Explanation
-----------
The problem here is flawed `poll()` implementation. When it tries to
see if pipe can be written without blocking, it eventually calls
`NtQueryInformationFile()` and tests `WriteQuotaAvailable`. However,
the meaning of quota was misunderstood. The value of quota is reduced
when either some data was written to a pipe, *or* there is a pending
read on the pipe. Therefore, if there is a pending read of size >= than
the pipe's buffer size, poll() will think that pipe is not writable and
will hang forever, usually that means deadlocking both pipe users.
I have studied the problem and found that Windows pipes track two values:
`QuotaUsed` and `BytesInQueue`. The code in `poll()` apparently wants to
know `BytesInQueue` instead of quota. Unfortunately, `BytesInQueue` can
only be requested from read end of the pipe, while `poll()` receives
write end.
The git's implementation of `poll()` was copied from gnulib, which also
contains a flawed implementation up to today.
I also had a look at implementation in cygwin, which is also broken in a
subtle way. It uses this code in `pipe_data_available()`:
fpli.WriteQuotaAvailable = (fpli.OutboundQuota - fpli.ReadDataAvailable)
However, `ReadDataAvailable` always returns 0 for the write end of the pipe,
turning the code into an obfuscated version of returning pipe's total
buffer size, which I guess will in turn have `poll()` always say that pipe
is writable. The commit that introduced the code doesn't say anything about
this change, so it could be some debugging code that slipped in.
These are the typical sizes used in git:
0x2000 - default read size in `strbuf_read()`
0x1000 - default read size in CRT, used by `strbuf_getwholeline()`
0x2000 - pipe buffer size in compat\mingw.c
As a consequence, as soon as child process uses `strbuf_read()`,
`poll()` in parent process will hang forever, deadlocking both
processes.
This results in two observable behaviors:
1) If parent process begins sending STDIN quickly (and usually that's
the case), then first `poll()` will succeed and first block will go
through. MAX_IO_SIZE_DEFAULT is 8MB, so if STDIN exceeds 8MB, then
it will deadlock.
2) If parent process waits a little bit for any reason (including OS
scheduler) and child is first to issue `strbuf_read()`, then it will
deadlock immediately even on small STDINs.
The problem is illustrated by `git stash push`, which will currently
read the entire patch into memory and then send it to `git apply` via
STDIN. If patch exceeds 8MB, git hangs on Windows.
Possible solutions
------------------
1) Somehow obtain `BytesInQueue` instead of `QuotaUsed`
I did a pretty thorough search and didn't find any ways to obtain
the value from write end of the pipe.
2) Also give read end of the pipe to `poll()`
That can be done, but it will probably invite some dirty code,
because `poll()`
* can accept multiple pipes at once
* can accept things that are not pipes
* is expected to have a well known signature.
3) Make `poll()` always reply "writable" for write end of the pipe
Afterall it seems that cygwin (accidentally?) does that for years.
Also, it should be noted that `pump_io_round()` writes 8MB blocks,
completely ignoring the fact that pipe's buffer size is only 8KB,
which means that pipe gets clogged many times during that single
write. This may invite a deadlock, if child's STDERR/STDOUT gets
clogged while it's trying to deal with 8MB of STDIN. Such deadlocks
could be defeated with writing less than pipe's buffer size per
round, and always reading everything from STDOUT/STDERR before
starting next round. Therefore, making `poll()` always reply
"writable" shouldn't cause any new issues or block any future
solutions.
4) Increase the size of the pipe's buffer
The difference between `BytesInQueue` and `QuotaUsed` is the size
of pending reads. Therefore, if buffer is bigger than size of reads,
`poll()` won't hang so easily. However, I found that for example
`strbuf_read()` will get more and more hungry as it reads large inputs,
eventually surpassing any reasonable pipe buffer size.
Chosen solution
---------------
Make `poll()` always reply "writable" for write end of the pipe.
Hopefully one day someone will find a way to implement it properly.
Reproduction
------------
printf "%8388608s" X >large_file.txt
git stash push --include-untracked -- large_file.txt
I have decided not to include this as test to avoid slowing down the
test suite. I don't expect the specific problem to come back, and
chances are that `git stash push` will be reworked to avoid sending the
entire patch via STDIN.
Signed-off-by: Alexandr Miloslavskiy <alexandr.miloslavskiy@syntevo.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>