A mutex and a condition variable is allocated for each thread and torn
down when the thread terminates. However, for certain workloads it can
happen that some threads are actually not started at all. In this case
we would leak the mutex and condition variable. Now we allocate them only
for those threads that are actually started.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In the threaded pack-objects code the main thread and the worker threads
must mutually signal that they have assigned a new pack of work or have
completed their work, respectively. Previously, the code used mutexes that
were locked in one thread and unlocked from a different thread, which is
bogus (and happens to work on Linux).
Here we rectify the implementation by using condition variables: There is
one condition variable on which the main thread waits until a thread
requests new work; and each worker thread has its own condition variable
on which it waits until it is assigned new work or signaled to terminate.
As a cleanup, the worker threads are spawned only after the initial work
packages have been assigned.
Signed-off-by: Johannes Sixt <johannes.sixt@telecom.at>
Acked-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code that splits the object list amongst work threads tries to do so
on "path" boundaries not to prevent good delta matches. However, in
some cases, a few paths may largely dominate the hash distribution and
it is not possible to have good load balancing without ignoring those
boundaries.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The current method consists of a master thread serving chunks of objects
to work threads when they're done with their previous chunk. The issue
is to determine the best chunk size: making it too large creates poor
load balancing, while making it too small has a negative effect on pack
size because of the increased number of chunk boundaries and poor delta
window utilization.
This patch implements a completely different approach by initially
splitting the work in large chunks uniformly amongst all threads, and
whenever a thread is done then it steals half of the remaining work from
another thread with the largest amount of unprocessed objects.
This has the advantage of greatly reducing the number of chunk boundaries
with an almost perfect load balancing.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is currently sorted and then walked backward. Not only this doesn't
feel natural for my poor brain, but it would make the next patch less
obvious as well.
So reverse the sort order, and reverse the list walking direction,
which effectively produce the exact same end result as before.
Also bring the relevant comment nearer the actual code and adjust it
accordingly, with minor additional clarifications.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The wrong value was substracted from delta_cache_size when replacing
a cached delta, as trg_entry->delta_size was used after the old size
had been replaced by the new size.
Noticed by Linus.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function mark_tree_uninteresting() assumed that the tree entries
are blob when they are not trees. This is not so. Since we do
not traverse into submodules (yet), the gitlinks should be ignored.
In general, we should try to start moving away from using the
"S_ISLNK()" like things for internal git state. It was a mistake to
just assume the numbers all were same across all systems in the first
place. This implementation converts to the "object_type", and then
uses a case statement.
Noticed by Ilari on IRC.
Test script taken from an earlier version by Dscho.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is a good idea to use pack index version 2 all the time since it has
proper protection against propagation of certain pack corruptions when
repacking which is not possible with index version 1, as demonstrated
in test t5302.
Hence this config option.
The default is still pack index version 1.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This one triggers only when git-pack-objects is called with
--all-progress and --stdout which is the combination used by
git-push.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since it is now OK to pass a null pointer to display_progress() and
stop_progress() resulting in a no-op, then we can simplify the code
and remove a bunch of lines by not making those calls conditional all
the time.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This allows for better management of progress "object" existence,
as well as making the progress display implementation more independent
from its callers.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Recently I was referred to the Grammar Police as the git-pack-objects
progress message 'Deltifying %u objects' is considered to be not
proper English to at least some small but vocal segment of the
English speaking population. Techncially we are applying delta
compression to these objects at this stage, so the new term is
slightly more acceptable to the Grammar Police but is also just
as correct.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Two functions, namely write_idx_file() and open_pack_file(), currently
return a const pointer. However that pointer is either a copy of the
first argument, or set to a malloc'd buffer when that first argument
is null. In the later case it is wrong to qualify that pointer as const
since ownership of the buffer is transferred to the caller to dispose of,
and obviously the free() function is not meant to be passed const
pointers.
Making the return pointer not const causes a warning when the first
argument is returned since that argument is also marked const.
The correct thing to do is therefore to remove the const qualifiers,
avoiding the need for ugly casts only to silence some warnings.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
To keep things well layered, sha1close() now returns the file descriptor
when it doesn't close the file.
An ugly cast was added to the return of write_idx_file() to avoid a
warning. A proper fix will come separately.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
... so don't even try in that case, and save another useless line of
progress display.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Each progress can be on a single line instead of two.
[sp: Changed "Checking files out" to "Checking out files" at
Johannes Sixt's suggestion as it better explains the
action that is taking place]
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
* jc/autogc:
git-gc --auto: run "repack -A -d -l" as necessary.
git-gc --auto: restructure the way "repack" command line is built.
git-gc --auto: protect ourselves from accumulated cruft
git-gc --auto: add documentation.
git-gc --auto: move threshold check to need_to_gc() function.
repack -A -d: use --keep-unreachable when repacking
pack-objects --keep-unreachable
Export matches_pack_name() and fix its return value
Invoke "git gc --auto" from commit, merge, am and rebase.
Implement git gc --auto
This new option is meant to be used in conjunction with the
options "git repack -a -d" usually invokes the underlying
pack-objects with. When this option is given, objects unreachable
from the refs in packs named with --unpacked= option are added
to the resulting pack, in addition to the reachable objects that
are not in packs marked with *.keep files.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This adds a --threads=<n> parameter to 'git pack-objects' with
documentation.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Try to keep object with the same name hash together.
Suggested by Martin Koegler.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With this, each thread get repeatedly assigned the next available chunk of
objects to process until the whole list is done. The idea is to have
reasonably small chunks so that all CPUs remain busy with a minimum
number of threads for as long as there is data to process.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
this is still rough, hence it is disabled by default. You need to compile
with "make THREADED_DELTA_SEARCH=1 ..." at the moment.
Threading is done on different portions of the object list to be
deltified. This is currently done by spliting the list into n parts and
then a thread is spawned for each of them. A better method would consist
of spliting the list into more smaller parts and have the n threads
pick the next part available.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is to help threadification of the delta search code, with a bonus
consistency check.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Not all objects are subject to deltification, so avoid carrying those
along, and provide the real count to progress display.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is based on Martin Koegler's idea to keep the object that
was successfully used as the base of the delta when it is about
to fall off the edge of the window. Instead of doing so only
for the objects at the edge of the window, this makes the window
a lru eviction mechanism. If an entry is used as a base, it is
moved to the last of the queue to be evicted.
This is a quick-and-dirty implementation, as it keeps the original
implementation of the data structure used for the window. This
originally was done as an array, not as an array of pointers,
because it was meant to be used as a cyclic FIFO buffer and a
plain array avoids an extra pointer indirection, while its FIFOness
eant that we are not "moving" the entries like this patch does.
The runtime from three versions were comparable. It seems to
make the resulting chain even shorter, which can only be good.
(stock "master") 15782196 bytes
chain length = 1: 2972 objects
chain length = 2: 2651 objects
chain length = 3: 2369 objects
chain length = 4: 2121 objects
chain length = 5: 1877 objects
...
chain length = 46: 490 objects
chain length = 47: 515 objects
chain length = 48: 527 objects
chain length = 49: 570 objects
chain length = 50: 408 objects
(with your patch) 15745736 bytes (0.23% smaller)
chain length = 1: 3137 objects
chain length = 2: 2688 objects
chain length = 3: 2322 objects
chain length = 4: 2146 objects
chain length = 5: 1824 objects
...
chain length = 46: 503 objects
chain length = 47: 509 objects
chain length = 48: 536 objects
chain length = 49: 588 objects
chain length = 50: 357 objects
(with this patch) 15612086 bytes (1.08% smaller)
chain length = 1: 4831 objects
chain length = 2: 3811 objects
chain length = 3: 2964 objects
chain length = 4: 2352 objects
chain length = 5: 1944 objects
...
chain length = 46: 327 objects
chain length = 47: 353 objects
chain length = 48: 304 objects
chain length = 49: 298 objects
chain length = 50: 135 objects
[jc: this is with code simplification follow-up from Nico]
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code favoring shallower deltas when size is equal was triggered
only when previous delta was also cached. There should be no relation
between cached deltas and same sized deltas.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a thin pack wants to send a tree object at "sub/dir", and
the commit that is common between the sender and the receiver
that is used as the base object has a subproject at that path,
we should not try to use the data at "sub/dir" of the base tree
as a tree object. It is not a tree to begin with, and more
importantly, the commit object there does not have to even
exist.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Not only are they unused, but the order in the function declaration
and the actual usage don't match.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
xmkstemp() performs error checking and prints a standard error message when
an error occur.
Signed-off-by: Luiz Fernando N. Capitulino <lcapitulino@mandriva.com.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 5a235b5e was missing this little detail. Otherwise your pack
will explode.
Problem noted by Brian Downing.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The delta depth doesn't have to be stored in the global object array
structure since it is only used during the deltification pass.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This adds an option (--window-memory=N) and configuration variable
(pack.windowMemory = N) to limit the memory size of the pack-objects
delta search window. This works by removing the oldest unpacked objects
whenever the total size goes above the limit. It will always leave
at least one object, though, so as not to completely eliminate the
possibility of computing deltas.
This is an extra limit on top of the normal window size (--window=N);
the window will not dynamically grow above the fixed number of entries
specified to fill the memory limit.
With this, repacking a repository with a mix of large and small objects
is possible even with a very large window.
Cleaner and correct circular buffer handling courtesy of Nicolas Pitre.
Signed-off-by: Brian Downing <bdowning@lavos.net>
Acked-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a new try_delta heuristic. Don't bother trying to make a delta if
the target object size is much smaller (currently 1/32) than the source,
as it's very likely not going to get a match. Even if it does, you will
have to read at least 32x the size of the new file to reassemble it,
which isn't such a good deal. This leads to a considerable performance
improvement when deltifying a mix of small and large files with a very
large window, because you don't have to wait for the large files to
percolate out of the window before things start going fast again.
Signed-off-by: Brian Downing <bdowning@lavos.net>
Acked-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We already apply a bias on the initial delta attempt with max_size being
a function of the base object depth. This has the effect of favoring
shallower deltas even if deeper deltas could be smaller, and therefore
creating a wider delta tree (see commits 4e8da195 and c3b06a69).
This principle should also be applied to all delta attempts for the same
object and not only the first attempt. With this the criteria for the
best delta is not only its size but also its depth, so that a shallower
delta might be selected even if it is larger than a deeper one. Even if
some deltas get larger, they allow for wider delta trees making the
depth limit less quickly reached and therefore better deltas can be
subsequently found, keeping the resulting pack size even smaller.
Runtime access to the pack should also benefit from shallower deltas.
Testing on different repositories showed slighter faster repacks,
smaller resulting packs, and a much nicer curve for delta depth
distribution with no more peak at the maximum depth level.
Improvements are even more significant with smaller depth limits.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change "try_delta" so that if it finds a delta that has the same size
but shallower depth than the existing delta, it will prefer the
shallower one. This makes certain delta trees vastly less deep.
Signed-off-by: Brian Downing <bdowning@lavos.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This uses "git-apply --whitespace=strip" to fix whitespace errors that have
crept in to our source files over time. There are a few files that need
to have trailing whitespaces (most notably, test vectors). The results
still passes the test, and build result in Documentation/ area is unchanged.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This patch unifies the write_index_file functions in
builtin-pack-objects.c and index-pack.c. As the name
"index" is overloaded in git, move in the direction of
using "idx" and "pack idx" when refering to the pack index.
There should be no change in functionality.
Signed-off-by: Geert Bosch <bosch@gnat.com>
Acked-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Two issues here:
1) git-repack -a --max-pack-size=10 on the GIT repo dies pretty quick.
There is a lot of confusion about deltas that were suposed to be
reused from another pack but that get stored undeltified due to pack
limit and object size doesn't match entry->size anymore. This test
is not really worth the complexity for determining when it is valid
so get rid of it.
2) If pack limit is reached, the object buffer is freed, including when
it comes from a cached delta data. In practice the object will be
stored in a subsequent pack undeltified, but let's make sure no
pointer to freed data subsists by clearing entry->delta_data.
I also reorganized that code a bit to make it more readable.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
Creating deltas between big blobs is a CPU and memory intensive task.
In the writing phase, all (not reused) deltas are redone.
This patch adds support for caching deltas from the deltifing phase, so
that that the writing phase is faster.
The caching is limited to small deltas to avoid increasing memory usage very much.
The implemented limit is (memory needed to create the delta)/1024.
Signed-off-by: Martin Koegler <mkoegler@auto.tuwien.ac.at>
Signed-off-by: Junio C Hamano <junkio@cox.net>
If builtin-pack-objects runs out of memory while finding
the best deltas, it bails out with an error.
If the delta index creation fails (because there is not enough memory),
we can downgrade the error message to a warning and continue with the
next object.
Signed-off-by: Martin Koegler <mkoegler@auto.tuwien.ac.at>
Signed-off-by: Junio C Hamano <junkio@cox.net>
