This is a follow up of issue 16853 (see comment http://bugs.python.org/issue16853#msg196984).

Why can't this be fixed in 3.4.1? It isn't an API change.

selectors.EpollSelector.modify() should use epoll.modify() instead of unregister()+register().

Here is a patch:

Issue #18932, selectors: Optimize the modify() method of selectors

Optimize also register() and unregister() methods of KqueueSelector: only call kqueue.control() once.

I tested SelectSelector, PollSelector, EpollSelector on Linux. I ran tests on FreeBSD, so also tested KqueueSelector. I didn't test DevpollSelector, but the code should be identical to PollSelector (the API is the same).

_BaseSelectorImpl.modify() still calls unregister() and register(). To my understanding the whole point of this proposal was to avoid that in order to obtain the speedup and (possibly) avoid race conditions.

Well, I'd like to see at least one benchmark.

Benchmark on Fedora 21 (Linux 3.18.3, glibc 2.20, Python 3.5 rev 7494f3972726).

Original:

haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.SelectSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 43.7 usec per loop
haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.PollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 45.1 usec per loop
haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.EpollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 48.4 usec per loop

Patched:

haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.SelectSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 37.2 usec per loop
haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.PollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 40.5 usec per loop
haypo@selma$ ./python -m timeit -s 'import os, selectors; s=selectors.EpollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
10000 loops, best of 3: 39.9 usec per loop

* SelectSelector: 14% faster (-6.5 us)
* PollSelector: 10% faster (-4.6 us)
* EpollSelector: 18% faster (-8.5 us)

--

> _BaseSelectorImpl.modify() still calls unregister() and register().

I chose to factorize code in modify() and only put the specific code in _modify(). The advantage is to pass directly oldkey and key to _modify().

We may even try to keep the selector consistent if _modify() fails. For example, start with unregister, and only register if _modify() succeed.

--

In a previous (local) patch, the default implementation of _modify() was:

    self.unregister(fileobj)
    return self.register(fileobj, events, data)

And each specialized _modify() started with:

    oldkey = self.unregister(fileobj)
    key = self.register(fileobj, events, data)

Do you prefer this?

Updated patch with a less surprising _modify() method:
- the default implementation simply calls unregister() + register(), as before
- drop SelectSelector._modify(): calling unregister() + register() is just fine

> - drop SelectSelector._modify(): calling unregister() + register() is just fine

I checked with strace: PollSelector.modify() doesn't require any syscall, so I propose to also drop it (to just call unregister + register). What do you think?

I would like to reduce the number of syscalls, calling a C function of the glibc is cheap.

> This is a follow up of issue 16853 (see comment http://bugs.python.org/issue16853#msg196984).

"(...) the problem with unregister() + register() vs a real
modify (e.g. EPOLL_CTL_MOD) is that it's subject to a race condition,
if an event fires during the short window where the FD isn't
registered anymore."

The race condition only occurs with selectors which have a state machine in the kernel: epoll, kqueue, devpoll. SelectSelector and PollSelector are state-less (in the kernel, the state is build for a single syscall and then destroyed). Am I right?

I mean something like this (for epoll), as in avoiding to call unregister() / register() at all:


diff -r 017e7391ab58 Lib/selectors.py
--- a/Lib/selectors.py  Wed Feb 04 08:37:02 2015 -0800
+++ b/Lib/selectors.py  Thu Feb 05 18:42:26 2015 +0100
@@ -412,6 +412,23 @@
                 pass
             return key
 
+        def modify(self, fileobj, events, data=None):
+            try:
+                key = self._fd_to_key[self._fileobj_lookup(fileobj)]
+            except KeyError:
+                raise KeyError("{!r} is not registered".format(fileobj)) from None
+            if data != key.data:
+                key = key._replace(data=data)
+                self._fd_to_key[key.fd] = key
+            if events != key.events:
+                epoll_events = 0
+                if events & EVENT_READ:
+                    epoll_events |= select.POLLIN
+                if events & EVENT_WRITE:
+                    epoll_events |= select.POLLOUT
+                self._epoll.modify(key.fd, epoll_events)
+            return key
+
         def select(self, timeout=None):
             if timeout is None:
                 timeout = -1


Before the patch:

~/svn/python/3.5$ ./python -m timeit -s 'import os, selectors; s=selectors.EpollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
100000 loops, best of 3: 14.7 usec per loop


After the patch:

~/svn/python/3.5$ ./python -m timeit -s 'import os, selectors; s=selectors.EpollSelector(); r,w=os.pipe(); s.register(r, selectors.EVENT_READ)' 's.modify(r, selectors.EVENT_WRITE); s.modify(r, selectors.EVENT_READ)'
100000 loops, best of 3: 3.07 usec per loop

That's equal to about a 4.5x (+450%) speedup if I'm not mistaken.

2015-02-05 18:46 GMT+01:00 Giampaolo Rodola' <report@bugs.python.org>:
> +        def modify(self, fileobj, events, data=None):
> +            ...
> +            if events != key.events:

You should update the SelectorKey in the case, no?

Yes, it appears so. I would suggest to come up with a patch which overrides modify() for all those selectors which can benefit from it (poll, epoll, kqueue) and refactor register() / unregister() (for select) in a separate issue / patch. Also, I suppose a couple of tests should be added (if not there already) in order to make sure that 'events' and 'data' arguments get actually changed in different situations.

@Giampaolo: I'm not sure that I understood your proposition? Can you please write a patch?

Attached is a patch (for epoll only) which updates the SelectorKey. This introduces a considerable slowdown compared to my first patch:

no patch:    16.2 usec per loop
your patch:  12.4 usec per loop
my patch:    10.9 usec per loop
first patch: 3.07 usec per loop

The culprit is the new nameduple's _replace() call. It's a shame a single line adds such a great slowdown. I think it makes sense to investigate whether we can introduce a faster replace mechanism for namedtuple.

Does anyone have a realistic use case where modify() is actually a
non-negligible part?

This was implemented in https://bugs.python.org/issue30014. Closing as duplicate.