socket sendmsg(), recvmsg() methods

[synapse]

BPO	6560
Nosy	@brettcannon, @jcea, @ncoghlan, @pitrou, @jackdied, @giampaolo, @brianmay
Superseder	bpo-12958: test_socket failures on Mac OS X
Files	srmsg.patch: sendmsg()/recvmsg() patch for rev 74186 sendrecvmsgtest.py: my tester script baikie-hwundram.diff: Different patch, with msg_name and scatter/gather baikie-hwundram-v2.diff: New version w/ docs and tests (work in progress) baikie-hwundram-v2.1.diff: Added test fix for non-IPv6 builds baikie-hwundram-v3.diff: New version; added RFC 3542 tests baikie-hwundram-v4.diff: New version; minor changes v4-replace-existing-classes.diff: Optional: replace SocketTCPTest, etc. with the classes from the sendmsg patch baikie-hwundram-v5.diff v5-replace-existing-classes.diff baikie-hwundram-v5-hg.diff baikie-hwundram-v6.diff baikie-hwundram-v6-for-2.x.diff issue6560_ssl_fixes_v1.diff: Attempt 1 at resolving SSL related buildbot failures pass_fds_osx.diff

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/ncoghlan'
closed_at = <Date 2011-09-11.05:46:03.256>
created_at = <Date 2009-07-24.08:30:17.314>
labels = ['type-feature', 'library']
title = 'socket sendmsg(), recvmsg() methods'
updated_at = <Date 2012-04-30.13:53:07.207>
user = 'https://bugs.python.org/synapse'

bugs.python.org fields:

activity = <Date 2012-04-30.13:53:07.207>
actor = 'python-dev'
assignee = 'ncoghlan'
closed = True
closed_date = <Date 2011-09-11.05:46:03.256>
closer = 'ncoghlan'
components = ['Library (Lib)']
creation = <Date 2009-07-24.08:30:17.314>
creator = 'synapse'
dependencies = []
files = ['14556', '14557', '15561', '16417', '16422', '17483', '17659', '17660', '19962', '19963', '22180', '22256', '22257', '23011', '23027']
hgrepos = []
issue_num = 6560
keywords = ['patch']
message_count = 55.0
messages = ['90875', '90876', '90877', '96400', '99842', '99951', '99970', '99972', '99976', '100232', '100317', '100351', '106684', '106797', '106925', '107745', '107746', '123501', '136496', '136595', '136688', '136721', '136725', '137216', '137344', '137376', '137688', '137711', '137719', '142656', '142685', '142690', '142695', '142778', '142780', '142787', '142804', '142809', '142812', '142814', '142815', '142816', '142822', '142874', '142885', '142897', '142901', '142982', '143082', '143412', '143714', '143855', '143933', '143947', '159693']
nosy_count = 17.0
nosy_names = ['brett.cannon', 'jcea', 'exarkun', 'ncoghlan', 'janssen', 'pitrou', 'therve', 'jackdied', 'baikie', 'giampaolo.rodola', 'schmichael', 'synapse', 'wiml', 'neologix', 'rosslagerwall', 'python-dev', 'brian']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = '12958'
type = 'enhancement'
url = 'https://bugs.python.org/issue6560'
versions = ['Python 3.3']

[synapse]

This is the rewritten-from-scratch implementation of the
sendmsg()/recvmsg() methods.
Any comments / suggestions / flames are very welcome. Currently it
supports what I need
and I'm only releasing it because I don't have much time to develop it
further in the
forseeable future (1-2 months). It is rewritten from scratch, using the
python c-api
documents. I've tried my best, but I wouldn't bet that it works
error-free. I'd be glad
if someone could give me a review on what I've might done wrong.

The features that are missing:

• using scatter/gather
• using it with non-stream oriented sockets (doesn't support addresses
  /msg_name/)

These should be very easy to implement. If no one takes up the task of
implementing the
missing features than I'll do it of course, you just have to wait a
little while. Of
course any errors present in the code right now will be fixed.

Thanks

Kalman Gergely

[synapse]

the tester application

[therve]

This is a duplicate (although updated patch) from bug bpo-1194378. It would
still need unit tests...

[baikie]

Hi, I'm afraid there may have been some duplication of effort
here - I set about reworking Heiko Wundram's original patch
(issue bpo-1194378) without knowing about this one. I don't have
unit tests yet either, but I saw this and thought I'd better post
what I had. It includes the items on Kalman's to-do list.

This patch is based on Heiko's original code, but changes the
interface, dropping the special processing of SCM_RIGHTS, etc. (I
think the encoding/decoding would be better handled with separate
functions/classes), and adds source address (msg_name) support
and scatter/gather I/O via sendmsg() and a new recvmsg_into()
method, as well as fixing various bugs and limitations.

Comments/flames welcome too. You'll see a few XXX comments in
the code. One in particular refers to the msg_name value from
recvmsg() on a connected socket; I've said in the docstring that
it is "unspecified" in this case, but it might or might not
contain a valid address, depending on the OS. These methods may
also need to be conditionally compiled if, say, CMSG_SPACE (or
sendmsg/recvmsg?) isn't available somewhere.

I didn't add a facility to receive into only part of a buffer in
recvmsg_into(), like the one in recv_into(), since memoryview
objects make it redundant AFAICT.

[jackdied]

I'm +1 on adding it but it requires unit tests. Also, recvmsg/sendmsg appear to be *Nix only functions so someone with a windows box would need to test this to see if it blows up.

[wiml]

I just ran across yet another implementation of sendmsg support for python sockets, whose feature set seems to complement Kalman Gergely's implementation:
http://www.pps.jussieu.fr/~ylg/PyXAPI by Yves Legrandgerard

(Out of curiosity, people have been submitting requests for, and patches with implementations of, this feature for five or more years now and they're never accepted. Is there some opposition to sendmsg support in Python? It's a missing feature that bites me pretty regularly.)

[jackdied]

I've been digging into the patch. Is there a reason sendmsg() wants an iterable of buffers instead of just accepting a str? The list-of-buffers more closely matches the underlying syscall but I'm not sure what the python benefit is, especially when recvmsg() only returns a single value (it only creates 1 iovec under the covers). Python doesn't have "readv" like methods so making sendmsg/recvmsg work like recv/send (straight strings) seems like the way to go.

Also, the "y*" format character for packing/unpacking tuples is no longer supported - I'm assuming it used to mean buffers.

Does anyone have a good reference for using recvmsg/sendmsg? I read the man pages and googled around but couldn't find anything. I have no experience with using the calls in-the-wild.

[jackdied]

Additional data point: the perl version takes a single scalar (instead of a list of scalars) for use with sendmsg()

http://search.cpan.org/~MJP/Socket-MsgHdr-0.01/MsgHdr.pm

[jackdied]

one of the other sprinters just pointed out that Modules/_multiprocessing.c (py3k branch) uses sendmsg/recvmsg internally to pass file descriptors back and forth. The code is very short and readable.

[baikie]

Thanks for your interest! I'm actually still working on the
patch I posted, docs and a test suite, and I'll post something
soon.

Yes, you could just use b"".join() with sendmsg() (and get
slightly annoyed because it doesn't accept buffers ;) ). I made
sendmsg() take multiple buffers because that's the way the system
call works, but also to match recvmsg_into(), which gives you the
convenience of being able to receive part of the message into a
bytearray and part into an array.array("i"), say, if that's how
the data is formatted.

As you might know, gather-write with sendmsg() can give a
performance benefit by letting the kernel assemble the message
while copying the data from userspace rather than having
userspace copy the data once to form the message and then having
the kernel copy it again when the system call is made. I suppose
with Python you just need a larger message to see the benefit :)
Since it can read from buffers, though, socket.sendmsg() can pull
a large chunk of data straight out of an mmap object, say, and
attach headers from a bytes object without the mmapped data being
touched by Python at all (or even entering userspace, in this
case).

The patch is for 3.x, BTW - "y*" is valid there (and does take a
buffer).

As for a good reference, I haven't personally seen one. There's
POSIX and RFC 3542, but they don't provide a huge amount of
detail. Perhaps the (updated) W. Richard Stevens networking
books? I've got the Stevens/Rago second edition of Advanced
Programming in the Unix Environment, which discusses FD and
credential passing with sendmsg/recvmsg, but not very well (it
misuses CMSG_LEN, for one thing). The networking books were
updated by different people though, so perhaps they do better.

The question of whether to use CMSG_NXTHDR() to step to the next
header when constructing the buffer for sendmsg() is a bit murky,
in particular. I've assumed that this is the way to do it since
the examples in RFC 3542 (and most of the code I've seen
generally) use CMSG_FIRSTHDR() to get the initial pointer, but
I've found that glibc's CMSG_NXTHDR() can (wrongly, I think)
return NULL if the buffer hasn't been zero-filled beforehand
(this causes segfaults with the patch I initially posted).

@wim:

Yes, the rfc3542 module from that package looks as if it would be
usable with these patches - although it's Python 2-only, GPL-only
and looks unmaintained. Those kind of ancillary data
constructors will actually be needed to make full portable use of
sendmsg() and recvmsg() for things like IPv6, SCTP, Linux's
socket error queues, etc. The same goes for data for the
existing get/setsockopt() methods, in fact - the present
suggestion to use the struct module is pretty inadequate when
there are typedefs involved and implementations might add and
reorder fields, etc.

The objects in that package seem a bit overcomplicated, though,
messing about with setter methods instead of just subclassing
"bytes" and having different constructors to create the object
from individual arguments or received bytes (say, ucred(1, 2, 3)
or ucred.from_bytes(...)).

Maybe the problem of testing patches well has been putting people
off so far? Really exercising the system's CMSG_*HDR() macros in
particular isn't entirely straightforward. I suppose there's
also a reluctance to write tests while still uncertain about how
to present the interface - that's another reason why I went for
the most general multiple-buffer form of sendmsg()!

[baikie]

OK, here's a new version as a work in progress. A lot of the new
stuff is uncommented (particularly the support code for the
tests), but there are proper docs this time and a fairly complete
test suite (but see below).

There are a couple of changes to the interface (hopefully the
last). The recvmsg() methods no longer receive ancillary data by
default, since calling them on an AF_UNIX socket with the old
default buffer could allow a malicious sender to send unwanted
file descriptors up to receiver's resource limit, and in a
multi-threaded program, another thread could then be prevented
from opening new file descriptors before the receiving thread had
a chance to close the unwanted ones.

Since the ancillary buffer size argument is now more likely to
need a value, I've moved it to second place; the basic argument
order is now the same as in Kalman Gergely's patch. CMSG_LEN()
and CMSG_SPACE() are now provided.

I've also used socket.error instead of ValueError when rejecting
some buffer object/array for being too big to handle, since the
system call itself might cause socket.error to be raised for a
smaller (oversized) object, failing with EMSGSIZE or whatever.

The code is now much more paranoid about checking the results of
the CMSG_*() macros, and will raise RuntimeError if it finds its
assumptions are not met.

I'd still like to add tests for receiving some of the RFC 3542
ancillary data items, especially since the SCM_RIGHTS tests can't
always (ever?) test recvmsg() with multiple items (if you send
two FD arrays, the OS can coalesce them into a single array
before delivering them).

[baikie]

I just found that the IPv6 tests don't get skipped when IPv6 is
available but disabled in the build - you can create IPv6
sockets, but not use them :/ This version fixes the problem.

[baikie]

Here is a new version of the patch; I've added some tests which
use the RFC 3542 interface (IPv6 advanced API) and am now quite
happy with it generally.

As well as Linux, I've tested it on an old (unsupported) FreeBSD
5.3 installation, which required a few changes in the tests and
the code, but is probably representative of many socket
implementations. Testing on other systems would be appreciated!

The main issue was that when truncating ancillary data, FreeBSD
seemed to just copy as much into the buffer as would fit and did
not adjust the last item's cmsg_len member to reflect the amount
of data that was actually present, so the item would appear to
extend past the end of the buffer. The last version of the patch
detected this and raised RuntimeError, which prevented any
truncated receives from succeeding. The new version instead
issues a warning and returns as much of the last item as is in
the buffer.

The warning could perhaps be disabled for systems like this,
given that it happens every time ancillary data is truncated, but
truncation generally shouldn't happen in a program's normal
operation, and on other platforms a bad cmsg_len value might
indicate that the returned data is actually incorrect in some
way.

After some investigation, I've stuck with using CMSG_FIRSTHDR()
and CMSG_NXTHDR() to step through the headers when assembling the
ancillary data in sendmsg().

The KAME IPv6 userspace utilities at [1] include several programs
which send multiple control messages at once, and these always
use CMSG_NXTHDR() to advance to the next uninitialized header,
while some (but not all) of them zero-fill the buffer beforehand,
suggesting they ran into the issue with glibc's macros returning
NULL (KAME developed the BSD IPv6 stack, and the zero-filling
isn't necessary with the BSD macros).

The alternative would be to add CMSG_SPACE(size) to the pointer
to get to the next header. Going by the diagram in RFC 3542,
that should be equivalent, but if some system defined
CMSG_SPACE(len) as, say, CMSG_LEN(len) + 3, instead of
(CMSG_LEN(len) + 3) & ~3, it would probably go unnoticed until
someone tried to use CMSG_SPACE() that way. So given the KAME
example, I think using CMSG_NXTHDR() with a zero-filled buffer is
the way to go.

[1] http://www.kame.net/dev/cvsweb2.cgi/kame/kame/kame/