Prior to c85c792 (pull --rebase: be cleverer with rebased upstream
branches, 2008-01-26), pull --rebase would run
git rebase $merge_head
which resulted in a call to
git format-patch ... --ignore-if-in-upstream $merge_head..$cur_branch
This resulted in patches from $merge_head..$cur_branch being applied, as
long as they did not already exist in $cur_branch..$merge_head.
Unfortunately, when upstream is rebased, $merge_head..$cur_branch also
refers to "old" commits that have already been rebased upstream, meaning
that many patches that were already fixed upstream would be reapplied.
This could result in many spurious conflicts, as well as reintroduce
patches that were intentionally dropped upstream.
So the algorithm was changed in c85c792 (pull --rebase: be cleverer with
rebased upstream branches, 2008-01-26) and d44e712 (pull: support rebased
upstream + fetch + pull --rebase, 2009-07-19). Defining $old_remote_ref to
be the most recent entry in the reflog for @{upstream} that is an ancestor
of $cur_branch, pull --rebase was changed to run
git rebase --onto $merge_head $old_remote_ref
which results in a call to
git format-patch ... --ignore-if-in-upstream $old_remote_ref..$cur_branch
The whole point of this change was to reduce the number of commits being
reapplied, by avoiding commits that upstream already has or had.
In the rebased upstream case, this change achieved that purpose. It is
worth noting, though, that since $old_remote_ref is always an ancestor of
$cur_branch (by its definition), format-patch will not know what upstream
is and thus will not be able to determine if any patches are already
upstream; they will all be reapplied.
In the non-rebased upstream case, this new form is usually the same as the
original code but in some cases $old_remote_ref can be an ancestor of
$(git merge-base $merge_head $cur_branch)
meaning that instead of avoiding reapplying commits that upstream already
has, it actually includes more such commits. Combined with the fact that
format-patch can no longer detect commits that are already upstream (since
it is no longer told what upstream is), results in lots of confusion for
users (e.g. "git is giving me lots of conflicts in stuff I didn't even
change since my last push.")
Cases where additional commits could be reapplied include forking from a
commit other than the tracking branch, or amending/rebasing after pushing.
Cases where the inability to detect upstreamed commits cause problems
include independent discovery of a fix and having your patches get
upstreamed by some alternative route (e.g. pulling your changes to a third
machine, pushing from there, and then going back to your original machine
and trying to pull --rebase).
Fix the non-rebased upstream case by ignoring $old_remote_ref whenever it
is contained in $(git merge-base $merge_head $cur_branch). This should
have no affect on the rebased upstream case.
Acked-by: Santi Béjar <santi@agolina.net>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Shell reports exit status only from the most downstream command
in a pipeline. In these tests, we want to make sure that the
command fails in a controlled way, and produces a correct error
message.
This issue was known by Jay who submitted the patch, and also was
pointed out by Hannes during the review process, but I forgot to
fix it up before applying. Sorry about that.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jc/maint-reflog-bad-timestamp:
t0101: use a fixed timestamp when searching in the reflog
Update @{bogus.timestamp} fix not to die()
approxidate_careful() reports errorneous date string
8424981: "Fix invalid read in quote_c_style_counted" introduced a test
that used "caractère spécial" as a directory name.
Git creates it as "caract\303\250re sp\303\251cial"
OS X stores it as "caracte\314\200re spe\314\201cial"
To work around this problem, use the already introduced $FN as the
directory name.
Signed-off-by: Brian Gernhardt <brian@gernhardtsoftware.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* maint:
blame: prevent a segv when -L given start > EOF
git-push: document all the status flags used in the output
Fix parsing of imap.preformattedHTML and imap.sslverify
git-add documentation: Fix shell quoting example
blame would segv if given -L <lineno> with <lineno> past the end of the file.
While we're fixing the bug, add test cases for an invalid <start> when called
as -L <start>,<end> or -L<start>.
Signed-off-by: Jay Soffian <jaysoffian@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This reverts most of commit a2430dde8c.
That commit made the situation better for repositories with relatively
small number of objects. However with many objects and a small pack size
limit, the time required to complete the repack tends towards O(n^2),
or even much worse with long delta chains.
Signed-off-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to guess an output archive's format consumed any --format
options and built a new one. Jonathan noticed that it does so in an
unsafe way, risking to overflow the static buffer fmt_opt.
Change the code to keep the existing --format options intact and to only
add a new one if a format could be guessed based on the output file name.
The new option is added as the first one, allowing the existing ones to
overrule it, i.e. explicit --format options given on the command line win
over format guesses, as before.
To simplify the code further, format_from_name() is changed to return the
full --format option, thus no potentially dangerous sprintf() calls are
needed any more.
Reported-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This function did not work on strings that were not NUL-terminated. It
reads through a length-bounded string, searching for characters in need of
quoting. After we find one, we output the quoted character, then advance
our pointer to find the next one. However, we never decremented the
length, meaning we ended up looking at whatever random junk was stored
after the string.
This bug was not found by the existing tests because most code paths feed
a NUL-terminated string. The notable exception is a directory name being
fed by ls-tree.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Revert the previous attempt to skip this test on platforms where we
currently cannot determine the system load. We want to make sure that
the max-load-limit codepath produces results cleanly, when gitweb is
updated and becomes capable of reading the load average by some other
method.
The code to check for load returns 0 if it doesn't know how to find
load. It also checks to see if the current load is higher than the
max load. So to force the script to quit early by setting the maxload
variable negative which should work for systems where we can detect
load (which should be a positive number) and systems where we can't
(where detected load is 0)
Signed-off-by: Brian Gernhardt <brian@gernhardtsoftware.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently gitweb only knows how to check for load using /proc/loadavg,
which isn't available on all systems. We shouldn't fail the test just
because we don't know how to check the system load.
Signed-off-by: Brian Gernhardt <brian@gernhardtsoftware.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It seems that we have bad interaction with the code related to
GIT_WORK_TREE and "grep --no-index", and broke running grep inside
the .git directory. For now, just revert it and resurrect it after
1.7.0 ships.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The value passed to --max-pack-size used to count in MiB which was
inconsistent with the corresponding configuration variable as well as
other command arguments which are defined to count in bytes with an
optional unit suffix. This brings --max-pack-size in line with the
rest of Git.
Also, in order not to cause havoc with people used to the previous
megabyte scale, and because this is a sane thing to do anyway, a
minimum size of 1 MiB is enforced to avoid an explosion of pack files.
Adjust and extend test suite accordingly.
Signed-off-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
First of all, trying to run 'git verify-pack' on packs produced by
the tests using pack.packSizeLimit always failed. After lots of digging
and head scratching, it turns out that the preceeding test simulating
a SHA1 collision did leave the repository quite confused, impacting
subsequent tests.
So let's move that destructive test last, and add tests to run
verify-pack on the output from those packSizeLimit tests to catch such
goofage.
Finally, group those packSizeLimit tests together.
Signed-off-by: Nicolas Pitre <nico@fluxnic.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* ms/filter-branch-submodule:
filter-branch: Add tests for submodules in tree-filter
filter-branch: Fix to allow replacing submodules with another content
* 'jh/gitweb-caching' (early part):
gitweb: Add optional extra parameter to die_error, for extended explanation
gitweb: add a "string" variant of print_sort_th
gitweb: add a "string" variant of print_local_time
gitweb: Check that $site_header etc. are defined before using them
gitweb: Makefile improvements
gitweb: Load checking
gitweb: Make running t9501 test with '--debug' reliable and usable
* sp/maint-fast-import-large-blob:
fast-import: Stream very large blobs directly to pack
bash: don't offer remote transport helpers as subcommands
Conflicts:
fast-import.c
If a blob is larger than the configured big-file-threshold, instead
of reading it into a single buffer obtained from malloc, stream it
onto the end of the current pack file. Streaming the larger objects
into the pack avoids the 4+ GiB memory footprint that occurs when
fast-import is processing 2+ GiB blobs.
Signed-off-by: Shawn O. Pearce <spearce@spearce.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This changes slightly the behavior of gitweb, so that it verifies
that the box isn't inundated with before attempting to serve gitweb.
If the box is overloaded, it basically returns a 503 Server Unavailable
until the load falls below the defined threshold. This helps dramatically
if you have a box that's I/O bound, reaches a certain load and you
don't want gitweb, the I/O hog that it is, increasing the pain the
server is already undergoing.
This behavior is controlled by $maxload configuration variable.
Default is a load of 300, which for most cases should never be hit.
Unset it (set it to undefined value, i.e. undef) to turn off checking.
Currently it requires that '/proc/loadavg' file exists, otherwise the
load check is bypassed (load is taken to be 0). So platforms that do
not implement '/proc/loadavg' currently cannot use this feature
(provisions are included for additional checks to be added by others).
There is simple test in t/t9501-gitweb-standalone-http-status.sh to
check that it correctly returns "503 Service Unavailable" if load is
too high, and also if there are any Perl warnings or errors.
Signed-off-by: John 'Warthog9' Hawley <warthog9@kernel.org>
Signed-off-by: Jakub Narebski <jnareb@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Remove test_debug lines after 'snapshots: tgz only default format
enabled' and 'snapshots: all enabled in default, use default disabled
value' tests. Those tests constitute of multiple gitweb_run
invocation, therefore outputting gitweb.output for the last gitweb_run
wouldn't help much in debugging test failure, and can only confuse.
For snapshot tests which check for "200 OK" status, change
test_debug 'cat gitweb.output'
to
test_debug 'cat gitweb.headers'
Otherwise when running this test with '--debug' option,
t/t9501-gitweb-standalone-http-status.sh would dump *binary data* (the
snapshot itself) to standard output, which can mess up state of terminal
due to term control characters which can be embedded in output.
Signed-off-by: Jakub Narebski <jnareb@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add tests to make sure that:
1) a submodule can be removed and its content replaced with regular files
('rewrite submodule with another content'). This test passes only with
the previous patch applied.
2) it is possible to replace submodule revision by direct index
manipulation ('replace submodule revision'). Although it would be
better to run such a filter in --index-filter, this test shows that
this functionality is not broken by the previous patch. This succeeds
both with and without the previous patch.
Signed-off-by: Michal Sojka <sojkam1@fel.cvut.cz>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Something like foo@{-1} is nonsensical, as the @{-N} syntax
is reserved for "the Nth last branch", and is not an actual
reflog selector. We should not feed such nonsense to
approxidate at all.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Previously interpret_branch_name would see @{-1} and stop
parsing, leaving the @{u} as cruft that provoked an error.
Instead, we should recurse if there is more to parse.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that we have several different types of @{} syntax, it
is a good idea to test them together, which reveals some
failures.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The original version used relative approxidates, which don't
reproduce as reliably as absolute ones. Commit 6c647a fixed
this for one case, but missed the "silly" case.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"diff --cc" output t4038 tests was fixed by b810cbb (diff --cc: a lost
line at the beginning of the file is shown incorrectly, 2009-07-22), which
was actually the commit that introduced this test..
An error in "git merge -s resolve" t6035 tests was fixed by 730f728
(unpack-trees.c: look ahead in the index, 2009-09-20).
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* jc/maint-reflog-bad-timestamp:
t0101: use a fixed timestamp when searching in the reflog
Update @{bogus.timestamp} fix not to die()
approxidate_careful() reports errorneous date string
* jl/diff-submodule-ignore:
Teach diff --submodule that modified submodule directory is dirty
git diff: Don't test submodule dirtiness with --ignore-submodules
Make ce_uptodate() trustworthy again
Using a dollar sign in double quotes isn't portable. Escape them with
a backslash or replace the double quotes with single quotes.
Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For a long time, the time based reflog syntax (e.g. master@{yesterday})
didn't complain when the "human readable" timestamp was misspelled, as
the underlying mechanism tried to be as lenient as possible. The funny
thing was that parsing of "@{now}" even relied on the fact that anything
not recognized by the machinery returned the current timestamp.
Introduce approxidate_careful() that takes an optional pointer to an
integer, that gets assigned 1 when the input does not make sense as a
timestamp.
As I am too lazy to fix all the callers that use approxidate(), most of
the callers do not take advantage of the error checking, but convert the
code to parse reflog to use it as a demonstration.
Tests are mostly from Jeff King.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When your current directory is not at the root of the working tree, and you
use the "-f" option with a relative path, the current code tries to read
from a wrong file, since argv[2] is now beyond the end of the rearranged
argument list.
This patch replaces the incorrect argv[2] with the variable holding the
given config file name.
The bug was introduced by d64ec16 (git config: reorganize to use parseopt).
[jc: added test]
Signed-off-by: Johan Herland <johan@herland.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The patch detection wants to inspect all the headers of a rfc2822 message
and ensure that they look like header fields. The headers are always
separated from the message body with a blank line. When Thunderbird saves
the message the blank line separating the headers from the body includes a
CR. The patch detection is failing because a CRLF doesn't match /^$/. Fix
this by allowing a CR to exist on the separating line.
Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We shouldn't have literal CR's in tests as they aren't portable.
Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
append_cr(), remove_cr(), q_to_nul() and q_to_cr() are defined in multiple
tests. Consolidate them into test-lib.sh so we can stop redefining them.
The use of remove_cr() in t0020 to test for a CR is replaced with a new
function has_cr() to accurately reflect what is intended (the output of
remove_cr() was being thrown away).
Signed-off-by: Stephen Boyd <bebarino@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
