If the deletion steps unexpectedly fail, it is less bad to leave a
reference without its reflog than it is to leave a reflog without its
reference, since the latter is an invalid repository state.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now the outside world interacts with the packed ref store only via the
generic refs API plus a few lock-related functions. This allows us to
delete some functions that are no longer used, thereby completing the
encapsulation of the packed ref store.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When processing a `files_ref_store` transaction, it is sometimes
necessary to delete some references from the "packed-refs" file. Do
that using a reference transaction conducted against the
`packed_ref_store`.
This change further decouples `files_ref_store` from
`packed_ref_store`. It also fixes multiple problems, including the two
revealed by test cases added in the previous commit.
First, the old code didn't obtain the `packed-refs` lock until
`files_transaction_finish()`. This means that a failure to acquire the
`packed-refs` lock (e.g., due to contention with another process)
wasn't detected until it was too late (problems like this are supposed
to be detected in the "prepare" phase). The new code acquires the
`packed-refs` lock in `files_transaction_prepare()`, the same stage of
the processing when the loose reference locks are being acquired,
removing another reason why the "prepare" phase might succeed and the
"finish" phase might nevertheless fail.
Second, the old code deleted the loose version of a reference before
deleting any packed version of the same reference. This left a moment
when another process might think that the packed version of the
reference is current, which is incorrect. (Even worse, the packed
version of the reference can be arbitrarily old, and might even point
at an object that has since been garbage-collected.)
Third, if a reference deletion fails to acquire the `packed-refs` lock
altogether, then the old code might leave the repository in the
incorrect state (possibly corrupt) described in the previous
paragraph.
Now we activate the new "packed-refs" file (sans any references that
are being deleted) *before* deleting the corresponding loose
references. But we hold the "packed-refs" lock until after the loose
references have been finalized, thus preventing a simultaneous
"pack-refs" process from packing the loose version of the reference in
the time gap, which would otherwise defeat our attempt to delete it.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently, a loose reference is deleted even before locking the
`packed-refs` file, let alone deleting any packed version of the
reference. This leads to two problems, demonstrated by two new tests:
* While a reference is being deleted, other processes might see the
old, packed value of the reference for a moment before the packed
version is deleted. Normally this would be hard to observe, but we
can prolong the window by locking the `packed-refs` file externally
before running `update-ref`, then unlocking it before `update-ref`'s
attempt to acquire the lock times out.
* If the `packed-refs` file is locked so long that `update-ref` fails
to lock it, then the reference can be left permanently in the
incorrect state described in the previous point.
In a moment, both problems will be fixed.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use a `packed_ref_store` transaction in the implementation of
`files_initial_transaction_commit()` rather than using internal
features of the packed ref store. This further decouples
`files_ref_store` from `packed_ref_store`.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
At least since v1.7, the elements of the `refs_to_prune` linked list
have been leaked. Fix the leak by teaching `prune_refs()` to free the
list elements as it processes them.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that the packed reference store supports transactions, we can use
a transaction to write the packed versions of references that we want
to pack. This decreases the coupling between `files_ref_store` and
`packed_ref_store`.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Implement `packed_delete_refs()` using a reference transaction. This
means that `files_delete_refs()` can use `refs_delete_refs()` instead
of `repack_without_refs()` to delete any packed references, decreasing
the coupling between the classes.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Implement the methods needed to support reference transactions for
the packed-refs backend. The new methods are not yet used.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
`packed_ref_store` is going to want to store some transaction-wide
data, so make a place for it.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The old code incremented the packed ref cache reference count when
acquiring the packed-refs lock, and decremented the count when
releasing the lock. This is unnecessary because:
* Another process cannot change the packed-refs file because it is
locked.
* When we ourselves change the packed-refs file, we do so by first
modifying the packed ref-cache, and then writing the data from the
ref-cache to disk. So the packed ref-cache remains fresh because any
changes that we plan to make to the file are made in the cache first
anyway.
So there is no reason for the cache to become stale.
Moreover, the extra reference count causes a problem if we
intentionally clear the packed refs cache, as we sometimes need to do
if we change the cache in anticipation of writing a change to disk,
but then the write to disk fails. In that case, `packed_refs_unlock()`
would have no easy way to find the cache whose reference count it
needs to decrement.
This whole issue will soon become moot due to upcoming changes that
avoid changing the in-memory cache as part of updating the packed-refs
on disk, but this change makes that transition easier.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This value, which is stashed in the last byte of an object_id hash,
gets handed around a lot. So add a sanity check before using it in
`load_subtree()`.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The original git-shortlog command parsed the output of
git-log, and the logic went something like this:
1. Read stdin looking for "author" lines.
2. Parse the identity into its name/email bits.
3. Apply mailmap to the name/email.
4. Reformat the identity into a single buffer that is our
"key" for grouping entries (either a name by default,
or "name <email>" if --email was given).
The first part happens in read_from_stdin(), and the other
three steps are part of insert_one_record().
When we do an internal traversal, we just swap out the stdin
read in step 1 for reading the commit objects ourselves.
Prior to 2db6b83d18 (shortlog: replace hand-parsing of
author with pretty-printer, 2016-01-18), that made sense; we
still had to parse the ident in the commit message.
But after that commit, we use pretty.c's "%an <%ae>" to get
the author ident (for simplicity). Which means that the
pretty printer is doing a parse/format under the hood, and
then we parse the result, apply the mailmap, and format the
result again.
Instead, we can just ask pretty.c to do all of those steps
for us (including the mailmap via "%aN <%aE>", and not
formatting the address when --email is missing).
And then we can push steps 2-4 into read_from_stdin(). This
speeds up "git shortlog -ns" on linux.git by about 3%, and
eliminates a leak in insert_one_record() of the namemailbuf
strbuf.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It's a common pattern in git commands to allocate some
memory that should last for the lifetime of the program and
then not bother to free it, relying on the OS to throw it
away.
This keeps the code simple, and it's fast (we don't waste
time traversing structures or calling free at the end of the
program). But it also triggers warnings from memory-leak
checkers like valgrind or LSAN. They know that the memory
was still allocated at program exit, but they don't know
_when_ the leaked memory stopped being useful. If it was
early in the program, then it's probably a real and
important leak. But if it was used right up until program
exit, it's not an interesting leak and we'd like to suppress
it so that we can see the real leaks.
This patch introduces an UNLEAK() macro that lets us do so.
To understand its design, let's first look at some of the
alternatives.
Unfortunately the suppression systems offered by
leak-checking tools don't quite do what we want. A
leak-checker basically knows two things:
1. Which blocks were allocated via malloc, and the
callstack during the allocation.
2. Which blocks were left un-freed at the end of the
program (and which are unreachable, but more on that
later).
Their suppressions work by mentioning the function or
callstack of a particular allocation, and marking it as OK
to leak. So imagine you have code like this:
int cmd_foo(...)
{
/* this allocates some memory */
char *p = some_function();
printf("%s", p);
return 0;
}
You can say "ignore allocations from some_function(),
they're not leaks". But that's not right. That function may
be called elsewhere, too, and we would potentially want to
know about those leaks.
So you can say "ignore the callstack when main calls
some_function". That works, but your annotations are
brittle. In this case it's only two functions, but you can
imagine that the actual allocation is much deeper. If any of
the intermediate code changes, you have to update the
suppression.
What we _really_ want to say is that "the value assigned to
p at the end of the function is not a real leak". But
leak-checkers can't understand that; they don't know about
"p" in the first place.
However, we can do something a little bit tricky if we make
some assumptions about how leak-checkers work. They
generally don't just report all un-freed blocks. That would
report even globals which are still accessible when the
leak-check is run. Instead they take some set of memory
(like BSS) as a root and mark it as "reachable". Then they
scan the reachable blocks for anything that looks like a
pointer to a malloc'd block, and consider that block
reachable. And then they scan those blocks, and so on,
transitively marking anything reachable from a global as
"not leaked" (or at least leaked in a different category).
So we can mark the value of "p" as reachable by putting it
into a variable with program lifetime. One way to do that is
to just mark "p" as static. But that actually affects the
run-time behavior if the function is called twice (you
aren't likely to call main() twice, but some of our cmd_*()
functions are called from other commands).
Instead, we can trick the leak-checker by putting the value
into _any_ reachable bytes. This patch keeps a global
linked-list of bytes copied from "unleaked" variables. That
list is reachable even at program exit, which confers
recursive reachability on whatever values we unleak.
In other words, you can do:
int cmd_foo(...)
{
char *p = some_function();
printf("%s", p);
UNLEAK(p);
return 0;
}
to annotate "p" and suppress the leak report.
But wait, couldn't we just say "free(p)"? In this toy
example, yes. But UNLEAK()'s byte-copying strategy has
several advantages over actually freeing the memory:
1. It's recursive across structures. In many cases our "p"
is not just a pointer, but a complex struct whose
fields may have been allocated by a sub-function. And
in some cases (e.g., dir_struct) we don't even have a
function which knows how to free all of the struct
members.
By marking the struct itself as reachable, that confers
reachability on any pointers it contains (including those
found in embedded structs, or reachable by walking
heap blocks recursively.
2. It works on cases where we're not sure if the value is
allocated or not. For example:
char *p = argc > 1 ? argv[1] : some_function();
It's safe to use UNLEAK(p) here, because it's not
freeing any memory. In the case that we're pointing to
argv here, the reachability checker will just ignore
our bytes.
3. Likewise, it works even if the variable has _already_
been freed. We're just copying the pointer bytes. If
the block has been freed, the leak-checker will skip
over those bytes as uninteresting.
4. Because it's not actually freeing memory, you can
UNLEAK() before we are finished accessing the variable.
This is helpful in cases like this:
char *p = some_function();
return another_function(p);
Writing this with free() requires:
int ret;
char *p = some_function();
ret = another_function(p);
free(p);
return ret;
But with unleak we can just write:
char *p = some_function();
UNLEAK(p);
return another_function(p);
This patch adds the UNLEAK() macro and enables it
automatically when Git is compiled with SANITIZE=leak. In
normal builds it's a noop, so we pay no runtime cost.
It also adds some UNLEAK() annotations to show off how the
feature works. On top of other recent leak fixes, these are
enough to get t0000 and t0001 to pass when compiled with
LSAN.
Note the case in commit.c which actually converts a
strbuf_release() into an UNLEAK. This code was already
non-leaky, but the free didn't do anything useful, since
we're exiting. Converting it to an annotation means that
non-leak-checking builds pay no runtime cost. The cost is
minimal enough that it's probably not worth going on a
crusade to convert these kinds of frees to UNLEAKS. I did it
here for consistency with the "sb" leak (though it would
have been equally correct to go the other way, and turn them
both into strbuf_release() calls).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Depending on the implementation of walks, limitted stack size may lead
to problems (for recursion).
Test name-rev and describe with deep repos and limitted stack size and
mark the former with known failure.
We add these tests (which add gazillions of commits) last so as to keep
the runtime of other subtests the same.
Signed-off-by: Michael J Gruber <git@grubix.eu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
t6120 breaks the repo state intentionally in the last tests.
Clean up the breakage afterwards (and before adding more tests).
Signed-off-by: Michael J Gruber <git@grubix.eu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
name-rev is used in a few tests, but tested only in t6120 along with
describe so far.
Add tests for name-rev with --all and --stdin.
Signed-off-by: Michael J Gruber <git@grubix.eu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The lazy prerequisite ULIMIT_STACK_SIZE is used only in t7004 so far.
Move it to test-lib.sh so that it can be used in other tests (which it will
be in a follow-up commit).
Signed-off-by: Michael J Gruber <git@grubix.eu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code was using two string_lists, one for the directories and
one for the files. The code never checks the lists independently
so we should be able to only use one list. The string_list also
is a O(log n) for lookup and insertion. Switching this to use a
hashmap will give O(1) which will save some time when there are
millions of paths that will be checked.
Signed-off-by: Kevin Willford <kewillf@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Similar to 65969d43d1 (merge: honor prepare-commit-msg hook, 2011-02-14)
merge should also honor the commit-msg hook: When a merge is stopped due
to conflicts or --no-commit, the subsequent commit calls the commit-msg
hook. However, it is not called after a clean merge. Fix this
inconsistency by invoking the hook after clean merges as well.
This change is motivated by Gerrit's commit-msg hook to install a ChangeId
trailer into the commit message. Without such a ChangeId, Gerrit refuses
to accept any commit by default, such that the inconsistency of (not)
running the commit-msg hook between commit and merge leads to confusion
and might block people from getting their work done.
As the githooks man page is very vocal about the possibility of skipping
the commit-msg hook via the --no-verify option, implement the option
in merge, too.
'git merge --continue' is currently implemented as calling cmd_commit
with no further arguments. This works for most other merge related options,
such as demonstrated via the --allow-unrelated-histories flag in the
test. The --no-verify option however is not remembered across invocations
of git-merge. Originally the author assumed an alternative in which the
'git merge --continue' command accepts the --no-verify flag, but that
opens up the discussion which flags are allows to the continued merge
command and which must be given in the first invocation.
Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
ce012deb98 ("read-cache: avoid allocating every ondisk entry when
writing", 2017-08-21) changed the way cache entries are written to the
index file. While previously it wrote the name to an struct that was
allocated using xcalloc(), it now uses ce_write() directly. Previously
ce_namelen - common bytes were written to the cache entry, which would
automatically make it nul terminated, as it was allocated using calloc.
Now we are writing ce_namelen - common + 1 bytes directly from the
ce->name to the index. If CE_STRIP_NAME however gets set in the split
index case ce->ce_namelen is set to 0 without changing the actual
ce->name buffer. When index-v4, this results in the first character of
ce->name being written out instead of just a terminating nul charcter.
As index-v4 requires the terminating nul character as terminator of
the name when reading it back, this results in a corrupted index.
Fix that by only writing ce_namelen - common bytes directly from
ce->name to the index, and adding the nul terminator in an extra call to
ce_write.
This bug was turned up by setting TEST_GIT_INDEX_VERSION = 4 in
config.mak and running the test suite (t1700 specifically broke).
Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git pull" supports a --recurse-submodules option but does not parse the
submodule.recurse configuration item to set the default for that option.
Meanwhile "git fetch" does support submodule.recurse, producing
confusing behavior: when submodule.recurse is enabled, "git pull"
recursively fetches submodules but does not update them after fetch.
Handle submodule.recurse in "git pull" to fix this.
Reported-by: Magnus Homann <magnus@homann.se>
Signed-off-by: Nicolas Morey-Chaisemartin <nicolas@morey-chaisemartin.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
pull parses first the cli options and then the config option.
The expected behavior is the other way around, so that config
options can not override the cli ones.
This patch changes the parsing order so config options are
parsed first.
Signed-off-by: Nicolas Morey-Chaisemartin <nicolas@morey-chaisemartin.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The function was deprecated in commit 89576613 ("treewide: deprecate
git_config_maybe_bool, use git_parse_maybe_bool", 2017-08-07) and has no
users.
Signed-off-by: Martin Ågren <martin.agren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is to address concerns raised by ThreadSanitizer on the mailing list
about threaded unprotected R/W access to map.size with my previous "disallow
rehash" change (0607e10009).
See:
https://public-inbox.org/git/adb37b70139fd1e2bac18bfd22c8b96683ae18eb.1502780344.git.martin.agren@gmail.com/
Add API to hashmap to disable item counting and thus automatic rehashing.
Also include API to later re-enable them.
When item counting is disabled, the map.size field is invalid. So to
prevent accidents, the field has been renamed and an accessor function
hashmap_get_size() has been added. All direct references to this
field have been been updated. And the name of the field changed
to map.private_size to communicate this.
Here is the relevant output from ThreadSanitizer showing the problem:
WARNING: ThreadSanitizer: data race (pid=10554)
Read of size 4 at 0x00000082d488 by thread T2 (mutexes: write M16):
#0 hashmap_add hashmap.c:209
#1 hash_dir_entry_with_parent_and_prefix name-hash.c:302
#2 handle_range_dir name-hash.c:347
#3 handle_range_1 name-hash.c:415
#4 lazy_dir_thread_proc name-hash.c:471
#5 <null> <null>
Previous write of size 4 at 0x00000082d488 by thread T1 (mutexes: write M31):
#0 hashmap_add hashmap.c:209
#1 hash_dir_entry_with_parent_and_prefix name-hash.c:302
#2 handle_range_dir name-hash.c:347
#3 handle_range_1 name-hash.c:415
#4 handle_range_dir name-hash.c:380
#5 handle_range_1 name-hash.c:415
#6 lazy_dir_thread_proc name-hash.c:471
#7 <null> <null>
Martin gives instructions for running TSan on test t3008 in this post:
https://public-inbox.org/git/CAN0heSoJDL9pWELD6ciLTmWf-a=oyxe4EXXOmCKvsG5MSuzxsA@mail.gmail.com/
Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When the RUNTIME_PREFIX compile-time knob isn't set, we
never look at the argv0_path we extract. We can push its
declaration inside the #ifdef to make it more clear that the
extract code is effectively a noop.
This also un-confuses leak-checking of the argv0_path
variable when RUNTIME_PREFIX isn't set. The compiler is free
to drop this static variable that we set but never look at
(and "gcc -O2" does so). But the compiler still must call
strbuf_detach(), since it doesn't know whether that function
has side effects; it just throws away the result rather than
putting it into the global.
Leak-checkers which work by scanning the data segment for
pointers to heap blocks would normally consider the block
as reachable at program end. But if the compiler removes the
variable entirely, there's nothing to find.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The system_path() function has an #ifdef in the middle of
it. Let's move the conditional logic into a sub-function.
This isolates it more, which will make it easier to change
and add to.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This new binary was introduced in commit 3921a0b ("perf: add test for
writing the index", 2017-08-21), but a .gitignore entry was not added
for it. Add that entry.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When setting the `eol` attribute, paths can change their dirty status
without any change in the working directory. This can cause confusion
and should at least be mentioned with a remedy.
Signed-off-by: Ben Boeckel <mathstuf@gmail.com>
Reviewed-by: Torsten Bögershausen <tboegi@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
strbuf_readlink() already frees the buffer for us on error. Clean up
if write_sha1_file() fails as well instead of returning early.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Clean up at the end and jump there instead of returning early.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
