The FileNotFoundException is typically raised in three conditions:
1. file doesn't exist
2. incompatible read vs. read/write open modes
3. filesystem locking
4. temporary lack of resources (e.g. too many open files)
1. is already managed, 2. would never happen as packs are not
overwritten while with 3. and 4. it is worth logging the exception and
retrying to read the pack again.
Log transient errors using an exponential backoff strategy to avoid
flooding the logs with the same error if consecutive retries to access
the pack fail repeatedly.
Bug: 513435
Change-Id: I03c6f6891de3c343d3d517092eaa75dba282c0cd
Signed-off-by: Luca Milanesio <luca.milanesio@gmail.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
In order to limit the number of directories we check for emptiness only
consider fanout directories which contained unreferenced loose objects
we deleted in the same gc run.
Change-Id: Idf8d512867ee1c8ed40bd55752122ce83a98ffa2
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
An orphan file is either a bitmap or an idx file in pack folder,
and its corresponding pack file is missing.
Change-Id: I3c4cb1f7aa99dd7b398bdb8d513f528d7761edff
Signed-off-by: Hongkai Liu <hongkai.liu@ericsson.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
We only need the tree id to add it to a TreeWalk so change tree's type
to AnyObjectId.
Bug: 509385
Change-Id: I98dd5fef15cd173fe1fd84273f0f48e64e12e608
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
When a repository is being GCed and a concurrent push is received, there
is the possibility of having a missing object. This is due to the fact
that after the list of objects to delete is built, there is a window of
time when an unreferenced and ready to delete object can be referenced
by the incoming push. In that case, the object would be deleted because
there is no way to know it is no longer unreferenced. This will leave
the repository in an inconsistent state and most of the operations fail
with a missing tree/object error.
Given the incoming push change the last modified date for the now
referenced object, verify this one is still a candidate to delete
before actually performing the delete operation.
Change-Id: Iadcb29b8eb24b0cb4bb9335b670443c138a60787
Signed-off-by: Hector Oswaldo Caballero <hector.caballero@ericsson.com>
FileSnapshot.isModified may have reported a file to be clean although it
was actually dirty.
Imagine you have a FileSnapshot on file f. lastmodified and lastread are
both t0. Now time is t1 and you
1) modify the file
2) update the FileSnapshot of the file (lastModified=t1, lastRead=t1)
3) modify the file again
4) wait 3 seconds
5) ask the Filesnapshot whether the file is dirty or not. It erroneously
answered it's clean.
Any file which has been modified longer than 2.5 seconds ago was
reported to be clean. As the test shows that's not always correct.
The real-world problem fixed by this change is the following:
* A gerrit server using JGit to serve git repositories is processing
fetch requests while simultaneously a native git garbage collection
runs on the repo.
* At time t1 native git writes temporary files in the pack folder
setting the mtime of the pack folder to t1.
* A fetch request causes JGit to search for new packfiles and JGit
remembers this scan in a Filesnapshot on the packs folder. Since the gc
is not finished JGit doesn't see any new packfiles.
* The fetch is processed and the gc ends while the filesystem timer is
still t1. GC writes a new packfile and deletes the old packfile.
* 3 seconds later another request arrives. JGit does not yet know about
the new packfile but is also not rescanning the pack folder because it
cached that the last scan happened at time t1 and pack folder's mtime is
also t1. Now JGit will not be able to resolve any object contained in
this new pack. This behavior may be persistent if objects referenced by
the ref/meta/config branch are affected so gerrit can't read permissions
stored in the refs/meta/config branch anymore and will not allow any
pushes anymore. The pack folder will not change its mtime and therefore
no rescan will take place.
Change-Id: I3efd0ccffeb97b01207dc3e7a6b85c6b06928fad
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
When the reply is already compressed (e.g. a packfile fetched using dumb
HTTP), "Content-Encoding: gzip" wastes bandwidth relative to sending the
content raw. So don't "Accept-Encoding: gzip" for such requests.
Change-Id: Id25702c0b0ed2895df8e9790052c3417d713572c
Signed-off-by: Zhen Chen <czhen@google.com>
The add(long) method was deprecated in favor of addWord(long) in
the 0.8.3 release of JavaEWAH [1].
[1] https://github.com/lemire/javaewah/commit/e443cf5e
Change-Id: I89c397ed02e040f57663d04504399dfdc0889626
Signed-off-by: David Pursehouse <david.pursehouse@gmail.com>
Fix JGits merge-base calculation in case of inconsistent commit times.
JGit was potentially failing to compute correct merge-bases when the
commit times where inconsistent (a parent commit was younger than a
child commit). The code in MergeBaseGenerator was aware of the fact that
sometimes the discovery of a merge base x can occur after the parents of
x have been seen (see comment in #carryOntoOne()). But in the light of
inconsistent commit times it was possible that these parents of a
merge-base have already been returned as a merge-base.
This commit fixes the bug by buffering all commits generated by
MergeBaseGenerator. It is expected that this buffer will be small
because the number of merge-bases will be small. Additionally a new
flag is used to mark the ancestors of merge-bases. This allows to filter
out the unwanted commits.
Bug: 507584
Change-Id: I9cc140b784c3231b972bd2c3de61a789365237ab
This does not address all cases where no message is specified, only
cases where Repository#isValidRefName returns false.
Change-Id: Ib88cdabfdcdf37be0053e06949b0e21ad87a9575
Signed-off-by: Grace Wang <gracewang92@gmail.com>
TransportHttp sets 'Accept-Encoding: gzip' to allow the server to
compress HTTP responses. When fetching a loose object over HTTP, it
uses the following code to read the response:
InputStream in = openInputStream(c);
int len = c.getContentLength();
return new FileStream(in, len);
If the content is gzipped, openInputStream decompresses it and produces
the correct content for the object. Unfortunately the Content-Length
header contains the length of the compressed stream instead of the
actual content length. Use a length of -1 instead since we don't know
the actual length.
Loose objects are already compressed, so the gzip encoding typically
produces a longer compressed payload. The value from the Content-Length
is too high, producing EOFException: Short read of block.
Change-Id: I8d5284dad608e3abd8217823da2b365e8cd998b0
Signed-off-by: Zhen Chen <czhen@google.com>
Helped-by: Jonathan Nieder <jrn@google.com>
The InputStream in FileStream in downloadPack is never closed.
Change-Id: I59975d0b8d51f4b3e3ba9d4496b254d508cb936d
Signed-off-by: Zhen Chen <czhen@google.com>
MonotonicClock can be implemented to provide more certainity about
time than the standard System.currentTimeMillis() can provide. This
can be used by classes such as PersonIdent and Ketch to rely on
more certainity about time moving in a strictly ascending order.
Gerrit Code Review can also leverage this interface through its
embedding of JGit and use MonotonicClock and ProposedTimestamp to
provide stronger assurance that NoteDb time is moving forward.
Change-Id: I1a3cbd49a39b150a0d49b36d572da113ca83a786
Use Oxygen M3 Orbit repository which provides the bundles built using
the new orbit-recipe based build.
CQ: 11658
Change-Id: I7f3dcc966732b32830c75d5daa55383bd028d182
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
The type parameters can now be inferred when creating
ConcurrentHashMap.
A for loop over the keys of a ConcurrentHashMap doesn't
need to use an Iterator<Map.Entry>; loop syntax handles
this just fine over keySet().
Change-Id: I1f85bb81b77f7cd1caec77197f2f0bf78e4a82a1
Java 8 fixed the silent flush during close issue by
FilterOutputStream (base class of BufferedOutputStream)
using try-with-resources to close the stream, getting a
behavior matching what JGit's SafeBufferedOutputStream
was doing:
try {
flush();
} finally {
out.close();
}
With Java 8 as the minimum required version to run JGit
it is no longer necessary to override close() or have
this class. Deprecate the class, and use the JRE's version
of close.
Change-Id: Ic0584c140010278dbe4062df2e71be5df9a797b3
This simplifies testing for Gerrit Code Review where
application code is updating the repository description
and the test harness uses InMemoryRepository.
Change-Id: I9fbcc028ae24d90209a862f5f4f03e46bfb71db0
This method pair allows the caller to read and modify the description
file that is traditionally used by gitweb and cgit when rendering a
repository on the web.
Gerrit Code Review has offered this feature for years as part of
its GitRepositoryManager interface, but its fundamentally a feature
of JGit and its Repository abstraction.
git-core typically initializes a repository with a default value
inside the description file. During getDescription() this string
is converted to null as it is never a useful description.
Change-Id: I0a457026c74e9c73ea27e6f070d5fbaca3439be5
Bazel runs ErrorProne by default and ErrorProne rightly complains that
allowing the user to specify any Cipher can lead to insecure code
(in particular, getCipher("AES") operates in ECB mode). Unfortunately
this is required to support existing repositories insecurely stored
on S3.
Extract the insecure factory code to its own class so this can be built
as a java_library() with this check disabled.
Change-Id: I34f381965bdaa25d5aa8ebf6d8d5271b238334e0
Because flush calls interrupt with writeLock held, it cannot interrupt
a write. Simplify by no longer defending against that.
Change-Id: Ib0b39b425335ff7b0ea1b1733562da5392576a15
StreamCopyThread#run consistently interrupts itself whenever it
discovers it has been interrupted by StreamCopyThread#flush while not
reading. The flushCount is not needed to avoid lost flushes.
All in-tree users of StreamCopyThread never flush. As a nice side
benefit, this avoids the expense of atomic operations that have no
purpose for those users.
Change-Id: I1afe415cd09a67f1891c3baf712a9003ad553062
Now that RepositoryCache have a time based eviction strategy, get rid
of the strategy to evict cache entries if heap memory is running low,
i.e. soft references. Main reason why time based eviction was
implemented was to offer an alternative to the unpredictable soft
references.
Relying on soft references is not working, especially in large heap. The
JVM GC will consider collecting soft references as last resort before
throwing an out of memory error. For example, an application like Gerrit
configured with a 128GB heap, GC will wait until all 128GB is filled
before collecting the soft references so the application will be
suffering long pauses caused by GC for a long time already. In other
words, you will have to restart application because it's unusable before
JVM eviction kicks in.
Keeping the SoftReference in RepositoryCache is causing more harm than
good. If you use the time based eviction (which is the default strategy)
and want to tune JVM to release soft references more aggressively, it
will release repositories from the cache even though they are not
expired which defeats the purpose of the repository cache.
Gerrit uses Lucene library which uses soft references and this is
causing a "memory leak" except if you configure JVM to release soft
references more aggressively which have the nasty side effect of
evicting non expired repositories from the cache.
Change-Id: I9940bd800464c7f007696d0ccde52ea617b2ebce
Signed-off-by: Hugo Arès <hugo.ares@ericsson.com>
Work around issues with JSch not handling interrupts by
isolating the JSch interactions onto another thread.
Run write and flush on a single threaded Executor using
simple Callable operations wrapping the method calls,
waiting on the future to determine the outcome before
allowing the caller to continue.
If any operation was interrupted the state of the stream
becomes fuzzy at close time. The implementation tries to
interrupt the pending write or flush, but this is very
likely to corrupt the stream object, so exceptions are
ignored during such a dirty close.
Change-Id: I42e3ba3d8c35a2e40aad340580037ebefbb99b53
In case a value is used which isn’t a power of 2 there will be a high
chance of java.lang.ArrayIndexOutBoundsException and
org.eclipse.jgit.errors.CorruptObjectException due to a mismatching
assumption for the DfsBlockCache#blockSizeShift parameter.
Change-Id: Ib348b3704edf10b5f93a3ffab4fa6f09cbbae231
Signed-off-by: Philipp Marx <smigfu@googlemail.com>
* GC.tooManyLooseObjects() always responded true since the loop missed
to advance the iterator so it always incremented until the threshold was
exceeded.
* Also fix loop exit criterion which was off by 1.
* Add some tests.
Change-Id: I70976dfaa026efbcf3c46bd45941f37277a18e04
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
StreamCopyThread.flush was introduced in
61645b938bc934fda3b0624c5bac1e3495634750 (Add timeouts to smart
transport clients, 2009-06-19) to support timeouts on write in JSch.
The commit message from that change explains:
JSch made a timeout on write difficult because they explicitly do
a catch for InterruptedException inside of their OutputStream. We
have to work around that by creating an additional thread that just
shuttles data between our own OutputStream and the real JSch stream.
The code that runs on that thread is structured as follows:
while (!done) {
int n = src.read(buf);
dst.write(buf, 0, n);
}
with src being a PipedInputStream representing the data to be written
to JSch. To add flush support, that change wanted to add an extra step
if (wantFlush)
dst.flush();
but to handle the case where the thread is blocked in the read() call
waiting for new input, it needs to interrupt the read. So that is how
it works: the caller runs
pipeOut.write(some data);
pipeOut.flush();
copyThread.flush();
to write some data and force it to flush by interrupting the read.
After the pipeOut.flush(), the StreamCopyThread reads the data that was
written and prepares to copy it out. If the copyThread.flush() call
interrupts the copyThread before it acquires writeLock and starts
writing, we throw away the data we just read to fulfill the flush.
Oops.
Noticed during the review of e67d59df3f
(StreamCopyThread: Do not let flush interrupt a write, 2016-11-04),
which introduced this bug.
Change-Id: I4aceb5610e1bfb251046097adf46bca54bc1d998
flush calls interrupt() to interrupt a pending read and trigger a
flush. Unfortunately that interrupt() call can also interrupt a
pending write, putting Jsch in a bad state and triggering "Short read
of block" errors. Add locking to ensure the flush only interrupts
reads as intended.
Change-Id: Ib105d9e107ae43549ced7e6da29c22ee41cde9d8
Luca Milanesio
Matthias Sohn
Hongkai Liu
Hector Oswaldo Caballero
Christian Halstrick
Zhen Chen
David Pursehouse
Grace Wang
Shawn Pearce
Dave Borowitz
Jonathan Nieder
Hugo Arès
Philipp Marx