BatchRefUpdate: repro racy atomic update, and fix it
PackedBatchRefUpdate was creating a new packed-refs list that was
potentially unsorted. This would be papered over when the list was
read back from disk in parsePackedRef, which detects unsorted ref
lists on reading, and sorts them. However, the BatchRefUpdate also
installed the new (unsorted) list in-memory in
RefDirectory#packedRefs.
With the timestamp granularity code committed to stable-5.1, we can
more often accurately decide that the packed-refs file is clean, and
will return the erroneous unsorted data more often. Unluckily timed
delays also cause the file to be clean, hence this problem was
exacerbated under load.
The symptom is that refs added by a BatchRefUpdate would stop being
visible directly after they were added. In particular, the Gerrit
integration tests uses BatchRefUpdate in its setup for creating the
Admin group, and then tries to read it out directly afterward.
The tests recreates one failure case. A better approach would be to
revise RefList.Builder, so it detects out-of-order lists and
automatically sorts them.
Fixes https://bugs.eclipse.org/bugs/show_bug.cgi?id=548716 and
https://bugs.chromium.org/p/gerrit/issues/detail?id=11373.
Bug: 548716
Change-Id: I613c8059964513ce2370543620725b540b3cb6d1
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Return a new instance from openSystemConfig and openUserConfig
Move the handling of cached user and system config to getSystemConfig
and getUserConfig methods and revert the implementation of
openSystemConfig and openUserConfig to the old stateless
implementation.
This ensures the open methods respect the passed-in parent config, which
may be different on each invocation. Additionally, returning a new
instance matches the behavior of the previous implementation of the
default system reader, which downstream callers may be depending on.
Move the implementation of the new caching methods getSystemConfig and
getUserConfig up to SystemReader. This avoids that we break the ABI for
subclasses of SystemReader.
Also see [1] which fixed a similar problem with Gerrit's custom
SystemReader.
[1] https://gerrit-review.googlesource.com/c/gerrit/+/225458
Change-Id: If54a2491932d8fc914d4649cb73c9e837c5b8ad0
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Use AtomicReferences to cache user and system level configs
This ensures that only one instance of user and one instance of system
config is set.
Change-Id: Idd00150f91d2d40af79499dd7bf8ad5940f87c4e
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
deleteChildren was called on directory instead of gitDir, leading to a
potential null pointer exception if the git directory existed initially.
Bug: 550340
Change-Id: Iafc3b2961253a99862a59e81c7371f7bc564b412
Signed-off-by: Adrien Bustany <adrien-xx-eclipse@bustany.org>
Avoid sign extension when comparing mtime with Instant#getEpochSecond
Ensure we use the same type when comparing seconds since the epoch.
This does not prevent that in 2038 timestamps in seconds since the epoch
stored in a 32 bit integer will overflow. Integer.MAX_VALUE translates
to 2038-01-19T03:14:07Z. After this date we'll have an issue since we
store seconds since the epoch in a 32 bit integer in some places.
Bug: 319142
Change-Id: If0c03003d40b480f044686e2f7a2f62c9f4e2fe1
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Fix deprecation in DirCache caused by Instant based DirCacheEntry
Replace the two int variables smudge_s and smudge_ns by an Instant and
use the new method DirCacheEntry.mightBeRacilyClean(Instant).
Change-Id: Id70adbb0856a64909617acf65da1bae8e2ae934a
Signed-off-by: Michael Keppler <Michael.Keppler@gmx.de>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Cache user global and system-wide git configurations
So far the git configuration and the system wide git configuration were
always reloaded when jgit accessed these global configuration files to
access global configuration options which are not in the context of a
single git repository. Cache these configurations in SystemReader and
only reload them if their file metadata observed using FileSnapshot
indicates a modification.
Change-Id: I092fe11a5d95f1c5799273cacfc7a415d0b7786c
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Signed-off-by: Thomas Wolf <thomas.wolf@paranor.ch>
The existing javadoc was copied from another method and not adapted.
Change-Id: I39a7e5d719b2c379de9bd1a4710a55a73700c6f0
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Improve retry handling when saving FileStoreAttributes fails
- fix handling of interrupts in FileStoreAttributes#saveToConfig
- increase retry wait time to 100ms
- don't wait after last retry
- dont retry if failure is caused by another exception than
LockFailedException
Change-Id: I108c012717d2bcce71f2c6cb9cf0879de704ebc2
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Make supportsAtomicCreateNewFile return true as default
The method org.eclipse.jgit.util.FS.supportsAtomicCreateNewFile()
should default to true as mentioned in docs [1]
org.eclipse.jgit.util.FS_POSIX.supportsAtomicCreateNewFile() method
will set the value to false if the git config
core.supportsatomiccreatenewfile is not set.
It should default to true if the configuration is undefined.
[1]
4169a95a65/org.eclipse.jgit/src/org/eclipse/jgit/util/FS_POSIX.java (L372)
Bug: 544164
Change-Id: I16ccf989a89da2cf4975c200b3228b25ba4c0d55
Signed-off-by: Vishal Devgire <vishaldevgire@gmail.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Remove FileBasedConfig.load(boolean) introduced in d45219ba
We can't add this method to the super class StoredConfig since that
abstracts from filesystem storage. MockSystemReader.MockConfig is a
StoredConfig and is also used by tests for dfs based storage. Hence
remove this leaky abstraction.
This implies we always use the fallback FileStoreAttributes which means
a config file modification is considered racy within the first 2
seconds. This should not be an issue since typically configs change
rarely and re-reading a config within the racy period is relatively
cheap since configs are small.
Change-Id: Ia2615addc24a7cadf3c566ee842c6f4f07e159a5
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Increase the safety factor to 2.5x for extra safety if max of measured
timestamp resolution and measured minimal racy threshold is < 100ms, use
1.25 otherwise since for large filesystem resolution values the
influence of finite resolution of the system clock should be negligible.
Before, not yet using the newly introduced minRacyThreshold measurement,
the threshold was 1.1x FS resolution, and we could issue the
following sequence of events,
start
create-file
read-file (currentTime)
end
which had the following timestamps:
create-file 1564589081
start 1564589082
read 1564589082
end 1564589082
In this case, the difference between create-file and read is 5ms,
which exceeded the 4ms FS resolution, even though the events together
took just 2ms of runtime.
Reproduce with:
bazel test --runs_per_test=100 \
//org.eclipse.jgit.test:org_eclipse_jgit_internal_storage_file_FileSnapshotTest
The file system timestamp resolution is 4ms in this case.
This code assumes that the kernel and the JVM use the same clock that
is synchronized with the file system clock. This seems plausible,
given the resolution of System.currentTimeMillis() and the latency for
a gettimeofday system call (typically ~1us), but it would be good to
justify this with specifications.
Also cover a source of flakiness: if the test runs under extreme load,
then we could have
start
create-file
<long delay>
read
end
which would register as an unmodified file. Avoid this by skipping the
test if end-start is too big.
[msohn]:
- downported from master to stable-5.1
- skip test if resolution is below 10ms
- adjust safety factor to 1.25 for resolutions above 100ms
Change-Id: I87d2cf035e01c44b7ba8364c410a860aa8e312ef
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
In LockFile#waitForStatChange wait in units of file time resolution
Since we now measure file time resolution we can use it to replace the
hard coded wait time of 25ms. FileSnapshot#equals will return true until
the mtime of the old (o) and the new FileSnapshot (n) differ by at least
one file time resolution.
Change-Id: Icb713a80ce9eb929242ed083406bfb6650c72223
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Cache FileStoreAttributeCache entries since looking up FileStore for a
file may be expensive on some platforms.
Implement a simple LRU cache based on ConcurrentHashMap using a simple
long counter to order access to cache entries.
Change-Id: I4881fa938ad2f17712c05da857838073a2fc4ddb
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Signed-off-by: Marc Strapetz <marc.strapetz@syntevo.com>
Also-By: Marc Strapetz <marc.strapetz@syntevo.com>
Fix FileSnapshot#save(long) and FileSnapshot#save(Instant)
Use the fallback timestamp resolution as already described in the
javadoc of these methods. Using zero file timestamp resolution doesn't
make sense.
Change-Id: Iaad2a0f99c3be3678e94980a0a368181b6aed38c
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Persist minimal racy threshold and allow manual configuration
To enable persisting the minimal racy threshold per FileStore add a
new config option to the user global git configuration:
- Config section is "filesystem"
- Config subsection is concatenation of
- Java vendor (system property "java.vendor")
- Java version (system property "java.version")
- FileStore's name, on Windows we use the attribute volume:vsn instead
since the name is not necessarily unique.
- separated by '|'
e.g.
"AdoptOpenJDK|1.8.0_212-b03|/dev/disk1s1"
The same prefix is used as for filesystem timestamp resolution, so
both values are stored in the same config section
- The config key for minmal racy threshold is "minRacyThreshold" as a
time value, supported time units are those supported by
DefaultTypedConfigGetter#getTimeUnit
- measure for 3 seconds to limit runtime which depends on hardware, OS
and Java version being used
If the minimal racy threshold is configured for a given FileStore the
configured value is used instead of measuring it.
When the minimal racy threshold was measured it is persisted in the user
global git configuration.
Rename FileStoreAttributeCache to FileStoreAttributes since this class
is now declared public in order to enable exposing all attributes in one
object.
Example:
[filesystem "AdoptOpenJDK|11.0.3|/dev/disk1s1"]
timestampResolution = 7000 nanoseconds
minRacyThreshold = 3440 microseconds
Change-Id: I22195e488453aae8d011b0a8e3276fe3d99deaea
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Also-By: Marc Strapetz <marc.strapetz@syntevo.com>
Measure minimum racy interval to auto-configure FileSnapshot
By running FileSnapshotTest#detectFileModified we found that the sum of
measured filesystem timestamp resolution and measured clock resolution
may yield a too small interval after a file has been modified which we
need to consider racily clean. In our tests we didn't find this behavior
on all systems we tested on, e.g. on MacOS using APFS and Java 8 and 11
this effect was not observed.
On Linux (SLES 15, kernel 4.12.14-150.22-default) we collected the
following test results using Java 8 and 11:
In 23-98% of 10000 test runs (depending on filesystem type and Java
version) the test failed, which means the effective interval which needs
to be considered racily clean after a file was modified is larger than
the measured file timestamp resolution.
"delta" is the observed interval after a file has been modified but
FileSnapshot did not yet detect the modification:
"resolution" is the measured sum of file timestamp resolution and clock
resolution seen in Java.
Java version filesystem failures resolution min delta max delta
1.8.0_212-b04 btrfs 98.6% 1 ms 3.6 ms 6.6 ms
1.8.0_212-b04 ext4 82.6% 3 ms 1.1 ms 4.1 ms
1.8.0_212-b04 xfs 23.8% 4 ms 3.7 ms 3.9 ms
1.8.0_212-b04 zfs 23.1% 3 ms 4.8 ms 5.0 ms
11.0.3+7 btrfs 98.1% 3 us 0.7 ms 4.7 ms
11.0.3+7 ext4 98.1% 6 us 0.7 ms 4.7 ms
11.0.3+7 xfs 98.5% 7 us 0.1 ms 8.0 ms
11.0.3+7 zfs 98.4% 7 us 0.7 ms 5.2 ms
Mac OS
1.8.0_212 APFS 0% 1 s
11.0.3+7 APFS 0% 6 us
The observed delta is not distributed according to a normal gaussian
distribution but rather random in the observed range between "min delta"
and "max delta".
Run this test after measuring file timestamp resolution in
FS.FileAttributeCache to auto-configure JGit since it's unclear what
mechanism is causing this effect.
In FileSnapshot#isRacyClean use the maximum of the measured timestamp
resolution and the measured "delta" as explained above to decide if a
given FileSnapshot is to be considered racily clean. Add a 30% safety
margin to ensure we are on the safe side.
Change-Id: I1c8bb59f6486f174b7bbdc63072777ddbe06694d
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
We should not list the complete cache but only show the cache entry at
hand.
Change-Id: I22be2a4dcbf0145155e23f2389bfcf5662cf23a6
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Repeat the test 10000 times to get statistics if measured
fsTimestampResolution is working in practice to detect racy git
situations.
Add a class to compute statistics for this test. Log delta between
lastModified and time when FileSnapshot failed to detect modification.
This happens if the racy git limit determined by measuring filesystem
timestamp resolution and clock resolution is too small. If it would be
correct FileSnapshot would always detect modification or mark it
modified if time since modification is smaller than the racy git limit.
Change-Id: Iabe7af1a7211ca58480f8902d4fa4e366932fc77
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Measure stored timestamp resolution instead of time to touch file
Measure granularity of timestamps stored in the filesystem by setting
and then getting lastModified timestamp until the read value changed.
Increase increment exponentially to limit number of iterations starting
with 1 microsecond since Java's FileTime (up to Java 12) truncates
timestamps to 1 microsecond resolution. The chosen algorithm yields 2000
steps between 1 ms and 2.5 s.
Also measure clock resolution and add that for the total timestamp
resolution. This avoids systematic measurement errors introduced by
doing IO to touch a file.
Change-Id: I9b37138619422452373e298d9d8c7cb2c384db3f
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
We should not use configuration when creating FileSnapshot when
accessing FileBasedConfig.
Change-Id: Ic521632870f18bb004751642b9d30648dd94049a
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Use Instant instead of milliseconds for filesystem timestamp handling
This enables higher file timestamp resolution on filesystems like ext4,
Mac APFS (1ns) or NTFS (100ns) providing high timestamp resolution on
filesystem level.
Note:
- on some OSes Java 8,9 truncate milliseconds, see
https://bugs.openjdk.java.net/browse/JDK-8177809, fixed in Java 10
- UnixFileAttributes truncates timestamp resolution to microseconds when
converting the internal representation to FileTime exposed in the API,
see https://bugs.openjdk.java.net/browse/JDK-8181493
- WindowsFileAttributes also provides only microsecond resolution
Change-Id: I25ffff31a3c6f725fc345d4ddc2f26da3b88f6f2
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Workaround SecurityException in FS#getFsTimestampResolution
On Android FS#getFsTimestampResolution always throws a
SecurityException, handle this by falling back to the fallback timestamp
resolution.
Bug: 548947
Change-Id: I0ee6cb3c20e189bdc8d488434a930427ad6f2df2
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
FS: ignore AccessDeniedException when measuring timestamp resolution
It seems on cygwin creating a file under a writable directory can fail
with AccessDeniedException. Log a warning in this case and fallback to
worst case timestamp resolution of 2 seconds.
Bug: 548648
Change-Id: Ic50c31ce9dc9ccadd4db5247df929418ac62d45c
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Checking lastModified is time critical hence debug trace is the only way
to analyze issues since debugging is impractical.
Also add configuration for buffering of log4j output to reduce runtime
impact when debug trace is on. Limit buffer to 1MiB and comment this
configuration out since we may not always want to use buffering.
Change-Id: Ib1a0537b67c8dc3fac994a77b42badd974ce6c97
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Use FileChannel.open to touch file and set mtime to now
Use options
- StandardOpenOption.CREATE to create touched file if not existing
- StandardOpenOption.SYNC to enforce synch of data and meta data changes
- StandardOpenOption.WRITE
Also set mtime explicitly in FileUtils#touch to the current system time.
This should fix that the previous implementation didn't work on
- locally cached Windows network share (CSC-CACHE filesystem) mapped as
a drive
- nfsv4 mounts on Linux
and that it didn't create unborn file like Linux command "touch".
Apache common's and Guava's touch() use the same approach.
Immediately after creating the probe file used to measure timestamp
resolution touch it. This ensures we always use the local system clock
when measuring filesystem timestamp resolution. This should prevent that
clock skew could influence the measured timestamp resolution in case of
a mounted network filesystem.
Bug: 548598
Change-Id: Iaeaf5967963f582395a195aa637b8188bfadac60
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Persist filesystem timestamp resolution and allow manual configuration
To enable persisting filesystem timestamp resolution per FileStore add a
new config section to the user global git configuration:
- Config section is "filesystem"
- Config subsection is concatenation of
- Java vendor (system property "java.vm.vendor")
- runtime version (system property "java.vm.version")
- FileStore's name
- separated by '|'
e.g.
"AdoptOpenJDK|1.8.0_212-b03|/dev/disk1s1"
The prefix is needed since some Java versions do not expose the full
timestamp resolution of the underlying filesystem. This may also
depend on the underlying operating system hence concrete key values
may not be portable.
- Config key for timestamp resolution is "timestampResolution" as a time
value, supported time units are those supported by
DefaultTypedConfigGetter#getTimeUnit
If timestamp resolution is already configured for a given FileStore
the configured value is used instead of measuring the resolution.
When timestamp resolution was measured it is persisted in the user
global git configuration.
Example:
[filesystem "AdoptOpenJDK|1.8.0_212-b03|/dev/disk1s1"]
timestampResolution = 1 seconds
If locking the git config file fails retry saving the resolution up to 5
times in order to workaround races with another thread.
In order to avoid stack overflow use the fallback filesystem timestamp
resolution when loading FileBasedConfig which creates itself a
FileSnapshot to help checking if the config changed.
Note:
- on some OSes Java 8,9 truncate to milliseconds or seconds, see
https://bugs.openjdk.java.net/browse/JDK-8177809, fixed in Java 10
- UnixFileAttributes up to Java 12 truncates timestamp resolution to
microseconds when converting the internal representation to FileTime
exposed in the API, see https://bugs.openjdk.java.net/browse/JDK-8181493
- WindowsFileAttributes also provides only microsecond resolution up to
Java 12
Hence do not attempt to manually configure a higher timestamp resolution
than supported by the Java version being used at runtime.
Bug: 546891
Bug: 548188
Change-Id: Iff91b8f9e6e5e2295e1463f87c8e95edf4abbcf8
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
In order to avoid blocking on the main thread during measurement
interactive applications like EGit may want to measure the filesystem
timestamp resolution asynchronously.
In order to enable measurement in the background call
FileStoreAttributeCache.setAsyncfileStoreAttrCache(true)
before the first access to cached FileStore attributes.
Bug: 548188
Change-Id: I8c9a2dbfc3f1d33441edea18b90e36b1dc0156c7
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Add a unittest.
In commit I5485db55 ("Fix FileSnapshot's consideration of file size"),
the special casing of UNKNOWN_SIZE was forgotten.
This change, together with I493f3b57b ("Measure file timestamp
resolution used in FileSnapshot") introduced a regression that would
occasionally surface in Gerrit integration tests marked UseLocalDisk,
with the symptom that creating the Admin user in NoteDb failed with a
LOCK_FAILURE.
Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
Change-Id: I7ffd972581f815c144f810481103c7985af5feb0
Timeout measuring file timestamp resolution after 2 seconds
It was reported that measuring file timestamp resolution may hang
indefinitely on nfs. Hence timeout this measurement at the known worst
filesystem timestamp resolution (FAT) of 2 seconds.
Bug: 548188
Change-Id: I17004b0aa49d5b0e76360a008af3adb911b289c0
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Fix non-deterministic hash of archives created by ArchiveCommand
Archives created by the ArchiveCommand didn't produce deterministic
archive hashes. For RevCommits RevWalk.parseTree returns the root tree
instead of the RevCommit hence retrieving the commit's timestamp didn't
work. Instead use RevWalk.parseAny and extract the tree manually.
Archive entries store timestamps with 1 second resolution hence we need
to wait longer when creating the same archive twice and compare archive
hashes. Otherwise hash comparison in tests wouldn't fail without this
patch.
Bug: 548312
Change-Id: I437d515de51cf68265584d28a8446cebe6341b79
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
As reported by Error Prone:
An inner class should be static unless it references members of its
enclosing class. An inner class that is made non-static unnecessarily
uses more memory and does not make the intent of the class clear.
See https://errorprone.info/bugpattern/ClassCanBeStatic
Change-Id: Ib99d120532630dba63cf400cc1c61c318286fc41
Signed-off-by: David Pursehouse <david.pursehouse@gmail.com>
(cherry picked from commit ee40efcea4)
Extend FileSnapshot for packfiles to also use checksum to detect changes
If the attributes of FileSnapshot don't detect modification of a
packfile read the packfile's checksum and compare it against the
checksum cached in the loaded packfile.
Since reading the checksum needs less IO than reloading the complete
packfile this may help to reduce the overhead to detect modficiation
when a gc completes while ObjectDirectory scans for packfiles in another
thread.
Bug: 546891
Change-Id: I9811b497eb11b8a85ae689081dc5d949ca8c4be5
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
Wait opening new packfile until it can't be racy anymore
If
- pack.waitPreventRacyPack = true (default is false)
- packfile size > pack.minSizePreventRacyPack (default is 100 MB)
wait after a new packfile was written and before it is opened until it
cannot be racy anymore.
If a new packfile is accessed while it's still racy at least the pack's
index will be reread by ObjectDirectory.scanPacksImpl(). Hence it may
save resources to wait one tick of the file system timer to avoid this
reloading. On filesystems with a coarse timestamp resolution it may be
beneficial to skip this wait for small packfiles.
Bug: 546891
Change-Id: I0e8bf3d7677a025edd2e397dd2c9134ba59b1a18
Signed-off-by: Matthias Sohn <matthias.sohn@sap.com>
David Pursehouse
Han-Wen Nienhuys
Matthias Sohn
Adrien Bustany
Michael Keppler
Vishal Devgire