Multi-ref transaction support we merged a few releases ago
unnecessarily kept many file descriptors open, risking to fail with
resource exhaustion. This is for 2.4.x track.
* mh/write-refs-sooner-2.4:
ref_transaction_commit(): fix atomicity and avoid fd exhaustion
ref_transaction_commit(): remove the local flags variable
ref_transaction_commit(): inline call to write_ref_sha1()
rename_ref(): inline calls to write_ref_sha1() from this function
commit_ref_update(): new function, extracted from write_ref_sha1()
write_ref_to_lockfile(): new function, extracted from write_ref_sha1()
t7004: rename ULIMIT test prerequisite to ULIMIT_STACK_SIZE
update-ref: test handling large transactions properly
ref_transaction_commit(): fix atomicity and avoid fd exhaustion
ref_transaction_commit(): remove the local flags variable
ref_transaction_commit(): inline call to write_ref_sha1()
rename_ref(): inline calls to write_ref_sha1() from this function
commit_ref_update(): new function, extracted from write_ref_sha1()
write_ref_to_lockfile(): new function, extracted from write_ref_sha1()
t7004: rename ULIMIT test prerequisite to ULIMIT_STACK_SIZE
update-ref: test handling large transactions properly
The old code was roughly
for update in updates:
acquire locks and check old_sha
for update in updates:
if changing value:
write_ref_to_lockfile()
commit_ref_update()
for update in updates:
if deleting value:
unlink()
rewrite packed-refs file
for update in updates:
if reference still locked:
unlock_ref()
This has two problems.
Non-atomic updates
==================
The atomicity of the reference transaction depends on all pre-checks
being done in the first loop, before any changes have started being
committed in the second loop. The problem is that
write_ref_to_lockfile() (previously part of write_ref_sha1()), which
is called from the second loop, contains two more checks:
* It verifies that new_sha1 is a valid object
* If the reference being updated is a branch, it verifies that
new_sha1 points at a commit object (as opposed to a tag, tree, or
blob).
If either of these checks fails, the "transaction" is aborted during
the second loop. But this might happen after some reference updates
have already been permanently committed. In other words, the
all-or-nothing promise of "git update-ref --stdin" could be violated.
So these checks have to be moved to the first loop.
File descriptor exhaustion
==========================
The old code locked all of the references in the first loop, leaving
all of the lockfiles open until later loops. Since we might be
updating a lot of references, this could result in file descriptor
exhaustion.
The solution
============
After this patch, the code looks like
for update in updates:
acquire locks and check old_sha
if changing value:
write_ref_to_lockfile()
else:
close_ref()
for update in updates:
if changing value:
commit_ref_update()
for update in updates:
if deleting value:
unlink()
rewrite packed-refs file
for update in updates:
if reference still locked:
unlock_ref()
This fixes both problems:
1. The pre-checks in write_ref_to_lockfile() are now done in the first
loop, before any changes have been committed. If any of the checks
fails, the whole transaction can now be rolled back correctly.
2. All lockfiles are closed in the first loop immediately after they
are created (either by write_ref_to_lockfile() or by close_ref()).
This means that there is never more than one open lockfile at a
time, preventing file descriptor exhaustion.
To simplify the bookkeeping across loops, add a new REF_NEEDS_COMMIT
bit to update->flags, which keeps track of whether the corresponding
lockfile needs to be committed, as opposed to just unlocked. (Since
"struct ref_update" is internal to the refs module, this change is not
visible to external callers.)
This change fixes two tests in t1400.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead, work directly with update->flags. This has the advantage that
the REF_DELETING bit, set in the first loop, can be read in the second
loop instead of having to be recomputed. Plus, it was potentially
confusing having both update->flags and flags, which sometimes had
different values.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
That was the last caller, so delete function write_ref_sha1().
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most of what it does is unneeded from these call sites.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is the first step towards separating the checking and writing of
the new reference value to committing the change.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Now that lock_ref_sha1_basic() gives us back its error messages via a
strbuf, incorporate its error message into our error message rather
than emitting two separate error messages.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
While we are in the area, let's remove a superfluous definite article
from the error message that is emitted when the reference cannot be
locked. This improves how it reads and makes it a bit shorter.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Now that lock_ref_sha1_basic() gives us back its error messages via a
strbuf, incorporate its error message into our error message rather
than emitting one error messages to stderr immediately and returning a
second to our caller.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Now that lock_ref_sha1_basic() gives us back its error messages via a
strbuf, incorporate its error message into our error message rather
than emitting two separate error messages.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
If there is a failure to lock a reference that is likely caused by a
D/F conflict (e.g., trying to lock "refs/foo/bar" when reference
"refs/foo" already exists), invoke verify_refname_available() to try
to generate a more helpful error message.
That function might not detect an error. For example, some
non-reference file might be blocking the deletion of an
otherwise-empty directory tree, or there might be a race with another
process that just deleted the offending reference. In such cases,
generate the strerror-based error message like before.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
For now, change the callers to spew the error to stderr like before.
But soon we will change them to incorporate the reason for the failure
into their own error messages.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
It shouldn't be spewing errors directly to stderr.
For now, change its callers to spew the errors to stderr.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Rename is_refname_available() to verify_refname_available() and change
its return value from 1 for success to 0 for success, to be consistent
with our error-handling convention. In a moment it will also get a
"struct strbuf *err" parameter.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
If two references that D/F conflict (e.g., "refs/foo" and
"refs/foo/bar") are created in a single transaction, the old code
discovered the problem only after the "commit" phase of
ref_transaction_commit() had already begun. This could leave some
references updated and others not, which violates the promise of
atomicity.
Instead, check for such conflicts during the "locking" phase:
* Teach is_refname_available() to take an "extras" parameter that can
contain extra reference names with which the specified refname must
not conflict.
* Change lock_ref_sha1_basic() to take an "extras" parameter, which it
passes through to is_refname_available().
* Change ref_transaction_commit() to pass "affected_refnames" to
lock_ref_sha1_basic() as its "extras" argument.
This change fixes a test case in t1404.
This code is a bit stricter than it needs to be. We could conceivably
allow reference "refs/foo/bar" to be created in the same transaction
as "refs/foo" is deleted (or vice versa). But that would be
complicated to implement, because it is not possible to lock
"refs/foo/bar" while "refs/foo" exists as a loose reference, but on
the other hand we don't want to delete some references before adding
others (because that could leave a gap during which required objects
are unreachable). There is also a complication that reflog files'
paths can conflict.
Any less-strict implementation would probably require tricks like the
packing of all references before the start of the real transaction, or
the use of temporary intermediate reference names.
So for now let's accept too-strict checks. Some reference update
transactions will be rejected unnecessarily, but they will be rejected
in their entirety rather than leaving the repository in an
intermediate state, as would happen now.
Please note that there is still one kind of D/F conflict that is *not*
handled correctly. If two processes are running at the same time, and
one tries to create "refs/foo" at the same time that the other tries
to create "refs/foo/bar", then they can race with each other. Both
processes can obtain their respective locks ("refs/foo.lock" and
"refs/foo/bar.lock"), proceed to the "commit" phase of
ref_transaction_commit(), and then the slower process will discover
that it cannot rename its lockfile into place (after possibly having
committed changes to other references). There appears to be no way to
fix this race without changing the locking policy, which in turn would
require a change to *all* Git clients.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Detect duplicates by storing the reference names in a string_list and
sorting that, instead of sorting the ref_updates directly.
* In a moment the string_list will be used for another purpose, too.
* This removes the need for the custom comparison function
ref_update_compare().
* This means that we can carry out the updates in the order that the
user specified them instead of reordering them. This might be handy
someday if, we want to permit multiple updates to a single reference
as long as they are compatible with each other.
Note: we can't use string_list_remove_duplicates() to check for
duplicates, because we need to know the name of the reference that
appeared multiple times, to be used in the error message.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
In the first loop (over prefixes of refname), use dirname to keep
track of the current prefix. This is not an improvement in itself, but
in a moment we will start using dirname for a role where a
NUL-terminated string is needed.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Now that we don't need a ref_entry to pass to
report_refname_conflict(), it is sufficient to store the refname of
the conflicting reference.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
It wasn't pulling its weight. And we are about to need code similar to
this where no ref_entry is available and with more diverse error
messages. Rather than try to generalize the function, just inline it.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
It wasn't pulling its weight. And in a moment we will need similar
tests that take a refname rather than a ref_entry as parameter, which
would have made entry_matches() even less useful.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
This change wouldn't be worth it by itself, but in a moment we will
use the strbuf for more string juggling.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
The function had a "dir" parameter that was shadowed by a local "dir"
variable within a code block. Use the former in place of the latter.
(This is consistent with "dir"'s use elsewhere in the function.)
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Change the comments to a running example of running the function with
refname set to "refs/foo/bar". Add some more explanation of the logic.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
"git prune" used to largely ignore broken refs when deciding which
objects are still being used, which could spread an existing small
damage and make it a larger one.
* jk/prune-with-corrupt-refs:
refs.c: drop curate_packed_refs
repack: turn on "ref paranoia" when doing a destructive repack
prune: turn on ref_paranoia flag
refs: introduce a "ref paranoia" flag
t5312: test object deletion code paths in a corrupted repository
When we delete a ref, we have to rewrite the entire
packed-refs file. We take this opportunity to "curate" the
packed-refs file and drop any entries that are crufty or
broken.
Dropping broken entries (e.g., with bogus names, or ones
that point to missing objects) is actively a bad idea, as it
means that we lose any notion that the data was there in the
first place. Aside from the general hackiness that we might
lose any information about ref "foo" while deleting an
unrelated ref "bar", this may seriously hamper any attempts
by the user at recovering from the corruption in "foo".
They will lose the sha1 and name of "foo"; the exact pointer
may still be useful even if they recover missing objects
from a different copy of the repository. But worse, once the
ref is gone, there is no trace of the corruption. A
follow-up "git prune" may delete objects, even though it
would otherwise bail when seeing corruption.
We could just drop the "broken" bits from
curate_packed_refs, and continue to drop the "crufty" bits:
refs whose loose counterpart exists in the filesystem. This
is not wrong to do, and it does have the advantage that we
may write out a slightly smaller packed-refs file. But it
has two disadvantages:
1. It is a potential source of races or mistakes with
respect to these refs that are otherwise unrelated to
the operation. To my knowledge, there aren't any active
problems in this area, but it seems like an unnecessary
risk.
2. We have to spend time looking up the matching loose
refs for every item in the packed-refs file. If you
have a large number of packed refs that do not change,
that outweighs the benefit from writing out a smaller
packed-refs file (it doesn't get smaller, and you do a
bunch of directory traversal to find that out).
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Most operations that iterate over refs are happy to ignore
broken cruft. However, some operations should be performed
with knowledge of these broken refs, because it is better
for the operation to choke on a missing object than it is to
silently pretend that the ref did not exist (e.g., if we are
computing the set of reachable tips in order to prune
objects).
These processes could just call for_each_rawref, except that
ref iteration is often hidden behind other interfaces. For
instance, for a destructive "repack -ad", we would have to
inform "pack-objects" that we are destructive, and then it
would in turn have to tell the revision code that our
"--all" should include broken refs.
It's much simpler to just set a global for "dangerous"
operations that includes broken refs in all iterations.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Various issues around "reflog expire", e.g. using --updateref when
expiring a reflog for a symbolic reference, have been corrected
and/or made saner.
* mh/expire-updateref-fixes:
reflog_expire(): never update a reference to null_sha1
reflog_expire(): ignore --updateref for symbolic references
reflog: improve and update documentation
struct ref_lock: delete the force_write member
lock_ref_sha1_basic(): do not set force_write for missing references
write_ref_sha1(): move write elision test to callers
write_ref_sha1(): remove check for lock == NULL
Currently, if --updateref is specified and the very last reflog entry
is expired or deleted, the reference's value is set to 0{40}. This is
an invalid state of the repository, and breaks, for example, "git
fsck" and "git for-each-ref".
The only place we use --updateref in our own code is when dropping
stash entries. In that code, the very next step is to check if the
reflog has been made empty, and if so, delete the "refs/stash"
reference entirely. Thus that code path ultimately leaves the
repository in a valid state.
But we don't want to the repository in an invalid state even
temporarily, and we don't want to leave an invalid state if other
callers of "git reflog expire|delete --updateref" don't think to do
the extra cleanup step.
So, if "git reflog expire|delete" leaves no more entries in the
reflog, just leave the reference unchanged.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If we are expiring reflog entries for a symbolic reference, then how
should --updateref be handled if the newest reflog entry is expired?
Option 1: Update the referred-to reference. (This is what the current
code does.) This doesn't make sense, because the referred-to reference
has its own reflog, which hasn't been rewritten.
Option 2: Update the symbolic reference itself (as in, REF_NODEREF).
This would convert the symbolic reference into a non-symbolic
reference (e.g., detaching HEAD), which is surely not what a user
would expect.
Option 3: Error out. This is plausible, but it would make the
following usage impossible:
git reflog expire ... --updateref --all
Option 4: Ignore --updateref for symbolic references.
We choose to implement option 4.
Note: another problem in this code will be fixed in a moment.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead, compute the value when it is needed.
Signed-off-by: Stefan Beller <sbeller@google.com>
Edited-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If a reference is missing, its SHA-1 will be null_sha1, which can't
possibly match a new value that ref_transaction_commit() is trying to
update it to. So there is no need to set force_write in this scenario.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
write_ref_sha1() previously skipped the write if the reference already
had the desired value, unless lock->force_write was set. Instead,
perform that test at the callers.
Two of the callers (in rename_ref()) unconditionally set force_write
just before calling write_ref_sha1(), so they don't need the extra
check at all. Nor do they need to set force_write anymore.
The last caller, in ref_transaction_commit(), still needs the test.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
None of the callers pass NULL to this function, and there doesn't seem
to be any usefulness to allowing them to do so.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git blame HEAD -- missing" failed to correctly say "HEAD" when it
tried to say "No such path 'missing' in HEAD".
* jk/blame-commit-label:
blame.c: fix garbled error message
use xstrdup_or_null to replace ternary conditionals
builtin/commit.c: use xstrdup_or_null instead of envdup
builtin/apply.c: use xstrdup_or_null instead of null_strdup
git-compat-util: add xstrdup_or_null helper
Add a docstring for update_ref(), emphasizing its similarity to
ref_transaction_update(). Rename its parameters to match those of
ref_transaction_update().
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If NULL is passed to ref_transaction_update()'s new_sha1 parameter,
then just verify old_sha1 (under lock) without trying to change the
new value of the reference.
Use this functionality to add a new function ref_transaction_verify(),
which checks the current value of the reference under lock but doesn't
change it.
Use ref_transaction_verify() in the implementation of "git update-ref
--stdin"'s "verify" command to avoid the awkward need to "update" the
reference to its existing value.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It makes no sense to delete a reference that is already known not to
exist.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Creating a reference requires a new_sha1 that is not NULL and not
null_sha1.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead, verify the reference's old value if and only if old_sha1 is
non-NULL.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead, verify the reference's old value if and only if old_sha1 is
non-NULL.
ref_transaction_delete() will get the same treatment in a moment.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead of having a separate have_old field, record this boolean value
as a bit in the "flags" field.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change the following functions' "flags" arguments from "int" to
"unsigned int":
* ref_transaction_update()
* ref_transaction_create()
* ref_transaction_delete()
* update_ref()
* delete_ref()
* lock_ref_sha1_basic()
Also change the "flags" member in "struct ref_update" to unsigned.
Suggested-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There is no reason to "reserve" a gap between the public and private
flags values.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It is only used internally now. Document it a little bit better, too.
Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Reviewed-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Restructure "reflog expire" to fit the reflogs better with the
recently updated ref API.
Looked reasonable (except that some shortlog entries stood out like
a sore thumb).
* mh/reflog-expire: (24 commits)
refs.c: let fprintf handle the formatting
refs.c: don't expose the internal struct ref_lock in the header file
lock_any_ref_for_update(): inline function
refs.c: remove unlock_ref/close_ref/commit_ref from the refs api
reflog_expire(): new function in the reference API
expire_reflog(): treat the policy callback data as opaque
Move newlog and last_kept_sha1 to "struct expire_reflog_cb"
expire_reflog(): move rewrite to flags argument
expire_reflog(): move verbose to flags argument
expire_reflog(): pass flags through to expire_reflog_ent()
struct expire_reflog_cb: a new callback data type
Rename expire_reflog_cb to expire_reflog_policy_cb
expire_reflog(): move updateref to flags argument
expire_reflog(): move dry_run to flags argument
expire_reflog(): add a "flags" argument
expire_reflog(): extract two policy-related functions
Extract function should_expire_reflog_ent()
expire_reflog(): use a lock_file for rewriting the reflog file
expire_reflog(): return early if the reference has no reflog
expire_reflog(): rename "ref" parameter to "refname"
...
"git blame HEAD -- missing" failed to correctly say "HEAD" when it
tried to say "No such path 'missing' in HEAD".
* jk/blame-commit-label:
blame.c: fix garbled error message
use xstrdup_or_null to replace ternary conditionals
builtin/commit.c: use xstrdup_or_null instead of envdup
builtin/apply.c: use xstrdup_or_null instead of null_strdup
git-compat-util: add xstrdup_or_null helper
* mh/reflog-expire: (24 commits)
refs.c: let fprintf handle the formatting
refs.c: don't expose the internal struct ref_lock in the header file
lock_any_ref_for_update(): inline function
refs.c: remove unlock_ref/close_ref/commit_ref from the refs api
reflog_expire(): new function in the reference API
expire_reflog(): treat the policy callback data as opaque
Move newlog and last_kept_sha1 to "struct expire_reflog_cb"
expire_reflog(): move rewrite to flags argument
expire_reflog(): move verbose to flags argument
expire_reflog(): pass flags through to expire_reflog_ent()
struct expire_reflog_cb: a new callback data type
Rename expire_reflog_cb to expire_reflog_policy_cb
expire_reflog(): move updateref to flags argument
expire_reflog(): move dry_run to flags argument
expire_reflog(): add a "flags" argument
expire_reflog(): extract two policy-related functions
Extract function should_expire_reflog_ent()
expire_reflog(): use a lock_file for rewriting the reflog file
expire_reflog(): return early if the reference has no reflog
expire_reflog(): rename "ref" parameter to "refname"
...
This replaces "x ? xstrdup(x) : NULL" with xstrdup_or_null(x).
The change is fairly mechanical, with the exception of
resolve_refdup, which can eliminate a temporary variable.
There are still a few hits grepping for "?.*xstrdup", but
these are of slightly different forms and cannot be
converted (e.g., "x ? xstrdup(x->foo) : NULL").
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Don't just reset, but release the resource held by the local
variable that is about to go out of scope.
Signed-off-by: Rene Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Junio C Hamano
Michael Haggerty
Jeff King
Stefan Beller
René Scharfe