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>
Windows update.
* js/mingw-use-utf8:
mingw: fix possible buffer overrun when calling `GetUserNameW()`
mingw: use Unicode functions explicitly
mingw: get pw_name in UTF-8 format
Many Win32 API functions actually exist in two variants: one with
the `A` suffix that takes ANSI parameters (`char *` or `const char *`)
and one with the `W` suffix that takes Unicode parameters (`wchar_t *`
or `const wchar_t *`).
The ANSI variant assumes that the strings are encoded according to
whatever is the current locale. This is not what Git wants to use on
Windows: we assume that `char *` variables point to strings encoded in
UTF-8.
There is a pseudo UTF-8 locale on Windows, but it does not work
as one might expect. In addition, if we overrode the user's locale, that
would modify the behavior of programs spawned by Git (such as editors,
difftools, etc), therefore we cannot use that pseudo locale.
Further, it is actually highly encouraged to use the Unicode versions
instead of the ANSI versions, so let's do precisely that.
Note: when calling the Win32 API functions _without_ any suffix, it
depends whether the `UNICODE` constant is defined before the relevant
headers are #include'd. Without that constant, the ANSI variants are
used. Let's be explicit and avoid that ambiguity.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The return type of the `GetProcAddress()` function is `FARPROC` which
evaluates to `long long int (*)()`, i.e. it cannot be cast to the
correct function signature by GCC 8.
To work around that, we first cast to `void *` and go on with our merry
lives.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The value of timeout starts as an int value, and for this reason it
cannot overflow unsigned long long aka ULONGLONG. The unsigned version
of this initial value is available in orig_timeout. The difference
(orig_timeout - elapsed) cannot wrap around because it is protected by
a conditional (as can be seen in the patch text). Hence, the ULONGLONG
difference can only have values that are smaller than the initial
timeout value and truncation to int cannot overflow.
Signed-off-by: Steve Hoelzer <shoelzer@gmail.com>
[j6t: improved both implementation and log message]
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Windows Vista (and later) actually have a working poll(), but we still
cannot use it because it only works on sockets.
So let's detect when we are targeting Windows Vista and undefine those
constants, and define `pollfd` so that we can declare our own pollfd
struct.
We also need to make sure that we override those constants *after*
`winsock2.h` has been `#include`d (otherwise we would not really
override those constants).
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The mailing address for the FSF has changed over the years. Rather than
updating the address across all files, refer readers to gnu.org, as the
GNU GPL documentation now suggests for license notices. The mailing
address is retained in the full license files (COPYING and LGPL-2.1).
The old address is still present in t/diff-lib/COPYING. This is
intentional, as the file is used in tests and the contents are not
expected to change.
Signed-off-by: Todd Zullinger <tmz@pobox.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Match what is done to pfd[i].revents when compute_revents() returns
0 to the upstream gnulib's commit d42461c3 ("poll: fixes for large
fds", 2015-02-20). The revents field is set to 0, without
incrementing the value rc to be returned from the function. The
original code left the field to whatever random value the field was
initialized to.
This fixes occasional hangs in git-upload-pack on HPE NonStop.
Signed-off-by: Randall S. Becker <randall.becker@nexbridge.ca>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
HANDLE is defined internally as a void *, but in many cases it is
actually guaranteed to be a 32-bit integer. In these cases, GCC should
not warn about a cast of a pointer to an integer of a different type
because we know exactly what we are doing.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ensure that when passing a pipe, the gnulib poll replacement will not
return 0 before the timeout has passed.
Not obeying the timeout (and merely returning 0) causes pathological
behavior when preparing a packfile for a repository and taking a
long time to do so. If poll were to return 0 immediately, this would
cause keep-alives to get sent as quickly as possible until the packfile
was created. Such deviance from the standard would cause megabytes (or
more) of keep-alive packets to be sent.
GetTickCount is used as it is efficient, stable and monotonically
increasing. (Neither GetSystemTime nor QueryPerformanceCounter have
all three of these properties.)
Signed-off-by: Edward Thomson <ethomson@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
SwitchToThread() only gives away the rest of the current time slice
to another thread in the current process. So if the thread that feeds
the file decscriptor we're polling is not in the current process, we
get busy-waiting.
I played around with this quite a bit. After trying some more complex
schemes, I found that what worked best is to just sleep 1 millisecond
between iterations. Though it's a very short time, it still completely
eliminates the busy wait condition, without hurting perf.
There code uses SleepEx(1, TRUE) to sleep. See this page for a good
discussion of why that is better than calling SwitchToThread, which
is what was used previously:
http://stackoverflow.com/questions/1383943/switchtothread-vs-sleep1
Note that calling SleepEx(0, TRUE) does *not* solve the busy wait.
The most striking case was when testing on a UNC share with a large repo,
on a single CPU machine. Without the fix, it took 4 minutes 15 seconds,
and with the fix it took just 1:08! I think it's because git-upload-pack's
busy wait was eating the CPU away from the git process that's doing the
real work. With multi-proc, the timing is not much different, but tons of
CPU time is still wasted, which can be a killer on a server that needs to
do bunch of other things.
I also tested the very fast local case, and didn't see any measurable
difference. On a big repo with 4500 files, the upload-pack took about 2
seconds with and without the fix.
[jc: this was first accepted in msysgit tree in May 2012 via a pull
request and Paolo Bonzini has also accepted the same fix to Gnulib
around the same time; see $gmane/247518 for a bit more detail]
Signed-off-by: Stepan Kasal <kasal@ucw.cz>
Acked-by: Johannes Sixt <j6t@kdbg.org>
Acked-by: Erik Faye-Lund <kusmabite@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With MinGW runtime version 4.0 this interferes with the previous
definition from sdkddkver.h.
Signed-off-by: Sebastian Schuberth <sschuberth@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Sparse issues an 'Using plain integer as NULL pointer' warning when
passing the constant '0' as the second parameter in the call to the
WSAEventSelect() function. The function parameter has a pointer type
(WSAEVENT, aka HANDLE, aka void *) so that, in order to suppress the
warning, we simply pass NULL for that parameter in the function call
expression.
Signed-off-by: Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This way it just got added to gnulib too the other day.
Signed-off-by: Joachim Schmitz <jojo@schmitz-digital.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If poll() is used as a milli-second sleep, like in help.c, by passing a NULL
in the 1st and a 0 in the 2nd arg, it exits with EFAULT.
As per Paolo Bonzini, the original author, this is a bug and to be fixed
Like in this commit, which is not to exit if the 2nd arg is 0. It got fixed
In gnulib in the same manner the other day.
Signed-off-by: Joachim Schmitz <jojo@schmitz-digital.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In order for non-win32 platforms to be able to use poll.c, #ifdef the
inclusion of two header files properly
Signed-off-by: Joachim Schmitz <jojo@schmitz-digital.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
move poll.[ch] out of compat/win32/ into compat/poll/ and adjust
Makefile with the changed paths. Adding comments to Makefile about
how/when to enable it and add logic for this
Signed-off-by: Joachim Schmitz <jojo@schmitz-digital.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Alexandr Miloslavskiy
Junio C Hamano
Johannes Schindelin
Steve Hoelzer
Todd Zullinger
Randall S. Becker
Edward Thomson
Theodore Leblond
Sebastian Schuberth
Ramsay Jones
Joachim Schmitz