There is a few system I/O calls in the mmap module that
can take some time. Would be be possible to put
Py_BEGIN_ALLOW_THREADS/Py_END_ALLOW_THREADS around
these to release the GIL and allow other Python threads
to do their work?

Logged In: YES 
user_id=341410

Can you come up with the listing of code blocks where it
would make sense to allow threads, or even provide a patch?

I don't know mmap myself but patches are welcome.

As soon as you're dealing with files (not anonymous mapping), you can get the same type of latency than when using open/read/write...
While it's probably not worth the trouble to release the GIL for every operation involving mmaped-files, there are a couple places where it should be considered, for example:
Modules/mmapmodule.c:new_mmap_object
#ifdef HAVE_FSTAT
#  ifdef __VMS
	/* on OpenVMS we must ensure that all bytes are written to the file */
	if (fd != -1) {
	        fsync(fd);
	}
#  endif

fsync() can take up to a couple seconds, so we should probably allow threads here (this is done only on VMS).
Another place worth looking at is when using msync(), like in mmap_object_dealloc():
	if (m_obj->data!=NULL) {
		msync(m_obj->data, m_obj->size, MS_SYNC);
		munmap(m_obj->data, m_obj->size);
	}

or in mmap_flush_method():
	if (-1 == msync(self->data + offset, size, MS_SYNC)) {
		PyErr_SetFromErrno(mmap_module_error);
		return NULL;
	}

msycn too can take quite a long time to complete.

I can write a patch if someone's willing to review it.

This shouldn't really care about VMS here. As for the rest, feel free to propose a patch :)
Please note that most non-trivial system calls, such as ftruncate(), mremap(), even mmap() itself, deserve to be enclosed in Py_BEGIN_ALLOW_THREADS/Py_END_ALLOW_THREADS pairs.

However, you'll have to be careful that the internal state of the mmap object remains consistent, such that using it from several threads doesn't crash the interpreter.
(you can even add multi-threaded unit tests if you want to be sure of this)

> This shouldn't really care about VMS here.

Correction: /we/ shouldn't really care about VMS here.
And the reason being, of course, that we have neither developers nor
known testers under VMS.
(we don't even know if the current trunk builds there or not)

Any news? If a patch is not ready, i can work on a patch too.

I'll add one more system I/O call that's not GIL-wrapped in the mmap module that can take some time:  mmap itself.

mmap on Linux with MAP_POPULATE (0x8000) as the flags can take quite a bit of time.  That's the flag that prefaults the memory range.  MAP_POPULATE has been around since Linux 2.5.46.

I know that MAP_POPULATE isn't explicitly supported in the module, but mmap.mmap does take arbitrary flags, so it isn't exactly unsupported either.