When fetching with the v0 protocol over ssh (or a local upload-pack with
pipes), the server closes the connection as soon as it is finished
sending the pack. So even though the client may still be operating on
the data via index-pack (e.g., resolving deltas, checking connectivity,
etc), the server has released all resources.
With the v2 protocol, however, the server considers the ssh session only
as a transport, with individual requests coming over it. After sending
the pack, it goes back to its main loop, waiting for another request to
come from the client. As a result, the ssh session hangs around until
the client process ends, which may be much later (because resolving
deltas, etc, may consume a lot of CPU).
This is bad for two reasons:
- it's consuming resources on the server to leave open a connection
that won't see any more use
- if something bad happens to the ssh connection in the meantime (say,
it gets killed by the network because it's idle, as happened in a
real-world report), then ssh will exit non-zero, and we'll propagate
the error up the stack.
The server is correct here not to hang up after serving the pack. The v2
protocol's design is meant to allow multiple requests like this, and
hanging up would be the wrong thing for a hypothetical client which was
planning to make more requests (though in practice, the git.git client
never would, and I doubt any other implementations would either).
The right thing is instead for the client to signal to the server that
it's not interested in making more requests. We can do that by closing
the pipe descriptor we use to write to ssh. This will propagate to the
server upload-pack as an EOF when it tries to read the next request (and
then it will close its half, and the whole connection will go away).
It's important to do this "half duplex" shutdown, because we have to do
it _before_ we actually receive the pack. This is an artifact of the way
fetch-pack and index-pack (or unpack-objects) interact. We hand the
connection off to index-pack (really, a sideband demuxer which feeds
it), and then wait until it returns. And it doesn't do that until it has
resolved all of the deltas in the pack, even though it was done reading
from the server long before.
So just closing the connection fully after index-pack returns would be
too late; we'd have held it open much longer than was necessary. And
teaching index-pack to close the connection is awkward. It's not even
seeing the whole conversation (the sideband demuxer is, but it doesn't
actually know what's in the packets, or when the end comes).
Note that this close() is happening deep within the transport code. It's
possible that a caller would want to perform other operations over the
same ssh transport after receiving the pack. But as of the current code,
none of the callers do, and there haven't been discussions of any plans
to change this. If we need to support that later, we can probably do so
by passing down a flag for "you're the last request on the transport;
it's OK to close" instead of the code just assuming that's true.
The description above all discusses v2 ssh, so it's worth thinking about
how this interacts with other protocols:
- in v0 protocols, we could do the same half-duplex shutdown (it just
goes into the v0 do_fetch_pack() instead). This does work, but since
it doesn't have the same persistence problem in the first place,
there's little reason to change it at this point.
- local fetches against git-upload-pack on the same machine will
behave the same as ssh (they are talking over two pipes, and see EOF
on their input pipe)
- fetches against git-daemon will run this same code, and close one of
the descriptors. In practice, this won't do anything, since there
our two descriptors are dups of each other, and not part of a
half-duplex pair. The right thing would probably be to call
shutdown(SHUT_WR) on it. I didn't bother with that here. It doesn't
face the same error-code problem (since it's just a TCP connection),
so it's really only an optimization problem. And git:// is not that
widely used these days, and has less impact on server resources than
an ssh termination.
- v2 http doesn't suffer from this problem in the first place, as our
pipes terminate at a local git-remote-https, which is passing data
along as individual requests via curl. Probably curl is keeping the
TCP/TLS connection open for more requests, and we might be able to
tell it manually "hey, we are done making requests now". But I think
that's much less important. It again doesn't suffer from the
error-code problem, and HTTP keepalive is pretty well understood
(importantly, the timeouts can be set low, because clients like curl
know how to reconnect for subsequent requests if necessary). So it's
probably not worth figuring out how to tell curl that we're done
(though if we do, this patch is probably the first step anyway;
fetch-pack closes the pipe back to remote-https, which would be the
signal that it should tell curl we're done).
The code is pretty straightforward. We close the pipe at the right
moment, and set it to -1 to mark it as invalid. I modified the later
cleanup code to avoid calling close(-1). That's not strictly necessary,
since close(-1) is a noop, but hopefully makes things a bit more obvious
to a reader.
I suspect that trying to call more transport functions after the close()
(e.g., calling transport_fetch_refs() again) would fail, as it's not
smart enough to realize we need to re-open the ssh connection. But
that's already true when v0 is in use. And no current callers want to do
that (and again, the solution is probably a flag in the transport code
to keep things open, which can be added later).
There's no test here, as the situation it covers is inherently racy (the
question is when upload-pack exits, compared to when index-pack finishes
resolving deltas and exits). The rather gross shell snippet below does
recreate the problematic situation; when run on a sufficiently-large
repository (git.git works fine), it kills an "idle" upload-pack while
the client is resolving deltas, leading to a failed clone.
(
git clone --no-local --progress . foo.git 2>&1
echo >&2 "clone exit code=$?"
) |
tr '\r' '\n' |
while read line
do
case "$done,$line" in
,Resolving*)
echo "hit resolving deltas; killing upload-pack"
killall -9 git-upload-pack
done=t
;;
esac
done
Reported-by: Greg Pflaum <greg.pflaum@pnp-hcl.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Git - fast, scalable, distributed revision control system
Git is a fast, scalable, distributed revision control system with an
unusually rich command set that provides both high-level operations
and full access to internals.
Git is an Open Source project covered by the GNU General Public
License version 2 (some parts of it are under different licenses,
compatible with the GPLv2). It was originally written by Linus
Torvalds with help of a group of hackers around the net.
Please read the file INSTALL for installation instructions.
Many Git online resources are accessible from https://git-scm.com/
including full documentation and Git related tools.
See Documentation/gittutorial.txt to get started, then see
Documentation/giteveryday.txt for a useful minimum set of commands, and
Documentation/git-<commandname>.txt for documentation of each command.
If git has been correctly installed, then the tutorial can also be
read with man gittutorial or git help tutorial, and the
documentation of each command with man git-<commandname> or git help <commandname>.
CVS users may also want to read Documentation/gitcvs-migration.txt
(man gitcvs-migration or git help cvs-migration if git is
installed).
Issues which are security relevant should be disclosed privately to
the Git Security mailing list git-security@googlegroups.com.
The maintainer frequently sends the "What's cooking" reports that
list the current status of various development topics to the mailing
list. The discussion following them give a good reference for
project status, development direction and remaining tasks.
The name "git" was given by Linus Torvalds when he wrote the very
first version. He described the tool as "the stupid content tracker"
and the name as (depending on your mood):
random three-letter combination that is pronounceable, and not
actually used by any common UNIX command. The fact that it is a
mispronunciation of "get" may or may not be relevant.
stupid. contemptible and despicable. simple. Take your pick from the
dictionary of slang.
"global information tracker": you're in a good mood, and it actually
works for you. Angels sing, and a light suddenly fills the room.
"goddamn idiotic truckload of sh*t": when it breaks
ae1a7eefff
