The code to produce progress output from "git commit-graph --write"
had a few breakages, which have been fixed.
* sg/commit-graph-progress-fix:
commit-graph: fix "Writing out commit graph" progress counter
commit-graph: fix progress of reachable commits
When an aliased command, whose output is piped to a pager by git,
gets killed by a signal, the pager got into a funny state, which
has been corrected (again).
* ta/wait-on-aliased-commands-upon-signal:
Wait for child on signal death for aliases to externals
Wait for child on signal death for aliases to builtins
The completion for diff command was added in fd0bc17557 but
missed the show command which also supports --color-moved[-ws].
This suffers from the very same problem [1] as the referenced
commit: no comma-separated list completion for --color-moved-ws.
[1]: https://github.com/scop/bash-completion/issues/240
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To display a progress line while reading commits from standard input
and looking them up, 5b6653e523 (builtin/commit-graph.c: dereference
tags in builtin, 2020-05-13) should have added a pair of
start_delayed_progress() and stop_progress() calls around the loop
reading stdin. Alas, the stop_progress() call ended up at the wrong
place, after write_commit_graph(), which does all the commit-graph
computation and writing, and has several progress lines of its own.
Consequently, that new
Collecting commits from input: 1234
progress line is overwritten by the first progress line shown by
write_commit_graph(), and its final "done" line is shown last, after
everything is finished:
$ { sleep 3 ; git rev-list -3 HEAD ; sleep 1 ; } | ~/src/git/git commit-graph write --stdin-commits
Expanding reachable commits in commit graph: 873402, done.
Writing out commit graph in 4 passes: 100% (3493608/3493608), done.
Collecting commits from input: 3, done.
Furthermore, that stop_progress() call was added after the 'cleanup'
label, where that loop reading stdin jumps in case of an error. In
case of invalid input this then results in the "done" line shown after
the error message:
$ { sleep 3 ; git rev-list -3 HEAD ; echo junk ; } | ~/src/git/git commit-graph write --stdin-commits
error: unexpected non-hex object ID: junk
Collecting commits from input: 3, done.
Move that stop_progress() call to the right place.
While at it, drop the unnecessary 'if (progress)' condition protecting
the stop_progress() call, because that function is prepared to handle
a NULL progress struct.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Reviewed-by: Derrick Stolee <stolee@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
An earlier "fix" to this script gave up updating it not to rely on
the current time because we cannot control what timestamp subversion
gives its commits. We however could solve the issue in a different
way and still use deterministic timestamps on Git commits.
One fix would be to sort the list of trees before removing duplicates,
but that loses information:
- we do care that the fetched history is in the same order
- there's a tree which appears twice in the history, and we'd want to
make sure that it's there both times
So instead, let's de-duplicate using a hash (preserving the order), and
drop only lines with identical trees and subjects (preserving the tree
which appears twice, since it has different subjects each time).
Signed-off-by: Jeff King <peff@peff.net>
Acked-by: Eric Wong <e@80x24.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We always set the name and email for committer and author idents to make
the test suite more deterministic, but not timestamps. Many scripts use
test_tick to get consistent and sensibly incrementing timestamps as they
create commits. But other scripts don't particularly care about the
timestamp, and are happy to use whatever the current system time is.
This non-determinism can be annoying:
- when debugging a test, comparing results between two runs can be
difficult, because the commit ids change
- this can sometimes cause tests to be racy. E.g., traversal order
depends on timestamp order. Even in a well-ordered set of commands,
because our timestamp granularity is one second, two commits might
sometimes have the same timestamp and sometimes differ.
Let's set a default timestamp for all scripts to use. Any that use
test_tick already will be unaffected (because their first test_tick call
will overwrite our default), but it will make things a bit more
deterministic for those that don't.
We should be able to choose any time we want here. I picked this one
because:
- it differs from the initial test_tick default, which may make it
easier to distinguish when debugging tests. I picked "April 1st
13:14:15" in the hope that it might stand out.
- it's slightly before the test_tick default. Some tests create some
commits before the first call to test_tick, so using an older
timestamps for those makes sense chronologically. Note that this
isn't how things currently work (where system times are usually more
recent than test_tick), but that also allows us to flush out a few
hidden timestamp dependencies (like the one recently fixed in
t5539).
- we could likewise pick any timezone we want. Choosing +0000 would
have required fixing up fewer tests, but we're more likely to turn
up interesting cases by not matching $TZ exactly. And since
test_tick already checks "-0700", let's try something in the "+"
zone range for variety.
It's possible that the non-deterministic times could help flush out bugs
(e.g., if something broke when the clock flipped over to 2021, our test
suite would let us know). But historically that hasn't been the case;
all time-dependent outcomes we've seen turned out to be accidentally
flaky tests (which we fixed by using test_tick). If we do want to cover
handling the current time, we should dedicate one script to doing so,
and have it unset GIT_COMMITTER_DATE explicitly.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The early part of t9100 creates an unusual "doubled" history in the
"git-svn" ref. When we get to t9100.17, it looks like this:
$ git log --oneline --graph git-svn
[...]
* efd0303 detect node change from file to directory #2
|\
* | 3e727c0 detect node change from file to directory #2
|/
* 3b00468 try a deep --rmdir with a commit
|\
* | b4832d8 try a deep --rmdir with a commit
|/
* f0d7bd5 import for git svn
Each commit we make with "git commit" is paired with one from "git svn
set-tree", with the latter as a merge of the first and its grandparent.
Later, t9100.17 wants to check that "git svn fetch" gets the same trees.
And it does, but just one copy of each. So it uses rev-list to get the
tree of each commit and pipes it to "uniq" to drop the duplicates. Our
input isn't sorted, but it will find adjacent duplicates. This works
reliably because the order of commits from rev-list always shows the
duplicates next to each other. For any one of those merges, we could
choose to show its duplicate or the grandparent first. But barring
clocks running backwards, the duplicate will always have a time equal to
or greater than the grandparent. Even if equal, we break ties by showing
the first-parent first, so the duplicates remain adjacent.
But this would break if the timestamps stopped moving in chronological
order. Normally we would rely on test_tick for this, but we have _two_
sources of time here:
- "git commit" creates one commit based on GIT_COMMITTER_DATE (which
respects test_tick)
- the "svn set-tree" one is based on subversion, which does not have
an easy way to specify a timestamp
So using test_tick actually breaks the test, because now the duplicates
are far in the past, and we'll show the grandparent before the
duplicate. And likewise, a proposed change to set GIT_COMMITTER_DATE in
all scripts will break it.
We _could_ fix this by sorting before removing duplicates, but
presumably it's a useful part of the test to make sure the trees appear
in the same order in both spots. Likewise, we could use something like:
perl -ne 'print unless $seen{$_}++'
to remove duplicates without impacting the order. But that doesn't work
either, because there are actually multiple (non-duplicate) commits with
the same trees (we change a file mode and then change it back). So we'd
actually have to de-duplicate the combination of subject and tree. Which
then further throws off t9100.18, which compares the tree hashes
exactly; we'd have to strip the result back down.
Since this test _isn't_ buggy, the simplest thing is to just work around
the proposed change by documenting our expectation that git-created
commits are correctly interleaved using the current time.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The description of `git diff` goes through several different invocations
(numbering added by me):
1. git diff [<options>] [--] [<path>...]
2. git diff [<options>] --no-index [--] <path> <path>
3. git diff [<options>] --cached [<commit>] [--] [<path>...]
4. git diff [<options>] <commit> [--] [<path>...]
5. git diff [<options>] <commit> <commit> [--] [<path>...]
6. git diff [<options>] <commit>..<commit> [--] [<path>...]
7. git diff [<options>] <commit> <commit>... <commit> [--] [<path>...]
8. git diff [<options>] <commit>...<commit> [--] [<path>...]
It then goes on to say that "all of the <commit> in the above
description, except in the last two forms that use '..' notations, can
be any <tree>". The "last two" actually refers to 6 and 8. This got out
of sync in commit b7e10b2ca2 ("Documentation: usage for diff combined
commits", 2020-06-12) which added item 7 to the mix.
As a further complication, after b7e10b2ca2 we also have some potential
confusion around "the '..' notation". The "..[.]" in items 6 and 8 are
part of the rev notation, whereas the "..." in item 7 is manpage
language for "one or more".
Move item 6 down, i.e., to between 7 and 8, to restore the ordering.
Because 6 refers to 5 ("synonymous to the previous form") we need to
tweak the language a bit.
An added bonus of this commit is that we're trying to steer users away
from `git diff <commit>..<commit>` and moving it further down probably
doesn't hurt.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit b7e10b2ca2 ("Documentation: usage for diff combined commits",
2020-06-12) modified the synopsis by adding an optional "[<commit>...]"
to
'git diff' [<options>] <commit> <commit> [--] [<path>...]
to effectively add
'git diff' [<options>] <commit> <commit>... <commit> [--] [<path>...]
as another valid invocation. Which makes sense.
Further down, in the description, it left the existing entry for
'git diff' [<options>] <commit> <commit> [--] [<path>...]
intact and added a new entry on
'git diff' [<options>] <commit> [<commit>...] <commit> [--] [<path>...]
where it says that "[t]his form is to view the results of a merge
commit" and details how "the first listed commit must be the merge
itself". But one possible instantiation of this form is `git diff
<commit> <commit>` for which the added text doesn't really apply.
Remove the brackets so that we lose this overlap between the two
descriptions. We can still use the more compact representation in the
synopsis.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Regarding reflog messages:
- We expect that a reflog message consists of a single line. The
file format used by the files backend may add a LF after the
message as a delimiter, and output by commands like "git log -g"
may complete such an incomplete line by adding a LF at the end,
but philosophically, the terminating LF is not a part of the
message.
- We however allow callers of refs API to supply a random sequence
of NUL terminated bytes. We cleanse caller-supplied message by
squashing a run of whitespaces into a SP, and by trimming trailing
whitespace, before storing the message. This is how we tolerate,
instead of erring out, a message with LF in it (be it at the end,
in the middle, or both).
Currently, the cleansing of the reflog message is done by the files
backend, before the log is written out. This is sufficient with the
current code, as that is the only backend that writes reflogs. But
new backends can be added that write reflogs, and we'd want the
resulting log message we would read out of "log -g" the same no
matter what backend is used, and moving the code to do so to the
generic layer is a way to do so.
An added benefit is that the "cleansing" function could be updated
later, independent from individual backends, to e.g. allow
multi-line log messages if we wanted to, and when that happens, it
would help a lot to ensure we covered all bases if the cleansing
function (which would be updated) is called from the generic layer.
Side note: I am not interested in supporting multi-line reflog
messages right at the moment (nobody is asking for it), but I
envision that instead of the "squash a run of whitespaces into a SP
and rtrim" cleansing, we can %urlencode problematic bytes in the
message *AND* append a SP at the end, when a new version of Git that
supports multi-line and/or verbatim reflog messages writes a reflog
record. The reading side can detect the presense of SP at the end
(which should have been rtrimmed out if it were written by existing
versions of Git) as a signal that decoding %urlencode recovers the
original reflog message.
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Both the git-bisect.sh as bisect--helper inspected the file system
directly.
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Adding tests for refs pointing to tree and blob shows that
we care about testing both positive ("see, my shiny new toy
does work") and negative ("and it won't do nonsensical
things when given an input it is not designed to work with")
cases.
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Let's avoid a big dense paragraph by using an unordered
list for the %(contents:XXXX) format specifiers.
While at it let's also make the following improvements:
- Let's not describe %(contents) using "complete message"
as it's not clear what an incomplete message is.
- Let's improve how the "subject" and "body" are
described.
- Let's state that "signature" is only available for
tag objects.
Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Christian Couder <chriscool@tuxfamily.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The test for "no shallow lines after receiving ACK ready" is very
sensitive to the timestamps of the commits we create. It's looking for
the fetch negotiation to send a "ready", which in turn depends on the
order in which we traverse commits during the negotiation.
It works reliably now because the base commit "7" is created without
test_commit, and thus gets a commit time matching the current system
clock. Whereas the new commits created in this test do use test_commit,
and get the usual test_tick time from 2005. So the fetch into the
"clone" repository results in a commit graph like this (I omitted some
of the "unrelated" commits for clarity; they're all just a sequence of
test_ticks):
$ git log --graph --format='%ct %s %d'
* 1112912953 new (origin/master, origin/HEAD)
* 1594322236 7 (grafted, master)
* 1112912893 unrelated15 (origin/unrelated15, unrelated15)
[...]
* 1112912053 unrelated1 (origin/unrelated1, unrelated1)
* 1112911993 new-too (HEAD -> newnew, tag: new-too)
The important things to see are:
- "7" is way in the future compared to the other commits
- "new-too" in the fetching repo is older than "new" (and its
"unrelated" ancestors) in the shallow repo
If we change our "setup shallow clone" step to use test_tick, too (and
get rid of the dependency on the system clock), then the test will fail.
The resulting graph looks like this:
$ git log --graph --format='%ct %s %d'
* 1112913373 new (origin/master, origin/HEAD)
* 1112912353 7 (grafted, master)
* 1112913313 unrelated15 (origin/unrelated15, unrelated15)
[...]
* 1112912473 unrelated1 (origin/unrelated1, unrelated1)
* 1112912413 new-too (HEAD -> newnew, tag: new-too)
Our "new-too" is still older than "new" and "unrelated", but now "7" is
older than all of them (because it advanced test_tick, which the other
tests built on top of). In the original, we advertised "7" as the first
"have" before anything else, but now "new-too" is more recent. You'd see
the same thing in the unlikely event that the system clock was set
before our test_tick default in 2005.
Let's make the timing requirements more explicit. The important thing is
that the client advertise all of its shared commits first, before
presenting its unique "new-too" commit. We can do that and get rid of
the system clock dependency at the same time by creating all of the
shared commits around time X (using test_tick), and then creating
"new-too" with some time long before X. The resulting graph looks like
this:
$ git log --graph --format='%ct %s %d'
* 1500001380 new (origin/master, origin/HEAD)
* 1500000420 7 (grafted, master)
* 1500001320 unrelated15 (origin/unrelated15, unrelated15)
[...]
* 1500000480 unrelated1 (origin/unrelated1, unrelated1)
* 1400000060 new-too (HEAD -> newnew, tag: new-too)
That also lets us get rid of the hacky test_tick added by f0e802ca20
(t5539: update a flaky test, 2014-07-14). That was clearly dancing
around the same problem, but only addressed the relationship between
commits created in the two subshells (which did use test_tick, but
overlapped because increments of test_tick in subshells are lost). Now
that we're using consistent and well-placed times for both lines of
history, we don't have to care about a one-tick difference between the
two sides.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A few of the perl tests in t9700 ask for the author and committer ident,
and then make sure we get something sensible. For the timestamp portion,
we just match [0-9]+, because the actual value will depend on when the
test is run. However, we do require that the timezone be "+0000". This
works reliably because we set $TZ in test-lib.sh. But in preparation for
changing the default timezone, let's be a bit more flexible. We don't
actually care about the exact value here, just that we were able to get
a sensible output from the perl module's access methods.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When using git clone with --separate-git-dir realgitdir and
realgitdir already exists, it's content is destroyed.
So, make sure we don't clone into an existing non-empty directory.
When d45420c1 (clone: do not clean up directories we didn't create,
2018-01-02) tightened the clean-up procedure after a failed cloning
into an empty directory, it assumed that the existing directory
given is an empty one so it is OK to keep that directory, while
running the clean-up procedure that is designed to remove everything
in it (since there won't be any, anyway). Check and make sure that
the $GIT_DIR is empty even cloning into an existing repository.
Signed-off-by: Ben Wijen <ben@wijen.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git checkout" failed to catch an error from fstat() after updating
a path in the working tree.
* mt/entry-fstat-fallback-fix:
entry: check for fstat() errors after checkout
"fetch.writeCommitGraph" was enabled when "feature.experimental" is
asked for, but it was found to be a bit too risky even for bold
folks in its current shape. The configuration has been ejected, at
least for now, from the "experimental" feature set.
* jn/eject-fetch-write-commit-graph-out-of-experimental:
experimental: default to fetch.writeCommitGraph=false
When "fetch.writeCommitGraph" configuration is set in a shallow
repository and a fetch moves the shallow boundary, we wrote out
broken commit-graph files that do not match the reality, which has
been corrected.
* tb/fix-persistent-shallow:
commit.c: don't persist substituted parents when unshallowing
Recent update to "git diff" meant as a code clean-up introduced a
bug in its error handling code, which has been corrected.
* ct/diff-with-merge-base-clarification:
diff: check for merge bases before assigning sym->base
"git log -Lx,y:path --before=date" lost track of where the range
should be because it didn't take the changes made by the youngest
commits that are omitted from the output into account.
* rs/line-log-until:
revision: disable min_age optimization with line-log
"git send-email --in-reply-to=<msg>" did not use the In-Reply-To:
header with the value given from the command line, and let it be
overridden by the value on In-Reply-To: header in the messages
being sent out (if exists).
* ra/send-email-in-reply-to-from-command-line-wins:
send-email: restore --in-reply-to superseding behavior
The command line completion support (in contrib/) used to be
prepared to work with "set -u" but recent changes got a bit more
sloppy. This has been corrected.
* vs/completion-with-set-u:
completion: nounset mode fixes
We don't give a "::" for the list separator, but just a single ":". This
ends up rendering literally, "--apply: Use applying strategies ...". As
a follow-on error, the list continuation, "+", also ends up rendering
literally (because we don't have a list).
This was introduced in 52eb738d6b ("rebase: add an --am option",
2020-02-15) and survived the rename in 10cdb9f38a ("rebase: rename the
two primary rebase backends", 2020-02-15).
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
76ffbca71a (commit-graph: write Bloom filters to commit graph file,
2020-04-06) added two delayed progress lines to writing the Bloom
filter index and data chunk. This is wrong, because a single common
progress is used while writing all chunks, which is not updated while
writing these two new chunks, resulting in incomplete-looking "done"
lines:
Expanding reachable commits in commit graph: 888679, done.
Computing commit changed paths Bloom filters: 100% (888678/888678), done.
Writing out commit graph in 6 passes: 66% (3554712/5332068), done.
Use the common 'struct progress' instance while writing the Bloom
filter chunks as well.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To display a progress line while iterating over all refs,
d335ce8f24 (commit-graph.c: show progress of finding reachable
commits, 2020-05-13) should have added a pair of
start_delayed_progress() and stop_progress() calls around a
for_each_ref() invocation. Alas, the stop_progress() call ended up at
the wrong place, after write_commit_graph(), which does all the
commit-graph computation and writing, and has several progress lines
of its own. Consequently, that new
Collecting referenced commits: 123
progress line is overwritten by the first progress line shown by
write_commit_graph(), and its final "done" line is shown last, after
everything is finished:
Expanding reachable commits in commit graph: 344786, done.
Computing commit changed paths Bloom filters: 100% (344786/344786), done.
Collecting referenced commits: 154, done.
Move that stop_progress() call to the right place.
While at it, drop the unnecessary 'if (data.progress)' condition
protecting the stop_progress() call, because that function is prepared
to handle a NULL progress struct.
Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In 11179eb311 ("entry.c: check if file exists after checkout",
2017-10-05) we started checking the result of the lstat() call done
after writing a file, to avoid writing garbage to the corresponding
cache entry. However, the code skips calling lstat() if it's possible
to use fstat() when it still has the file descriptor open. And when
calling fstat() we don't do the same error checking. To fix that, let
the callers of fstat_output() know when fstat() fails. In this case,
write_entry() will try to use lstat() and properly report an error if
that fails as well.
Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The `git log` synopsis mentions `<revision range>`, and the description
of this option links to gitrevisions(7), but a nice explanation of
how a revision range can be constructed from individual commits,
optionnally prefixed with `^`, also exists in `rev-list-description.txt`.
Include this description in the man page for `git log`.
Add Asciidoc 'ifdef's to `rev-list-description.txt` so that either `git
rev-list` or `git log` appears in the respective man pages.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A following commit will reuse the description of the `git rev-list`
command in the `git log` manpage.
Move this description to a separate file.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Tweak a sentence to make it a little more readable.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Using '{caret}' inside double quotes and immediately following with a
single quoted word does not create the desired output: '<commit1>'
appears verbatim instead of being emphasized.
Use a litteral caret ('^') instead.
Also, remove the leading tabs in shell examples to bring them more in
line with the rest of the documentation.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The "Specifying ranges" section does not mention explicitly that
several commits can be specified to form a range.
Add a mention to that effect.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add links to the documentation for `git rev-list` and `git diff`
instead of simply mentioning them, to make it easier for readers to reach
these documentation pages. Let's link to `git diff` as this is the
porcelain command, and the rest of the family (`diff-index`, `diff-tree` and
`diff-files`) are mentioned in the "Raw output format" section of the
`git diff` documentation.
Signed-off-by: Philippe Blain <levraiphilippeblain@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The fetch.writeCommitGraph feature makes fetches write out a commit
graph file for the newly downloaded pack on fetch. This improves the
performance of various commands that would perform a revision walk and
eventually ought to be the default for everyone. To prepare for that
future, it's enabled by default for users that set
feature.experimental=true to experience such future defaults.
Alas, for --unshallow fetches from a shallow clone it runs into a
snag: by the time Git has fetched the new objects and is writing a
commit graph, it has performed a revision walk and r->parsed_objects
contains information about the shallow boundary from *before* the
fetch. The commit graph writing code is careful to avoid writing a
commit graph file in shallow repositories, but the new state is not
shallow, and the result is that from that point on, commands like "git
log" make use of a newly written commit graph file representing a
fictional history with the old shallow boundary.
We could fix this by making the commit graph writing code more careful
to avoid writing a commit graph that could have used any grafts or
shallow state, but it is possible that there are other pieces of
mutated state that fetch's commit graph writing code may be relying
on. So disable it in the feature.experimental configuration.
Google developers have been running in this configuration (by setting
fetch.writeCommitGraph=false in the system config) to work around this
bug since it was discovered in April. Once the fix lands, we'll
enable fetch.writeCommitGraph=true again to give it some early testing
before rolling out to a wider audience.
In other words:
- this patch only affects behavior with feature.experimental=true
- it makes feature.experimental match the configuration Google has
been using for the last few months, meaning it would leave users in
a better tested state than without it
- this should improve testing for other features guarded by
feature.experimental, by making feature.experimental safer to use
Reported-by: Jay Conrod <jayconrod@google.com>
Helped-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since 37b9dcabfc (shallow.c: use '{commit,rollback}_shallow_file',
2020-04-22), Git knows how to reset stat-validity checks for the
$GIT_DIR/shallow file, allowing it to change between a shallow and
non-shallow state in the same process (e.g., in the case of 'git fetch
--unshallow').
However, when $GIT_DIR/shallow changes, Git does not alter or remove any
grafts (nor substituted parents) in memory.
This comes up in a "git fetch --unshallow" with fetch.writeCommitGraph
set to true. Ordinarily in a shallow repository (and before 37b9dcabfc,
even in this case), commit_graph_compatible() would return false,
indicating that the repository should not be used to write a
commit-graphs (since commit-graph files cannot represent a shallow
history). But since 37b9dcabfc, in an --unshallow operation that check
succeeds.
Thus even though the repository isn't shallow any longer (that is, we
have all of the objects), the in-core representation of those objects
still has munged parents at the shallow boundaries. When the
commit-graph write proceeds, we use the incorrect parentage, producing
wrong results.
There are two ways for a user to work around this: either (1) set
'fetch.writeCommitGraph' to 'false', or (2) drop the commit-graph after
unshallowing.
One way to fix this would be to reset the parsed object pool entirely
(flushing the cache and thus preventing subsequent reads from modifying
their parents) after unshallowing. That would produce a problem when
callers have a now-stale reference to the old pool, and so this patch
implements a different approach. Instead, attach a new bit to the pool,
'substituted_parent', which indicates if the repository *ever* stored a
commit which had its parents modified (i.e., the shallow boundary
prior to unshallowing).
This bit needs to be sticky because all reads subsequent to modifying a
commit's parents are unreliable when unshallowing. Modify the check in
'commit_graph_compatible' to take this bit into account, and correctly
avoid generating commit-graphs in this case, thus solving the bug.
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: Jonathan Nieder <jrnieder@gmail.com>
Reported-by: Jay Conrod <jayconrod@google.com>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In symdiff_prepare(), we iterate over the set of parsed objects to pick
out any symmetric differences, including the left, right, and base
elements. We assign the results into pointers in a "struct symdiff", and
then complain if we didn't find a base, like so:
sym->left = rev->pending.objects[lpos].name;
sym->right = rev->pending.objects[rpos].name;
sym->base = rev->pending.objects[basepos].name;
if (basecount == 0)
die(_("%s...%s: no merge base"), sym->left, sym->right);
But the least lines are backwards. If basecount is 0, then basepos will
be -1, and we will access memory outside of the pending array. This
isn't usually that big a deal, since we don't do anything besides a
single pointer-sized read before exiting anyway, but it does violate the
C standard, and of course memory-checking tools like ASan complain.
Let's put the basecount check first. Note that we haveto split it from
the other assignments, since the die() relies on sym->left and
sym->right having been assigned (this isn't strictly necessary, but is
easier to read than dereferencing the pending array again).
Reported-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The first two commits created in t6000 are done without test_tick,
meaning they use the current system clock. After that, we create one
with test_tick, which means it uses a deterministic time in the past.
The result of the "symleft flag bit is propagated down from tag" test
relies on the output order of commits from git-log, which in turn
depends on these timestamps. So this test is technically dependent on
the system clock time, though in practice it would only matter if your
system clock was set before test_tick's default time (which is in 2005).
However, let's use test_tick consistently for those early commits (and
update the expected output to match). This makes the test deterministic,
which is in turn easier to reason about and debug.
Note that there's also a fourth commit here, and it does not use
test_tick. It does have a deterministic timestamp because of the prior
use of test_tick in the script, but it will always be the same time as
the third commit. Let's use test_tick here, too, for consistency. The
matching timestamps between the third and fourth commit are not an
important part of the test.
We could also use test_commit in all of these cases, as it runs
test_tick under the hood. But it would be awkward to do so, as these
tests diverge from the usual test_commit patterns (e.g., by creating
multiple files in a single commit).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
