When we are limiting a rev-list traversal due to
UNINTERESTING refs, we have to walk down the tips (both
interesting and uninteresting) to find where they intersect.
We keep a queue of commits to examine, pop commits off
the queue one by one, and potentially add their parents. The
size of the queue will naturally fluctuate based on the
"width" of the history graph; i.e., the number of
simultaneous lines of development. But for the most part it
will stay in the same ballpark as the initial number of tips
we fed, shrinking over time (as we hit common ancestors of
the tips). So roughly speaking, if we start with `N` tips,
we'll spend much of the time with a queue around `N` items.
For each UNINTERESTING commit we pop, we call
still_interesting to check whether marking its parents as
UNINTERESTING has made the whole queue uninteresting (in
which case we can quit early). Because the queue is stored
as a linked list, this is `O(N)`, where `N` is the number of
items in the queue. So processing a queue with `N` commits
marked UNINTERESTING (and one or more interesting commits)
will take `O(N^2)`.
If you feed a lot of positive tips, this isn't a problem.
They aren't UNINTERESTING, so they don't incur the
still_interesting check. It also isn't a problem if you
traverse from an interesting tip to some UNINTERESTING
bases. We order the queue by recency, so the interesting
commits stay at the front of the queue as we walk down them.
The linear check can exit early as soon as it sees one
interesting commit left in the queue.
But if you want to know whether an older commit is reachable
from a set of newer tips, we end up processing in the
opposite direction: from the UNINTERESTING ones down to the
interesting one. This may happen when we call:
git rev-list $commits --not --all
in check_everything_connected after a fetch. If we fetched
something much older than most of our refs, and if we have a
large number of refs, the traversal cost is dominated by the
quadratic behavior.
These commands simulate the connectivity check of such a
fetch, when you have `$n` distinct refs in the receiver:
# positive ref is 100,000 commits deep
git rev-list --all | head -100000 | tail -1 >input
# huge number of more recent negative refs
git rev-list --all | head -$n | sed s/^/^/ >>input
time git rev-list --stdin <input
Here are timings for various `n` on the linux.git
repository. The `n=1` case provides a baseline for just
walking the commits, which lets us see the still_interesting
overhead. The times marked with `+` subtract that baseline
to show just the extra time growth due to the large number
of refs. The `x` numbers show the slowdown of the adjusted
time versus the prior trial.
n | before | after
--------------------------------------------------------
1 | 0.991s | 0.848s
10000 | 1.120s (+0.129s) | 0.885s (+0.037s)
20000 | 1.451s (+0.460s, 3.5x) | 0.923s (+0.075s, 2.0x)
40000 | 2.731s (+1.740s, 3.8x) | 0.994s (+0.146s, 1.9x)
80000 | 8.235s (+7.244s, 4.2x) | 1.123s (+0.275s, 1.9x)
Each trial doubles `n`, so you can see the quadratic (`4x`)
behavior before this patch. Afterwards, we have a roughly
linear relationship.
The implementation is fairly straightforward. Whenever we do
the linear search, we cache the interesting commit we find,
and next time check it before doing another linear search.
If that commit is removed from the list or becomes
UNINTERESTING itself, then we fall back to the linear
search. This is very similar to the trick used by fce87ae
(Fix quadratic performance in rewrite_one., 2008-07-12).
I considered and rejected several possible alternatives:
1. Keep a count of UNINTERESTING commits in the queue.
This requires managing the count not only when removing
an item from the queue, but also when marking an item
as UNINTERESTING. That requires touching the other
functions which mark commits, and would require knowing
quickly which commits are in the queue (lookup in the
queue is linear, so we would need an auxiliary
structure or to also maintain an IN_QUEUE flag in each
commit object).
2. Keep a separate list of interesting commits. Drop items
from it when they are dropped from the queue, or if
they become UNINTERESTING. This again suffers from
extra complexity to maintain the list, not to mention
CPU and memory.
3. Use a better data structure for the queue. This is
something that could help the fix in fce87ae, because
we order the queue by recency, and it is about
inserting quickly in recency order. So a normal
priority queue would help there. But here, we cannot
disturb the order of the queue, which makes things
harder. We really do need an auxiliary index to track
the flag we care about, which is basically option (2)
above.
The "cache" trick is simple, and the numbers above show that
it works well in practice. This is because the length of
time it takes to find an interesting commit is proportional
to the length of time it will remain cached (i.e., if we
have to walk a long way to find it, it also means we have to
pop a lot of elements in the queue until we get rid of it
and have to find another interesting commit).
The worst case is still quadratic, though. We could have `N`
uninteresting commits at the front of the queue, followed by
`N` interesting commits, where commit `i` has parent `i+N`.
When we pop commit `i`, we will notice that the parent of
the next commit, `i+1+N` is still interesting and cache it.
But then handling commit `i+1`, we will mark its parent
`i+1+N` uninteresting, and immediately invalidate our cache.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When commit fe8e3b7 refactored type_from_string to allow
input that was not NUL-terminated, it switched to using
strncmp instead of strcmp. But this means we check only the
first "len" bytes of the strings, and ignore any remaining
bytes in the object_type_string. We should make sure that it
is also "len" bytes, or else we would accept "comm" as
"commit", and so forth.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The name (not user) and email setting should be in config section
"user" and not in "core" as documented in Documentation/config.txt.
Signed-off-by: Ossi Herrala <oherrala@gmail.com>
Reviewed-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The old description is rather clearly a wrong cut-and-paste from
t2016-checkout-patch.sh.
Signed-off-by: Matthieu Moy <Matthieu.Moy@imag.fr>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The test passes if one replaces the 'e' command with a 'y' command in
the 'add -p' session.
Reported-by: Tanky Woo <wtq1990@gmail.com>
Signed-off-by: Matthieu Moy <Matthieu.Moy@imag.fr>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The second argument to test_path_is_file and test_path_is_dir
must be $2 and not $*, which instead would repeat the file
name in the error message.
Signed-off-by: Elia Pinto <gitter.spiros@gmail.com>
Reviewed-by: Matthieu Moy <Matthieu.Moy@grenoble-inp.fr>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Because the function reads one character at the time, unfortunately
we cannot use the easier skip_utf8_bom() helper, but at least we do
not have to duplicate the constant string this way.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With the recent change to ignore the UTF8 BOM at the beginning of
.gitignore files, we now have two codepaths that do such a skipping
(the other one is for reading the configuration files).
Introduce utf8_bom[] constant string and skip_utf8_bom() helper
and teach .gitignore code how to use it.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Even though the previous step shifts where the "entry" begins, we
still iterate over the original buf[], which may begin with the
UTF-8 BOM we are supposed to be skipping. At the end of the first
line, the code grabs the contents of it starting at "entry", so
there is nothing wrong per-se, but the logic looks really confused.
Instead, move the buf pointer and shrink its size, to truly
pretend that UTF-8 BOM did not exist in the input.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some text editors like Notepad or LibreOffice write an UTF-8 BOM in
order to indicate that the file is Unicode text rather than whatever the
current locale would indicate.
If someone uses such an editor to edit a gitignore file, we are left
with those three bytes at the beginning of the file. If we do not skip
them, we will attempt to match a filename with the BOM as prefix, which
won't match the files the user is expecting.
Signed-off-by: Carlos Martín Nieto <cmn@elego.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In t1430, we check whether deleting the branch "../../foo"
will delete ".git/foo". However, this is not that
interesting a test; the precious file ".git/foo" does not
look like a ref, so even if we did not notice the "escape"
from the "refs/" hierarchy, we would fail for that reason
(i.e., if you turned refname_is_safe into a noop, the test
still passes).
Let's add an additional test for the same thing, but with a
file that actually looks like a ref. That will make sure we
are exercising the refname_is_safe code. While we're at it,
let's also make the code work a little harder by adding some
extra paths and some empty path components.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Prior to d0f810f (refs.c: allow listing and deleting badly
named refs, 2014-09-03), read_packed_refs would barf on any
malformed refnames by virtue of calling create_ref_entry
with the "check" parameter set to 1. That commit loosened
our reading so that we call check_refname_format ourselves
and just set a REF_BAD_NAME flag.
We then call create_ref_entry with the check parameter set
to 0. That function learned to do an extra safety check even
when the check parameter is 0, so that we don't load any
dangerous refnames (like "../../../etc/passwd"). This is
implemented by calling refname_is_safe() in
create_ref_entry().
However, we can observe that refname_is_safe() can only be
true if check_refname_format() also failed. So in the common
case of a sanely named ref, we perform _both_ checks, even
though we know that the latter will never trigger. This has
a noticeable performance impact when the packed-refs file is
large.
Let's drop the refname_is_safe check from create_ref_entry(),
and make it the responsibility of the caller. Of the three
callers that pass a check parameter of "0", two will have
just called check_refname_format(), and can check the
refname-safety only when it fails. The third case,
pack_if_possible_fn, is copying from an existing ref entry,
which must have previously passed our safety check.
With this patch, running "git rev-parse refs/heads/does-not-exist"
on a repo with a large (1.6GB) packed-refs file went from:
real 0m6.768s
user 0m6.340s
sys 0m0.432s
to:
real 0m5.703s
user 0m5.276s
sys 0m0.432s
for a wall-clock speedup of 15%.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We spend a lot of time in strbuf_getwholeline in a tight
loop reading characters from a stdio handle into a buffer.
The libc getdelim() function can do this for us with less
overhead. It's in POSIX.1-2008, and was a GNU extension
before that. Therefore we can't rely on it, but can fall
back to the existing getc loop when it is not available.
The HAVE_GETDELIM knob is turned on automatically for Linux,
where we have glibc. We don't need to set any new
feature-test macros, because we already define _GNU_SOURCE.
Other systems that implement getdelim may need to other
macros (probably _POSIX_C_SOURCE >= 200809L), but we can
address that along with setting the Makefile knob after
testing the feature on those systems.
Running "git rev-parse refs/heads/does-not-exist" on a repo
with an extremely large (1.6GB) packed-refs file went from
(best-of-5):
real 0m8.601s
user 0m8.084s
sys 0m0.524s
to:
real 0m6.768s
user 0m6.340s
sys 0m0.432s
for a wall-clock speedup of 21%.
Based on a patch from Rasmus Villemoes <rv@rasmusvillemoes.dk>.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As with the recent speedup to strbuf_addch, we can avoid
calling strbuf_grow() in a tight loop of single-character
adds by instead checking strbuf_avail.
Note that we would instead call strbuf_addch directly here,
but it does more work than necessary: it will NUL-terminate
the result for each character read. Instead, in this loop we
read the characters one by one and then add the terminator
manually at the end.
Running "git rev-parse refs/heads/does-not-exist" on a repo
with an extremely large (1.6GB) packed-refs file went from
(best-of-5):
real 0m10.948s
user 0m10.548s
sys 0m0.412s
to:
real 0m8.601s
user 0m8.084s
sys 0m0.524s
for a wall-clock speedup of 21%.
Helped-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We mark strbuf_addch as inline, because we expect it may be
called from a tight loop. However, the first thing it does
is call the non-inline strbuf_grow(), which can handle
arbitrary-sized growth. Since we know that we only need a
single character, we can use the inline strbuf_avail() to
quickly check whether we need to grow at all.
Our check is redundant when we do call strbuf_grow(), but
that's OK. The common case is that we avoid calling it at
all, and we have made that case faster.
On a silly pathological case:
perl -le '
print "[core]";
print "key$_ = value$_" for (1..1000000)
' >input
git config -f input core.key1
this dropped the time to run git-config from:
real 0m0.159s
user 0m0.152s
sys 0m0.004s
to:
real 0m0.140s
user 0m0.136s
sys 0m0.004s
for a savings of 12%.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We read config files character-by-character from a stdio
handle using fgetc(). This incurs significant locking
overhead, even though we know that only one thread can
possibly access the handle. We can speed this up by taking
the lock ourselves, and then using getc_unlocked to read
each character.
On a silly pathological case:
perl -le '
print "[core]";
print "key$_ = value$_" for (1..1000000)
' >input
git config -f input core.key1
this dropped the time to run git-config from:
real 0m0.263s
user 0m0.260s
sys 0m0.000s
to:
real 0m0.159s
user 0m0.152s
sys 0m0.004s
for a savings of 39%. Most config files are not this big,
but the savings should be proportional to the size of the
file (i.e., we always save 39%, just of a much smaller
number).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
strbuf_getwholeline calls getc in a tight loop. On modern
libc implementations, the stdio code locks the handle for
every operation, which means we are paying a significant
overhead. We can get around this by locking the handle for
the whole loop and using the unlocked variant.
Running "git rev-parse refs/heads/does-not-exist" on a repo
with an extremely large (1.6GB) packed-refs file went from:
real 0m18.900s
user 0m18.472s
sys 0m0.448s
to:
real 0m10.953s
user 0m10.384s
sys 0m0.580s
for a wall-clock speedup of 42%. All times are best-of-3,
and done on a glibc 2.19 system.
Note that we call into strbuf_grow while holding the lock.
It's possible for that function to call other stdio
functions (e.g., printing to stderr when dying due to malloc
error); however, the POSIX.1-2001 definition of flockfile
makes it clear that the locks are per-handle, so we are fine
unless somebody else tries to read from our same handle.
This doesn't ever happen in the current code, and is
unlikely to be added in the future (we would have to do
something exotic like add a die_routine that tried to read
from stdin).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
POSIX.1-2001 specifies some functions for optimizing the
locking out of tight getc() loops. Not all systems are
POSIX, though, and even not all POSIX systems are required
to implement these functions. We can check for the
feature-test macro to see if they are available, and if not,
provide a noop implementation.
There's no Makefile knob here, because we should just detect
this automatically. If there are very bizarre systems, we
may need to add one, but it's not clear yet in which
direction:
1. If a system defines _POSIX_THREAD_SAFE_FUNCTIONS but
these functions are missing or broken, we would want a
knob to manually turn them off.
2. If a system has these functions but does not define
_POSIX_THREAD_SAFE_FUNCTIONS, we would want a knob to
manually turn them on.
We can add such a knob when we find a real-world system that
matches this.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
strbuf_getwholeline calls fgetc in a tight loop. Using the
getc form, which can be implemented as a macro, should be
faster (and we do not care about it evaluating our argument
twice, as we just have a plain variable).
On my glibc system, running "git rev-parse
refs/heads/does-not-exist" on a file with an extremely large
(1.6GB) packed-refs file went from (best of 3 runs):
real 0m19.383s
user 0m18.876s
sys 0m0.528s
to:
real 0m18.900s
user 0m18.472s
sys 0m0.448s
for a wall-clock speedup of 2.5%.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This reverts commit 2bf15a3330, whose
intention was good, but the verbosity levels used in merge-recursive
turns out to be rather uneven. For example, a merge of two branches
with conflicting submodule updates used to report CONFLICT: output
with --quiet but no longer (which *is* desired), while the final
"Automatic merge failed; fix conflicts and then commit" message is
still shown even with --quiet (which *is* inconsistent).
Originally reported by Bryan Turner; it is too early to declare what
the concensus is, but it seems that we would need to level the
verbosity levels used in merge strategy backends before we can go
forward. In the meantime, we'd revert to the old behaviour until
that happens.
cf. $gmane/267245
Windows does not have process groups. It is, therefore, the simplest
to pretend that each process is in its own process group.
While here, move the getppid() stub from its old location (between
two sync related functions) next to the two new functions.
Signed-off-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Disable the display of the progress if stderr is not the
current foreground process.
Still display the final result when done.
Signed-off-by: Luke Mewburn <luke@mewburn.net>
Acked-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function parses the input to compute the broken-down time in
"struct tm", and the GMT timezone offset. If the timezone offset
does not exist in the input, the broken-down time is turned into the
number of seconds since epoch both in the current timezone and in
GMT and the offset is computed as their difference.
However, we forgot to make sure tm.tm_isdst is set to -1 (i.e. let
the system figure out if DST is in effect in the current timezone
when turning the broken-down time to the number of seconds since
epoch); it is done so at the beginning of the function, but a call
to match_digit() in the function can lead to a call to gmtime_r() to
clobber the field.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Diagnosed-by: Eric Sunshine <sunshine@sunshineco.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the input does not have GMT timezone offset, the code computes
it by computing the local and GMT time for the given timestamp. But
there is no point doing so if the given timestamp is known to be a
bogus one.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"diff-highlight" (in contrib/) used to show byte-by-byte
differences, which meant that multi-byte characters can be chopped
in the middle. It learned to pay attention to character boundaries
(assuming the UTF-8 payload).
* jk/colors:
diff-highlight: do not split multibyte characters
"git merge --quiet" did not squelch messages from the underlying
merge-recursive strategy.
* jk/merge-quiet:
merge: pass verbosity flag down to merge-recursive
A push into an unborn branch, with "receive.denyCurrentBranch" set
to "updateInstead", did not check out the working tree as expected.
* jc/update-instead-into-void:
push-to-deploy: allow pushing into an unborn branch and updating it
Changed inaccurate count of "rough rules" from three to the more
generic 'a few'.
Signed-off-by: Julian Gindi <juliangindi@gmail.com>
Reviewed-by: Eric Sunshine <sunshine@sunshineco.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "help-all" option is being initialized with a wrong value.
While being semantically wrong this can also cause a segmentation
fault in gcc on ARMv7 hardfloat platforms with a hardened
toolchain. Fix this by initializing with a NULL value.
Signed-off-by: Patrick Steinhardt <ps@pks.im>
Reviewed-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
By using a tree with multiple identical files and allowing copy detection to
choose any one of them, the check in the test is unnecessarily complex. We can
simplify by:
* Modify source file (file2) before copying the file.
* Check that only file2 is the source in the output of "p4 filelog".
* Remove all "case" statements and replace them with simple tests to check
that source is "file2".
Signed-off-by: Vitor Antunes <vitor.hda@gmail.com>
Acked-by: Luke Diamand <luke@diamand.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ignore an extra ':' at the end of the hostname in URL's like
"ssh://example.com:/path/to/repo"
The colon is meant to separate a port number from the hostname.
If the port is empty, the colon should be ignored, see RFC 3986.
It had been working for URLs with ssh:// scheme, but was unintentionally
broken in 86ceb3, "allow ssh://user@[2001:db8::1]/repo.git"
Reported-by: Reid Woodbury Jr. <reidw@rawsound.com>
Signed-off-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
