In a server scenario such as Gerrit Code Review, there may be many
atomic BatchRefUpdates contending for locks on both the packed-refs file
and some subset of loose refs. We already retry lock acquisition to
improve this situation slightly, but we can do better by using an
in-process lock. This way, instead of retrying and potentially exceeding
their timeout, different threads sharing the same Repository instance
can wait on a fair lock without having to touch the disk lock. Since a
server is probably already using RepositoryCache anyway, there is a high
likelihood of reusing the Repository instance.
Change-Id: If5dd1dc58f0ce62f26131fd5965a0e21a80e8bd3
If a repo frequently uses PackedBatchRefUpdates, there is likely to be
contention on the packed-refs file, so it's not appropriate to fail
immediately the first time we fail to acquire a lock. Add some logic to
RefDirectory to support general retrying of lock acquisition.
Currently, there is a hard-coded wait starting at 100ms and backing off
exponentially to 1600ms, for about 3s of total wait. This is no worse
than the hard-coded backoff that JGit does elsewhere, e.g. in
FileUtils#delete. One can imagine a scheme that uses per-repository
configuration of backoff, and the current interface would support this
without changing any callers.
Change-Id: I4764e11270d9336882483eb698f67a78a401c251
Make sure all objects referenced by references are reachable. Stop at
the first missing object.
Change-Id: Ifcd7392c4321b17d9290bd87f038bc62bc10dabb
Signed-off-by: Zhen Chen <czhen@google.com>
JGit already had some fsck-like classes like ObjectChecker which can
check for an individual object.
The read-only FsckPackParser which will parse all objects within a pack
file and check it with ObjectChecker. It will also check the pack index
file against the object information from the pack parser.
Change-Id: Ifd8e0d28eb68ff0b8edd2b51b2fa3a50a544c855
Signed-off-by: Zhen Chen <czhen@google.com>
On-disk reflogs are not stored in the packed-refs file, so we cannot
ensure atomic updates. We choose the lesser evil of dropping failed
reflog updates on the floor, rather than throwing an exception even
though the underlying ref updates succeeded.
Add tests for reflogs to BatchRefUpdateTest.
Change-Id: Ia456ba9e36af8e01fde81b19af46a72378e614cd
* Factor out helpers for setting up and executing updates.
* Use common assert methods, with a special enum type that papers over
the fact that there is no ReceiveCommand.Result for transaction
aborted.
* Static import ReceiveCommand.Type constants.
* Add blank lines to separate repo setup, update execution, and asserts.
Change-Id: Ic3717f94331abfc7ae3e92065f3fe32026bf7cea
Run with @Parameterized, so we don't have to duplicate test setup for
each atomic/non-atomic test. We still have to have two different sets of
asserts for the cases where the behavior is different. In fact, this is
a readability win: it emphasizes that performing the exact same setup
except for the atomic setting will have different behavior.
Change-Id: I78a8214075e204732a423341f14c09de273a7854
The existing packed-refs file provides a mechanism for implementing
atomic multi-ref updates without any changes to the on-disk format or
lockfile protocol. We just need to make sure that there are no loose
refs involved in the transaction, which we can achieve by packing the
refs while holding locks on all loose refs. Full details of the
algorithm are in the PackedBatchRefUpdate javadoc.
This change does not implement reflog support, which will come in a
later change.
Change-Id: I09829544a0d4e8dbb141d28c748c3b96ef66fee1
ReceiveCommand.Result has a slightly richer set of possibilities, so it
makes sense for RefUpdate.Result to have more values in order to match.
In particular, this allows us to return REJECTED_MISSING_OBJECT from
RefUpdate when an object is missing.
The comment in RefUpdate#safeParse about expecting some old objects to be
missing is only applicable to the old ID, not the new ID. A missing new
ID is a bug or programmer error, and we should not update a ref to point
to one.
Fix various tests that started failing because they depended for no good
reason on setting refs to point to nonexistent objects; it's always easy
to create a real object when necessary.
It is possible that some downstream users of RefUpdate.Result might
choose to handle one of the new statuses differently, for example by
providing a more user-readable error message; that is not done in this
change.
Change-Id: I734b1c32d5404752447d9e20329471436ffe05fc
Fix patch matching for patterns of form a/b/** : this should not match
paths like a/b but still match a/b/ and a/b/c.
Change-Id: Iacbf496a43f01312e7d9052f29c3f9c33807c85d
Signed-off-by: Dmitry Pavlenko <pavlenko@tmatesoft.com>
Signed-off-by: Andrey Loskutov <loskutov@gmx.de>
Signed-off-by: David Pursehouse <david.pursehouse@gmail.com>
* changes:
Add tests for updating single refs to missing objects
Fix deleting symrefs
RefDirectory: Throw exception if CAS of packed ref list fails
ReceiveCommand: Explicitly check constructor preconditions
BatchRefUpdate: Document when getPushOptions is null
This change makes it possible to configure the 'inCoreLimit' of LocalFile
used in ResolveMerger#insertMergeResult. Since LocalFile itself has some
risks, e.g. it may be left behind as garbage in case of failure. It should
be good to be able to control the size limit for using LocalFile.
Change-Id: I3dc545ade370b2bbdb7c610ed45d5dd4d39b9e8e
Signed-off-by: Changcheng Xiao <xchangcheng@google.com>
Allow a DFS implementation to report blockSize to DfsPackFile,
bypassing alignment errors and corrections in the DfsBlockCache when
the blockSize of a specific file differs from the cache's configured
blockSize.
Change-Id: Ic376314d4a86a0bd528c033e169d93eef035b233
When a file uses a different block size (e.g. 500) than the cache
(e.g. 512), and the DfsPackFile's blockSize field has not been
initialized, the cache misaligns block loads. The cache uses its
default of 512 to compute the block alignment instead of the file's
500.
This causes DfsReader try to set an empty range into an Inflater,
resulting in an object being unable to load.
Change-Id: I7d6352708225f62ef2f216d1ddcbaa64be113df6
Holding the current DfsBlock in a local variable 'b' may prevent the
Java GC from reclaiming it while loading the next block. Remove the
local variable and rely only on the field.
Change-Id: Ibfc8394cac717b485fdc94d5c8479c3f8ca78ee4
Simple test to verify two DfsRepository instances will reuse the same
DfsBlocks in the DfsBlockCache, even though the DfsStreamKey instance
is now different between their DfsPackFile instances.
Change-Id: I409c109142dea488d189b9ac0d3c319755dce7b4
In the future with reftable a DFS implementation may choose to create
a PackDescription that contains only a REFTABLE extension. Filter
these out by only creating a DfsPackFile if the PackDescription as the
expected PackExt.PACK.
Change-Id: I4c831622378156ae6b68f82c1ee1db5e150893be
Not all DFS implementations use globally unique pack names in the
DfsPackDescription. Most require the DfsRepositoryDescription to
qualify the pack. Include DfsRepositoryDescription in the default
DfsStreamKey implementation, to prevent cache collisions.
Change-Id: I9ebf0c76bf2b414a702ae050b32e42588067bc44
Using a HashMap is overkill for this storage. PackExt is a
constrained type that permits no more than 32 unique values in the JVM.
Each is assigned a unique index (getPosition), which can be used as
indexes in a simple long[].
Change-Id: Ib8e3b2db15d3fde28989b6f4b9897f8a7bb36f3b
When 07f98a8b71 ("Derive DfsStreamKey from DfsPackDescription")
stopped caching DfsPackFile in the DfsBlockCache, the DfsPackFile began
to always load the idx, bitmap, or compute reverse index, as the cache
handles were no longer populated by prior requests.
Rework caching to lookup the objects from the DfsBlockCache if the
local DfsPackFile handle is invalid. This allows the DfsPackFile to
be more of a flyweight instance across requests.
Change-Id: Ic7b42ce2d90692cccea36deb30c2c76ccc81638b
While implementing a custom subclass of DfsStreamKey it became obvious
the required derive(String) was making it impossible to construct an
efficient key in all cases.
Instead, use a special wrapper type ForReverseIndex around the INDEX's
own DfsStreamKey to denote the reverse index stream in the
DfsBlockCache. This adds a smaller layer of boxing, but eliminates
weird issues for DFS implementors using specialized DfsStreamKey
implementations for space efficiency reasons.
Now that DfsStreamKey is reasonably light-weight, avoid allocating the
index and reverse index keys until necessary. DfsPackFile mostly
holds the DfsBlockCache.Ref handle to the object, and only needs the
DfsStreamKey when its looking up the handle.
Change-Id: Icea78e8f7f1514087b94ef5f525d9573ea2913f2
By making this a deterministic function, DfsBlockCache can stop
retaining a map of every DfsPackDescription it has ever seen. This
fixes a long standing memory leak in DfsBlockCache.
This refactoring also simplifies the idea of setting up more
lightweight objects around streams.
Change-Id: I051e7b96f5454c6b0a0e652d8f4a69c0bed7f6f4
The reader may find it surprising that this succeeds without incident
unless there is peeling or a fast-forward check involved. This behavior
may be changed in the future, but for now, just document the current
behavior.
Change-Id: I348b37e93e0264dc0905c4d58ce881852d1dfe5e
The RefDirectory implementation of doDelete never considered whether to
delete a symref or its leaf, because the detachingSymbolicRef bit was
never exposed from RefUpdate. The behavior was thus incorrectly to
always delete the symref, never the leaf.
There was no test for this behavior. The only thing that attempted to be
a test was testDeleteHeadInBareRepo, but this test was broken for
reasons unrelated to this bug. Specifically, it set the leaf to point to
a completely nonexistent object, and then asserted that deleting HEAD
resulted in NO_CHANGE. The only reason this test ever passed is because
of a quirk of updateImpl, which treats a missing object as the same as
null. This quirk aside, the test wasn't really testing the right thing.
Turn this into a real test by writing out a real object and pointing the
leaf at that.
Also, add a test for the detachingSymbolicRef case, i.e. deleting the
symref and leaving the leaf alone.
Change-Id: Ib96d2a35b4f99eba0734725486085fc6f9d78aa5
The contents of the packedRefList AtomicReference should never differ
from what we expect prior to writing, because this segment of the code
is protected by the packed-refs lock file on disk. If it does happen,
whether due to programmer error or a rogue process not respecting the
locking protocol, it's better to let the caller know than to silently
drop the whole commit operation on the floor.
The existing concurrentOnlyOneWritesPackedRefs test is inherently
nondeterministic as written, and was already about 6% flaky as measured
by bazel:
$ bazel test --runs_per_test=200 //org.eclipse.jgit.test:org_eclipse_jgit_internal_storage_file_GcPackRefsTest
...
INFO: Elapsed time: 42.608s, Critical Path: 10.35s
//org.eclipse.jgit.test:org_eclipse_jgit_internal_storage_file_GcPackRefsTest FAILED in 12 out of 200 in 1.6s
Stats over 200 runs: max = 1.6s, min = 1.1s, avg = 1.3s, dev = 0.1s
This flakiness was caused by the assumption that exactly one of the 2
threads would fail, when both might actually succeed in practice due to
racing on the compare-and-swap.
For whatever reason, this change affected the interleaving behavior in
such a way that the flakiness jumped to around 50%. Making the
interleaving of the test fully deterministic is beyond the scope of this
change, but a simple tweak to the assertion is enough to make it pass
consistently 200+ times both before and after this change.
Change-Id: I5ff4dc39ee05bda88d47909acb70118f3d0c8f74
Some downstream code checks whether a ReceiveCommand is a create or a
delete based on the type field. Other downstream code (in particular a
good chunk of Gerrit code I wrote) checks the same thing by comparing
oldId/newId to zeroId. Unfortunately, there were no strict checks in the
constructor that ensures that zeroId is only set for oldId/newId if the
type argument corresponds, so a caller that passed mismatched IDs and
types would observe completely undefined behavior as a result. This is
and always has been a misuse of the API; throw IllegalArgumentException
so the caller knows that it is a misuse.
Similarly, throw from the constructor if oldId/newId are null. The
non-nullness requirement was already documented. Fix RefDirectoryTest to
not do the wrong thing.
Change-Id: Ie2d0bfed8a2d89e807a41925d548f0f0ce243ecf
This new base class has the minimum set of properties and methods
necessary for DfsBlockCache to manage blocks of a file in the cache.
Subclasses can use DfsBlockCache for any content.
This refactoring opens the door for additional PackExt types other
than PACK to be stored on a block-by-block basis by the DfsBlockCache.
Change-Id: I307228fc805c3ff0c596783beb24fd52bec35ba8
Instead of overloading the pack's DfsStreamKey with negative positions
for the idx, reverse idx and bitmap, assign a unique DfsStreamKey for
each of these related streams.
Change-Id: Ie048036c74a1d1bbf5ea7e888452dc0c1adf992f
This renaming supports reusing DfsStreamKey in a future commit
to index other PackExt type streams inside of the DfsBlockCache.
Change-Id: Ib52d374e47724ccb837f4fbab1fc85c486c5b408
Dave Borowitz
David Pursehouse
Christian Halstrick
Han-Wen Nienhuys
Zhen Chen
Joan Goyeau
Dmitry Pavlenko
Shawn Pearce
Changcheng Xiao
Matthias Sohn