When adding many objects to a repo with `core.fsync=loose-object`,
the cost of fsync'ing each object file can become prohibitive.
One major source of the cost of fsync is the implied flush of the
hardware writeback cache within the disk drive. This commit introduces
a new `core.fsyncMethod=batch` option that batches up hardware flushes.
It hooks into the bulk-checkin odb-transaction functionality, takes
advantage of tmp-objdir, and uses the writeout-only support code.
When the new mode is enabled, we do the following for each new object:
1a. Create the object in a tmp-objdir.
2a. Issue a pagecache writeback request and wait for it to complete.
At the end of the entire transaction when unplugging bulk checkin:
1b. Issue an fsync against a dummy file to flush the log and hardware
writeback cache, which should by now have seen the tmp-objdir writes.
2b. Rename all of the tmp-objdir files to their final names.
3b. When updating the index and/or refs, we assume that Git will issue
another fsync internal to that operation. This is not the default
today, but the user now has the option of syncing the index and there
is a separate patch series to implement syncing of refs.
On a filesystem with a singular journal that is updated during name
operations (e.g. create, link, rename, etc), such as NTFS, HFS+, or XFS
we would expect the fsync to trigger a journal writeout so that this
sequence is enough to ensure that the user's data is durable by the time
the git command returns. This sequence also ensures that no object files
appear in the main object store unless they are fsync-durable.
Batch mode is only enabled if core.fsync includes loose-objects. If
the legacy core.fsyncObjectFiles setting is enabled, but core.fsync does
not include loose-objects, we will use file-by-file fsyncing.
In step (1a) of the sequence, the tmp-objdir is created lazily to avoid
work if no loose objects are ever added to the ODB. We use a tmp-objdir
to maintain the invariant that no loose-objects are visible in the main
ODB unless they are properly fsync-durable. This is important since
future ODB operations that try to create an object with specific
contents will silently drop the new data if an object with the target
hash exists without checking that the loose-object contents match the
hash. Only a full git-fsck would restore the ODB to a functional state
where dataloss doesn't occur.
In step (1b) of the sequence, we issue a fsync against a dummy file
created specifically for the purpose. This method has a little higher
cost than using one of the input object files, but makes adding new
callers of this mechanism easier, since we don't need to figure out
which object file is "last" or risk sharing violations by caching the fd
of the last object file.
_Performance numbers_:
Linux - Hyper-V VM running Kernel 5.11 (Ubuntu 20.04) on a fast SSD.
Mac - macOS 11.5.1 running on a Mac mini on a 1TB Apple SSD.
Windows - Same host as Linux, a preview version of Windows 11.
Adding 500 files to the repo with 'git add' Times reported in seconds.
object file syncing | Linux | Mac | Windows
--------------------|-------|-------|--------
disabled | 0.06 | 0.35 | 0.61
fsync | 1.88 | 11.18 | 2.47
batch | 0.15 | 0.41 | 1.53
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make it clearer in the naming and documentation of the plug_bulk_checkin
and unplug_bulk_checkin APIs that they can be thought of as
a "transaction" to optimize operations on the object database. These
transactions may be nested so that subsystems like the cache-tree
writing code can optimize their operations without caring whether the
top-level code has a transaction active.
Add a flush_odb_transaction API that will be used in update-index to
make objects visible even if a transaction is active. The flush call may
also be useful in future cases if we hold a transaction active around
calling hooks.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit prepares for adding batch-fsync to the bulk-checkin
infrastructure.
The bulk-checkin infrastructure is currently used to batch up addition
of large blobs to a packfile. When a blob is larger than
big_file_threshold, we unconditionally add it to a pack. If bulk
checkins are 'plugged', we allow multiple large blobs to be added to a
single pack until we reach the packfile size limit; otherwise, we simply
make a new packfile for each large blob. The 'unplug' call tells us when
the series of blob additions is done so that we can finish the packfiles
and make their objects available to subsequent operations.
Stated another way, bulk-checkin allows callers to define a transaction
that adds multiple objects to the object database, where the object
database can optimize its internal operations within the transaction
boundary.
Batched fsync will fit into bulk-checkin by taking advantage of the
plug/unplug functionality to determine the appropriate time to fsync
and make newly-added objects available in the primary object database.
* Rename 'state' variable to 'bulk_checkin_packfile', since we will
later be adding 'bulk_fsync_objdir'. This also makes the variable
easier to find in the debugger, since the name is more unique.
* Rename finish_bulk_checkin to flush_bulk_checkin_packfile and call it
unconditionally from unplug_bulk_checkin. Internally it will
conditionally do a flush if there's any work to do.
* Move the 'plugged' data member of 'bulk_checkin_state' into a separate
static variable. Doing this avoids resetting the variable in
finish_bulk_checkin when zeroing the 'bulk_checkin_state'. As-is, we
seem to unintentionally disable the plugging functionality the first
time a new packfile must be created due to packfile size limits. While
disabling the plugging state only results in suboptimal behavior for
the current code, it would be fatal for the bulk-fsync functionality
later in this patch series.
The net effect of these changes is to make a clear separation between
the portion of the bulk-checkin infrastructure that is related to the
packfile (nearly all of it at present) and the part that is related to
other future optimizations of the ODB.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If sync_file_range is not available when building the configure script,
there is a cosmetic bug when running that script reporting
"HAVE_SYNC_FILE_RANGE: command not found". Remove that error message by
defining HAVE_SYNC_FILE_RANGE to an empty string, rather than generating
a script where that appears as a bare command.
Signed-off-by: Adam Dinwoodie <adam@dinwoodie.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The warning for an unrecognized fsyncMethod was not
camel-cased.
Reported-by: Jiang Xin <worldhello.net@gmail.com>
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit b9f5d035 (core.fsync: documentation and user-friendly
aggregate options, 2022-03-15) introduced an incorrect value for
FSYNC_COMPONENTS_DEFAULT. We need an AND-NOT rather than OR-NOT.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit adds aggregate options for the core.fsync setting that are
more user-friendly. These options are specified in terms of 'levels of
safety', indicating which Git operations are considered to be sync
points for durability.
The new documentation is also included here in its entirety for ease of
review.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Plug random memory leaks.
* ab/plug-random-leaks:
repository.c: free the "path cache" in repo_clear()
range-diff: plug memory leak in read_patches()
range-diff: plug memory leak in common invocation
lockfile API users: simplify and don't leak "path"
commit-graph: stop fill_oids_from_packs() progress on error and free()
commit-graph: fix memory leak in misused string_list API
submodule--helper: fix trivial leak in module_add()
transport: stop needlessly copying bundle header references
bundle: call strvec_clear() on allocated strvec
remote-curl.c: free memory in cmd_main()
urlmatch.c: add and use a *_release() function
diff.c: free "buf" in diff_words_flush()
merge-base: free() allocated "struct commit **" list
index-pack: fix memory leaks
Newer version of GPGSM changed its output in a backward
incompatible way to break our code that parses its output. It also
added more processes our tests need to kill when cleaning up.
Adjustments have been made to accommodate these changes.
* fs/gpgsm-update:
t/lib-gpg: kill all gpg components, not just gpg-agent
t/lib-gpg: reload gpg components after updating trustlist
gpg-interface/gpgsm: fix for v2.3
Check the return value from parse_tree_indirect() to turn segfaults
into calls to die().
* gc/parse-tree-indirect-errors:
checkout, clone: die if tree cannot be parsed
Align the level of verbose output from the ort backend during inner
merge to that of the recursive backend.
* en/merge-ort-align-verbosity-with-recursive:
merge-ort: exclude messages from inner merges by default
Makefile refactoring with a bit of suffixes rule stripping to
optimize the runtime overhead.
* ab/make-optim-noop:
Makefiles: add and use wildcard "mkdir -p" template
Makefile: add "$(QUIET)" boilerplate to shared.mak
Makefile: move $(comma), $(empty) and $(space) to shared.mak
Makefile: move ".SUFFIXES" rule to shared.mak
Makefile: define $(LIB_H) in terms of $(FIND_SOURCE_FILES)
Makefile: disable GNU make built-in wildcard rules
Makefiles: add "shared.mak", move ".DELETE_ON_ERROR" to it
scalar Makefile: use "The default target of..." pattern
"git fetch" can make two separate fetches, but ref updates coming
from them were in two separate ref transactions under "--atomic",
which has been corrected.
* ps/fetch-atomic:
fetch: make `--atomic` flag cover pruning of refs
fetch: make `--atomic` flag cover backfilling of tags
refs: add interface to iterate over queued transactional updates
fetch: report errors when backfilling tags fails
fetch: control lifecycle of FETCH_HEAD in a single place
fetch: backfill tags before setting upstream
fetch: increase test coverage of fetches
This commit introduces the new ability for the user to harden
the index. In the event of a system crash, the index must be
durable for the user to actually find a file that has been added
to the repo and then deleted from the working tree.
We use the presence of the COMMIT_LOCK flag and absence of the
alternate_index_output as a proxy for determining whether we're
updating the persistent index of the repo or some temporary
index. We don't sync these temporary indexes.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This change introduces code to parse the core.fsync setting and
configure the fsync_components variable.
core.fsync is configured as a comma-separated list of component names to
sync. Each time a core.fsync variable is encountered in the
configuration heirarchy, we start off with a clean state with the
platform default value. Passing 'none' resets the value to indicate
nothing will be synced. We gather all negative and positive entries from
the comma separated list and then compute the new value by removing all
the negative entries and adding all of the positive entries.
We issue a warning for components that are not recognized so that the
configuration code is compatible with configs from future versions of
Git with more repo components.
Complete documentation for the new setting is included in a later patch
in the series so that it can be reviewed once in final form.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit introduces the infrastructure for the core.fsync
configuration knob. The repository components we want to sync
are identified by flags so that we can turn on or off syncing
for specific components.
If core.fsyncObjectFiles is set and the core.fsync configuration
also includes FSYNC_COMPONENT_LOOSE_OBJECT, we will fsync any
loose objects. This picks the strictest data integrity behavior
if core.fsync and core.fsyncObjectFiles are set to conflicting values.
This change introduces the currently unused fsync_component
helper, which will be used by a later patch that adds fsyncing to
the refs backend.
Actual configuration and documentation of the fsync components
list are in other patches in the series to separate review of
the underlying mechanism from the policy of how it's configured.
Helped-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This commit introduces the `core.fsyncMethod` configuration
knob, which can currently be set to `fsync` or `writeout-only`.
The new writeout-only mode attempts to tell the operating system to
flush its in-memory page cache to the storage hardware without issuing a
CACHE_FLUSH command to the storage controller.
Writeout-only fsync is significantly faster than a vanilla fsync on
common hardware, since data is written to a disk-side cache rather than
all the way to a durable medium. Later changes in this patch series will
take advantage of this primitive to implement batching of hardware
flushes.
When git_fsync is called with FSYNC_WRITEOUT_ONLY, it may fail and the
caller is expected to do an ordinary fsync as needed.
On Apple platforms, the fsync system call does not issue a CACHE_FLUSH
directive to the storage controller. This change updates fsync to do
fcntl(F_FULLFSYNC) to make fsync actually durable. We maintain parity
with existing behavior on Apple platforms by setting the default value
of the new core.fsyncMethod option.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Including NTSecAPI.h in git-compat-util.h causes build errors in any
other file that includes winternl.h. NTSecAPI.h was included in order to
get access to the RtlGenRandom cryptographically secure PRNG. This
change scopes the inclusion of ntsecapi.h to wrapper.c, which is the only
place that it's actually needed.
The build breakage is due to the definition of UNICODE_STRING in
NtSecApi.h:
#ifndef _NTDEF_
typedef LSA_UNICODE_STRING UNICODE_STRING, *PUNICODE_STRING;
typedef LSA_STRING STRING, *PSTRING ;
#endif
LsaLookup.h:
typedef struct _LSA_UNICODE_STRING {
USHORT Length;
USHORT MaximumLength;
#ifdef MIDL_PASS
[size_is(MaximumLength/2), length_is(Length/2)]
#endif // MIDL_PASS
PWSTR Buffer;
} LSA_UNICODE_STRING, *PLSA_UNICODE_STRING;
winternl.h also defines UNICODE_STRING:
typedef struct _UNICODE_STRING {
USHORT Length;
USHORT MaximumLength;
PWSTR Buffer;
} UNICODE_STRING;
typedef UNICODE_STRING *PUNICODE_STRING;
Both definitions have equivalent layouts. Apparently these internal
Windows headers aren't designed to be included together. This is
an oversight in the headers and does not represent an incompatibility
between the APIs.
Signed-off-by: Neeraj Singh <neerajsi@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Updates to how command line options to "git help" are handled.
* ab/help-fixes:
help: don't print "\n" before single-section output
help: add --no-[external-commands|aliases] for use with --all
help: error if [-a|-g|-c] and [-i|-m|-w] are combined
help: correct usage & behavior of "git help --all"
help: note the option name on option incompatibility
help.c: split up list_all_cmds_help() function
help tests: test "git" and "git help [-a|-g] spacing
help.c: use puts() instead of printf{,_ln}() for consistency
help doc: add missing "]" to "[-a|--all]"
Remove the escape hatch we added when we introduced the weather
balloon to use variadic macros unconditionally, to make it official
that we now have a hard dependency on the feature.
* ab/c99-variadic-macros:
C99: remove hardcoded-out !HAVE_VARIADIC_MACROS code
git-compat-util.h: clarify GCC v.s. C99-specific in comment
General clean-up in reftable implementation, including
clarification of the API documentation, tightening the code to
honor documented length limit, etc.
* hn/reftable-no-empty-keys:
reftable: rename writer_stats to reftable_writer_stats
reftable: add test for length of disambiguating prefix
reftable: ensure that obj_id_len is >= 2 on writing
reftable: avoid writing empty keys at the block layer
reftable: add a test that verifies that writing empty keys fails
reftable: reject 0 object_id_len
Documentation: object_id_len goes up to 31
"git cat-file" learns "--batch-command" mode, which is a more
flexible interface than the existing "--batch" or "--batch-check"
modes, to allow different kinds of inquiries made.
* jc/cat-file-batch-commands:
cat-file: add --batch-command mode
cat-file: add remove_timestamp helper
cat-file: introduce batch_mode enum to replace print_contents
cat-file: rename cmdmode to transform_mode
Improve failure case behaviour of xdiff library when memory
allocation fails.
* pw/xdiff-alloc-fail:
xdiff: handle allocation failure when merging
xdiff: refactor a function
xdiff: handle allocation failure in patience diff
xdiff: fix a memory leak
In sparse-checkouts, files mis-marked as missing from the working tree
could lead to later problems. Such files were hard to discover, and
harder to correct. Automatically detecting and correcting the marking
of such files has been added to avoid these problems.
* en/present-despite-skipped:
repo_read_index: add config to expect files outside sparse patterns
Accelerate clear_skip_worktree_from_present_files() by caching
Update documentation related to sparsity and the skip-worktree bit
repo_read_index: clear SKIP_WORKTREE bit from files present in worktree
unpack-trees: fix accidental loss of user changes
t1011: add testcase demonstrating accidental loss of user modifications
Many output modes of "ls-files" do not work with its
"--recurse-submodules" option, but the "-s" mode has been taught to
work with it.
* jt/ls-files-stage-recurse:
ls-files: support --recurse-submodules --stage
"git checkout -b branch/with/multi/level/name && git stash" only
recorded the last level component of the branch name, which has
been corrected.
* gc/stash-on-branch-with-multi-level-name:
stash: strip "refs/heads/" with skip_prefix
The error message given by "git switch HEAD~4" has been clarified
to suggest the "--detach" option that is required.
* ah/advice-switch-requires-detach-to-detach:
switch: mention the --detach option when dying due to lack of a branch
Use designated initializers we started using in mid 2017 in more
parts of the codebase that are relatively quiescent.
* ab/c99-designated-initializers:
fast-import.c: use designated initializers for "partial" struct assignments
refspec.c: use designated initializers for "struct refspec_item"
convert.c: use designated initializers for "struct stream_filter*"
userdiff.c: use designated initializers for "struct userdiff_driver"
archive-*.c: use designated initializers for "struct archiver"
object-file: use designated initializers for "struct git_hash_algo"
trace2: use designated initializers for "struct tr2_dst"
trace2: use designated initializers for "struct tr2_tgt"
imap-send.c: use designated initializers for "struct imap_server_conf"
When "index-pack" dies due to incoming data exceeding the maximum
allowed input size, include the value of the limit in the error
message.
* mc/index-pack-report-max-size:
index-pack: clarify the breached limit
Tighten the language around "working tree" and "worktree" in the
docs.
* ds/worktree-docs:
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: use 'worktree' over 'working tree'
worktree: extract checkout_worktree()
worktree: extract copy_sparse_checkout()
worktree: extract copy_filtered_worktree_config()
worktree: combine two translatable messages
A not-so-common mistake is to write a script to feed "git bisect
run" without making it executable, in which case all tests will
exit with 126 or 127 error codes, even on revisions that are marked
as good. Try to recognize this situation and stop iteration early.
* rs/bisect-executable-not-found:
bisect--helper: double-check run command on exit code 126 and 127
bisect: document run behavior with exit codes 126 and 127
bisect--helper: release strbuf and strvec on run error
bisect--helper: report actual bisect_state() argument on error
Further polishing of "git sparse-checkout".
* en/sparse-checkout-fixes:
sparse-checkout: reject arguments in cone-mode that look like patterns
sparse-checkout: error or warn when given individual files
sparse-checkout: pay attention to prefix for {set, add}
sparse-checkout: correctly set non-cone mode when expected
sparse-checkout: correct reapply's handling of options
Test modernization.
* cg/t3903-modernize:
tests: make the code more readable
tests: allow testing if a path is truly a file or a directory
t/t3903-stash.sh: replace test [-d|-f] with test_path_is_*
The "struct path_cache" added in 102de880d24 (path.c: migrate global
git_path_* to take a repository argument, 2018-05-17) is only used
directly by code in repository.[ch] (but populated in path.[ch]).
Let's move this code to repository.[ch], and stop leaking this memory
when we run repo_clear(). To avoid the cast change it from a "const
char *" to a "char *".
This also removes the "PATH_CACHE_INIT" macro, which has never been
used for anything. For the "struct repository" we already make a hard
assumption that it (and "the_repository") can be identically
initialized by making it a "static" variable, so making use of a
"PATH_CACHE_INIT" somewhere would have been confusing.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Amend code added in d9c66f0b5bf (range-diff: first rudimentary
implementation, 2018-08-13) to use a "goto cleanup" pattern. This
makes for less code, and frees memory that we'd previously leak.
The reason for changing free(util) to FREE_AND_NULL(util) is because
at the end of the function we append the contents of "util" to a
"struct string_list" if it's non-NULL.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Create a public release_patch() version of the private free_patch()
function added in 13b5af22f39 (apply: move libified code from
builtin/apply.c to apply.{c,h}, 2016-04-22). Unlike the existing
function this one doesn't free() the "struct patch" itself, so we can
use it for variables on the stack.
Use it in range-diff.c to fix a memory leak in common range-diff
invocations, e.g.:
git -P range-diff origin/master origin/next origin/seen
Would emit several errors when compiled with SANITIZE=leak, but now
runs cleanly.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix a memory leak in code added in 6c622f9f0bb (commit-graph: write
commit-graph chains, 2019-06-18). We needed to free the "lock_name" if
we encounter errors, and the "graph_name" after we'd run unlink() on
it.
For the case of write_commit_graph_file() refactoring the code to free
the "lock_name" after we were done using the "struct lock_file lk"
would have made the control flow more complex. Luckily we can free the
"lock_file" right after the hold_lock_file_for_update() call, if it
makes use of "path" at all it'll have copied its contents to a "struct
strbuf" of its own.
While I'm at it let's fix code added in fb10ca5b543 (sparse-checkout:
write using lockfile, 2019-11-21) in write_patterns_and_update() to
avoid the same complexity that I thought I needed when I wrote the
initial fix for write_commit_graph_file(). We can free the
"sparse_filename" right after calling hold_lock_file_for_update(), we
don't need to wait until we're exiting the function to do so.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix a bug in fill_oids_from_packs(), we should always stop_progress(),
but did not do so if we returned an error here. This also plugs a
memory leak in those cases by releasing the two "struct strbuf"
variables the function uses.
While I'm at it stop hardcoding "-1" here and just use the return
value of error() instead, which happens to be "-1".
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When this code was migrated to the string_list API in
d88b14b3fd6 (commit-graph: use string-list API for input, 2018-06-27)
it was made to use used both STRING_LIST_INIT_NODUP and a
strbuf_detach() pattern.
Those should not be used together if string_list_clear() is expected
to free the memory, instead we need to either use STRING_LIST_INIT_DUP
with a string_list_append_nodup(), or a STRING_LIST_INIT_NODUP and
manually fiddle with the "strdup_strings" member before calling
string_list_clear(). Let's do the former.
Since "strdup_strings = 1" is set now other code might be broken by
relying on "pack_indexes" not to duplicate it strings, but that
doesn't happen. When we pass this down to write_commit_graph() that
code uses the "struct string_list" without modifying it. Let's add a
"const" to the variable to have the compiler enforce that assumption.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
