This gives us a significant speedup when adding, committing and stat'ing files.
Also, since Windows doesn't really handle symlinks, we let stat just uses lstat.
We also need to replace fstat, since our implementation and the standard stat()
functions report slightly different timestamps, possibly due to timezones.
We simply report UTC in our implementation, and do our FILETIME to time_t
conversion based on the document at http://support.microsoft.com/kb/167296.
With Moe's repo structure (100K files in 100 dirs, containing 2-4 bytes)
mkdir bummer && cd bummer; for ((i=0;i<100;i++)); do
mkdir $i && pushd $i;
for ((j=0;j<1000;j++)); do echo "$j" >$j; done;
popd;
done
We get the following performance boost:
With normal lstat & stat Custom lstat/fstat
------------------------ ------------------------
Command: git init Command: git init
------------------------ ------------------------
real 0m 0.047s real 0m 0.063s
user 0m 0.031s user 0m 0.015s
sys 0m 0.000s sys 0m 0.015s
------------------------ ------------------------
Command: git add . Command: git add .
------------------------ ------------------------
real 0m19.390s real 0m12.031s 1.6x
user 0m 0.015s user 0m 0.031s
sys 0m 0.030s sys 0m 0.000s
------------------------ ------------------------
Command: git commit -a.. Command: git commit -a..
------------------------ ------------------------
real 0m30.812s real 0m16.875s 1.8x
user 0m 0.015s user 0m 0.015s
sys 0m 0.000s sys 0m 0.015s
------------------------ ------------------------
3x Command: git-status 3x Command: git-status
------------------------ ------------------------
real 0m11.860s real 0m 5.266s 2.2x
user 0m 0.015s user 0m 0.015s
sys 0m 0.015s sys 0m 0.015s
real 0m11.703s real 0m 5.234s
user 0m 0.015s user 0m 0.015s
sys 0m 0.000s sys 0m 0.000s
real 0m11.672s real 0m 5.250s
user 0m 0.031s user 0m 0.015s
sys 0m 0.000s sys 0m 0.000s
------------------------ ------------------------
Command: git commit... Command: git commit...
(single file) (single file)
------------------------ ------------------------
real 0m14.234s real 0m 7.735s 1.8x
user 0m 0.015s user 0m 0.031s
sys 0m 0.000s sys 0m 0.000s
Signed-off-by: Marius Storm-Olsen <mstormo_git@storm-olsen.com>
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
The problem with Windows's own implementation is that it tries to be
clever when a console program is invoked from a GUI application: In this
case it sometimes automatically allocates a new console window. As a
consequence, the IO channels of the spawned program are directed to the
console, but the invoking application listens on channels that are now
directed to nowhere.
In this implementation we use the lowlevel facilities of CreateProcess(),
which offers a flag to tell the system not to open a console. As a side
effect, only stdin, stdout, and stderr channels will be accessible from
C programs that are spawned. Other channels (file handles, pipe handles,
etc.) are still inherited by the spawned program, but it doesn't get
enough information to access them.
Johannes Schindelin integrated path quoting and unified the various
*execv* and *spawnv* helpers. Eric Raible suggested to also quote '{'.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
gethostbyname() is the first function that calls into the Winsock library,
and it is wrapped only to initialize the library.
socket() is wrapped for two reasons:
- Windows's socket() creates things that are like low-level file handles,
and they must be converted into file descriptors first.
- And these handles cannot be used with plain ReadFile()/WriteFile()
because they are opened for "overlapped IO". We have to use WSASocket()
to create non-overlapped IO sockets.
connect() must be wrapped because Windows's connect() expects the low-level
sockets, not file descriptors, and we must first unwrap the file descriptor
before we can pass it on to Windows's connect().
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
This emulation of poll() is by far not general. It assumes that the
fds that are to be waited for are connected to pipes. The pipes are
polled in a loop until data becomes available in at least one of them.
If only a single fd is waited for, the implementation actually does
not wait at all, but assumes that a subsequent read() will block.
In order not to needlessly burn CPU time, the CPU is yielded to other
processes before the next round in the poll loop using Sleep(0). Note that
any sleep timeout greater than zero will reduce the efficiency by a
magnitude.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
On Windows, we have spawnv() variants to run a child process instead of
fork()/exec(). In order to attach pipe ends to stdin, stdout, and stderr,
we have to use this idiom:
save1 = dup(1);
dup2(pipe[1], 1);
spawnv();
dup2(save1, 1);
close(pipe[1]);
assuming that the descriptors created by pipe() are not inheritable.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
On Unix the idiom to use a pipe is as follows:
pipe(fd);
pid = fork();
if (!pid) {
dup2(fd[1], 1);
close(fd[1]);
close(fd[0]);
...
}
close(fd[1]);
i.e. the child process closes the both pipe ends after duplicating one
to the file descriptors where they are needed.
On Windows, which does not have fork(), we never have an opportunity to
(1) duplicate a pipe end in the child, (2) close unused pipe ends. Instead,
we must use this idiom:
save1 = dup(1);
pipe(fd);
dup2(fd[1], 1);
spawn(...);
dup2(save1, 1);
close(fd[1]);
i.e. save away the descriptor at the destination slot, replace by the pipe
end, spawn process, restore the saved file.
But there is a problem: Notice that the child did not only inherit the
dup2()ed descriptor, but also *both* original pipe ends. Although the one
end that was dup()ed could be closed before the spawn(), we cannot close
the other end - the child inherits it, no matter what.
The solution is to generate non-inheritable pipes. At the first glance,
this looks strange: The purpose of pipes is usually to be inherited to
child processes. But notice that in the course of actions as outlined
above, the pipe descriptor that we want to inherit to the child is
dup2()ed, and as it so happens, Windows's dup2() creates inheritable
duplicates.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
When an external git command is invoked, it can be a Bourne shell script.
This patch looks into the command file to see whether it is one.
In this case, the command line is rearranged to invoke the shell
with the proper arguments.
With this change, scripted git commands work. Command line arguments
to those scripts cannot be complex (contain spaces or double-quotes), yet.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
The timer is implemented using a thread that calls the signal handler
at regular intervals.
We also replace Windows's signal() function because we must intercept
that SIGALRM is set (which is used when a timer is canceled).
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
Windows's rename() is based on the MoveFile() API, which fails if the
destination exists. Here we work around the problem by using MoveFileEx().
Furthermore, the posixly correct error is returned if the destination is
a directory.
The implementation is still slightly incomplete, however, because of the
missing error code translation: We assume that the failure is due to
permissions.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
getpwuid() is implemented just enough that GIT does not issue errors.
Since the information that it returns is not very useful, users are
required to set up user.name and user.email configuration.
All uses of getpwuid() are like getpwuid(getuid()), hence, the return value
of getuid() is irrelevant and the uid parameter is not even looked at.
Side note: getpwnam() is only used to resolve '~' and '~username' paths,
which is an idiom not known on Windows, hence, we don't implement it.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
The wrapper does two things:
- Requests to open /dev/null are redirected to open the nul pseudo file.
- A request to open a file that currently exists as a directory on
Windows fails with EACCES; this is changed to EISDIR.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
GIT's guts work with a forward slash as a path separators. We do not change
that. Rather we make sure that only "normalized" paths enter the depths
of the machinery.
We have to translate backslashes to forward slashes in the prefix and in
command line arguments. Fortunately, all of them are passed through
functions in setup.c.
A macro has_dos_drive_path() is defined that checks whether a path begins
with a drive letter+colon combination. This predicate is always false on
Unix. Another macro is_dir_sep() abstracts that a backslash is also a
directory separator on Windows.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
With this change GIT can be compiled and linked using MinGW. Builtins
that only read the repository such as the log family and grep already
work.
Simple stubs are provided for a number of functions that the Windows C
runtime does not offer. They will be completed in later patches.
However, a fix for the snprintf/vsnprintf replacement is applied here
to avoid buffer overflows.
Dmitry Kakurin pointed out that access(..., X_OK) would always fails on
Vista and suggested the -D__USE_MINGW_ACCESS workaround.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
