The Java GC doesn't always clear these before running out of memory
and failing allocations. In practice OpenJDK 7 is leaving these live,
removing any advantage of the SoftReference to attempt to shed memory
when the GC is unable to continue allocating.
Instead follow the pattern of the DfsBlockCache and use hard refs
to the object data. Require applications to configure the cache
size more accurately given expected memory usage.
Change-Id: I87586b3e71b1cba0308a6a278d42e971be4bccd3
The LRU chain management code was broken leading to situations where
the chain was incomplete. This prevented the cache from removing
items when it exceeded its memory target, causing a leak.
One case was repeated hit on the head of the chain. moveToHead(e)
was invoked linking the head back to itself in a cycle orphaning
the rest of the table.
Add some unit tests to cover this and a few other paths.
Change-Id: Ib27486eaa1b1d2bf1c745a56d0a5832bfb029322
As Chris pointed out change I822721c76c64e614f87a080ced2457941f53adcd
slowed down merge since ANY_DIFF filter is much less efficient than the
manual detection of diffs done in ResolveMerger.processEntry() since it
avoids unnecessary filesystem calls using the git index. Hence only set
the ANY_DIFF filter on bare repositories which don't have a working tree
to scan.
To test performance I used the setup described in Chris' comment on
change I822721c76c64e614f87a080ced2457941f53adcd and modified
ResolveMerger.mergeTrees() to not add the working tree in order to
simulate merging in a bare repository.
At least on Mac I couldn't detect a speedup, with and without the
ANY_DIFF filter merge test takes an average 0.67sec.
Change-Id: I17b3a06f369cee009490f54ad1a2deb6c145c7cf
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Avoid always calling `sh -c umask` on startup, instead deferring
the invocation until the first time a working tree file needs to
use the execute bit. This allows servers using bare repos to avoid
a costly fork+exec for a value that is never used.
Store the umask as an int instead of two Boolean. This is slightly
smaller memory (one int vs. two references) and makes it easier for
an application to force setting the umask to a value that overrides
whatever the shell told JGit.
Simplify the code to bail by returning early when canExecute is
false, which is the common case for working tree files.
Change-Id: Ie713647615bc5bdf5d71b731a6748c28ea21c900
Applications that use a commit message once and do not
need it again can free the body to save memory. Expose
the disposeBody() methods to support this and use it in
pgm.Log which only visits each commit once.
Change-Id: I4142a0749c24f15386ee7fb119934a0432234de3
This was added a very long time ago to support the failed
DHT storage implementation. Since then no storage system
was able to make use of this API, but it pollutes internals
of the walkers.
Kill the API on ObjectReader and drop the invocations from
the walker code.
Change-Id: I36608afdac13a6c3084d7c7e0af5e0cb22900332
Previously using an ObjectWalk meant uninteresting commits may keep
their commit message buffers in memory just in case they were found to
be on the boundary and were output as UNINTERESTING for the caller.
This was incorrect inside StartGenerator. ObjectWalk hides these
internal UNINTERESTING cases from its caller unless RevSort.BOUNDARY
was explicitly set, and its false by default. Callers never see one
of these saved uninteresting commits.
Change the test to allow early dispose unless the application has
explicitly asked for RevSort.BOUNDARY. This allows uninteresting
commit buffers to be discarded and garbage collected in ObjectWalks
when the caller will never be given the RevCommit.
Change-Id: Ic1419cc1d9ee95f4d09386dd0730d54c12dcc157
Despite being the primary author of RevWalk and ObjectWalk I still
fail to remember to setRetainBody(false) in application code using
an ObjectWalk to examine the graph.
Document the default for RevWalk is setRetainBody(true), where the
application usually wants the commit bodies to display or inspect.
Change the default for ObjectWalk to setRetainBody(false), as nearly
all callers want only the graph shape and do not need the larger text
inside a commit body. This allows some code in JGit to be simplified.
Change-Id: I367e42209e805bd5e1f41b4072aeb2fa98ec9d99
That more or less defeats the purpose of using a StringBuilder.
Change-Id: I519f7bf1c9b6670e63c3714210f834ee845dc69f
Signed-off-by: Sebastian Schuberth <sschuberth@gmail.com>
Change-Id: I3145f74ecee9f5b368e7f4b9fd7cb906f407eff5
Signed-off-by: Sebastian Schuberth <sschuberth@gmail.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
If UploadPack or ReceivePack has an exception record an identifier
associated with the repository as part of the log message. This can
help the HTTP admin track down the offending repository and take
action to repair the root cause.
Change-Id: I58f22b33cdb40994f044a26fba9fe965b45be51d
When DirCacheTree.contains() is called and 'aOff' is greater than 'aLen'
an ArrayIndexOutOfBoundsException was thrown. This fix makes
DirCacheTree.contains() more robust and allows parsing such index files
without throwing AIOOB.
I couldn't create a test case leading to this situation but I have seen
such situations while inspecting Bug: 465393. It seems that such
situations are created on Windows when there are invalid pathes in the
index. There may be a not yet known bug leading to such situations in
combination with invalid pathes.
Bug: 465393
Change-Id: I6535d924a22cba9a05df0ccd7e6dc2c9ddc42375
Since git-core ff5effd (v1.7.12.1) the native wire protocol transmits
the server and client implementation and version strings using
capability "agent=git/1.7.12.1" or similar.
Support this in JGit and hang the implementation data off UploadPack
and ReceivePack. On HTTP transports default to the User-Agent HTTP
header until the client overrides this with the optional capability
string in the first line.
Extract the user agent string into a UserAgent class under transport
where it can be specified to a different value if the application's
build process has broken the Implementation-Version header in the
JGit package.
Change-Id: Icfc6524d84a787386d1786310b421b2f92ae9e65
A larger than expected number of real-world repositories found on
the Internet contain invalid author, committer and tagger lines
in their history. Many of these seem to be caused by users misusing
the user.name and user.email fields, e.g.:
[user]
name = Au Thor <author@example.com>
email = author@example.com
that some version of Git (or a reimplementation thereof) copied
directly into the object header. These headers are not valid and
are rejected by a strict fsck, making it impossible to transfer
the repository with JGit/EGit.
Another form is an invalid committer line with double negative for
the time zone, e.g.
committer Au Thor <a@b> 1288373970 --700
The real world is messy. :(
Allow callers and users to weaken the fsck settings to accept these
sorts of breakages if they really want to work on a repo that has
broken history. Most routines will still function fine, however
commit timestamp sorting in RevWalk may become confused by a corrupt
committer line and sort commits out of order. This is mostly fine if
the corrupted chain is shorter than the slop window.
Change-Id: I6d529542c765c131de590f4f7ef8e7c1c8cb9db9
This error happens on nfs file system when you try to read a file that
was deleted or replaced.
When the error happens because the file was deleted, removing it from
the list is the proper way to handle the error, same use case as
FileNotFoundException. When the error happens because the file was
replaced, removing the file from the list will cause the file to be
re-read so it will get the latest version of the file.
Bug: 462868
Change-Id: I368af61a6cf73706601a3e4df4ef24f0aa0465c5
Signed-off-by: Hugo Arès <hugo.ares@ericsson.com>
Pack not found and pack corrupted/invalid are handled by the code (pack
is removed from the list) so logging an error and the stacktrace is
misleading because it implies that there is an action to take to fix the
error.
Lower the log level to warn and remove the stacktrace for those 2 types
of errors and keep the error log statement for any other.
Change-Id: I2400fe5fec07ac6d6c244b852cce615663774e6e
Signed-off-by: Hugo Arès <hugo.ares@ericsson.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
The Iff2de881 tried to fix missing tree ..." but introduced severe
performance degradation (>10x in some cases) when acting as server
(git push) and as client (replication). IOW cure is worse than the
disease.
This reverts commit c4797fe986.
Change-Id: I4e6056eb352d51277867f857a0cab380eca153ac
Signed-off-by: David Ostrovsky <david@ostrovsky.org>
Cached packs are only used when writing over the network or to
a bundle file and reuse validation is always disabled in these
two contexts. The client/consumer of the stream will be SHA-1
checksumming every object.
Reuse validation is most critical during local GC to avoid silently
ignoring corruption by stopping as soon as a problem is found and
leaving everything alone for the end-user to debug and salvage.
Cached packs are not supported during local GC as the bitmap rebuild
logic does not support including a cached pack in the result.
Strip out the validation and force PackWriter to always disable the
cached pack feature if reuseValidation is enabled.
Change-Id: If0d7baf2ae1bf1f7e71bf773151302c9f7887039
Sensible suggestion from Terry Parker as a late comment on
commit f2efcdc6f769d59722b17e9274932d585035cfb6.
Change-Id: I225775bfb6d3d91ae066ff00f9d80a9c02a422c2
This hint allows an underlying implementation to read more bytes when
possible and buffer them locally for future read calls to consume.
Change-Id: Ia986a1bb8640eecb91cfbd515c61fa1ff1574a6f
When a large pack (> 30% of the block cache) is being reused by
copying it pollutes the block cache with noise by storing blocks
that are never referenced again.
Avoid this by streaming the file directly from its channel onto
the output stream.
Change-Id: I2e53de27f3dcfb93de68b1fad45f75ab23e79fe7
The clone or fetch depth is a valuable bit of information
for access logging. Create a public getter to faciliate access.
A precondition check prevents unintentional misuse when the
data isn't valid yet.
Change-Id: I4603d5fd3bd4a767e3e2419b0f2da3664cfbd7f8
Signed-off-by: David Pletcher <dpletcher@google.com>
JGit hit IllegalArgumentException: invalid content length
when pushing large packs to S3.
Bug: 463015
Change-Id: Iddf50d90c7e3ccb15b9ff71233338c6b204b3648
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Shawn Pearce
Matthias Sohn
Sebastian Schuberth
Christian Halstrick
Hugo Arès
David Ostrovsky
David Pletcher