"git gc --auto" opens file descriptors for the packfiles before
spawning "git repack/prune", which would upset Windows that does
not want a process to work on a file that is open by another
process. The issue has been worked around.
* kg/gc-auto-windows-workaround:
gc --auto: release pack files before auto packing
Teach gc --auto to release pack files before auto packing the repository
to prevent failures when removing them.
Also teach the test 'fetching with auto-gc does not lock up' to complain
when it is no longer triggering an auto packing of the repository.
Fixes https://github.com/git-for-windows/git/issues/500
Signed-off-by: Kim Gybels <kgybels@infogroep.be>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The commit-graph file is a very helpful feature for speeding up git
operations. In order to make it more useful, make it possible to
write the commit-graph file during standard garbage collection
operations.
Add a 'gc.commitGraph' config setting that triggers writing a
commit-graph file after any non-trivial 'git gc' command. Defaults to
false while the commit-graph feature matures. We specifically do not
want to have this on by default until the commit-graph feature is fully
integrated with history-modifying features like shallow clones.
Helped-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Code clean-up to adjust to a more recent lockfile API convention that
allows lockfile instances kept on the stack.
* ma/lockfile-cleanup:
lock_file: move static locks into functions
lock_file: make function-local locks non-static
refs.c: do not die if locking fails in `delete_pseudoref()`
refs.c: do not die if locking fails in `write_pseudoref()`
t/helper/test-write-cache: clean up lock-handling
"git gc" in a large repository takes a lot of time as it considers
to repack all objects into one pack by default. The command has
been taught to pretend as if the largest existing packfile is
marked with ".keep" so that it is left untouched while objects in
other packs and loose ones are repacked.
* nd/repack-keep-pack:
pack-objects: show some progress when counting kept objects
gc --auto: exclude base pack if not enough mem to "repack -ad"
gc: handle a corner case in gc.bigPackThreshold
gc: add gc.bigPackThreshold config
gc: add --keep-largest-pack option
repack: add --keep-pack option
t7700: have closing quote of a test at the beginning of line
Placing `struct lock_file`s on the stack used to be a bad idea, because
the temp- and lockfile-machinery would keep a pointer into the struct.
But after 076aa2cbd (tempfile: auto-allocate tempfiles on heap,
2017-09-05), we can safely have lockfiles on the stack. (This applies
even if a user returns early, leaving a locked lock behind.)
These `struct lock_file`s are local to their respective functions and we
can drop their staticness.
For good measure, I have inspected these sites and come to believe that
they always release the lock, with the possible exception of bailing out
using `die()` or `exit()` or by returning from a `cmd_foo()`.
As pointed out by Jeff King, it would be bad if someone held on to a
`struct lock_file *` for some reason. After some grepping, I agree with
his findings: no-one appears to be doing that.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git gc --prune=nonsense" spent long time repacking and then
silently failed when underlying "git prune --expire=nonsense"
failed to parse its command line. This has been corrected.
* jc/parseopt-expiry-errors:
parseopt: handle malformed --expire arguments more nicely
gc: do not upcase error message shown with die()
A few commands that parse --expire=<time> command line option behave
sillily when given nonsense input. For example
$ git prune --no-expire
Segmentation falut
$ git prune --expire=npw; echo $?
129
Both come from parse_opt_expiry_date_cb().
The former is because the function is not prepared to see arg==NULL
(for "--no-expire", it is a norm; "--expire" at the end of the
command line could be made to pass NULL, if it is told that the
argument is optional, but we don't so we do not have to worry about
that case).
The latter is because it does not check the value returned from the
underlying parse_expiry_date().
This seems to be a recent regression introduced while we attempted
to avoid spewing the entire usage message when given a correct
option but with an invalid value at 3bb0923f ("parse-options: do not
show usage upon invalid option value", 2018-03-22). Before that, we
didn't fail silently but showed a full usage help (which arguably is
not all that better).
Also catch this error early when "git gc --prune=<expiration>" is
misspelled by doing a dummy parsing before the main body of "gc"
that is time consuming even begins. Otherwise, we'd spend time to
pack objects and then later have "git prune" first notice the error.
Aborting "gc" in the middle that way is not harmful but is ugly and
can be avoided.
Helped-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
pack-objects could be a big memory hog especially on large repos,
everybody knows that. The suggestion to stick a .keep file on the
giant base pack to avoid this problem is also known for a long time.
Recent patches add an option to do just this, but it has to be either
configured or activated manually. This patch lets `git gc --auto`
activate this mode automatically when it thinks `repack -ad` will use
a lot of memory and start affecting the system due to swapping or
flushing OS cache.
gc --auto decides to do this based on an estimation of pack-objects
memory usage, which is quite accurate at least for the heap part, and
whether that fits in half of system memory (the assumption here is for
desktop environment where there are many other applications running).
This mechanism only kicks in if gc.bigBasePackThreshold is not configured.
If it is, it is assumed that the user already knows what they want.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This config allows us to keep <N> packs back if their size is larger
than a limit. But if this N >= gc.autoPackLimit, we may have a
problem. We are supposed to reduce the number of packs after a
threshold because it affects performance.
We could tell the user that they have incompatible gc.bigPackThreshold
and gc.autoPackLimit, but it's kinda hard when 'git gc --auto' runs in
background. Instead let's fall back to the next best stategy: try to
reduce the number of packs anyway, but keep the base pack out. This
reduces the number of packs to two and hopefully won't take up too
much resources to repack (the assumption still is the base pack takes
most resources to handle).
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The --keep-largest-pack option is not very convenient to use because
you need to tell gc to do this explicitly (and probably on just a few
large repos).
Add a config key that enables this mode when packs larger than a limit
are found. Note that there's a slight behavior difference compared to
--keep-largest-pack: all packs larger than the threshold are kept, not
just the largest one.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This adds a new repack mode that combines everything into a secondary
pack, leaving the largest pack alone.
This could help reduce memory pressure. On linux-2.6.git, valgrind
massif reports 1.6GB heap in "pack all" case, and 535MB in "pack
all except the base pack" case. We save roughly 1GB memory by
excluding the base pack.
This should also lower I/O because we don't have to rewrite a giant
pack every time (e.g. for linux-2.6.git that's a 1.4GB pack file)..
PS. The use of string_list here seems overkill, but we'll need it in
the next patch...
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Refactoring the internal global data structure to make it possible
to open multiple repositories, work with and then close them.
Rerolled by Duy on top of a separate preliminary clean-up topic.
The resulting structure of the topics looked very sensible.
* sb/object-store: (27 commits)
sha1_file: allow sha1_loose_object_info to handle arbitrary repositories
sha1_file: allow map_sha1_file to handle arbitrary repositories
sha1_file: allow map_sha1_file_1 to handle arbitrary repositories
sha1_file: allow open_sha1_file to handle arbitrary repositories
sha1_file: allow stat_sha1_file to handle arbitrary repositories
sha1_file: allow sha1_file_name to handle arbitrary repositories
sha1_file: add repository argument to sha1_loose_object_info
sha1_file: add repository argument to map_sha1_file
sha1_file: add repository argument to map_sha1_file_1
sha1_file: add repository argument to open_sha1_file
sha1_file: add repository argument to stat_sha1_file
sha1_file: add repository argument to sha1_file_name
sha1_file: allow prepare_alt_odb to handle arbitrary repositories
sha1_file: allow link_alt_odb_entries to handle arbitrary repositories
sha1_file: add repository argument to prepare_alt_odb
sha1_file: add repository argument to link_alt_odb_entries
sha1_file: add repository argument to read_info_alternates
sha1_file: add repository argument to link_alt_odb_entry
sha1_file: add raw_object_store argument to alt_odb_usable
pack: move approximate object count to object store
...
The reason callers have to call this is to make sure either packed_git
or packed_git_mru pointers are initialized since we don't do that by
default. Sometimes it's hard to see this connection between where the
function is called and where packed_git pointer is used (sometimes in
separate functions).
Keep this dependency internal because now all access to packed_git and
packed_git_mru must go through get_xxx() wrappers.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
See previous patch for explanation.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a repository argument to allow prepare_packed_git callers to be
more specific about which repository to handle. See commit "sha1_file:
add repository argument to link_alt_odb_entry" for an explanation of
the #define trick.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In a process with multiple repositories open, packfile accessors
should be associated to a single repository and not shared globally.
Move packed_git and packed_git_mru into the_repository and adjust
callers to reflect this.
[nd: while at there, wrap access to these two fields in get_packed_git()
and get_packed_git_mru(). This allows us to lazily initialize these
fields without caller doing that explicitly]
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Teach parse-options API an option to help the completion script,
and make use of the mechanism in command line completion.
* nd/parseopt-completion: (45 commits)
completion: more subcommands in _git_notes()
completion: complete --{reuse,reedit}-message= for all notes subcmds
completion: simplify _git_notes
completion: don't set PARSE_OPT_NOCOMPLETE on --rerere-autoupdate
completion: use __gitcomp_builtin in _git_worktree
completion: use __gitcomp_builtin in _git_tag
completion: use __gitcomp_builtin in _git_status
completion: use __gitcomp_builtin in _git_show_branch
completion: use __gitcomp_builtin in _git_rm
completion: use __gitcomp_builtin in _git_revert
completion: use __gitcomp_builtin in _git_reset
completion: use __gitcomp_builtin in _git_replace
remote: force completing --mirror= instead of --mirror
completion: use __gitcomp_builtin in _git_remote
completion: use __gitcomp_builtin in _git_push
completion: use __gitcomp_builtin in _git_pull
completion: use __gitcomp_builtin in _git_notes
completion: use __gitcomp_builtin in _git_name_rev
completion: use __gitcomp_builtin in _git_mv
completion: use __gitcomp_builtin in _git_merge_base
...
Teach gc to stop traversal at promisor objects, and to leave promisor
packfiles alone. This has the effect of only repacking non-promisor
packfiles, and preserves the distinction between promisor packfiles and
non-promisor packfiles.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git gc" tries to avoid running two instances at the same time by
reading and writing pid/host from and to a lock file; it used to
use an incorrect fscanf() format when reading, which has been
corrected.
* aw/gc-lockfile-fscanf-fix:
gc: call fscanf() with %<len>s, not %<len>c, when reading hostname
Earlier in this codepath, we (ab)used "%<len>c" to read the hostname
recorded in the lockfile into locking_host[HOST_NAME_MAX + 1] while
substituting <len> with the actual value of HOST_NAME_MAX.
This turns out to be incorrect, as it is an instruction to read
exactly the specified number of bytes. Because we are trying to
read at most that many bytes, we should be using "%<len>s" instead.
Helped-by: A. Wilcox <awilfox@adelielinux.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit taught the tempfile code to give up
ownership over tempfiles that have been renamed or deleted.
That makes it possible to use a stack variable like this:
struct tempfile t;
create_tempfile(&t, ...);
...
if (!err)
rename_tempfile(&t, ...);
else
delete_tempfile(&t);
But doing it this way has a high potential for creating
memory errors. The tempfile we pass to create_tempfile()
ends up on a global linked list, and it's not safe for it to
go out of scope until we've called one of those two
deactivation functions.
Imagine that we add an early return from the function that
forgets to call delete_tempfile(). With a static or heap
tempfile variable, the worst case is that the tempfile hangs
around until the program exits (and some functions like
setup_shallow_temporary rely on this intentionally, creating
a tempfile and then leaving it for later cleanup).
But with a stack variable as above, this is a serious memory
error: the variable goes out of scope and may be filled with
garbage by the time the tempfile code looks at it. Let's
see if we can make it harder to get this wrong.
Since many callers need to allocate arbitrary numbers of
tempfiles, we can't rely on static storage as a general
solution. So we need to turn to the heap. We could just ask
all callers to pass us a heap variable, but that puts the
burden on them to call free() at the right time.
Instead, let's have the tempfile code handle the heap
allocation _and_ the deallocation (when the tempfile is
deactivated and removed from the list).
This changes the return value of all of the creation
functions. For the cleanup functions (delete and rename),
we'll add one extra bit of safety: instead of taking a
tempfile pointer, we'll take a pointer-to-pointer and set it
to NULL after freeing the object. This makes it safe to
double-call functions like delete_tempfile(), as the second
call treats the NULL input as a noop. Several callsites
follow this pattern.
The resulting patch does have a fair bit of noise, as each
caller needs to be converted to handle:
1. Storing a pointer instead of the struct itself.
2. Passing the pointer instead of taking the struct
address.
3. Handling a "struct tempfile *" return instead of a file
descriptor.
We could play games to make this less noisy. For example, by
defining the tempfile like this:
struct tempfile {
struct heap_allocated_part_of_tempfile {
int fd;
...etc
} *actual_data;
}
Callers would continue to have a "struct tempfile", and it
would be "active" only when the inner pointer was non-NULL.
But that just makes things more awkward in the long run.
There aren't that many callers, so we can simply bite
the bullet and adjust all of them. And the compiler makes it
easy for us to find them all.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We run an early part of "git gc" that deals with refs before
daemonising (and not under lock) even when running a background
auto-gc, which caused multiple gc processes attempting to run the
early part at the same time. This is now prevented by running the
early part also under the GC lock.
* jk/gc-pre-detach-under-hook:
gc: run pre-detach operations under lock
We normally try to avoid having two auto-gc operations run
at the same time, because it wastes resources. This was done
long ago in 64a99eb47 (gc: reject if another gc is running,
unless --force is given, 2013-08-08).
When we do a detached auto-gc, we run the ref-related
commands _before_ detaching, to avoid confusing lock
contention. This was done by 62aad1849 (gc --auto: do not
lock refs in the background, 2014-05-25).
These two features do not interact well. The pre-detach
operations are run before we check the gc.pid lock, meaning
that on a busy repository we may run many of them
concurrently. Ideally we'd take the lock before spawning any
operations, and hold it for the duration of the program.
This is tricky, though, with the way the pid-file interacts
with the daemonize() process. Other processes will check
that the pid recorded in the pid-file still exists. But
detaching causes us to fork and continue running under a
new pid. So if we take the lock before detaching, the
pid-file will have a bogus pid in it. We'd have to go back
and update it with the new pid after detaching. We'd also
have to play some tricks with the tempfile subsystem to
tweak the "owner" field, so that the parent process does not
clean it up on exit, but the child process does.
Instead, we can do something a bit simpler: take the lock
only for the duration of the pre-detach work, then detach,
then take it again for the post-detach work. Technically,
this means that the post-detach lock could lose to another
process doing pre-detach work. But in the long run this
works out.
That second process would then follow-up by doing
post-detach work. Unless it was in turn blocked by a third
process doing pre-detach work, and so on. This could in
theory go on indefinitely, as the pre-detach work does not
repack, and so need_to_gc() will continue to trigger. But
in each round we are racing between the pre- and post-detach
locks. Eventually, one of the post-detach locks will win the
race and complete the full gc. So in the worst case, we may
racily repeat the pre-detach work, but we would never do so
simultaneously (it would happen via a sequence of serialized
race-wins).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Convert code that divides and rounds up to use DIV_ROUND_UP to make the
intent clearer and reduce the number of magic constants.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix configuration codepath to pay proper attention to commondir
that is used in multi-worktree situation, and isolate config API
into its own header file.
* bw/config-h:
config: don't implicitly use gitdir or commondir
config: respect commondir
setup: teach discover_git_directory to respect the commondir
config: don't include config.h by default
config: remove git_config_iter
config: create config.h
Stop including config.h by default in cache.h. Instead only include
config.h in those files which require use of the config system.
Signed-off-by: Brandon Williams <bmwill@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some platforms have ulong that is smaller than time_t, and our
historical use of ulong for timestamp would mean they cannot
represent some timestamp that the platform allows. Invent a
separate and dedicated timestamp_t (so that we can distingiuish
timestamps and a vanilla ulongs, which along is already a good
move), and then declare uintmax_t is the type to be used as the
timestamp_t.
* js/larger-timestamps:
archive-tar: fix a sparse 'constant too large' warning
use uintmax_t for timestamps
date.c: abort if the system time cannot handle one of our timestamps
timestamp_t: a new data type for timestamps
PRItime: introduce a new "printf format" for timestamps
parse_timestamp(): specify explicitly where we parse timestamps
t0006 & t5000: skip "far in the future" test when time_t is too limited
t0006 & t5000: prepare for 64-bit timestamps
ref-filter: avoid using `unsigned long` for catch-all data type
Git's source code assumes that unsigned long is at least as precise as
time_t. Which is incorrect, and causes a lot of problems, in particular
where unsigned long is only 32-bit (notably on Windows, even in 64-bit
versions).
So let's just use a more appropriate data type instead. In preparation
for this, we introduce the new `timestamp_t` data type.
By necessity, this is a very, very large patch, as it has to replace all
timestamps' data type in one go.
As we will use a data type that is not necessarily identical to `time_t`,
we need to be very careful to use `time_t` whenever we interact with the
system functions, and `timestamp_t` everywhere else.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
gethostname(2) may not NUL terminate the buffer if hostname does
not fit; unfortunately there is no easy way to see if our buffer
was too small, but at least this will make sure we will not end up
using garbage past the end of the buffer.
* dt/xgethostname-nul-termination:
xgethostname: handle long hostnames
use HOST_NAME_MAX to size buffers for gethostname(2)
If the full hostname doesn't fit in the buffer supplied to
gethostname, POSIX does not specify whether the buffer will be
null-terminated, so to be safe, we should do it ourselves. Introduce
new function, xgethostname, which ensures that there is always a \0
at the end of the buffer.
Signed-off-by: David Turner <dturner@twosigma.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
POSIX limits the length of host names to HOST_NAME_MAX. Export the
fallback definition from daemon.c and use this constant to make all
buffers used with gethostname(2) big enough for any possible result
and a terminating NUL.
Inspired-by: David Turner <dturner@twosigma.com>
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: David Turner <dturner@twosigma.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We probe the "17/" loose object directory for auto-gc, and
use a local buffer to format the path. We can just use
git_path() for this. It handles paths of any length
(reducing our error handling). And because we feed the
result straight to a system call, we can just use the static
variant.
Note that git_path also knows the string "objects/" is
special, and will replace it with git_object_directory()
when necessary.
Another alternative would be to use sha1_file_name() for the
pretend object "170000...", but that ends up being more
hassle for no gain, as we have to truncate the final path
component.
Signed-off-by: Jeff King <peff@peff.net>
The experimental "split index" feature has gained a few
configuration variables to make it easier to use.
* cc/split-index-config: (22 commits)
Documentation/git-update-index: explain splitIndex.*
Documentation/config: add splitIndex.sharedIndexExpire
read-cache: use freshen_shared_index() in read_index_from()
read-cache: refactor read_index_from()
t1700: test shared index file expiration
read-cache: unlink old sharedindex files
config: add git_config_get_expiry() from gc.c
read-cache: touch shared index files when used
sha1_file: make check_and_freshen_file() non static
Documentation/config: add splitIndex.maxPercentChange
t1700: add tests for splitIndex.maxPercentChange
read-cache: regenerate shared index if necessary
config: add git_config_get_max_percent_split_change()
Documentation/git-update-index: talk about core.splitIndex config var
Documentation/config: add information for core.splitIndex
t1700: add tests for core.splitIndex
update-index: warn in case of split-index incoherency
read-cache: add and then use tweak_split_index()
split-index: add {add,remove}_split_index() functions
config: add git_config_get_split_index()
...
This function will be used in a following commit to get the expiration
time of the shared index files from the config, and it is generic
enough to be put in "config.c".
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A server can end up in a state where there are lots of unreferenced
loose objects (say, because many users are doing a bunch of rebasing
and pushing their rebased branches). Running "git gc --auto" in
this state would cause a gc.log file to be created, preventing
future auto gcs, causing pack files to pile up. Since many git
operations are O(n) in the number of pack files, this would lead to
poor performance.
Git should never get itself into a state where it refuses to do any
maintenance, just because at some point some piece of the maintenance
didn't make progress.
Teach Git to ignore gc.log files which are older than (by default)
one day old, which can be tweaked via the gc.logExpiry configuration
variable. That way, these pack files will get cleaned up, if
necessary, at least once per day. And operators who find a need for
more-frequent gcs can adjust gc.logExpiry to meet their needs.
There is also some cleanup: a successful manual gc, or a
warning-free auto gc with an old log file, will remove any old
gc.log files.
It might still happen that manual intervention is required
(e.g. because the repo is corrupt), but at the very least it won't
be because Git is too dumb to try again.
Signed-off-by: David Turner <dturner@twosigma.com>
Helped-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When git gc --auto does an incremental repack of loose objects, we do
not expect to be able to write a bitmap; it is very likely that
objects in the new pack will have references to objects outside of the
pack. So we shouldn't try to write a bitmap, because doing so will
likely issue a warning.
This warning was making its way into gc.log. When the gc.log was
present, future auto gc runs would refuse to run.
Patch by Jeff King.
Bug report, test, and commit message by David Turner.
Signed-off-by: David Turner <dturner@twosigma.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git gc --aggressive" used to limit the delta-chain length to 250,
which is way too deep for gaining additional space savings and is
detrimental for runtime performance. The limit has been reduced to
50.
* jk/reduce-gc-aggressive-depth:
gc: default aggressive depth to 50
"git gc --aggressive" used to limit the delta-chain length to 250,
which is way too deep for gaining additional space savings and is
detrimental for runtime performance. The limit has been reduced to
50.
* jk/reduce-gc-aggressive-depth:
gc: default aggressive depth to 50
This commit message is long and has lots of background and
numbers. The summary is: the current default of 250 doesn't
save much space, and costs CPU. It's not a good tradeoff.
Read on for details.
The "--aggressive" flag to git-gc does three things:
1. use "-f" to throw out existing deltas and recompute from
scratch
2. use "--window=250" to look harder for deltas
3. use "--depth=250" to make longer delta chains
Items (1) and (2) are good matches for an "aggressive"
repack. They ask the repack to do more computation work in
the hopes of getting a better pack. You pay the costs during
the repack, and other operations see only the benefit.
Item (3) is not so clear. Allowing longer chains means fewer
restrictions on the deltas, which means potentially finding
better ones and saving some space. But it also means that
operations which access the deltas have to follow longer
chains, which affects their performance. So it's a tradeoff,
and it's not clear that the tradeoff is even a good one.
The existing "250" numbers for "--aggressive" come
originally from this thread:
http://public-inbox.org/git/alpine.LFD.0.9999.0712060803430.13796@woody.linux-foundation.org/
where Linus says:
So when I said "--depth=250 --window=250", I chose those
numbers more as an example of extremely aggressive
packing, and I'm not at all sure that the end result is
necessarily wonderfully usable. It's going to save disk
space (and network bandwidth - the delta's will be re-used
for the network protocol too!), but there are definitely
downsides too, and using long delta chains may
simply not be worth it in practice.
There are some numbers in that thread, but they're mostly
focused on the improved window size, and measure the
improvement from --depth=250 and --window=250 together.
E.g.:
http://public-inbox.org/git/9e4733910712062006l651571f3w7f76ce64c6650dff@mail.gmail.com/
talks about the improved run-time of "git-blame", which
comes from the reduced pack size. But most of that reduction
is coming from --window=250, whereas most of the extra costs
come from --depth=250. There's a link in that thread showing
that increasing the depth beyond 50 doesn't seem to help
much with the size:
https://vcscompare.blogspot.com/2008/06/git-repack-parameters.html
but again, no discussion of the timing impact.
In an earlier thread from Ted Ts'o which discussed setting
the non-aggressive default (from 10 to 50):
http://public-inbox.org/git/20070509134958.GA21489%40thunk.org/
we have more numbers, with the conclusion that going past 50
does not help size much, and hurts the speed of normal
operations.
So from that, we might guess that 50 is actually a sweet
spot, even for aggressive, if we interpret aggressive to
"spend time now to make a better pack". It is not clear that
"--depth=250" is actually a better pack. It may be slightly
_smaller_, but it carries a run-time penalty.
Here are some more recent timings I did to verify that. They
show three things:
- the size of the resulting pack (so disk saved to store,
bandwidth saved on clones/fetches)
- the cost of "rev-list --objects --all", which shows the
effect of the delta chains on trees (commits typically
don't delta, and the command doesn't touch the blobs at
all)
- the cost of "log -Sfoo", which will additionally access
each blob
All cases were repacked with "git repack -adf --depth=$d
--window=250" (so basically, what would happen if we tweaked
the "gc --aggressive" default depth).
The timings are all wall-clock best-of-3. The machine itself
has plenty of RAM compared to the repositories (which is
probably typical of most workstations these days), so we're
really measuring CPU usage, as the whole thing will be in
disk cache after the first run.
The core.deltaBaseCacheLimit is at its default of 96MiB.
It's possible that tweaking it would have some impact on the
tests, as some of them (especially "log -S" on a large repo)
are likely to overflow that. But bumping that carries a
run-time memory cost, so for these tests, I focused on what
we could do just with the on-disk pack tradeoffs.
Each test is done for four depths: 250 (the current value),
50 (the current default that tested well previously), 100
(to show something on the larger side, which previous tests
showed was not a good tradeoff), and 10 (the very old
default, which previous tests showed was worse than 50).
Here are the numbers for linux.git:
depth | size | % | rev-list | % | log -Sfoo | %
-------+-------+-------+----------+--------+-----------+-------
250 | 967MB | n/a | 48.159s | n/a | 378.088 | n/a
100 | 971MB | +0.4% | 41.471s | -13.9% | 342.060 | -9.5%
50 | 979MB | +1.2% | 37.778s | -21.6% | 311.040s | -17.7%
10 | 1.1GB | +6.6% | 32.518s | -32.5% | 279.890s | -25.9%
and for git.git:
depth | size | % | rev-list | % | log -Sfoo | %
-------+-------+-------+----------+--------+-----------+-------
250 | 48MB | n/a | 2.215s | n/a | 20.922s | n/a
100 | 49MB | +0.5% | 2.140s | -3.4% | 17.736s | -15.2%
50 | 49MB | +1.7% | 2.099s | -5.2% | 15.418s | -26.3%
10 | 53MB | +9.3% | 2.001s | -9.7% | 12.677s | -39.4%
You can see that that the CPU savings for regular operations improves as we
decrease the depth. The savings are less for "rev-list" on a smaller repository
than they are for blob-accessing operations, or even rev-list on a larger
repository. This may mean that a larger delta cache would help (though setting
core.deltaBaseCacheLimit by itself doesn't).
But we can also see that the space savings are not that great as the depth goes
higher. Saving 5-10% between 10 and 50 is probably worth the CPU tradeoff.
Saving 1% to go from 50 to 100, or another 0.5% to go from 100 to 250 is
probably not.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"gc.autoPackLimit" when set to 1 should not trigger a repacking
when there is only one pack, but the code counted poorly and did
so.
* ew/gc-auto-pack-limit-fix:
gc: fix off-by-one error with gc.autoPackLimit
"gc.autoPackLimit" when set to 1 should not trigger a repacking
when there is only one pack, but the code counted poorly and did
so.
* ew/gc-auto-pack-limit-fix:
gc: fix off-by-one error with gc.autoPackLimit
This matches the documentation and allows gc.autoPackLimit=1
to maintain a single pack without attempting a repack on every
"git gc --auto" invocation.
Signed-off-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Having a leftover .idx file without corresponding .pack file in
the repository hurts performance; "git gc" learned to prune them.
* dk/gc-idx-wo-pack:
gc: remove garbage .idx files from pack dir
t5304: test cleaning pack garbage
prepare_packed_git(): refactor garbage reporting in pack directory
Add a custom report_garbage handler to collect and remove
garbage .idx files from the pack directory.
Signed-off-by: Doug Kelly <dougk.ff7@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
