When updating the skip-worktree bits in the index to align with new
values in a sparse-checkout file, Git scans the entire working
directory with lstat() calls. In a sparse-checkout, many of these
lstat() calls are for paths that do not exist.
Enable the fscache feature during this scan. Since enable_fscache()
calls nest, the disable_fscache() method decrements a counter and
would only clear the cache if that counter reaches zero.
In a local test of a repo with ~2.2 million paths, updating the index
with git read-tree -m -u HEAD with a sparse-checkout file containing
only /.gitattributes improved from 2-3 minutes to ~6 seconds.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Using FindFirstFileExW() requires the OS to allocate a 64K buffer for each
directory and then free it when we call FindClose(). Update fscache to call
the underlying kernel API NtQueryDirectoryFile so that we can do the buffer
management ourselves. That allows us to allocate a single buffer for the
lifetime of the cache and reuse it for each directory.
This change improves performance of 'git status' by 18% in a repo with ~200K
files and 30k folders.
Documentation for NtQueryDirectoryFile can be found at:
https://docs.microsoft.com/en-us/windows-hardware/drivers/ddi/content/ntifs/nf-ntifs-ntquerydirectoryfilehttps://docs.microsoft.com/en-us/windows/desktop/FileIO/file-attribute-constantshttps://docs.microsoft.com/en-us/windows/desktop/fileio/reparse-point-tags
To determine if the specified directory is a symbolic link, inspect the
FileAttributes member to see if the FILE_ATTRIBUTE_REPARSE_POINT flag is
set. If so, EaSize will contain the reparse tag (this is a so far
undocumented feature, but confirmed by the NTFS developers). To
determine if the reparse point is a symbolic link (and not some other
form of reparse point), test whether the tag value equals the value
IO_REPARSE_TAG_SYMLINK.
The NtQueryDirectoryFile() call works best (and on Windows 8.1 and
earlier, it works *only*) with buffer sizes up to 64kB. Which is 32k
wide characters, so let's use that as our buffer size.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The recent change to make fscache thread specific relied on fscache_enable()
being called first from the primary thread before being called in parallel
from worker threads. Make that more robust and protect it with a critical
section to avoid any issues.
Helped-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Now that the fscache is single threaded, take advantage of the mem_pool as
the allocator to significantly reduce the cost of allocations and frees.
With the reduced cost of free, in future patches, we can start freeing the
fscache at the end of commands instead of just leaking it.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
The threading model for fscache has been to have a single, global cache.
This puts requirements on it to be thread safe so that callers like
preload-index can call it from multiple threads. This was implemented
with a single mutex and completion events which introduces contention
between the calling threads.
Simplify the threading model by making fscache thread specific. This allows
us to remove the global mutex and synchronization events entirely and instead
associate a fscache with every thread that requests one. This works well with
the current multi-threading which divides the cache entries into blocks with
a separate thread processing each block.
At the end of each worker thread, if there is a fscache on the primary
thread, merge the cached results from the worker into the primary thread
cache. This enables us to reuse the cache later especially when scanning for
untracked files.
In testing, this reduced the time spent in preload_index() by about 25% and
also reduced the CPU utilization significantly. On a repo with ~200K files,
it reduced overall status times by ~12%.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Update enable_fscache() to take an optional initial size parameter which is
used to initialize the hashmap so that it can avoid having to rehash as
additional entries are added.
Add a separate disable_fscache() macro to make the code clearer and easier
to read.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Add tracing around initializing and discarding mempools. In discard report
on the amount of memory unused in the current block to help tune setting
the initial_size.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Track fscache hits and misses for lstat and opendir requests. Reporting of
statistics is done when the cache is disabled for the last time and freed
and is only reported if GIT_TRACE_FSCACHE is set.
Sample output is:
11:33:11.836428 compat/win32/fscache.c:433 fscache: lstat 3775, opendir 263, total requests/misses 4052/269
Signed-off-by: Ben Peart <benpeart@microsoft.com>
At the end of the status command, disable and free the fscache so that we
don't leak the memory and so that we can dump the fscache statistics.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
Use FindFirstFileExW with FindExInfoBasic to avoid forcing NTFS to look up
the short name. Also switch to a larger (64K vs 4K) buffer using
FIND_FIRST_EX_LARGE_FETCH to minimize round trips to the kernel.
In a repo with ~200K files, this drops warm cache status times from 3.19
seconds to 2.67 seconds for a 16% savings.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
On file systems that support it, this can dramatically speed up operations
like add, commit, describe, rebase, reset, rm that would otherwise have to
lstat() every file to "re-match" the stat information in the index to that
of the file system.
On a synthetic repo with 1M files, "git reset" dropped from 52.02 seconds to
14.42 seconds for a savings of 72%.
Signed-off-by: Ben Peart <benpeart@microsoft.com>
This is retry of #1419.
I added flush_fscache macro to flush cached stats after disk writing
with tests for regression reported in #1438 and #1442.
git checkout checks each file path in sorted order, so cache flushing does not
make performance worse unless we have large number of modified files in
a directory containing many files.
Using chromium repository, I tested `git checkout .` performance when I
delete 10 files in different directories.
With this patch:
TotalSeconds: 4.307272
TotalSeconds: 4.4863595
TotalSeconds: 4.2975562
Avg: 4.36372923333333
Without this patch:
TotalSeconds: 20.9705431
TotalSeconds: 22.4867685
TotalSeconds: 18.8968292
Avg: 20.7847136
I confirmed this patch passed all tests in t/ with core_fscache=1.
Signed-off-by: Takuto Ikuta <tikuta@chromium.org>
When I do git fetch, git call file stats under .git/objects for each
refs. This takes time when there are many refs.
By enabling fscache, git takes file stats by directory traversing and that
improved the speed of fetch-pack for repository having large number of
refs.
In my windows workstation, this improves the time of `git fetch` for
chromium repository like below. I took stats 3 times.
* With this patch
TotalSeconds: 9.9825165
TotalSeconds: 9.1862075
TotalSeconds: 10.1956256
Avg: 9.78811653333333
* Without this patch
TotalSeconds: 15.8406702
TotalSeconds: 15.6248053
TotalSeconds: 15.2085938
Avg: 15.5580231
Signed-off-by: Takuto Ikuta <tikuta@chromium.org>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Fix regression described in:
https://github.com/git-for-windows/git/issues/1392
which was introduced in:
b2353379bb
Problem Symptoms
================
When the user has a .gitignore file that is a symlink, the fscache
optimization introduced above caused the stat-data from the symlink,
rather that of the target file, to be returned. Later when the ignore
file was read, the buffer length did not match the stat.st_size field
and we called die("cannot use <path> as an exclude file")
Optimization Rationale
======================
The above optimization calls lstat() before open() primarily to ask
fscache if the file exists. It gets the current stat-data as a side
effect essentially for free (since we already have it in memory).
If the file does not exist, it does not need to call open(). And
since very few directories have .gitignore files, we can greatly
reduce time spent in the filesystem.
Discussion of Fix
=================
The above optimization calls lstat() rather than stat() because the
fscache only intercepts lstat() calls. Calls to stat() stay directed
to the mingw_stat() completly bypassing fscache. Furthermore, calls
to mingw_stat() always call {open, fstat, close} so that symlinks are
properly dereferenced, which adds *additional* open/close calls on top
of what the original code in dir.c is doing.
Since the problem only manifests for symlinks, we add code to overwrite
the stat-data when the path is a symlink. This preserves the effect of
the performance gains provided by the fscache in the normal case.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Make fscache_enabled() function public rather than static.
Remove unneeded fscache_is_enabled() function.
Change is_fscache_enabled() macro to call fscache_enabled().
is_fscache_enabled() now takes a pathname so that the answer
is more precise and mean "is fscache enabled for this pathname",
since fscache only stores repo-relative paths and not absolute
paths, we can avoid attempting lookups for absolute paths.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach read_directory_recursive() and add_excludes() to
be aware of optional fscache and avoid trying to open()
and fstat() non-existant ".gitignore" files in every
directory in the worktree.
The current code in add_excludes() calls open() and then
fstat() for a ".gitignore" file in each directory present
in the worktree. Change that when fscache is enabled to
call lstat() first and if present, call open().
This seems backwards because both lstat needs to do more
work than fstat. But when fscache is enabled, fscache will
already know if the .gitignore file exists and can completely
avoid the IO calls. This works because of the lstat diversion
to mingw_lstat when fscache is enabled.
This reduced status times on a 350K file enlistment of the
Windows repo on a NVMe SSD by 0.25 seconds.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach "add" to use preload-index and fscache features
to improve performance on very large repositories.
During an "add", a call is made to run_diff_files()
which calls check_remove() for each index-entry. This
calls lstat(). On Windows, the fscache code intercepts
the lstat() calls and builds a private cache using the
FindFirst/FindNext routines, which are much faster.
Somewhat independent of this, is the preload-index code
which distributes some of the start-up costs across
multiple threads.
We need to keep the call to read_cache() before parsing the
pathspecs (and hence cannot use the pathspecs to limit any preload)
because parse_pathspec() is using the index to determine whether a
pathspec is, in fact, in a submodule. If we would not read the index
first, parse_pathspec() would not error out on a path that is inside
a submodule, and t7400-submodule-basic.sh would fail with
not ok 47 - do not add files from a submodule
We still want the nice preload performance boost, though, so we simply
call read_cache_preload(&pathspecs) after parsing the pathspecs.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Teach FSCACHE to remember "not found" directories.
This is a performance optimization.
FSCACHE is a performance optimization available for Windows. It
intercepts Posix-style lstat() calls into an in-memory directory
using FindFirst/FindNext. It improves performance on Windows by
catching the first lstat() call in a directory, using FindFirst/
FindNext to read the list of files (and attribute data) for the
entire directory into the cache, and short-cut subsequent lstat()
calls in the same directory. This gives a major performance
boost on Windows.
However, it does not remember "not found" directories. When STATUS
runs and there are missing directories, the lstat() interception
fails to find the parent directory and simply return ENOENT for the
file -- it does not remember that the FindFirst on the directory
failed. Thus subsequent lstat() calls in the same directory, each
re-attempt the FindFirst. This completely defeats any performance
gains.
This can be seen by doing a sparse-checkout on a large repo and
then doing a read-tree to reset the skip-worktree bits and then
running status.
This change reduced status times for my very large repo by 60%.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
If multiple threads access a directory that is not yet in the cache, the
directory will be loaded by each thread. Only one of the results is added
to the cache, all others are leaked. This wastes performance and memory.
On cache miss, add a future object to the cache to indicate that the
directory is currently being loaded. Subsequent threads register themselves
with the future object and wait. When the first thread has loaded the
directory, it replaces the future object with the result and notifies
waiting threads.
Signed-off-by: Karsten Blees <blees@dcon.de>
Checking the work tree status is quite slow on Windows, due to slow
`lstat()` emulation (git calls `lstat()` once for each file in the
index). Windows operating system APIs seem to be much better at scanning
the status of entire directories than checking single files.
Add an `lstat()` implementation that uses a cache for lstat data. Cache
misses read the entire parent directory and add it to the cache.
Subsequent `lstat()` calls for the same directory are served directly
from the cache.
Also implement `opendir()`/`readdir()`/`closedir()` so that they create
and use directory listings in the cache.
The cache doesn't track file system changes and doesn't plug into any
modifying file APIs, so it has to be explicitly enabled for git functions
that don't modify the working copy.
Note: in an earlier version of this patch, the cache was always active and
tracked file system changes via ReadDirectoryChangesW. However, this was
much more complex and had negative impact on the performance of modifying
git commands such as 'git checkout'.
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Add a macro to mark code sections that only read from the file system,
along with a config option and documentation.
This facilitates implementation of relatively simple file system level
caches without the need to synchronize with the file system.
Enable read-only sections for 'git status' and preload_index.
Signed-off-by: Karsten Blees <blees@dcon.de>
Emulating the POSIX lstat API on Windows via GetFileAttributes[Ex] is quite
slow. Windows operating system APIs seem to be much better at scanning the
status of entire directories than checking single files. A caching
implementation may improve performance by bulk-reading entire directories
or reusing data obtained via opendir / readdir.
Make the lstat implementation pluggable so that it can be switched at
runtime, e.g. based on a config option.
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Emulating the POSIX `dirent` API on Windows via
`FindFirstFile()`/`FindNextFile()` is pretty staightforward, however,
most of the information provided in the `WIN32_FIND_DATA` structure is
thrown away in the process. A more sophisticated implementation may
cache this data, e.g. for later reuse in calls to `lstat()`.
Make the `dirent` implementation pluggable so that it can be switched at
runtime, e.g. based on a config option.
Define a base DIR structure with pointers to `readdir()`/`closedir()`
that match the `opendir()` implementation (similar to vtable pointers in
Object-Oriented Programming). Define `readdir()`/`closedir()` so that
they call the function pointers in the `DIR` structure. This allows to
choose the `opendir()` implementation on a call-by-call basis.
Make the fixed-size `dirent.d_name` buffer a flex array, as `d_name` may
be implementation specific (e.g. a caching implementation may allocate a
`struct dirent` with _just_ the size needed to hold the `d_name` in
question).
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
We will use them in the upcoming "FSCache" patches (to accelerate
sequential lstat() calls).
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
This is the branch thicket of patches in Git for Windows that are
considered ready for upstream. To keep them in a ready-to-submit shape,
they are kept as close to the beginning of the branch thicket as
possible.
If `feature.experimental` and `feature.manyFiles` are set and the user
has not explicitly turned off the builtin FSMonitor, we now start
the built-in FSMonitor by default.
Only forcing it when UNSET matches the behavior of UPDATE_DEFAULT_BOOL()
used for other repo settings.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Create unit tests to move a directory. Verify that `git status`
gives the same result with and without FSMonitor enabled.
NEEDSWORK: This test exposes a bug in the untracked-cache on
Windows when FSMonitor is disabled. These are commented out
for the moment.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach Git to perform binary search over the cache-entries for a directory
notification and then linearly scan forward to find the immediate children.
Previously, when the FSMonitor reported a modified directory Git would
perform a linear search on the entire cache-entry array for all
entries matching that directory prefix and invalidate them. Since the
cache-entry array is already sorted, we can use a binary search to
find the first matching entry and then only linearly walk forward and
invalidate entries until the prefix changes.
Also, the original code would invalidate anything having the same
directory prefix. Since a directory event should only be received for
items that are immediately within the directory (and not within
sub-directories of it), only invalidate those entries and not the
whole subtree.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Create 2x2 test matrix with the untracked-cache and fsmonitor--daemon
features and a series of edits and verify that status output is
identical.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Set the `FSMONITOR_CHANGED` bit on `istate->cache_changed` when
FSMonitor returns a very large repsonse to ensure that the index is
written to disk.
Normally, when the FSMonitor response includes a tracked file, the
index is always updated. Similarly, the index might be updated when
the response alters the untracked-cache (when enabled). However, in
cases where neither of those cause the index to be considered changed,
the FSMonitor response is wasted. Subsequent Git commands will make
requests with the same token and receive the same response.
If that response is very large, performance may suffer. It would be
more efficient to force update the index now (and the token in the
index extension) in order to reduce the size of the response received
by future commands.
This was observed on Windows after a large checkout. On Windows, the
kernel emits events for the files that are changed as they are
changed. However, it might delay events for the containing
directories until the system is more idle (or someone scans the
directory (so it seems)). The first status following a checkout would
get the list of files. The subsequent status commands would get the
list of directories as the events trickled out. But they would never
catch up because the token was not advanced because the index wasn't
updated.
This list of directories caused `wt_status_collect_untracked()` to
unnecessarily spend time actually scanning them during each command.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach the listener thread to shutdown the daemon if the spelling of the
worktree root directory changes.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach fsmonitor--daemon client threads to create a cookie file
inside the .git directory and then wait until FS events for the
cookie are observed by the FS listener thread.
This helps address the racy nature of file system events by
blocking the client response until the kernel has drained any
event backlog.
This is especially important on MacOS where kernel events are
only issued with a limited frequency. See the `latency` argument
of `FSeventStreamCreate()`. The kernel only signals every `latency`
seconds, but does not guarantee that the kernel queue is completely
drained, so we may have to wait more than one interval. If we
increase the frequency, the system is more likely to drop events.
We avoid these issues by having each client thread create a unique
cookie file and then wait until it is seen in the event stream.
Co-authored-by: Kevin Willford <Kevin.Willford@microsoft.com>
Co-authored-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Teach fsmonitor--daemon to periodically truncate the list of
modified files to save some memory.
Clients will ask for the set of changes relative to a token that they
found in the FSMN index extension in the index. (This token is like a
point in time, but different). Clients will then update the index to
contain the response token (so that subsequent commands will be
relative to this new token).
Therefore, the daemon can gradually truncate the in-memory list of
changed paths as they become obsolete (older than the previous token).
Since we may have multiple clients making concurrent requests with a
skew of tokens and clients may be racing to the talk to the daemon,
we lazily truncate the list.
We introduce a 5 minute delay and truncate batches 5 minutes after
they are considered obsolete.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Create framework in Win32 version of the "health" thread to
periodically inspect the system and shutdown if warranted.
This version just include the setup for the timeout in
WaitForMultipleObjects() and calls (currently empty) table
of functions.
A later commit will add functions to the table to actually
inspect the system.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Repeat all of the fsmonitor perf tests using `git fsmonitor--daemon` and
the "Simple IPC" interface.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Create another thread to watch over the daemon process and
automatically shut it down if necessary.
This commit creates the basic framework for a "health" thread
to monitor the daemon and/or the file system. Later commits
will add platform-specific code to do the actual work.
The "health" thread is intended to monitor conditions that
would be difficult to track inside the IPC thread pool and/or
the file system listener threads. For example, when there are
file system events outside of the watched worktree root or if
we want to have an idle-timeout auto-shutdown feature.
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
