or lldb) and print statements ;)

Yeah. I took an initial stab at that, but it was too minimal. I'll take
another stab.

In this particular case, could you please tell me the values of addr and
ssl?

ssl is usually, but not always None. I just hacked out SSL-related tests
and ssl support, and still hanging. So for the sake of this discussion,
SSL is always None.

The address is almost always ('127.0.0.1', 0). I just did some test
cleanup to make this so. :) Previously, it often had a host of
'localhost' or '' and a randomly selected port. There may also be a
test or 2 that uses unix-domain sockets. (If not, I should add one. :) ).

Does the loop have any other scheduled tasks?

No.

Does it run in the main thread?

No. (I don't currently call set_event_loop, because I pass the loop around
and nothing was relying on the default event loop. Previously, I was
calling set_event_loop, and the server was hanging then as well.

I'm 70% sure this is a resource exhaustion problem of some sort.

Also worth noting:

• I tried just using uvloop for clients and still got a test failure, one
  of which had to do with running the client within a multiprocessing Process.
• The test servers almost always are run in a multi-Processing process.

Jim

Jim Fulton
http://jimfulton.info

[jimfulton]

On Thu, Jul 7, 2016 at 6:05 PM, Yury Selivanov notifications@github.com
wrote:

Hm, actually, there is only one await in Loop.create_server, and it waits
on getaddrinfo. So it might be a bug in the DNS resolver. What's the
value of addr?

Almost always ('127.0.0.1', 0)

[jimfulton]

On Thu, Jul 7, 2016 at 6:45 PM, Yury Selivanov notifications@github.com
wrote:

And there's also an option to make a debug build with make debug. When
built in debug mode, loop has a print_debug_info() method which outputs a
bunch of useful debug info.

OK, I'll try that. Would you suggest print_debug_info() here:

https://github.com/zopefoundation/ZEO/blob/debug-uvloop/src/ZEO/asyncio/server.py#L226

[1st1]

OK, I'll try that. Would you suggest print_debug_info() here:

The idea is to create a debug task that will periodically print debug info on the screen. This is how I use it in uvloop benchmarks:
https://github.com/MagicStack/uvloop/blob/master/examples/bench/echoserver.py#L78

I'm thinking about adding a manhole or something, that would let you attach to a running event loop from the outside.

[jimfulton]

The program was running out of file descriptors. Not an issue with uvloop at all. Sorry for the noise.

[1st1]

Thanks a lot for confirming this!

[jimfulton]

I was wrong. It has nothing to do with running out of file descriptors. (I was able to make uvloop fail when having too many open files, but that's not what made the tests fail.)

Many of the ZEO tests are integration tests. Typically, these tests run a server in a subprocess using multiprocessing. A few tests run the server in a thread, which makes it possible to introspect server state. Mixing thread-based and multiprocessing-based tests leads to hanging in the test runner when using uvloop.

Here's a minimal script that demonstrates the problem:

https://gist.github.com/jimfulton/59c02a96ebe4720d512b2b0ee426c7ed

I'm not sure if this issue is specific to multiprocessing (I suspect so because magic :) ) or would also be an issue with other ways of running a server in a subprocess.

[1st1]

Interesting. Thanks a lot for the script -- will investigate tomorrow.

[1st1]

Looks like this bug is in libuv. A call to uv_getaddrinfo never calls back the passed callback function. I suspect that there is a lock or something that breaks on fork() + starting a new event loop in the forked process.

Options:

1. Start using Python's getaddrinfo in a Python threadpool.
2. Fix the bug in libuv.
3. Switch to c-ares or getdns (I wanted to do this sometime anyways).

Not sure how to even estimate which option is more preferable here.

cc @saghul

[saghul]

I suspect that there is a lock or something that breaks on fork() + starting a new event loop in the forked process.

This is the case, yes. The threadpool has some global state which needs to be adjusted in the child. Since uv_getaddrinfo runs on the threadpool it will fail. Basically using any loop before fork() and trying to use them (or a new one) is not currently supported.

There is, however, a pull request which addresses this: libuv/libuv#846 more eyes are always welcome!

Start using Python's getaddrinfo in a Python threadpool.

You might want to do this for another reason though: Python's getaddrinfo is not quite like getaddrinfo(3), it makes some normalizations and other stuff I no longer remember but it might be a good idea to use it, for consistency.

[1st1]

I suspect that there is a lock or something that breaks on fork() + starting a new event loop in the forked process.

This is the case, yes. The threadpool has some global state which needs to be adjusted in the child. Since uv_getaddrinfo runs on the threadpool it will fail. Basically using any loop before fork() and trying to use them (or a new one) is not currently supported.

There is, however, a pull request which addresses this: libuv/libuv#846 more eyes are always welcome!

I've seen that PR, it looks good to me (I spent quite a bit of time looking and playing with libuv internals). But I'm not a core libuv committer to say the LGTM :)

Start using Python's getaddrinfo in a Python threadpool.

You might want to do this for another reason though: Python's getaddrinfo is not quite like getaddrinfo(3), it makes some normalizations and other stuff I no longer remember but it might be a good idea to use it, for consistency.

Yes, but I already have all the code to normalize libuv results into what Python code is expecting to see. All that code would be useful for c-ares or getdns though.

@jimfulton How serious is this bug? Does it preclude you from using ZODB with uvloop?

[jimfulton]

The bug is mainly an inconvenience at this point.

Informal performance measurements indicate that most of the win is on the client, so for the moment, I'm not using uvloop on the server. I could use it on the server if I rearrange some tests, which I think I'll probably do. (If it was easy for you to fix, I'd have waited, but I'll go ahead and work around it.)

(Lack of connect_accepted_socket is also preventing using on the server for the main project I'm working on. :) )

[1st1]

I'll try to add connect_accepted_socket to uvloop today :)

[saghul]

I've seen that PR, it looks good to me (I spent quite a bit of time looking and playing with libuv internals). But I'm not a core libuv committer to say the LGTM :)

If you don't mind, please leave a comment, it helps.

All that code would be useful for c-ares or getdns though.

I'm not sure about that. c-ares does not implement getaddrinfo, and I don't know about getdns. You can make shim using A and AAAA queries, but mimicking the system getaddrinfo will be tricky.

[1st1]

@saghul I've left a comment for libuv/libuv#846 -- it solves this particular bug and other bugs we have with os.fork calls. When do you think it can be merged? (I really, really don't want to use patched libuv).

[saghul]

@1st1 thanks! I can't really give an estimate, it's a (relatively) big change, so we are cautious.

[1st1]

@jimfulton FWIW uvloop v0.5.1 has loop.connect_accepted_socket()!

[jimfulton]

Great! I'll try it today.

BTW, I've updated the ZEO tests so they pass with uvloop (when run a certain way that runs groups of tests in separate processes).

[jimfulton]

connect_accepted_socket works great. :) ZEO is now using it in it's multi-threaded-server configuration. Thanks.

[1st1]

Finally some good news ;)

[1st1]

Fixed in master. Will be in the next uvloop release (0.9.0) soon.

[1st1]

Please try uvloop v0.9.0.