What’s new in dCache 2.10
The release notes
AuthorsPaul Millar <paul.millar@desy.de>
Executive summary
Improved SRM scheduling and protection of other components.
Reimplemented webadmin active transfers page.
Improved support for HTTP and HTTPS 3rd-party transfers.
Added support for 3rd-party transfers through WebDAV.
Improved support for HTTP 3rd-party transfers through SRM.
Improved default HTML rendering in WebDAV.
Added JSON support for info.
Incompatibilities
The NFS namespace cache configuration parameter nfs.namespace-cache.unit is renamed to nfs.namespace-cache.time.unit; this makes it compliant with dCache configuration guidelines.
Kerberos uses values in /etc/krb5.conf by default. If both kerberos.realm and kerberos.key-distribution-center-list are set then custom values are used.
In billing log files, transfer-specific information is written as a colon-separated list of items. For HTTP and xrootd, the format has changed, with a colon now separating the protocol version information and the client IP address. The output for HTTP and xrootd is now consistent with information recorded for other protocol transfers.
Release 2.10.62
poolmanager
PoolManager was updated to properly handle the dcache.authz.staging.pep
and dcache.authz.staging
parameters. This allows to enable stage protection properly.
Changelog 2.10.61..2.10.62
- 29cbe36
- [maven-release-plugin] prepare release 2.10.62
- 4247d9a
- PoolManager : stage protection, fix error in stage.fragment
- 332e0ff
- [maven-release-plugin] prepare for next development iteration
Release 2.10.61
pnfsmanager
Fixed a problem in pnfsmanager that prevented the use of chimera.db.password.file.
Changelog 2.10.60..2.10.61
- 53a6ba9
- [maven-release-plugin] prepare release 2.10.61
- 4700ae2
- pnfsmanager: Fix chimera.db.password.file support
- a381b7e
- [maven-release-plugin] prepare for next development iteration
Release 2.10.60
Changes affecting multiple services
Fixed an issue with the dcache heap dump command
when called with a simple file name as the output path.
In this case the dump could in some cases be written to a different directory while the script claimed the dump had failed.
The dcache dump heap command has a --force
option for cases in which the
JVM is unresponsive. This option was ignored for processes not running
as root.
This is fixed now.
cells
A bouncing message bug in System cell is fixed.
pool
Fix race condition in request scheduler.
Changelog 2.10.59..2.10.60
- 40e95fd
- [maven-release-plugin] prepare release 2.10.60
- 1a19fdd
- dcache: fix heap dump to simple file names
- 619d06f
- script: Make dump heap –force work for non-root processes
- ead548d
- srm: Do not expose TURL before request is ready
- 6196710
- pool: Fix race condition in request scheduler
- b560f4e
- billing: additional fixes to insert triggers
- 1aba4b3
- system-test: update disposable-CA generated credentials
- fef3b09
- cells: Avoid bouncing message on no-route errors in System cell
- 770e269
- [maven-release-plugin] prepare for next development iteration
Release 2.10.59
many
When representing checksums in the admin interface and configuration files, checksums are now presented in an improved format.
Changelog 2.10.58..2.10.59
- 6f9dd3f
- [maven-release-plugin] prepare release 2.10.59
- 78df3e6
- common: fix ChecksumType.toString()
- 967f410
- [maven-release-plugin] prepare for next development iteration
Release 2.10.58
cells
LoginManager would occasionally generate error messages similar to “Discarding listening on $LOCATION 53684’ because its age of 18721640 ms exceeds its time to live of 4500 ms.”. This was due to erroneous reuse of old message envelopes in the internal messaging. This change fixes that problem.
This change addresses a potential problem in which messages sent between cells in the same domain could appear older than they are and thus would risk being discarded due to the time-to-live being expired.
srm
A Tier–1 site reported problems with a major WLCG VO’s read requests. Investigating the source of the problems showed that the srm_ifce library, used by the (outdated) GFAL v1 and the (supported) GFAL v2 SRM libraries, drastically limits the permitted lifetime of requests without providing admins any way to configure this.
For sites seeing errors related to desiredTotalRequestTime being exceeded, this change provides the new configuration option srm.request.maximum-client-assumed-bandwidth in srm.properties as a work-around.
Sites not observing such errors do not need to change anything with regard to this value.
Changelog 2.10.57..2.10.58
- a73a6ed
- [maven-release-plugin] prepare release 2.10.58
- 831dad1
- srm: add short request lifetime work-around
- 838e4f9
- chimera: Fix regression in inheriting ACLs on directory creation (HSQLDB)
- 9cf030e
- cells: Improve robustness of message time to live
- bd4bcbc
- cells: Fix erroneous reuse of message envelope in location manager registration
- 337ba79
- [maven-release-plugin] prepare for next development iteration
Release 2.10.57
pnfsmanager
Billing entries for SRM uploads recently lost the storage class part of the entry. This update fixes that issue.
poolmanager
This modification fixes a potential race condition in pool manager.
webdav
This modification corrects the error reporting under WebDAV. When attempting to delete a non-existing file, unauthenticated users receive a 401 Unauthorized response, while authenticated users receive a 404 Not Found response.
Changelog 2.10.56..2.10.57
- 66254ce
- [maven-release-plugin] prepare release 2.10.57
- b3aa010
- pnfsmanager: Fix regression in SRM billing entries
- 18d6a05
- poolmanager: Fix race condition in pool selection unit
- bd0791c
- webdav: fix 404 error if attempting to delete a nonexistent file
- aa6eef6
- [maven-release-plugin] prepare for next development iteration
Release 2.10.56
pnfsmanager
With this release, PnfsManager adds safety checks rejecting invalid upload paths that SRM might erroneously supply. This release hardens an installation against possible bugs triggering data loss.
This release adds a check that detects failed or incomplete SRM uploads and prevents the file from being committed to its final path. Common symptoms of this bug were zero sized files that experiment catalogues registered as successfully uploaded.
srm
This release adds a check to detect broken uploads using SRM during the final stage of file transmission. While it causes transfers to take a little more time, resilience against upload failures is increased.
Changelog 2.10.55..2.10.56
- 60d6f92
- [maven-release-plugin] prepare release 2.10.56
- 0f3e6d7
- srm: Check for broken files during srmPutDone
- f9b2635
- pnfsmanager: Check file size and upload completion when committing temporary upload paths
- a99b248
- pnfsmanager: Protect against erroneous upload paths
- f5b9fc4
- [maven-release-plugin] prepare for next development iteration
Release 2.10.55
Changes affecting multiple services
Several cases of slow performance were reported while deleting directory in Chimera. This is now fixed.
pool
When command execution to migrate files between pools (e.g. migration
concurrency
or migration copy
) is interrupted due to the failure to
find migration job the returned error message is considered as a bug.
This is now fixed so that a new message is returned indicating that
the job being requested does not exist.
srm
When file upload is cancelled the value of temporary upload path tracked by SRM could be a value different from a regular path, either because it was changed outside of dCache, or it contains entries from a very old version of dCache. This could result in data loss while canceling upload. The current release fixed a potential data loss scenario.
statistics
The statistics service creates static HTML pages that describe dCache
usage over time as simple files that the webadmin service can
serve. This includes information about pools and store-units. The
problem is that the statistics webpages do not show information about
any pool or store-unit that contains a /
in the name. This is now
fixed. A side-effect is that the history of any pool or store-unit
containing a ^
in the name is lost.
Changelog 2.10.54..2.10.55
- 940029f
- [maven-release-plugin] prepare release 2.10.55
- f3a19ff
- chimera: Alter statistics target for t_tags(itagid)
- 6a2f876
- srm: Add safe-guard against invalid file ID in put requests
- e05f2b0
- pool: Don’t consider failure to find migration job a bug
- d9503fe
- statistics: encode ‘/’ in filenames
- 194d2af
- [maven-release-plugin] prepare for next development iteration
Release 2.10.54
nfs
Pinning files is now a non-blocking operation. For files stored on tape, this should result in a more responsive system behaviour, avoiding NFS blocking in situations with many concurrent pin requests.
Changelog 2.10.53..2.10.54
- c27ed41
- [maven-release-plugin] prepare release 2.10.54
- 7b3c317
- adapted commit 3eab402754b814f681a14d296d808031b05f2737 for 2.10 branch
- e83c140
- nfs: use noitify instead of blocking sendAndWait when sending pin/unpin messages via touch “.(get)(<file_name>)(pin)” command
- c00a14b
- [maven-release-plugin] prepare for next development iteration
Release 2.10.53
Changes affecting multiple services
Sometimes when a cell start up was interrupted an error message was logged as a bug. This is now fixed.
info-provider
The GLUE infomation provider supplies information about the dCache instance, which is important for the clients in WLCG area. Because in dCache different doors can have different roots, clients may need to adjust their path when accessing dCache through different doors. The info-provider is updated so that a new path root property is provided. This allows clients to modify paths, as necessary. Note that the SRM door already supports this translation when redirecting clients for transfers.
Changelog 2.10.52..2.10.53
- 627dfda
- [maven-release-plugin] prepare release 2.10.53
- 1e40f5f
- info-provider: publish door root path
- 902503e
- cells: Suppress illegal state exception during initialization
- 8e7fbb8
- [maven-release-plugin] prepare for next development iteration
Release 2.10.52
pool
Revert Netty version back to v3.9.9. A previous release upgraded the version of Netty, but this appears to have introduced problems for some HTTP transfers where the pools run out of memory.
spacemanager
Spacemanager backs off when it encounters a problem writing to the database. Previously, if the problem was due to deadlocks then the two tasks involved are delayed by the same amount, which means it is possible that subsequent attempt will also deadlock. This release randomises the delay to reduce the likelihood of this problem occuring.
Changelog 2.10.51..2.10.52
- 770e638
- [maven-release-plugin] prepare release 2.10.52
- 985cbae
- pool: revert Netty back to v3.9.9
- b3edbfc
- spacemanager: Randomize backoff in case of transient errors
- 82c9902
- [maven-release-plugin] prepare for next development iteration
Release 2.10.51
xrootd
Fix dCache handling of open requests where uploads were considered downloads.
Changelog 2.10.50..2.10.51
- cf8c162
- [maven-release-plugin] prepare release 2.10.51
- bd1bf66
- xrootd: Fix classification of uploads
- 94f6494
- [maven-release-plugin] prepare for next development iteration
Release 2.10.50
Changes affecting multiple services
Don’t log Error while reading from tunnel:
java.nio.channels.AsynchronousCloseException
when a domain shuts
down.
Changelog 2.10.49..2.10.50
- a13edb1
- [maven-release-plugin] prepare release 2.10.50
- 2e55290
- cells: Don’t log AsynchronousCloseException when tunnel closes
- 3baf68f
- [maven-release-plugin] prepare for next development iteration
Release 2.10.49
Changes affecting multiple services
Update the Spring, Milton, AspectJ, Jetty and DataNucleus-core libraries to latest version. All dCache services are affected.
pool
If a 3rd-party transfer fails then the pool may log and report incomplete information on why this happened. This release fixes this problem.
Changelog 2.10.48..2.10.49
- b107659
- [maven-release-plugin] prepare release 2.10.49
- 52b890a
- 2.10: upgrade third party dependencies
- 8fa4aa4
- http–3rd-party: ensure IOException logged with toString
- f5bba65
- info: fix test to be less critical on timing
- 5047e84
- [maven-release-plugin] prepare for next development iteration
Release 2.10.48
nfs
Report EOF when client undertakes mixed read/write workload and attempts to read beyond currently written data.
poolmanager
Previous releases of dCache contained a bug where replicas generated by pool-to-pool copies failed to include the access latency and retention policy. While not directly affecting dCache operations, the result is that this information is no longer reliable.
spacemanager
Fix listing by PNFS-ID. Glob support is removed as it was non-functional.
Changelog 2.10.47..2.10.48
- 6cd514d
- [maven-release-plugin] prepare release 2.10.48
- 4422057
- poolmanager: Fix missing access latency and retention policy on pool to pool copy
- 05f9430
- spacemanager: Fix listing by pnfs id
- f6aada3
- [maven-release-plugin] prepare for next development iteration
- 761dd30
- nfs-proxy: keep track of written bytes
Release 2.10.47
Changes affecting multiple services
This release fixes a caching issue where changes to inode metadata
(e.g., ownership or permissions) for /
(the root directory of
Chimera) are not visible until the service is restarted. This affects
NFS doors and pnfsmanager service.
Changelog 2.10.46..2.10.47
- 69ef1b6
- [maven-release-plugin] prepare release 2.10.47
- c54b8b5
- chimera: Prevent filling of stat cache of root inode
- 15012ac
- [maven-release-plugin] prepare for next development iteration
Release 2.10.46
Changes affecting multiple services
The chimera library, used by PnfsManager and NFS, contains possible race conditions that can lead to a NullPointerException. These are updated so that Chimera gives the correct error message under these circumstances.
The chimera library, used by the pnfsmanager and nfs services,
contains a bug where two near-simultaneous attempts to delete a
hitherto empty directory and write a file into the same directory will
both succeed but leave an orphaned file: it exists in the t_dirs
table but the parent does not exist in t_inodes
table. This seems
to be triggered when an ftp door fails with no write pool
configured
. This release fixes this problem.
pnfsmanager
Fix the error message (logged by the domain hosting pnfsmanager) if an attempt to finalise an SRM upload fails within pnfsmanager, or if an attempt to cancel an SRM upload fails within pnfsmanager.
webdav
Update to the latest version of milton.
xrootd
Update the alice-token
plugin to allow the host name check to
succeed on dual-stack (IPv4 and IPv6) machines.
Changelog 2.10.45..2.10.46
- 7b4db8b
- [maven-release-plugin] prepare release 2.10.46
- 06416d1
- xrootd: Update alice token plugin to fix IPv6 compatibility
- 6540fed
- chimera: Detect races in directory deletion
- 5e4d00b
- chimera: Detect races during move
- 6d150cb
- webdav: update to latest milton
- 5b065e8
- PnfsManager: remove copy-n-paste error in error message
- 96a2000
- system-test: Fix grid-security settings
- 2386c80
- rpm: enforce SL5 compatibility when building RPM packages
- 948023b
- [maven-release-plugin] prepare for next development iteration
Release 2.10.45
Changes affecting multiple services
Eliminate race condition that can lead to a NullPointerException if a cell does not shut down cleanly.
The dcache
script no longer checks whether the hostkey.pem
file is in
PKCS#8 format when dcache stop
is invoked. Previously this could lead
to orphaned dCache domains; for example, when upgrading dCache RPM.
admin
Fix potential IndexOutOfBoundsException should the response from the
acm
cell be malformed.
pnfsmanager
Contribution from Kurchatov: update the ChimeraCleaner to work-around the Java compiler’s inability to produce compatible Java 7 binaries when compiling with JDK 8. Note: sites using official dCache.org packages do NOT need to upgrade as a result of this change.
pool
Fix an intended pool-to-pool transfer optimisation: the receiving pool failed to reuse a delayed mover, should the pool-to-pool request timeout and be retried.
Fix pools so that they do not log NullPointerException if a pool receives a request to restore a file from an HSM to which it has no access.
Scripts
Fix how the writedata
command in the chimera
shell accepted data:
the command-line argument was ignored if supplied and data was taken
from stdin, if no argument was supplied then the command would fail
with NullPointerException. Note: this command does NOT write
data into dCache, but into Chimera.
srm
Fix the context information included when logging failures to write job information to the database.
Changelog 2.10.44..2.10.45
- f0cd764
- [maven-release-plugin] prepare release 2.10.45
- 03207f3
- srm-client, dcache: fixed passing incompatible arguments to functions
- 049be02
- dcache: removed unecessary use of non-short-circuit logic
- 6212e25
- pool: Fix NPE when restoring file
- 21b618f
- ChimeraCleaner: reallow to be compiled/run on different Java versions
- bc2c0da
- scripts: do not check for PKCS#8 formatted hostkey.pem on shutdown
- 3f10a11
- chimera: Null value passed to non-null parameter in org.dcache.chimera.cli.Shell$WriteCommand.call()
- 434b345
- srm: Use correct logging context when saving jobs
- 5d75167
- cells: Fix NPE during shutdown
- 3d7d49e
- [maven-release-plugin] prepare for next development iteration
Release 2.10.44
Changes affecting multiple services
When starting up, all doors (dcap, ftp, nfs, srm, webdav, xrootd) and pools advertise their presence to other dCache components before they are able to handle incoming requests. This can lead to subsequent queries timing out as the service finishes starting up. With this version of dCache, doors and pools only advertise their presence once they can handle incoming requests.
pool
This release updates how dCache configures the Berkeley DB when used for storing pool metadata. In addition, dCache will now no longer disable the pool when suffering a Berkeley DB-related problem if the Berkeley DB environment is still valid. Combined, these two changed should greatly reduce the occurances of pools disabling themselves when under heavy IO load.
Changelog 2.10.43..2.10.44
- 92d00ca
- [maven-release-plugin] prepare release 2.10.44
- 56cd33a
- pool: Refine Berkeley DB failure handling
- 7253888
- Don’t announce cells to other services until they have started
- 2e2680a
- [maven-release-plugin] prepare for next development iteration
Release 2.10.43
Changes affecting multiple services
The ftp
, webdav
and xrootd
doors will delete the target file if
an upload was unsuccessful. The copy manager (part of the
transfermanagers
service) has a similar behaviour if an internal
copy is unsuccessful. If this delete was unsuccessful (e.g., the
client deleted the file itself) previous dCache versions would log
this at ERROR
level. With this dCache version, such occurrences are
logged at DEBUG
level.
nfs
Fix race condition that can occur when a pool is first accepting pNFS transfers if multiple requests are processed almost simultaneously.
webdav
The webdav door has separate configuration allowing the admin to configure the door-local path that contains site-local files and the URI prefix to access those files. Earlier versions of dCache mistakenly used the former for the latter, which this release fixes.
Changelog 2.10.42..2.10.43
- 1d3e3ad
- [maven-release-plugin] prepare release 2.10.43
- f74cb79
- Revert “webdav: Add robots.txt”
- 1b30b12
- webdav: Respect webdav.static-content.uri property
- cf20217
- doors: Do not log failure to delete absent files on upload failures:
- d4d8283
- nfs4: fix race in request processing
- afddd5b
- webdav: Add robots.txt
- e066291
- [maven-release-plugin] prepare for next development iteration
Release 2.10.42
dcap
Doors describe their root path to SRM so it can calculate appropriate TURLs. Previous versions of dCache had dcap doors register incorrect paths, which this release fixes.
dCache configuration allows an admin to control if certain ciphers are allowed. In particular, this allows sites to remove support for problematic ciphers or hashing algorithms. This release fixes a problem where the GSI-dcap door failed to honour such settings.
pool
The replica-manager periodically requests a list of file replicas that a pool is hosting. In previous versions of dCache, if the pool finds a broken file then an error is returned to replica-manager. The replica-manager then considered the entire pool as being offline. With this version od dCache, such errors are logged on the pool. The replica-manager will not consider the pool as hosting that file’s data, but will otherwise consider the pool online.
The different HSM operations (flush, stage and remove) have internal timeouts after which the pool considers the request as failed. In previous versions of dCache, the default pool setup includes a four hour timeout for flush and stage but neglected to set a default for delete. This omission caused delete operations to time-out very quickly. With this release, delete operations also have a default of four hours.
Fix the Ruby implementation of the hsmcp script (hsmcp.rb
) so it can
parse new command-line arguments that include concurrency options.
The concurrency for active HSM operations is configurable and may be adjusted dynamically. In earlier versions of dCache, decreasing the concurrency only became effective when that operation started to idle. This has been fixed so the limits start to have an effect as operations complete.
Each movers can have one of three priority (LOW, MEDIUM, HIGH) and a
selection discipline (FIFO or LIFO). The documented behaviour was for
queue with names that start with -
have LIFO discipline and those
that start with any other character have FIFO discipline. Due to a
bug, the order was wrong, with the priorities inverted and the two
disciplines swapped, so LOW priority movers were started before MEDIUM
level and MEDIUM were started before HIGH. This release fixes this so
HIGH priority movers are selected preferentially over MEDIUM and,
MEDIUM priority movers are chosen over LOW priority; however, it was
decided to keep the disciplines as in previous versions and updated
the documentation accordingly. There are several reasons for this:
first, there is no difference between LIFO and FIFO when movers are
not queued; second, neither discipline will help if the pool is
persistently overloaded; third, LIFO discipline (although unfair) is
documented as providing a better overall throughput during a
time-limited overload; fourth, by default dCache has been running with
LIFO discipline since v1.9.11 (released 2011–01–13) without any
apparent problems.
HTTP third-party transfers report back if there was a problem verifying that the transfer was successful. One possible problem is that the remote server failed to supply checksum information. Reporting of such situations is now fixed.
webdav
In previous versions of dCache, should a user cancel a transfer shortly after a mover is created then there was a risk that the mover is abandoned. This is fixed with this release.
xrootd
In earlier versions of dCache, if the xrootd door times out for a write request while waiting for the mover to send the redirection information then the mover is abandoned. This is fixed with this release.
Changelog 2.10.41..2.10.42
- ade3c9b
- [maven-release-plugin] prepare release 2.10.42
- 4710de1
- hsmcp: update to match new HSM interface.
- abf2356
- (2.10) old replica manager: prevent pool being listed as offline when there are files with corrupt metadata
- 524026d
- webdav: Kill abandoned movers
- 64b14ef
- xrootd: Kill mover on aborted write
- 7d25a32
- pool: Fix transfer prioritization
- f3082be
- pool: Add nearline storage default timeouts
- b9ab9bd
- pool: Let script nearline storage provider scale down when lowering limits
- 9b110dc
- dcap: Fix broken argument parsing
- a07edc5
- dcap: Fix socket factory argument parsing
- 75c5e85
- pool: Fix error reporting in remote HTTP mover
- 1c08f8c
- prepare for next development iteration
Release 2.10.41
srm
Fix security vulnerability in srm EGI-SVG–2015–9495 (restricted).
Release 2.10.40
Changes affecting multiple services
Fix a performance regression when deleting directories; the fix affects the pnfsmanager and nfs services.
In many cases, poolmanager would timeout after ten seconds when asked which pool to use for a transfer. This behaviour was not intended. The consequence of this bug is protocol specific: for some protocols, the door retries internally while other doors propagate this error to the client. Another consequence was the increased risk of the domain hosting poolmanager running out of memory, particularly when staging files. This release fixes the underlying problem. It is recommended that all doors be upgraded.
spacemanager
Fix bug that can result in leaked entries in space-manager file management from failed uploads. The problem is most likely triggered when a client cancels an FTP upload at the same time as the correponding SRM upload request expires. The problem may also be triggered by communication failure with PnfsManager and the user deleting the failed upload before the pool retries.
Changelog 2.10.39..2.10.40
- e420596
- [maven-release-plugin] prepare release 2.10.40
- c67c041
- spacemanager: Fix race condition leading to leaked reservation entries
- 5dbfa08
- chimera: Resolve performance regression in directory deletion
- b022089
- doors: Fix pool selection timeout handling
- fc881e4
- Fix timeout math to avoid overflow
- b77a047
- Preparing for next release cycle
Release 2.10.39
ftp
Fix security vulnerability in gsi
and kerberos
authenticated ftp.
Release 2.10.38
Changes affecting multiple services
The System
cell of each domain contains a version
command that
allows discovery of which dCache version is running. This release
fixes this command. Note: there is no problem with the dcache
script’s version
command.
nfs
The NFS protocol provides access to additional infomation through dot
commands. This release fixes the nameof
and pathof
commands for
non-ASCII filenames.
pnfsmanager
Fix that dCache respects the setgid bit on a parent directory when the user uploads a file via the SRM protocol. Important: the srm node should be updated at the same time.
srm
Enforce authorisation of requests to finalise or cancel an upload.
When initiating an upload, the user’s uid and primary gid are taken as
the request owner-uid and owner-gid respectively. Only users that
have the same uid as the request’s owner-uid or are a member of the
request’s owner-gid are allowed to cancel or finalise an upload.
Important: the srm
must be updated if the pnfsmanager
is
updated.
webadmin
This release fixes a bug with the periodic building of billing plots. Previously, if the billing service took too long to reply then there would be no further updates to the billing plots.
Changelog 2.10.37..2.10.38
- a2c9f0c
- [maven-release-plugin] prepare release 2.10.38
- 5feb45b
- rpm: remove “commented out” macros lines from spec file
- 511e4fe
- chimera: fix nameof and pathof for paths containing unicode
- dacd102
- chimera: Let SRM respect setgid on upload
- ec811ad
- srm: Add authorization to put done and abort requests
- f2f1f44
- module: cells
- bc0f510
- (2.10) dcache-webadmin: add TimeoutCacheException to catch clause in billing service
- ecc0aed
- [maven-release-plugin] prepare for next development iteration
Release 2.10.37
Changes affecting multiple services
Specifying the DISABLE_BROKEN_DH
flag in the dcache.authn.ciphers
configuration property disables all Diffie-Hellman ciphers if Java 7
is used; if dCache is run with Java 8 then this flag has no effect.
Disabling DH ciphers is necessary because Java 7 contains a broken
implementation of Diffie-Hellman, which was fixed with the release of
Java 7 update 51. This dCache release updates the behaviour of the
DISABLE_BROKEN_DH
flag to allow Diffie-Hellman ciphers if dCache is
run within Java 7 update 51 or later.
The description of the DISABLE_EC
and DISABLE_RC4
flags have been
expanded and updated.
infoprovider
An earlier patch added support for publishing a single dCache instance with multiple SRM endpoints. This provided incompatible with sites that use a DNS alias for their official SRM endpoint, so that change is reverted with this release. Support for publishing multiple SRM endpoints is available with dCache v2.13.
nfs
Add support for the TEST_STATEID
and FREE_STATEID
RPC methods.
These are used by the Linux kernel during recovery procedure. The
previous lack of support for these methods can lead to the leaking of
stateids, which can lead to NFS4ERR_RESOURCE : Too many states
being
logged.
pool
Improve the error message (logged by the pool and in billing) should the FTP mover fail to connect to the client.
This release updates the error the pool reports to an NFS client when the client attempts to read or write and the pool cannot find the mover. This situation is mostly likely caused by restarting the pool. When it receives the modified response, the client will fall back to using proxy-IO. This allows NFS clients that were reading a file to survive a pool restart.
Changelog 2.10.36..2.10.37
- 09b0d9d
- [maven-release-plugin] prepare release 2.10.37
- 22bc8dd
- pool: fix ftp mover to provide better logging when failing to connect
- 364d627
- crypto: refine handling of broken ciphers
- 2a76e9a
- Revert “infoprovider: remove single SRM instance limitation”
- 2d3dfa2
- libs: update to nfs4j–0.9.7
- 354a7d6
- pool: report IO error if we cant find NFS mover
- 7957f75
- [maven-release-plugin] prepare for next development iteration
Release 2.10.36
Changes affecting multiple services
The event logger records when messages are received and sent by cells. Some cell messages send string commands; if so, the log entry contains that string. In previous releases, such string commands were mistakenly double-escaped. This is fixed with this release.
ftp
Fix debug output to include the flavour of GSS implementation; for
example, GssFtpDoorV1::secure_reply: going to authorize using k5
pnfsmanager
When a user attempts to delete a symbolic link using a non-NFS door, previous versions of dCache would resolve the symbolic link to determine whether the user is allowed to delete the symblic link: only if the user is allowed to delete the symblic link’s target would the symbolic link be removed. With this release, the check verifies if the user is allowed to delete the symbolic link instead.
pool
This release updates the JVM command-line to make it explicit that compressed object references are in use. This allows the Berkeley DB library to calculate a more accurate cache size, potentially improving pool performance.
In previuos releases, any attempt is made to query a pool’s info (e.g., via the admin interface) while the pool is initialising will block until the initialisation has completed. This has a knock-on effect of blocking all subsequent messages. With this release, requesting information about a pool will not block during initialisation.
Fix high memory usage during pool initialisation if pool has any precious files.
xrootd
Earlier releases will record a stack-trace if xrootd recieves a malformed request. This is now fixed.
Changelog 2.10.35..2.10.36
- 46cc3fd
- [maven-release-plugin] prepare release 2.10.36
- 052822f
- cell: Don’t quote string command in event logger
- ca11c50
- pool: Explicitly enable compressed oops to calculate correct cache size
- 70a70e1
- pool: Fix locking bug causing high memory usage during pool initialization
- c781e18
- chimera: Fix path resolution on delete
- f3007c3
- xrootd: Fix ‘xrootd logs stack-trace on malformed request’
- 1b70957
- move execution of the superclass method before any concrete class initializations
- b87523b
- pool: do not list a repository during initialization
- a5c220a
- [maven-release-plugin] prepare for next development iteration
Release 2.10.35
pool
The Berkeley DB based metadata storage can sometimes fail. Should this happen, the pool must be restarted. In previously releases, such problems were logged with an unclear message and a stack-trace; the pool would continue to operate but would fail all subsequent transfers; nothing made it clear the pool must be restarted. With this release, such Berkeley DB problems will be logged with a concise error message and the pool will be disable, making the restart requirement explicit.
srm
In previous versions of dCache, the ls
operation in SRM occationally
returned incomplete or inconsistent results. This is now fixed.
Changelog 2.10.34..2.10.35
- 472479e
- [maven-release-plugin] prepare release 2.10.35
- 70b12c0
- srm: fix race condition in ls response
- de527be
- pool: Disable pool on meta data failures
- 1f7f733
- [maven-release-plugin] prepare for next development iteration
Release 2.10.34
Changes affecting multiple services
In previous releases, dCache required a layout file be present, even
if that file was empty. This has two negative impacts: the dcache stop
(typically invoked automatically when updating a package) would not work,
nodes where only scripts (e.g., info-provider) are used would require unecessary
configuration. With this release, a missing file generates a warning but
does not prevent the dcache stop
command or scripts from working. This
warning may be suppressed by setting dcache.layout.uri
to an empty string.
httpd
The tape related queues (flush and restore) have no maximum limit, yet both the old web information and new webadmin show a dummy maximum value for these queues. This meaningless maximum value is no longer shown.
info
The info service collects information about, amongst other things, the interface(s) a door listens on. This is made available in different formats. The URL-formatted version, used by info-provider, always used an IP address even when a name was known. This is now fixed.
infoprovider
It is possible to run multiple SRM endpoints in dCache, provided certain restrictions are upheld. With this release, the info-provider publishes multiple SRM endpoints correctly.
nfs
Some shells, when attempt to overwrite a tag’s content using NFS, do so in a way that Chimera previously failed to support. This failure was reported back to the user as a remote I/O error. This release fixes this problem.
pool
Reduce latency when a pool processes a request to start a read mover. This improves dCache responsiveness when a client opens a file for reading.
The NFS mover uses the file’s size when processing read and write requests. For read operations, the file size cannot change. This release takes advantage of this to reduce the load on the underlying filesystem.
poolmanager
Poolmanager may attempt to create additional copies of a file, only to discover
such attempts fail because of other constraints. This leads to the log file
containing entries like P2P denied: already too many copies
and P2P denied:
No pool candidates available/configured/left for p2p or file already everywhere
.
With this release, such entries are logged at info level: they no longer appear
in the log file, but are available via the poolmanager’s pinboard.
webdav
Fix proxied uploads when the client does not send the file size, either directly or via SRM; in particular, this fixes compatibility with ARC.
Changelog 2.10.33..2.10.34
- ea62128
- [maven-release-plugin] prepare release 2.10.34
- 136a382
- systemtest: fix install command in credentials command
- 085ef11
- infoprovider: remove single SRM instance limitation
- 6dcb36e
- info: fix url-name to publish hostname
- 3d69717
- chimera: throw FileExistChimeraException if tag already exists
- 7f5b622
- (2.10) webadmin: do not display numerical value for max restores or stores
- 700915f
- boot: Don’t fail on missing layout file
- c7fc663
- httpd: Do now show maximum for restore and flush queues
- 2291a98
- poolmanager: Reduce log level of p2p denial
- 63d5f7b
- webdav: Fix proxied upload with chunked encoding
- c981400
- pool: simplify duplicate request handling
- 85cf92e
- [maven-release-plugin] prepare for next development iteration
- efaff21
- pool: reduce load on back-end file system
Release 2.10.33
Changes affecting multiple services
A new security configuration option allows the dCache admin to ban all
SSL/TLS ciphers that use the RC4 cipher. RFC–7465 states services
MUST NOT accept an RC4-based cipher suite. Adding the DISABLE_RC4
option to the dcache.authn.ciphers
makes dCache compliant with
RFC–7465. This option is not enabled by default to avoid possible
regression with clients that require the RC4 cipher. This property
affects dcap (with GSI), ftp (with GSI), srm, webadmin, webdav (with
SSL/TLS), xrootd (GSI plugin) services.
pnfsmanager
Fixed ACL inheritance when uploading data through SRM. In earlier versions of dCache, a file uploaded through SRM failed to inherit any inheritable ACEs from the parent directory.
This release brings some modest performance improvements when creating
upload directories. This improvement is available automatically only
to sites that have not yet upgraded to 2.10 (or later). Sites already
running 2.10 or later can enjoy the same improvements by deleting the
upload directory (/upload
by default) to allow dCache to recreate
it. Important: deleting the upload directory will fail any
current SRM uploads; it is recommended to do this during down-time.
pool
The pool’s migration module may be invoked with different pool
selection modes: the -select
option. The random selection option
(-select=random
) excludes pools that are full, but mistakenly
considers replicas that could be deleted (i.e., non-sticky cached
replicas) as part of the used space; this treats a pool as full even
when the pool has removable files. With this release, pools that are
full but contain some cached files are potential targets for random
pool selection.
In earlier releases of dCache, the save
command failed to record the
stage, flush and remove timeouts for nearline storage (rh set
timeout
, st set timeout
, rm set timeout
respectively). This is
now fixed.
This release introduces the pool.mover.nfs.port.min
and
pool.mover.nfs.port.max
configuration properties. Previously, pools
listened on a random port between dcache.net.lan.port.min
and
dcache.net.lan.port.max
; the two new configuration properties take
the two dcache...
configuration properties as default values. Once
the pool is listening on a particular port, it will try to listen on
the same port after restarting. If this proves impossible, another
port is selected and will be used subsequently. Important: using
the same port is important as the pool listening on another TCP port
can trigger high load on the client machine.
Scripts
Fix JAR selection when a short-lived java process is started. This is typically done when using one of the scripts.
spacemanager
Fix writing into a reservation when using a protocol that does not
provide a username or FQAN; for example, when writing into dCache
using NFS and with the WriteToken
directory tag set. Previously
writing would fail with a Message processing failed: null group
message.
srm
The srm service can generate RemoteException
stack traces when
dCache is behaving correctly. These are now logged at debug level and
without a stack trace.
Changelog 2.10.32..2.10.33
- d443035
- [maven-release-plugin] prepare release 2.10.33
- 52eac62
- chimera: Fix merge conflict and Java 7 compatibility
- 1e39414
- pool: introduce unique port number for nfs mover
- f1ec21a
- pool: dedicated port range for nfs
- 61caf5a
- pool: Store nearline storage timeouts to pool setup file
- e98b145
- pnfsmanager: Create base upload directories without tags and acls
- b5a9992
- pnfsmanager: Inherit ACLs on upload with SRM
- dcb821e
- chimera: Add ACL insert triggers for HSQLDB
- 528d5ed
- Fix limited class path generation
- 0ed000f
- pool: Let random pool selection select pools with removable files
- 9248f25
- crypto: allow banning of RC4 cipher suites
- 73c07e0
- spacemanager: Allow unowned files in reservations
- 54a6df2
- srm: Don’t log erroneous stack trace
- 9ae8661
- [maven-release-plugin] prepare for next development iteration
Release 2.10.32
Changes affecting multiple services
This release upgrades the BouncyCastle version from v1.45 to v1.46.
The main motivation is to improve concurrency, so obtaining better
performance on multi-core machines. The update affects the xacml
and voms
plugins to gPlazma and any door configured to use (or that
always uses) GSI authentication: dcap, ftp and srm. There is no
cross-dependency; domains hosting these service may be updated
independently.
pnfsmanager
Fix possible leaked upload directory if PnfsManager takes too long to create the upload directory.
scripts
The chimera
script provides both an interactive shell and the
ability to run a Chimera operation as a single command-line
invocation. This release fixes an problem where the single Chimera
operation fails to provide the output if it is too short.
srm
If a client releases a reservation using the SRM protocol while another SRM client is querying for information about that reservation then there is a risk that the reservation will appear to exist for some 30 seconds, despite the reservation being successfully released. This release fixes this problem.
Changelog 2.10.31..2.10.32
- c2452f6
- [maven-release-plugin] prepare release 2.10.32
- 4751d27
- srm: Fix cache invalidation of space meta data
- 7d4d98d
- libs: Update to voms-api-java 2.0.10.1
- eeaf8cc
- libs: use bouncycastle–1.46
- f33549d
- libs: update jglobus to 2.0.6.9d
- 19baddc
- chimera: Fix single command invocation of chimera utility
- 277cde1
- pnfsmanager: Resolve upload directory leak caused by missing reply flag
- dd3b1ac
- [maven-release-plugin] prepare for next development iteration
Release 2.10.31
Changes affecting multiple services
dCache uses a standard format to monitor the performance of various
components: in the srm door to record how quickly SRM requests are
processed (print srm counters
command), in the nfs door and pool to
monitor NFS performance (stats
and nfs stats
commands,
respectively), the generic cell message monitoring (monitoring info
command), and pnfsmanager service (the “Statistics” section in
info
). This release fixes a rounding error that prevents these
statistics from including long-lived requests.
pinmanager
The pinmanager service has the ls
command that allows the admin to
limit the results to a specific pin or all pins against some PNFS-ID.
This release fixes listing by pin id.
pnfsmanager
With this release, the path of automatically generated upload
directories has changed slightly to improve SRM upload performance.
Previously, these generated paths had the form <upload>/<unique-ID>
,
where <upload>
is the value of the dcache.upload-directory
configuration property (/upload
by default) and <unique-ID>
is
some unique value (a UUID). With this release, these directories have
the form <upload>/<processor-ID>/<unique-ID>
, where <processor-ID>
is some small integer value. Standard-conforming clients are unaffect
by this change and no action are needed by the admin from this change.
Update pnfsmanager to avoid creating temporary directories if the srm has already discarded the request. This helps dCache recover more quickly when it is overloaded from SRM uploads.
pool
Fixed pool’s erroneous interpretation of the HSM timeout (4 hours, by default) as being from when a staging request was initially received, rather than from when it started processing the request (by starting the HSM script or via the new plug-in mechanism).
This release fixes how a pool’s invokates the HSM script. Previously, the pool mistakely omitted the additional arguments that an admin may configure the pool to include.
replicamanager
This release fixes the Can't clear the tables
error reported when
starting replicamanager service with certain PosgreSQL versions.
spacemanager
Fix the update space
admin command so it can remove any ownership
from a reservation. The reservation’s owner is allowed to release the
reservation via the SRM protocol. If the reservation has no owner, it
may only be released through the admin interface.
srm
The info
admin command provides details about the srm service,
including information specifically about SRM activity that can be
processed synchronously or asynchronously: get, put, reserve-space,
ls, and bring-onling. This information describes how many requests
are in each of the possible states, one of which is Waiting for CPU
.
This release fixes the output of the info
command to show the
correct Waiting for CPU
values.
This release drastically improve srm service performance when there are a large number (e.g., thousands) of queued requests; for example, this brings considerable improve for bulk bring-online requests.
This release also brings improved performance when the SRM has only a
few queued requests; more specifically, when the number of queued
requests is less than or equal the number of currently unoccupied
max-inprogress slots. For example, if dCache is not processing any
GET requests then the first srm.request.get.max-inprogress
GET
requests are processed more quickly.
Fix the srm service switching from synchronous to asynchronous processing when processing a bulk-requests with many files; in earlier releases, such large requests could prevent a request from falling back to asynchronous processing.
This release fixes the srm service so it cancels corresponding pinmanager requests when an SRM client aborts a bring-online request.
Changelog 2.10.30..2.10.31
- a660869
- [maven-release-plugin] prepare release 2.10.31
- 2ccd1b1
- (2.10) replica manager: fix table truncation
- 27f8e91
- common: Fix division by zero regression in gauges
- 4bec44b
- common: Fix rounding error in request gauge
- 0208029
- pnfsmanager: Discard upload path creation request on TTL expiration
- a4ef57a
- pnfsmanager: Use per-thread upload directory to reduce lock contention
- 6a51067
- srm: Optimize scheduler performance
- 7f750bd
- srm: Further optimize SRM scheduler
- 7638d21
- srm: Abort pinning when cancelling bring-online requests
- 3a8298d
- system-test: Enable MVCC and logging for HSQLDB
- 37ee6d7
- srm: Fix sync to async mode timeout
- b737a5e
- pinmanager: Fix listing by id
- fdcb0fa
- pool: Add HSM options to hsm script remove callout
- 6511528
- srm: Fix queue size reporting
- 3098128
- pool: Fix timeout behavior of HSM requests
- 857e8b1
- spacemanager: Allow spaces to become unowned
- 501eb58
- [maven-release-plugin] prepare for next development iteration
Release 2.10.30
nfs
There was an error in how ACLs were interpreted by the nfs door where multiple ACEs for the same user were compacted ignoring the flags. With this release, ACEs against the same user but with different flags are honoured.
Changelog 2.10.29..2.10.30
- 8eeaaf3
- [maven-release-plugin] prepare release 2.10.30
- cd071b1
- libs: update to nfs4j–0.9.6
- 57787cc
- [maven-release-plugin] prepare for next development iteration
Release 2.10.29
Changes affecting multiple services
This release fixes an NFS problem triggered when a second process opens a file that is already opened by another process on the same computer. The NFS specification allows the second open to reuse the “layout” (== mover) from the first open. With this release, the door will detect this and reuse the existing mover; the mover is updated to allow this. The nfs door and all pools accessible via the NFS protocol must be updated to this release or newer.
admin
Restore compatibility with loginbroker information for pcells; in particular, the problem affected information provided by the srm door.
dcap
No longer log a stack-trace when transfering a file if the client is
expected to connect to the pool (the “active client mode” option; the
-A
in the dccp command) and the pool is shutdown before the client
connects.
srm
Some SRM requests support bulk requests, where multiple SURLs are processed in the same fashion. The assumption was that the SURLs in a bulk request are distinct. Bulkd requests have been observed that violate this assumption: a request with a SURL appears more than once. This causes problems when staging files from tape. With this release, a SURL that appears multiple times in a request is processed exactly once.
Changelog 2.10.28..2.10.29
- 6008a13
- [maven-release-plugin] prepare release 2.10.29
- d9eb548
- dcap: fix stack-trace when shutting down pool waiting for connection
- 3d261ec
- srm: Remove duplicate SURLs in bringonline and get requests
- 0097928
- admin: Restore pcells compatibility with loginrbroker
- 8756386
- nfs: share mover for the same client
- 79a956c
- nfs: use NFSv4MoverHandler instead of Map in embedded NFS server
- ea49803
- system-test: add missing dCache disposible CA certificate
- cb794a7
- system-test: add regenerated host and user credentials
- 6275545
- [maven-release-plugin] prepare for next development iteration
Release 2.10.28
Changes affecting multiple services
In previous versions of dCache deleting a directory with the last
reference to a tag did not remove that tag. With the introduction of
upload directories this problem became acute, as many such directories
are created and deleted. With this release, deleting the directory
with a tag’s last reference is very likely (but not guaranteed) to
remove that tag’s inode. Further information about this will be made
available via user-forum. The update affects both the nfs door and
pnfsmanager; however, most sites will see the most benefit from
updating pnfsmanager. Important #1 an updated nfs door will wait
for database changes that updating pnfsmanager will enact.
Important #2 the database changes enacted by pnfsmanager can take
awhile (an hour or so for large dCache instances) as the change adds
an index to the t_inodes
database table.
The “standard” Linux flag to mark an ACE as inherit-only is i
, yet
previous versions of dCache accepted only ‘o’ as the inherit-only
flag. With this release, both o
and i
are accepted. The chimera
shell, pnfsmanager and nfs door should be updated.
dcap
Previously the dcap door provided the dcap client with the wrong errno
(error number) should the client attempt to operate on a nonexisting
file or directory: EIO
was returned instead of ENOENT
. This is
now fixed.
nfs
Information about the filesystem, as reported by the df
command, is
expensive to calculate. Previously, the result was generated when a
client requests the value and the result cached for an hour. This
proved awkward as all clients blocked until the calculation completes.
With this release, updating the cached value is a background activity:
the cached value is returned until the update completes.
Additionally, the cache lifetime is configurable via the
nfs.fs-stat-cache.time
and nfs.fs-stat-cache.time.unit
properties.
pool
The file integrety checking (single file, entire pool one-off check and background checking) produced ambiguous output; for example, not being able to scan a file because it is still being uploaded counted towards an “error” count, but no corresponding log message is included. With this release, the output is less ambiguous and distinguishes between corruption, temporary and more permanent problems.
replicamanager
Added an admin command to query pool-manager for an updated list of pools that are a member of the resilient pool group. This allows adjustments to the set of pools participating in replication without restarting the replicamanager service.
Changelog 2.10.27..2.10.28
- 698724f
- [maven-release-plugin] prepare release 2.10.28
- 1021708
- replicamanager: add an admin command to re-fetch resilient pool group
- f00a040
- acl: fix compatibility with linux ace
- af7b0ad
- chimera: Delete unreferences tag inodes
- 6487ea2
- pool: update scrubber messages to be less ambigous
- a143cfa
- dcap: fileAttributesNotAvailable must set pass ENOENT to the client
- 969eb14
- [maven-release-plugin] prepare for next development iteration
- f94ed28
- chimera: do not maintain time-based cached value of FsStat
- 0f93f35
- chimera: squash usedFiles() and usedSize() into a single query
Release 2.10.27
Changes affecting multiple services
In prior versions of dCache, the path
field that billing logs
for transfers contained the actual transfer path; i.e., for
SRM-initiated uploads this was the auto-generated path from the
TURL and not the user-supplied path from the SURL. This proved
confusing so, with this release, the path
field now has the
user-supplied path (i.e., from the SURL). An additional field
(transferPath
) has the transfer path. While this is
not logged by default, billing configuration (the
billing.text.format.mover-info-message
and similar properties)
may be updated to include it. All doors used by srm clients and
the billing service should be updated.
dcap
This version of dCache uses the connected socket when discovering the IP address of the client for “channel binding”, rather than a constant value. This is important for dual-stack machines (with both IPv4 and IPv6 addresses) that host dcap doors with either gsi- or kerberos-based authentication.
spacemanager
In earlier versions of dCache, if a file is deleted while being uploaded the space-manager will consider the transfer as successful (and so using some of the reservation’s capacity) whereas the pool would delete the file immediately after the upload is complete. Such “leaks” results in a reservation being reported as having more used capacity and less free capacity than it should.
srm
The srm service maintains a counter of the number of requests for each different type. With earlier versions of dCache, if the srm encounters jobs that have timed out while the service was no running, these counters became inaccurate. This is now fixed.
When an srm request times out, dCache may need to take some action to “clean up”; for example, removing the upload directory. If srm is configured to discard all requests on start-up these cleanup operations did not happen. This is fixed with this release. As a consequence, startup times will be longer.
If srm.persistence.enable.store-transient-state
is set
to false
(the default value) then transfer requests
do not survive an srm service restart if the limit on concurrent
transfers prevented handing out the TURLs. With this
release, such requests survive an srm restart.
Previously, if srm.persistence.enable.store-transient-state
is set
to false
(as is the default) then information needed to clean up a job
might not be stored in the database. If, after restarting the srm,
the job times out or is aborted then there is insufficient information
to clean up, resulting in upload directories not being cleaned, pins
(for bring-online requests) not being cleared, copy requests not being
cancelled and lifetime extensions being lost. These problems are
fixed with this release.
Changelog 2.10.25..2.10.27
- a94ac75
- [maven-release-plugin] prepare release 2.10.27
- eeace29
- all: Fix several NPEs when submitting billing messages
- daf84a3
- [maven-release-plugin] prepare for next development iteration
- 12c07e1
- [maven-release-plugin] prepare release 2.10.26
- 7c4c24d
- srm: Fix scheduler counter initialization on restart
- 66f8951
- doors: Log real path in billing
- 475fa58
- srm: Allow the SRM to take action upon cleaned requests
- 4789735
- srm: Force save jobs when adding information needed for cancellation
- bb4acc2
- srm: Force save when job becomes RQUEUED
- d1ee02c
- javatunnel: use connected socket to discover local inet address
- 7512d5c
- [maven-release-plugin] prepare for next development iteration
- c4c8640
- spacemanager: Fix race condition leading to leaked files
Release 2.10.25
Changes affecting multiple services
When dCache suffers a sufficiently high inrush of requests that a cell’s request queue is exhausted, new requests to that cell are rejected. While overall dCache handles this situation correctly (degrading gracefully), the internal message counting and event logging were not updated correctly. This is now fixed.
Fix a bug that triggers a NullPointerException in webadmin. Although the problem is logged in webadmin, the cause is in the doors supplying the information. Therefore, the dcap and ftp doors are to be updated.
When publishing an IPv6 interface, do not include the zone information. Zone information is an extension to IPv6, which appends a ‘%’ plus some opaque identifier to the regular address. Not all clients understand this extension and reject the address as invalid. As zone information is not useful anyway, with this release zones are no longer published. This update is for all doors.
The gplazma, pinmanager pool and srm services can process several dCache-internal messages concurrently. Previously, shutting down these services failed to wait for concurrent activity to finish (within a reasonable deadline), which risked partially committed activity and logging errors when shutting down a busy system. This is now fixed.
Fix possible NullPointerException in httpd and admin services.
In most cases dCache will publish door URLs with a hostname; if the address cannot be resolved then the address is used instead. Previously IPv6 addresses were written incorrectly: without the square brackets. This is now fixed if the info service and the info-provider scripts are updated.
admin
Fix division-by-zero error when the SSH client reports zero width or
height. This happens forcing allocation of a pseudo TTY without
having a real TTY (see the -t
option to OpenSSH client).
pinmanager
Fix pinmanager so it does not log a stack-trace if it failed to fetch fresh pool status information from poolmanager while unpinning a file.
pool
Previously, the metadata reconstruction for files where the upload was
not completed and written without the client specifying a retention
policy would trigger a stack-trace java.lang.IllegalStateException:
Attribute is not defined: RETENTION_POLICY
. This is now fixed.
Sites have reported pools becoming stuck with messages like
UNEXPECTED_STATE_FATAL: Unexpected internal state, unable to
continue.
, which is due to a problem with the Berkeley DB. This
release includes an updated version that should fix these problems.
spacemanager
A number of sites have reported problems with the database resolving deadlocks. While such reports are expected and dCache behaves correctly when they happen, this release should reduce the likelihood of them appearing.
Changelog 2.10.24..2.10.25
- 52b4482
- [maven-release-plugin] prepare release 2.10.25
- 026bd03
- info/info-provider: publish valid IPv6 addresses
- df0a450
- loginbroker: strip off zone off published interface name
- 4aa0d5b
- admin: Avoid division by zero when the client reports a zero sized terminal
- 42a1929
- spacemanager: Optimize space record deletion
- 69aca22
- Update netty and berkeley DB to latest version
- 3a92a16
- dcache: Make cell communication use the correct timeout
- 186166c
- httpd,admin: Fix NPE in transfer collectors
- 955ae65
- doors: Fix race condition that causes NPE in webadmin
- de754a4
- cells: Orderly shutdown of multi-threaded cells
- 1f58c6f
- cells: Fix event queue counting bug
- d869b04
- pinmanager: Don’t log stack trace when unable to fetch pool monitor
- 6f4358b
- pool: Fix meta data reconstruction
- 74f8190
- [maven-release-plugin] prepare for next development iteration
Release 2.10.24
Changes affecting multiple services
Avoid potential message loop when both sender and receipient domains are restarted while the message is in-flight.
admin
pcells distinguishes between a failure to send a message to a cell and that cell taking too long to respond. Support for that distinction was lost with 2.10; with this release, it is restored.
The tab-completion feature of admin parses the help-hint to discover what expansions are available. This has been updated to support more commands.
httpd
Fix compatibility with pcells. This requires a corresponding update to pcells.
Changelog 2.10.23..2.10.24
- 0a220c4
- [maven-release-plugin] prepare release 2.10.24
- 9bab859
- admin: Fix command completion
- 25eff9b
- admin: Propagate NoRouteToCellException to pcells
- ff3541f
- cells: Restore CellExceptionMessage encoding
- dadae07
- httpd: Fix pcells compatibility
- 9f44f34
- [maven-release-plugin] prepare for next development iteration
Release 2.10.23
admin
Restore support for pcells to gracefully handle timeouts.
Restore compatibility for pcells when querying space-manager.
cleaner
Fix bug reported as java.lang.ClassCastException: java.lang.String
cannot be cast to [Ljava.lang.String;
.
pool
Fix pool reconstruction. If, on startup, the pool detects it is in an inconsistent state it will attempt to recover from that. With this release, this should work.
srm
Update SRM default property values. From various reports, it is clear that the current default values are not ideal for many sites. The following properties are adjusted:
srm.limits.jetty-connector.backlog
increased to support bursts of activity; a larger value may be appropriate but not available due to default Linux configuration.srm.request.threads
reduced as processing is asynchronous.srm.request.ls.threads
change default to besrm.request.ls.max-in-progress
asls
requests are blocking.srm.request.max-requests
increased to satisfy user-demand; needed as clients typically request more concurrent TURLs that they make concurrent requests.srm.request.max-transfers
increased to same value assrm.request.max-requests
. This way (by default) dCache never blocks requests pending a client returning a TURL.srm.persistence.remove-expired-period
increased to 10 minutes to reduce stress on the database.srm.service.pnfsmanager.timeout
decrease to two minutes as PnfsManager should respond within that time and clients will likely disconnect if they don’t hear a response within that time.srm.service.spacemanager.timeout
decrease to 30 seconds as the service should respond quickly.srm.protocols.disallowed.get
andsrm.protocols.disallowed.put
now includefile
protocol by default.srm.protocols.loginbroker.timeout
decrease to 20 seconds as this is a very light-weight service.
Support catching and logging some bugs that previously would be silently ignored.
Increase default value for srm.limits.db.queue
to survive database
activity when faced with bulk staging requests.
Fix a ConcurrentModificationException caused when SRM processes two requests close tegether.
Changelog 2.10.22..2.10.23
- 32230d6
- [maven-release-plugin] prepare release 2.10.23
- d817f6f
- admin: Restore pcells compatibility
- 4d51df9
- srm: log more SRM bugs
- b453bb7
- admin: Restore timeout semantics for pcells compatibility
- 25b076e
- cleaner: Fix class cast exception
- 3702d41
- srm: Increase default database queue
- ba0fd18
- srm: Use more sensible default values
- 431e4e0
- srm: Fix ConcurrentModificationException in Axis
- e3b85a2
- pool: Fix pool entry reconstruction
- c082039
- [maven-release-plugin] prepare for next development iteration
Release 2.10.22
Changes affecting multiple services
This release improves the quality of information recorded when logging problems and fixing synchronization of lease time. This affects the nfs door and pools.
pool
When a client reads a file, the pool reads blocks of data from the local filesystem. When reading such a block, the pool could receive fewer bytes than requested. Previously, the pool assumed that this only happens when the end-of-file is reached; however, this is not guaranteed. Should this assumption be violated then the data sent to the client will be corrupt. In practise, the pool’s assumption is true for Linux and local filesystems; however, the code has been updated to remove this theoretical cause of corruption.
Fix a problem where a door can trigger the pool to post-process a file many times. Each trigger starts a new thread, resulting in very large system load.
Fix that, during startup, the pool would fail when recovering a broken file that has access-latency and retention-policy determined by spacemanager.
Extend the nearline SPI so it provides plugins with the file’s path and provides an easier way of report errors.
spacemanager
Improve the logging and handling of transitory errors.
Changelog 2.10.21..2.10.22
- 70fa3b0
- [maven-release-plugin] prepare release 2.10.22
- 2b734fd
- pool: Fix and align pool meta data recovery with current pnfs manager
- 58ca239
- pool: ignore duplicated mover kill requests
- 48b3a47
- spacemanager: Making logging and handling of transient errors more robust
- 0a95f67
- pool: Fix read corruption in HTTP mover
- c484e1a
- pool: Extend nearline storage SPI with path and custom error codes
- ed56243
- libs: update to nfs4j–0.9.5
- 0e8edbd
- [maven-release-plugin] prepare for next development iteration
Release 2.10.21
Changes affecting multiple services
If a bug is found in dCache then it should be logged. For the older admin commands, an uninformative message was logged.
The LocationManager service is started automatically in the broker
domain (typically dCacheDomain
). This answers requests from other
domains with instructions on how to connect to the dCache cluster
(typically a star topology, other domains connecting to the
dCacheDomain
). With this release, the FQDN is sent rather than the
hostname name. The node hosting the broker domain should be updated.
The version of HikariCP has been updated to 2.0.1. This is to correct for a bug in v1.3.9 that triggers logging a stack-trace with:
Internal accounting inconsistency, totalConnections=-1
Previously, if a client attempts to write a file with Access Latency and Retention Policy that conflicts with the selected reservation a stack-trace was logged. This is now fixed: both the FTP door and spacemanager should be updated.
Previously, if a bug was discovered when starting up a service dCache would abort starting up with a non-informative log message. Now, the stack-trace is logged.
nfs
Log abandoned movers with the corresponding stateid.
pnfsmanager
When writing into dCache with SRM, the Access Latency (AL) and Retential Policy (RP) may be specified or omitted. Additionally, the client may specify a space reservation into which the file should be written. If the client specifies both, they must match.
In dCache, there are three mechanisms to support a client that
specifies neither AL/RP nor space reservation: the directory can have
AccessLatency
and RetentialPolicy
tags, the directory can have the
WriteToken
tag, there are a system-wide default AL/RP values.
Previously, dCache would reject uploads where the user-supplied AL/RP information does not match the AccessLatency/RetentialPolicy tags, despite the latter being intended as default values.
With this release, if the client specifies neither space token, AL or
RP then the directory tags will be used. If a directory specifies
both WriteToken
and AccessLatency
/RetentionPolicy
tags, then
these have to be consistent. If the directory conains a WriteToken
tag and the client specifies AL/RP, then the client specified values
have to be consistent with the WriteToken
tag.
pool
When attempting to upgrade non-precious and non-cached files (e.g. a
file marked broken
), the receiving end of the migration module would
answer twice: first (correctly) with a failure and then (incorrectly)
with a success. This is now fixed.
spacemanager
With RDBMs, transactional deadlock rollbacks are normal behaviour, which happen when the database must choose between two conflicting and concurrent changes. Previously, spacemanager would aggressively retry when this happens. This has been observed to trigger performance degradation. This release includes several strategies to minimise the impact of this.
Previously, the default number of space-manager threads was the same as the number of database connections. This does not take into account that there is background activity that also needs access to the database. With this release, the number of threads has been lowered; having a large number of threads also increases the likelihood of seeing transactional deadlock rollbacks.
Fix the shutdown sequence of spacemanager. Previously, shutting down a busy spacemanager could lead to attempts to modify the database after all database connections were closed; such failures were logged.
Reduce logging on various DB errors; generic transient errors are now warnings and transactional deadlock rollbacks are logged at debug level.
Changelog 2.10.20..2.10.21
- 9ce8df8
- [maven-release-plugin] prepare release 2.10.21
- 2f41134
- LocationManager: Use fqdn instead of hostname
- a4b8812
- spacemanager: Controlled shutdown
- 7fc3600
- spacemanager: Make request processing more robust
- 41305ce
- spacemanager: Reduce log level on various transient DB errors
- e05fb97
- spacemanager: Minor simplification to link group updates
- 7d325d8
- spacemanager: Don’t log stack-trace on AL/RP/Reservation conflict
- d5a9928
- spacemanager: Lower default for number of threads
- 40efb76
- 2.10: Upgrade HikariCP to 2.0.1
- 8608a2d
- nfs4: log abandoned movers with WARN
- 5e35b46
- cells: log bugs found by CellShell
- bd53579
- cells: fix how bugs are reported from ac_ command.
- 4c6eb11
- pool: Fix bug in migration module upgrade logic
- 4f1c69b
- pnfsmanager: Fix upload to space token that conflicts with AL and RP tags
- 740b40c
- [maven-release-plugin] prepare for next development iteration
Release 2.10.20
Changes affecting multiple services
Various scripts, including the dcache
command, invoke the java
command with a list of directories in which Java should look for
support libraries. Previously, the current working directory was
(mistakenly) included in that list. This could lead to odd behaviour;
one particular example is running a dcache database
command from the
/etc/dcache
directory. This release fixes this problem by excluding
the current directory.
Normally, if running an admin command in some cell triggers a bug then
the log file of the domain hosting that cell will contain a
stack-trace. Previously, for certain admin commands (spacemanager and
sweeper ls
in the pool) this did not happen. This is fixed with
this release of dCache.
When there is some problem in the communication between domains and error message is logged. Previously the explanation for the problem was logged as “null”. With this release, a more descriptive explanation is provided.
The help text within the chimera CLI and for the pool’s migration commands was badly formatted; this release fixes this.
spacemanager
By default, when logging messages some contextual information is included (in square brackets). This typically includes the cell name and the kind of activity that triggered the log message. Previously, some messages omitted all contextual information, which is fixed with this release of dCache.
Fixed the ls spaces
and ls files
commands so they do not fail if
there is a reservation without an owner.
All space-reservations may have an owner: a username or an VOMS group; ownership may be further restricted by VOMS role. When created through the admin interface, a reservation’s ownership is optional. With previous versions of dCache, if a space-reservation has no owner then anyone can release it. With this release of dCache, a reservation without an owner may only be released through the admin interface.
Changelog 2.10.19..2.10.20
- 1cdaf24
- [maven-release-plugin] prepare release 2.10.20
- 07b031f
- spacemanager: Change release authorization for unowned reservations
- 275bbf8
- Fix valueSpec help parser
- a8ca6ed
- spacemanager: Fix NPE in listing space reservations
- e85713a
- Maintain CDC of threads in decorated thread pool
- 806e8c0
- spacemanager: Log unexpected exceptions
- 8d37d0a
- tunnel: use toString if IOException#getMessage returns null
- 67d047c
- Exclude cwd from classpath
- 9feb238
- [maven-release-plugin] prepare for next development iteration
Release 2.10.19
Changes affecting multiple services
Fix potential database connection leak for nfs door and pnfsmanager
service. This was triggered when attempting to create an already
existing non-zero level via the .(access)
or .(use)
dot commands.
This patch fixes a race condition in Chimera that affects the nfs door and the pnfsmanager service. The effect is that, if two clients attempt to delete the same target (a file, link or directory) at the same time then the nlink count for the parent directory is decreased twice. “At the same time” means within the time taken to process the deletion; this is instance-specific but should be much less than 1 ms for well-configured systems. Sites can repare any incorrect nlinks with the following SQL:
UPDATE t_inodes SET inlink = (
SELECT COUNT(*) FROM t_dirs WHERE t_inodes.ipnfsid = t_dirs.iparent
) WHERE itype = 16384;
This is safe to run on a running production instance, but may take some time and will affect dCache’s responsiveness while running.
gplazma
The description for how to migrate away from using the forbidden
useGPlazmaAuthorizationModule
and useGPlazmaAuthorizationCell
properties had caused confusion. The description has now been updated
to be more explicit.
httpd
Fix filtering boxes and sorting on Pool Admin, Pool Usage, Poolgroups, Space Tokens and Tape Transfer Queue.
pnfsmanager
When a user uploads a file via SRM, a directory is created under the
update directory (/upload
by default). Should this fail, the upload
will fail; however, this was not logged. With this release,
pnfsmanager now logs why such failures happened.
pool
Previously, the NFS mover stopped when the client disconnected. This had two problems: a client that never connects to the pool leaves a mover that never dies and file transfers are not robust against transitory networking problems, despite the client attempting to reconnect. The latter problem is particularly bad when writing data as falling back to the door is not supported and such failures are not handled well by the Linux kernel.
With this release of dCache, if the client is not connected then the mover queries the door every 7.5 minutes to check it is still needed. The mover dies only when the client closes the file, the door declares the client is lost or the nfs door is stopped or restarted.
IMPORTANT Any pool upgraded to this version of dCache that is used for NFS transfers requires all NFS doors to be upgraded to this version or newer.
Currently, if there is a problem while a file is being uploaded using HTTP chunked transfer encoding then dCache will contain an incomplete file with no error mentioned in any log. With this release, such partially uploaded files will still exist (due to the partial upload semantics of pools) but the problem is logged on the pool and with billing.
spacemanager
Add a -blocking
option to the update link groups
admin command to
ensure all subsequent commands will use the updated information.
srm
When uploading files, the client may choose not to specify an access
latency, a retention policy or a space reservation for these files.
Likewise, when reserving space, the client may choose not to specify
an access latency for the reservation. Previously, the detailed view
of the admin ls
command showed null
for these fields if the client
did not specify them. With this release, those fields are omitted if
they have no useful data.
Uploading a file with SRM involves three steps: preparing for the
upload (srmPrepareToPut
), uploading the file, marking the upload
finished (srmPutDone
). The third step can fail but previously the
response from dCache is always Upload failed.
. With this release, a
meaningful error message is returned.
Changelog 2.10.18..2.10.19
- 53b9c39
- [maven-release-plugin] prepare release 2.10.19
- 67f3a8b
- pool: Propagate HTTP mover failures to pool
- 1e29206
- chimera: fix potential transaction leak on error path§
- 7f17d90
- gplazma: update error message for forbidden properties
- 84ad1e2
- spacemanager: add blocking option to update link groups command
- b45ebdd
- srm: don’t list absent information in ls
- 6cc547a
- fix broken merge
- c550606
- pnfsmanager: log problems when creating upload directory
- e03d1f9
- (2.10) webadmin: make jquery selector specific to individual tables
- 38628b4
- (2.10) webadmin: restore missing components to respect jquery script options
- adf3ae1
- srm: include the reason why upload failed
- b74d4a4
- chimera: fix race condition on remove
- 044c33a
- webadmin: ensure unique id attributes for all (currently) tested UI elements
- c6c6e87
- pool: update NFS transfer service to validate inactive movers
- 6e47601
- [maven-release-plugin] prepare for next development iteration
- a5634b0
- libs: use nfs 0.9.x
Release 2.10.18
ftp
Fix default value for ftp.authz.readonly
for plain (unencrypted)
doors. This restores the default value to the dCache v2.6 default
value of true
.
Add the DN of a user in the access log file for “Grid FTP” access.
The response from the plain (unencrypted) ftp door if the user specifies the wrong password is badly formed. Althogh it is possible that some clients are robust against such incorrect responses, with this release the ftp door responds correctly.
nfs
Previously, when enabling the nfs door’s namespace cache, querying information about a freshly created file could provide stale information; for example, querying a recently uploaded file’s length could mistakenly show the file length is zero. With this release, the close after a successful write will invalidate that file’s entry in the nfs door’s cache, allowing it to report up-to-date information.
Prevent leaking memory if a client’s data transfer is proxied (i.e., no use of pNFS) and there is a communication error between the pool and the nfs door.
webdav
Fixes a bug where, if a double-slash is present, all parts of the path
leading up to the double-slash are ignored; for example, with the bug,
a path like /a/b//c/d
is handled as if /c/d
was specified. With
this release, double-slashes are treated like single slashes; the
above example is handled as if /a/b/c/d
was specified.
Fix the NullPointerException triggered if client attempts to upload a file as a child of some existing file.
Changelog 2.10.17..2.10.18
- ef63f5d
- [maven-release-plugin] prepare release 2.10.18
- c928fff
- ftp: fix response if user fails to authenticate to weak FTP door
- 7064cd3
- nfs-proxy: remove proxy adapter on IO errors
- 1d069d7
- ftp: fix invalid default for ftp.authz.readonly property
- 4d54186
- webdav: fix double-slash bug by upgrading to patched milton
- 7c754df
- webdav: fix NullPointerException when PUT as a child of a file
- 7ade469
- ftp: add user to access log
- 48178df
- nfs4: invalidate vfs cache on successful write
- fecbf9c
- nfs4: add NFS file handle into NfsTransfer class
- aaa07fe
- libs: update to nfs4j–0.8.6
- 2f4a5e2
- [maven-release-plugin] prepare for next development iteration
Release 2.10.17
httpd
Fix two minor issues when authenticating with the webadmin interface: “unauthorised access” and being redirected to the home page. The unauthorised access error can occur when selecting “Login” under the bird logo (top right corner); this is now fixed. The redirection problem occurs when selecting a tab that requires administrative privileges while not logged in; this redirects the browser to the login page. Previously, after a successful login, the browser was redirected to the home page. Now the browser is redirected to the selected tab.
info-provider
The previous bug-fix release of dCache included a regression in the
info-provider. The SRM endpoint URL (which starts httpg://
) omitted
the port number. This is now fixed.
In previous releases, the info-provider assumed the broker domain is
dCacheDomain
. This assumption has been removed.
pool
Fix logging that the HTTP mover returned an error so that a generic message (“An unexpected server error has occurred.”) is logged if no more concrete message is available, rather than “null”.
srm
The info
command in the srm
service shows the current number of
requests in each state for each request-type, along with the maximum
allowed. Previously, the total failed to include requests in READY
state.
webdav
A previous bug-fix release fixed how dCache responds when the client
attempts to DELETE
a non-existent file. Unfortunately, this
triggered a different problem where such activity results in a
stack-trace that starts java.lang.ClassCastException:
java.lang.String cannot be cast to javax.security.auth.Subject
. This
second problem is now fixed.
Changelog 2.10.16..2.10.17
- ba2064d
- [maven-release-plugin] prepare release 2.10.17
- 6cfd529
- Revert “info-provider: fix publishing SRM port number”
- fb9e33d
- info-provider: remove dCacheDomain assumption.
- c77a2e2
- pool: Fix logging in HTTP mover
- 9d617f6
- srm: Fix calculation of total number of requests
- 23f966a
- webdav: Alternative to fixing return code of DELETE of absent file
- 6767c95
- (2.10) webadmin: fix login redirect bug
- 73d9d2d
- [maven-release-plugin] prepare for next development iteration
Release 2.10.16
Changes affecting multiple services
Reduce memory usage by avoiding multiple instances of common strings. Although this improves many services, pools are most affected.
Previous dCache releases included some of the code necessary for pcells, even though dCache made no use of this code. As part of on-going consolidation effort, this code has now been removed from regular dCache releases.
The cell.name
is a core property within dCache. Its omission (most
easily, by forgetting to specify pool.name
) prevents the dcache
script for working. The script is now robust against such errors.
If the dcache check-config
command discovers that a deprecated
property is being configured, it looks for the alternative property
that should be used instead. In earlier releases, under certain
circumstances, the wrong alternative property is selected. This is
now fixed.
It is possible that a sudden burst of activity from many clients exceeds dCache capacity to queue such requests. Although dCache is designed to degrade gracefully under such circumstances, there existed the possibility of certain requests becoming stuck or memory leaking. This is now fixed.
nfs
Upgrade to nfs4j v0.8.5. This fixes export file parsing if a host is
mentioned multiple times; note localhost
must now have an explicit
entry in the exports file. This also fixes a deadlock when closing a
file on a busy dCache instance.
Allow an nfs door to avoid being published to loginbroker. Not being published in loginbroker means the door is unavailable to SRM and is not published by info-provider.
pinmanager
If the maximum concurrency (pinmanager.cell.threads.max
) exceeds
the number of database connections
(‘pinmanager.db.connections.max’) then there is a risk of pinmanager
becoming deadlocked under heavy load. This release reduces the
default maximum concurrency to avoid this.
pnfsmanager
Fix NullPointerException when a file is stored in a directory with an empty tag.
pool
Upon reloading the pool configuration an error is produced when an nearline storage was already defined in the existing configuration. This is now fixed.
A recent release fixed a bug that caused
pool.mover.ftp.allow-incoming-connections
to be ignored. Fixing
that bug revealed another that caused the property to have the
opposite effect. This is now fixed.
dCache pool configuration allows passing (fixed) arguments to the HSM script. In earlier releases, these arguments were supplied to the script in an arbitrary order. While HSM scripts should not depend on the order, they might; so, with this release, the order is preserved.
webdav
Although dCache behaves correctly if the client interrupts a proxied transfer; however, this is logged as a bug. This is now fixed.
Make HTTP third-party copy feedback and detecting when client disconnects more robust against internal dCache problems.
Changelog 2.10.14..2.10.16
- e747823
- [maven-release-plugin] prepare release 2.10.16
- 0f760f6
- info-provider: fix publishing SRM port number
- 6d67a58
- [maven-release-plugin] prepare for next development iteration
- 7306099
- [maven-release-plugin] prepare release 2.10.15
- dcf8d77
- webdav: don’t log a stack-track when proxy transfer is interrupted
- 1144520
- fix broken commit
- bee43f4
- webdav: make progress markers more robust
- 4932840
- Fix compilation with Java 7
- 2b5cdda
- cells: Fix message timeout in case of thread pool overflows
- 5516165
- chimera: handle NULL field of directory tags
- 6277cd0
- pinmanager: Adjust pinmanager defaults
- e766e8e
- Internalize common strings in StorageInfo
- 634bbe1
- bootloader: Don’t fail on missing cell name
- 0ebd0af
- check-config: fix warning for deprecated properties
- 6ef87e1
- pool: Fix regression preventing configuration to be reloaded
- 2b27ad6
- (2.10) gitignore additions for IntelliJ
- 6fd254d
- libs: update to nfs4j–0.8.5
- 7e52c73
- pool: Fix regression causing FTP movers to default to proxy mode
- 477f0c8
- nfs: Allow not to publish to loginbroker
- cd8b59e
- pool: Preserve HSM option ordering
- bbd5570
- [maven-release-plugin] prepare for next development iteration
- bf26db9
- cells: Delete pcells related classes
Release 2.10.14
Changes affecting multiple services
Although dCache system configuration property names do not contain
spaces, it is possible to define such properties. Previously, doing
so breaks the dcache
command. This is now fixed.
Update for the nfs door and the pools that allows faster restart. The door is also updated to be more responsive to namespace operations when suffering large number of proxy reads.
If the number of concurrent connections is set to -1
then the ftp
and dcap doors will leak memory, eventually triggering an
out-of-memory error that will restart the domain. This is now fixed.
admin
Fix erronous reporting of bugs when the user supplies incorrect arguments to an admin command.
dcap
Fix the permissions check for bring-online with plain dcap and a NFSv3 mounted dCache.
Fix regression against v2.6 and earlier dCache in how gsidcap doors are known to SRM and how they are published in BDII/GLUE.
ftp
The ftp door supports the HELP
command as some clients use this
commands output to discover if certain optional functionality is
available. This release fixes the output from this command.
The ftp door’s access log records some responses without the
corresponding command-line from the client; subsequent responses (if
the response is multiple-lined) and all responses due to STOR
and
RETR
commands were affected. This is now fixed.
nfs
The info
command now lists the proxy adaptors. A proxy adaptor is
created when the client rejects the pool and falls back to reading
data from the door.
Improve shutdown of nfs door when the embedded portmapper is used.
pool
Improve which interface the ftp mover selects when the client is redirected to the pool. In particular, the door will only redirect the client if the IP protocol matches (IPv4 vs IPv6); for example, if a client connects with IPv6 and the pool has no IPv6 address then the door will now proxy the data connection.
The pool.mover.ftp.allow-incoming-connections
property had no
effect. This is now fixed.
Add the IoMode
(whether the mover is accepting new data, or
supplying existing data) to the status line describing the mover.
Improve shutdown of nearline storage subsystem; now dCache will
attempt to cancel all ongoing activity before shutting down the pool.
The previous asynchronous approach could lead to
IllegalStateException
being logged if the pool was busy at that
time.
spacemanager
Adjust spacemanager schema migration to be aware of earlier dCache bugs and to work-around site-local indexes that clash with new indexes that dCache needs.
In previous versions of dCache, both spacemanager and pnfsmanager had default access-latency and retention-policy settings. The spacemanager default retention-policy is only used if the admin doesn’t specify a value when creating a reservation through the admin interface. The spacemanager default access-latency is also used as a default for the admin command; additionally, it is used when processing an SRM reserveSpace request that omits the (optional) access-latency information — the retention-policy is mandatory in such SRM requests.
With this release, the spacemanager no longer has default
access-latency or retention-policy properties. Creating a
space-reservation in the admin interface now requires specifying both
the access-latency and retention-policy. When processing an SRM
reserve space request without an access-latency, the spacemanager will
check from which linkgroups the user is authorised to reserve space.
It will choose an access-latency based on the request’s
retention-policy. If the user is authorised to reserve space from
link-groups such that both access-latency options (ONLINE
and
NEARLINE
) are possible, space-manager will prefer to reserve
ONLINE
if the retention-policy is REPLICA
and NEARLINE
if the
retention-policy is CUSTODIAL
.
srm
Previously, if the SRM client requests listing a directory, specifies a non-zero offset and does not limit the response size then dCache would fail this request with an IllegalArgumentException. This is now fixed.
webdav
With the recent upgrade of the Milton library some new behaviour was
introduced. One example is that, under certain circumstances (and to
support certain clients) the Milton library returns a 401
(not
authorised) when attempting to delete a non-existing file.
Unfortunately, this change then broke ATLAS clients. This patch
updates dCache so it returns 404
(not found) under these
circumstances.
Fix the response when a client requests a byte-range beyond the end of a file. This is necessary for compatibly with ARC clients.
Changelog 2.10.13 to 2.10.14
- e35c5ab
- [maven-release-plugin] prepare release 2.10.14
- 37af85b
- webdav: Fix reported content length for partial GETs
- 87eb009
- ftp: ensure client command line is available in access log
- 70afdaa
- webdav: Fix return code on DELETE of absent file
- 409228b
- srm: fix semi-infinite ls range with non-zero offset
- ab86627
- pool: Partially fix interface selection for FTP mover
- bc3e762
- pool: Fix typo that breaks pool.mover.ftp.allow-incoming-connections
- 7b19902
- shell: fix shell oracle for configuration keys with a space
- 23715ac
- nfs: show current proxy-io transfers in the door
- cfaa0e9
- chimera-cli: fix chown
- 7540735
- dcap: Fix regression in published protocol family
- 0aac485
- ftp: fix help output
- 5175403
- loginmanager: Fix leak caused by absent child limit
- f540180
- pool: Fix shutdown of nearline storage subsystem
- 83608c6
- spacemanager: Fix null constraints and other schema migration issues
- fe1da05
- dcap: check for url in some commands
- 3901c80
- admin: Fix error reporting
- 553fdce
- pool: add Io mode into mover’s status line
- a275f1e
- spacemanager: Get rid of access latency and retention policy defaults
- b079809
- nfs: stop embedded portmap on shutdown
- 0063089
- libs: update to nfs4j–0.8.4
- 0312ac0
- [maven-release-plugin] prepare for next development iteration
- 907f1da
- dcap: refactor PnfsSessionHandler to unify permission check and url handling
Release 2.10.13
ftp
Relax requirements for the EPSV and EPRT commands. Previous versions of dCache would reject all RFC 2428 commands from an IPv4 client. This has been relaxed with this release. Now, IPv4 clients can use EPRT. IPv4 clients can also use EPSV if delayed passive is enabled.
info
Info service includes a safety feature that prevents it from bombarding the rest of dCache with too many messages. With this release, the safety limit is reduced allowing info to send messages more often.
Reduce the delay between subsequent messages of the same type. This allows the info service to recover information more quickly after being restarted.
pool
The NFS specification allows the server to specify multiple addresses when telling the client where to connect; for example, specifying both an IPv4 and an IPv6 address, or both addresses for multi-homed machines. This requires the client to choose the appropriate interface. For Scientific Linux 6, the kernel client will always use the first supplied address in the list and fail if it cannot access the pool with that address. With this release, pools will order the list, using heuristics to select which IPv4 address is “correct” and list it first.
Prior to flushing a file to tape, the pool checks if the name-space entry still exists and deletes the replica (rather than flushing) if not. Version 2.10.7 introduced a regression where, if this happens, there is a IllegalStateException stack-trace. This is fixed with this release.
poolmanager
With previous dCache versions, the WAAS selection algorithm had a bug where it could (mistakenly) consider pools full if all pools had very fresh files. This is fixed with this release.
spacemanager
Fix various issues to improve the robustness of spacemanager: handle failed uploads correctly, handle files deleted during upload correctly, increase robustness against dCache loosing (internal) messages.
webdav
Upgrade to Milton v2.6. This fixes the buffering problem where a
proxied vector read request results in the entire file being written
to a tmp
directory and not deleted. With this release, requests for
100 kiB or less data result in no data being written to disk; requests
for more than 100 kiB are still written to disk, but only the data
needed to satisfy the request is stored and the file is deleted once
the response has been sent. Some issues persist: data isn’t deleted
if there is a failure sending it to the client and the whole file is
requested from the pool.
Changelog 2.10.12 to 2.10.13
- 04efb0b
- [maven-release-plugin] prepare release 2.10.13
- 4431ddf
- ftp: Relax requirements for EPSV and EPRT
- 1f7afdb
- poolmanager: Fix full pool detection for WASS
- 2e1ecd5
- ftp: Relax requirements for EPSV and EPRT
- 27783b3
- Upgrade to Milton 2.6
- b74db9d
- info: Reduce safety limit to 50 ms
- e4fe223
- pool: reorder ip addresses returned to NFS client
- a157ef8
- pool: Fix ISE when flushing deleted file
- 53bbe92
- info: Reduce delay between messages
- c509fbc
- spacemanager: Fix various error recovery scenarios
- 39dadfa
- [maven-release-plugin] prepare for next development iteration
Release 2.10.12
Changes affecting multiple services
Restarting a domain within an active dCache instance can lead to a
domain receiving messages for a cell as it is starting. Strict
control is require to avoid the cell attempting to process messages
before it is ready. This release fixes one place where this control
was missed, which could lead to a NullPointerException
. While this
problem can affect any core dCache service, it was noticed with the
spacemanager service.
httpd
dCache versions including and after 2.11.0, 2.10.9, 2.9.12, 2.8.16,
2.7.21 and 2.6.36 required sites to delete existing RRD files when
upgrading; i.e., run the command rm -f /var/lib/dcache/plots/*.rrd
when the domain hosting the httpd
service is stopped. This release
reverts that change, but requires sites that have already upgraded to
repeat the rm
command. Sites upgrading from an earlier dCache
version do not need to delete anything.
Fix a potential NullPointerException
in poolCollector
.
nfs
Protect against a NullPointerException
if the client attempts to
read the contents of a file’s level where that level exists in the
database but contains a Nil value. This does not happen under normal
circumstances.
pool
Previously, any problems found when the xrootd client is writing to a pool were silently ignored: neither the pool, the client, or billing appreciated there was a problem. With this release, problems are logged in billing, reported back to the client. The client is also disconnected.
The Berkeley DB, which may be used to store file metadata on the pool, does not like being interrupted. The pool tries hard to avoid interrupting reading or writing; this release fixes one place that slipped through.
Fix regression in the output from the info command: it did not include statistics about the number of HSM requests and HSM timeouts. Also fixes how active HSM jobs are counted: cancelled jobs are still active until the underlying job as ended.
Improve the error message reported back to the xrootd client should its request have an invalid or missing UUID.
Changelog 2.10.11 to 2.10.12
- 9de625c
- [maven-release-plugin] prepare release 2.10.12
- 4af8b23
- imera: protect against NPE in FsSqlDriver#read
- df5bab0
- xrootd: Propagate mover errors to dCache
- fb61ae4
- pool: Restore flush and stage stats in info
- 3f9044d
- pool: Improve xrootd error message on missing UUID
- 6762e04
- Fix NPE in cell initialization
- 46940a6
- Fix NPE in httpd service
- 036e80f
- pool: Avoid interrupting Berkeley DB in migration module server
- 8257601
- dcache-webadmin: revert rrd data source names
- 9c9bc38
- [maven-release-plugin] prepare for next development iteration
Release 2.10.11
Changes affecting multiple services
dCache will discard queued internal messages where the sender is no longer expecting a reply (due to internal time out); this helps an overloaded system recover quickly. Such discarded messages are logged. This release provides better logging when this happens: string commands are logged correctly and both the time-to-live and the age have been added.
The xrootd door and, on the pool, the HTTP and xrootd movers logging included incorrect context information, always showing the first connection. This has been fixed.
alarms
Fix minor corruption at beginning of the alarms defaults file.
httpd
In previous versions of dCache, the webadmin war file is automatically
unpacked. This has been problematic as a dCache upgrade did not
always trigger updating the unpacked webadmin, resulting in dCache
running the older webadmin. With this release, dCache no longer
unpacks the war file; the webadminWarunpackdir
and
httpd.container.webapps.tmp-dir
properties are now obsolete.
Changelog from 2.10.10 to 2.10.11
- 8448a04
- [maven-release-plugin] prepare release 2.10.11
- 3cbd3f8
- cells: Refine TTL discard message
- f436cad
- xrootd,pool: Fix xrootd and http logging context
- 5ca7b4b
- alarms.properties fix accidental header corruption
- 6f37b1c
- [maven-release-plugin] prepare for next development iteration
- cca38b7
- (2.10) dcache-webadmin: change Jetty setting so .war is not unpacked
Release 2.10.10
Changes affecting multiple services
Doors report to billing when a file is deleted. Previously, many doors neglect to include the PNFS-ID and sent only the path. This can be ambiguous so, with this release, all doors send the PNFS-ID in addition to the path. Some doors also failed to send the file size (if known) and client IP address. These, too, have been fixed. In summary: * dcap additionally sends: PNFS-ID, file-size and client address. * ftp additionally sends: PNFS-ID * srm additionally sends: client address. * webdav additionally sends: PNFS-ID and file size * xrootd additionally sends: PNFS-ID.
A bug was discovered that resulted in the nlink count becoming negative. While wrong, this had a knock-on effect that prevented pnfsmanager and nfs services from listing the affected directory. With this release, directory listing are robust against such problems.
alarms
Fix alarm definition that triggers when the pool discovers a file with the wrong checksum. Previously it mistakenly triggered if an upload was incomplete.
dcap
Create a session identifier for the duration of a client connecting. This session identifier is available in log messages and in the billing entries.
gplazma
Update ldap plugin to support both RFC 2307 and RFC 2307bis LDAP schema types. Additionally, the memory footprint of this plugin has been reduced and its performance improved.
nfs
Update nfs4j to v0.9.2. This brings some small performance benefits when listing a directory and for reading data through NFS v3.
Fix bug where, when a client read or write activity is proxied, the corresponding dcap mover wasn’t removed if client didn’t wait for the queued mover to start. Previously, such movers would accumulate on the pool.
pnfsmanager
Prevent attempts to remove the .
and ..
directories.
This release improves the shutdown procedure. It’s now faster and does not generate error messages if shutdown while dCache is active.
spacemanager
When uploading a file outside a space reservation (i.e., using implicit space reservation) on a loaded dCache instance, the spacemanager may report a database deadlock. These messages do not indicate a problem: dCache behaves correctly and the upload is not affected; however, there is an impact on performance. With this release, there should be no further deadlock messages and spacemanager should place less load on the database, with corresponding increase in performance.
webadmin
Display the command’s output in the cell admin page to use a monospace font.
Remove a javascript error due to missing clojure dependency.
Sending commands in cell admin page is fixed.
webdav
Fix webdav door so it no longer reports a stack-trace when the client
issues a USERINFO
request. The davfs2 client, in particular, issues
these requests.
xrootd
Add a logging plugin. This allows an admin to investigate problems between an xrootd client and either the xrootd door or the pool. It is not enabled by default: sites must explicitly include the plugin in their configuration to achieve additional logging.
Create a session identifier for the duration of a client connecting. This session identifier is available in log messages and in the billing entries.
Changelog from 2.10.9 to 2.10.10
- 4c162a3
- [maven-release-plugin] prepare release 2.10.10
- 592236f
- doors: Add pnfs ID to billing remove entries
- d4d868e
- webadmin: Fix ClassCastException in cell admin
- bec30a9
- alarms: fix regex for checksum alarm
- 1ef0e2a
- Fix compilation for Java 7
- d741823
- webadmin: remove redundant head element in alarms panel html
- a46f2f1
- dcache-webadmin: change output field of cell admin page to monospace font
- 7274f12
- (2.11) dcache-webadmin: eliminate clojure dependency
- 525962e
- libs: update to nfs4j–0.8.3
- 115cfe1
- dcache-xrootd: Extend and improve xrootd access log plugin
- c5c4775
- gplazma-ldap: pull only required attribute
- 8e55ee2
- gplazma-ldap: support uniqueMember based group membership query
- 6c51061
- nfs-proxy: kill mover on close
- 5d7820b
- chimera: protect list initialization from FS inconsistencies
- 8757c04
- webadmin: tidy up unavailable page slightly
- 320324b
- xrootd: Add access-log plugin
- 3e58f42
- xrootd: Fix compilation of backported session initialization
- d176e4c
- fix WebDAV door logging stacktrace on USERINFO request
- 15ec5c8
- dcap: Initialize CDC session
- a6215e6
- xrootd: Initialize session
- 3dbecd4
- chimera: prevent attempts to remove ‘.’ and ‘..’
- 077c6d8
- spacemanager: Avoid a transaction deadlock and reduce DB overhead
- 0b54294
- pnfsmanager: Controlled shutdown of processing threads
- fe21ad9
- srm-client: Report file level errors
- 393fad1
- [maven-release-plugin] prepare for next development iteration
Release 2.10.9
Fixes affecting multiple services
Update dcache check-config
to print an error if site-local
configuration contains scoped properties: one that contain a ‘/’
character.
Fix that, under certain circumstances, the FTP and dcap doors could log error messages against the wrong cell. Also makes shutting down the doors a more orderly process.
billing
Make shutdown quieter and faster.
httpd
Fix pool queue plots in webadmin.
nfs
Performance improvements.
pnfsmanager
Fix the message reported when moving (or renaming) a file or directory and certain problems were found: the destination directory isn’t a directory, the source doesn’t exist, or overwriting with different types (e.g., overwriting a directory with a file).
Fail requests earlier that plan to upload a file if that file’s path already exists as a directory. Previously, dCache would fail such uploads after the file was upload (during srmPutDone). With this release, dCache will fail during srmPrepareToPut instead.
poolmanager
If a site using WAAS specifies too large a Space Cost Factor then the algorithm used for write pool selection breaks down, logging “Unreachable statement.” Starting with this release, a warning is logged indicating the cause of the problem.
Made shutting down poolmanager more orderly and faster.
srm
The srm.enable.space-reservation.implicit
property is now marked
forbidden. This means a dCache instance with this property configured
will not start. Consider the
spacemanager.enable.unreserved-uploads-to-linkgroups
property
instead.
webdav
The third-party copy implementation supports the Credential HTTP header. If an incorrect value is specified then the error returned to the user provides incorrect list of valid values. This is now fixed.
Changelog from 2.10.8 to 2.10.9
- bd39c7a
- [maven-release-plugin] prepare release 2.10.9
- cdb2028
- webdav: correct error message for incorrect Credential header value
- 6cf3a1d
- srm-client: delegate by default for https 3rd-party transfers
- 8c061b0
- chimera: Add messages to JdbcFs#move exceptions
- f37fb2a
- poolmanager: Warn when spacecostfactor is too big
- e7589b2
- poolmanager: Warn when spacecostfactor is too big
- ac6cd39
- pnfsmanager: Check file type before overwrite
- cf33a8b
- srm-client: Report srmPutDone failure
- d07c1a7
- check-config: Produce error when using scoped properties
- ec26c94
- srm: Mark srm.enable.space-reservation.implicit forbidden
- 9a8fa95
- loginmanager: Associate ScheduledExecutor with the cell’s thread group
- c59f668
- pom: update to nfs4j–0.8.2
- 0cf0d3f
- (2.10) pool queue plots - fix partial package refactoring and neglected removal of jndi arguments
- 0f82d09
- billing: Improve shutdown for db backend
- 80186eb
- poolmanager: Shut down request container thread pool during shutdown
- 7bf0578
- [maven-release-plugin] prepare for next development iteration
Release 2.10.8
Fixes affecting multiple services
Running the eval
admin command so it returns a non-zero return-code
would respond indicating that a bug had been found. This has been
fixed.
An named environment may be executed and can contain a reference to itself. If this recursion is unchecked, eventually memory is exhausted and the domain would restart. Now an error is logged and the domain is not killed.
The RPM package now includes the file /etc/logrotate.d/dcache
. This
is a default configuration for the logrotate application so that
domain logs are rotated weekly, with compression, retaining one year
history.
Previously, some exceptions were logged with the wrong cell context or without any context. This has been fixed.
Relative upload directories
This release adds support for upload directories that are relative to
the user’s root directory. If the dcache.user-directory
configuration property is defined as a relative path, it is now
interpreted relative to the account’s root directory.
Some sites use account-specific roots to guarantee isolation of users from each other. Previously, such sites had to run additional doors to allow SRM uploads. With this release, sites now have an alternative: to configure dCache so that uploads are make into user-specific upload directories.
There are both benefits and disadvantages to both approaches
(deploying extra doors, user-specific upload directories). These are
documented in the dcache.properties
default file (see the
dcache.upload-directory
property). Sites are encouraged to read
this carefully before configuring a relative upload directory.
Sites that do not use user-specific roots remain unaffected by this issue.
All doors and pnfsmanager needs to be updated for the change to be effective. If the feature is not used, then there is no limitation on upgrades.
admin
Under certain circumstances the admin interface would hide useful information about a bug that some admin command had triggered. That information is made available.
billing
Fix indexer’s inability to parse billing PoolHitInfoMessage events.
Improve error message if the indexer cannot parse a date argument.
Billing records issued by doors never included the file’s size. This is now fixed for all doors except NFS.
The indexer includes a work-around for the lack of 4-digit year in the billing records. Unfortunately, this work-around was broken and could lead to badly formatted dates being output. This is now fixed.
The indexer will skip unrecognized records. When the output format is YAML the raw billing line is included as a comment.
Fixes an issue where index would output incorrect JSON output for multiple days.
nfs
Improve client cache consistency when creating files at rates of 1 kHz or greater.
Fix reported rdev
value.
pool
Fix internal copying of files (triggered by the SRM copy
command) so
they respect the LAN port range.
Fix migration module’s random pool selection so that it does not select pools that are full.
Fix NullPointerException triggered when using migration module’s proportional pool selection and all pools are full.
Make cancelling a migration task more robust against bugs: a cancelled task timer is started first so, if there is a problem, the task will timeout.
The rep ls
command can calculate per-storage-class statistics. This
release fixes a problem where files without a storage-info would
trigger an IllegalStateException when calculating these statistics.
Fix sweeping of files without a storage-info. Previously an IllegalStateException was thrown; this could be triggered by the admin interface or by the sweeper.
Fix a problem where the pool would register free space before actually deleting the file from the file system; for a brief moment, the pool would appear to have more free space that is actually available.
Periodically, pools check that dCache’s internal accounting of total
and free capacity does not exceed the OS supplied values for the
partition (i.e., the output from the df
command); if they do then
the dCache internal accounting is adjusted to match. The dCache.org
team have observed that, after a file is deleted, some seconds may
elapse before the corresponding extra free capacity is reported by the
OS. To protect pools from this effect, the total and free capacity
check is suppressed for 60 seconds after a file was deleted.
spacemanager
If a transient deadlock resolution error occurs, unconditionally retry the operation and do not log it. Such errors are part of RDBMS design when resolving concurrent updates.
srm
Fix the srm service’s admin interface ls
command so that listed jobs
can optionally include only those that have failed, have completed or
were cancelled.
webadmin
Webadmin periodically creates billing graphs by querying the billing
service for the information it needs. Previously, if this query
failed (e.g., the billing service was restarted or failed to start
first) then no further billing graphs are generated and a domain
restart is needed. As a result of this work, the properties
poolqplots.refresh-interval
, poolqplots.refresh-interval-unit
,
httpd.plots.pool-queue.refresh
and
httpd.plots.pool-queue.refresh.unit
are no longer supported. The
httpd.plots.pool-queue.min-time-step
and
httpd.plots.pool-queue.min-time-step.unit
properties now also cover
this configuration.
Changelog from 2.10.7 to 2.10.8
- 4ef77da
- [maven-release-plugin] prepare release 2.10.8
- e100b34
- chimera: fix unit tests
- 97fa65a
- pool: Fix pool size health check in case of asynchronous release of space
- a11e7c6
- spacemanager: Suppress transient deadlock resolution errors
- 7710255
- pool: Fix race leading to false positices in pool size health checks
- 5673c4b
- pool: Fix ISE in CacheEntryImpl#toString
- d5851e8
- pool: Fix ISE in ‘rep ls’
- c6280fa
- Made PoolQueuePlotData enum java 8 compatible.
- e25597b
- (2.10) webadmin: fix exit login in billing refresh loop
- aa873e3
- Removed comments.
- 0efdd4d
- Marked refresh properties as obsolete.
- f665f17
- pool: Make migration task cancellation more robust
- 787be04
- cell: Log exceptions within the correct cell context
- d85629d
- pool: Fix pool selection bugs in migration module
- cf8768a
- admin: Propopage NPE to user
- 86ca482
- chimera: query for inode generation on readdir
- c65b7fe
- billing: Fix date handling for json and yaml output
- 261183b
- billing: Add file size to request records
- 61002c8
- billing: Improve error message on failur to parse date arguments
- c0392c8
- billing: Add billing format for PoolHitInfoMessage
- 1eb71fe
- (2.10) webadmin: minor improvements to rrd4j-based pool-queue plots
- 403a709
- pnfsmanager,doors: Add user relative upload directories
- 48b8e7b
- rpm: include logrotate config file
- 921e569
- srm-client: Output file level errors for sync requests
- c7dd7c0
- pool: Respect LAN port range for internal srmcp transfers
- 3d51f29
- srm: Fix listing of failed, done and cancelled jobs
- 950f726
- cell: Prevent interpreter stack overflow from killing the domain
- 42d2907
- cell: Declassify eval failure as a bug
- 73ee6d8
- [maven-release-plugin] prepare for next development iteration
Release 2.10.7
Fixes affecting multiple services
The nfs and srm both cache replies from gPlazma to increase the speed of authenticating and identifying the user. Previously transient errors when communicating with gPlazma were also cached, delaying the recovery time from such errors. Now, such errors are not cached.
Fix a problem where show pinboard
command can trigger the error:
java.lang.IllegalArgumentException: number to skip cannot be
negative
.
Update the comments within the configuration property files to reflect the new configuration property names.
Fix OpenMQ communications to use new configuration property names.
dcap
Fix opening a file when space-manager enabled.
ftp
Fix SRM uploads when the ftp door has a static root (the ftp.root
property) and this static root is not an ancestor (i.e., a prefix) of
the user’s root.
nfs
No longer trigger an NullPointerException when opening the
.(parent)()
dot-command file in Chimera’s root directory.
The dcache ports
command now shows the NFS port based on the new
configuration properties.
pool
Fix a race-condition when a user starts uploading a file that should go to tape, deletes the file before the write is completed, then cancels the file upload. Previously, there was a risk that the pool fails to delete the uploaded data.
Fix Berkley DB usage when an HSM operation is cancelled to avoid an error logged as:
InterruptedException may cause incorrect internal state, unable to
continue. Environment is invalid and must be closed.
Fix xrootd support for vector read. The problem was discovered with
ROOT v6, which reports errors like Single readv transfer is too
large
.
Fix the dcap mover to use the new pool.mover.dcap.port
configuration
property.
httpd
Fix the statistics location to use statistics.location
configuration
property.
srm
Fix how durations are reported when reporting information about jobs
(e.g., the ls
admin command).
Fix cleaning of upload directories for expired uploads after service restart.
xrootd
Fix problems when calculating a response to the kXR_set request. If encountered, the problem results in the following being logged, where ‘nnn’ in ‘xrootd-disk-nnn’ is some integer number:
Uncaught exception in thread xrootd-disk-nnn java.lang.IllegalStateException: null
Changelog from 2.10.4 to 2.10.7
- be5d6ea
- [maven-release-plugin] prepare release 2.10.7
- 45afb2a
- ftp: Fix root path validation for upload directory
- 9912145
- [maven-release-plugin] prepare for next development iteration
- b28e2f7
- [maven-release-plugin] prepare release 2.10.6
- 9d1b4a8
- xrootd: Upgrade to xrootd4j 1.3.5
- d2185fa
- [maven-release-plugin] prepare for next development iteration
- 66b5982
- [maven-release-plugin] prepare release 2.10.5
- e947426
- pool: Prevent interruption of replica deletion during HSM flush
- 5c3f6b6
- pool: Fix HSM cancellation
- 6bfdfa8
- The gPlazma cache used by NFS and SRM caches both positive and negative replies. Unfortunately it also caches failures to communicate with gPlazma. This means that a transient timeout would be cached too, thus increasing the effect of the transient error.
- 6ddd40b
- cells: fix IAE in ‘show pinboard’
- c9c84ce
- commons: fix TimeUtils nanos to duration string conversion
- 42e2a90
- pool: Fix xrootd vector read limits
- 0e216e4
- chimera: fix NPE on ‘.(parent)()’ for root inode
- f5852f9
- dcap: fix interaction with Spacemanager
- d6a4e2a
- srm-client: Fix logback configuration
- af8e546
- srm: Fix asynchroneous job storage leak
- e513209
- configuration: adjust references to deprecated properties
- 235237b
- unittests: increase timeouts and fix race in DiskSpaceAllocatorTest
- 8d749c0
- [maven-release-plugin] prepare for next development iteration
Release 2.10.4
Fixes affecting multiple services
In some cases, dCache would log the wrong cause of an internal failure. This has been fixed.
The pinboard has been improved in two ways:
- The timestamp now includes AM/PM
- The memory and CPU usage has been decreased.
Two problems are fixed that affected those services that make use of the grid trust store (/etc/grid-security/certificates): gplazma, srm, gsiftp, webdav and xrootd:
When a CA is found not to have a signing-policy file, subsequent attempts to discover a signing-policy file will automatically fail for one hour. Previously, if any CA was removed since the service started then certain CAs (those using UTF–8 encoded subject RDNs) would be placed in this not-found cache by mistake. This resulted users with certificates from such CAs would succeed the first time they used dCache but all subsequent attempts would fail until the cache expires or the trust-store is refreshed.
A minor memory leak is fixed. The amount of memory leaked was proportial to the number of different CAs in use.
Two problems are fixed that affected services providing GSI or plain SSL authentication using the jGlobus library:
dCache now supports TLS v1.2 clients when using GSI or plain SSL using the jGlobus library.
Signal end-of-stream when remote sends a CLOSE notification. The remote party normally sends a CLOSE notification before terminating the TCP connection, typically send from server to client. Previously, receiving such a notification was treated as an error.
Fixed thread safetly of timestamp formatting, principally used in logging and generating the output of admin commands.
billing
Fixes a bug where the bytes read graph can show inflated values.
httpd
Update service to use httpd.service.loginbroker
configuration
property rather than the generic dcache.service.loginbroker
.
Introduce the httpd.setup
configuration property to expose and allow
configuration of the existing mechanism to support site-specific
customisation of the httpd service batch file.
nfs
Fix permission check when client requests a list of some directory’s contents.
Add support for mounting when the server-side mount-point is a sym-link.
srm
Fix the lifetime of the response when client makes a successful request for a reservation with infinite lifetime.
Update the info
admin command to produce scheduler information that
is easier to understand and better reflects the new scheduling.
Fix incomplete restore of “ready queued” jobs. After restarting the srm service, any jobs in this state RQUEUED would not be handled correctly. This has been fixed.
Allow credentials to be reloaded; in previous versions of 2.10 this was ineffective and resulted in a null-pointer exception being logged instead.
Allow srm to start if there were active bring-online requests when the service was stopped.
Fix the set max ready get
command in the admin interface.
Fix a problem where, if a client aborts an upload shortly after initiating the upload then the server blocks all activity for a timeout period. The SRM will eventually recover but, during this period, it appears deadlocked.
The SRM checks whether there is a door that statisfieds a users upload request, taking into account the door’s root directory. It could be that doors are configured so that the upload directory is not visible, preventing any uploads. This problem is logged so it is more obvious.
chimera
Speed up deletes by Chimera no longer explicitly removing non-zero levels when a file is deleted. The database handles this automatically.
The chimera shell failed to interpret the creation time flag when choosing which time should be displayed in directory listing.
Changelog from 2.10.3 to 2.10.4
- 5169a28
- [maven-release-plugin] prepare release 2.10.4
- 72dedf9
- srm: Log common misconfiguration preventing upload
- 4fa394e
- common: Fix race condition in AtomicCounter unit test
- eca3c50
- srm: Fix deadlock like bug
- a1f3d96
- billing plots: use ‘transferred’ rather than ‘size’ for bytes
- dc67a13
- httpd: Use service specific property for list of login brokers
- 8595102
- srm: Fix reporting of infinite space lifetime
- 06d205a
- chimera: Fix creation time flag in ls command
- a65d0e5
- Fix CacheException references in error messages
- adbae8e
- srm: Align SRM info output with recent changes to scheduler configuration
- 0d0bbd6
- chimera: do not explicitly remove enties in t_level_x
- 95165df
- cells: Improve pinboard
- 7085fa2
- libs: update jglobus to 2.0.6-rc8.d
- a75a70d
- srm: Fix incomplete restore of “ready queued” jobs
- 65de44e
- Avoid thread-unsafe use of SimpleDateFormatter
- 916b322
- srm: Fix credential store registration
- a83e308
- srm: Fix restore of bring-online requests on restart
- c6c3e8f
- srm: Fix set max ready get command
- e6245c9
- libs: update to jglobus–2.0.6-rc7.d
- 8b3c169
- libs: update to nfs4j–0.8.1
- 04d1b1e
- [maven-release-plugin] prepare for next development iteration
- 6d94f34
- build: Update to findbugs 3
Release 2.10.3
Utilities
Add support for an alternative ACL format for the setfacl
command in
the chimera command. Now, an ACL may be specified either like
USER:7:rlwfx:o:ALLOW:FFFF
or like USER:7:i+rlwfx:o:FFFF
.
pnfsmanager
Previously, the automatic schema management for Chimera failed to
create an index on the iparent
column of t_dirs
table. The lack
of this index results in deletions becoming progressively slower as
dCache stores more files. If your dCache instance was created with
dCache v2.3.0 or newer then it is likely affected by this problem.
Instances created with an earlier version are not affected. The
pnfsmanager service will check for the index when it starts and add
the index if it is missing. Alternatively, the check may be performed
manually using the dcache database update
command.
pool
This release fixes several issues with the sweeper:
Previously, the sweeper consumed considerable amount of memory while freeing up space; in the worse case, this could result in the domain running out of memory and restarting; this has been fixed.
The
sweeper ls
command would make the pool unresponsive while the output is generated; now, running this command no longer prevents the pool from taking on more work.As the
sweeper purge
,sweeper free
andrep rmclass
commands can take a long time to finish, the admin interface does not wait for them to finish; instead, a quick response is returned and the command continues in the background. Previously there was no indication when such commands had finished; now, the pool will log when the command has completed.The pool logs regular sweeper operations at lower priority; this makes the pool logs less noisy when operating under normal conditions.
poolmanager
In previous versions, poolmanager would log a stack-trace if a user attempts to open a file for reading and no read pool could be found because a unit failed to match. Now the failure is logged as a normal message.
xrootd
The xrootd door now informs billing when a user uses the xrootd protocol to delete a file.
gplazma
The ldap plugin allows a door to request converting a uid or gid back to a username; currently, only the nfs door makes use of this functionality, when generating a directory list. In previous dCache releases, the plugin behaved incorrectly if the uid or gid is unknown; this is now fixed.
Changelog from 2.10.2 to 2.10.3
- ab651cb
- [maven-release-plugin] prepare release 2.10.3
- 577f16f
- chimera: mark as-run if i_dirs_ipnfsid exists
- 0f1cea4
- solaris: fix solaris package script and add work-around for pkgmk bug
- e1dd257
- acl: add heuristic to detect acl format
- 7f9ef6d
- chimera: create missing index i_dirs_iparent
- 0bfaa7e
- pool: Resolve high memory usage and other issues in sweeper
- befad25
- info: Fix HashMap ordering assumption in unit test
- 08f7865
- Make JDK byte code verification bug workaround Java 8 compatible
- fc19b7a
- gplazma: Fix JVM implementation dependency in unit test
- 639653f
- poolmanager: Suppress stack trace in case of unmatch units
- 92fe17a
- xrootd: Add billing entry on delete
- ca6c026
- bugfix: fix reverseMap not throwing NoSuchPrincipalExceptions
- 8e7edbc
- [maven-release-plugin] prepare for next development iteration
Release 2.10.2
Cell startup
Fix potential dead-lock when a dCache domain is starting. The problem is triggered when a cell receives a message while initialising. If dCache is started “from cold” then this cannot happen. However, if a domain that contains well-known cell restarts while the rest of dCache is operational then the problem could be triggered. For example, if a door domain ran out of memory then there is a small risk that the domain would not restart correctly. The problem was present since dCache v2.8.0 and is fixed with this release.
Changelog from 2.10.1 to 2.10.2
- 44daad1
- [maven-release-plugin] prepare release 2.10.2
- 333a819
- cells: Fix deadlock during startup
- fa3b2f1
- [maven-release-plugin] prepare for next development iteration
Release 2.10.1
cell communication
The decision on which cell communication system to initialise was
based on the (old) broker.scheme
configuration property and ignored
the (new) dcache.broker.scheme
configuration property; this has been
fixed. Note that configuration using the old broker.scheme
continues to be supported for backwards compatibility.
billing
Update to use the billing.text.dir
configuration property instead of
billingLogsDir
; the latter is still supported for backwards
compatibility.
dcap
When the client authenticates via GSI, the dcap door will check the
user certificate (generally, a voms-proxy certificate) to extract the
DN and primary FQAN. To do this, it needs to know where the trusted
CAs and trusted VOMS server identities are stored; these are typically
/etc/grid-security/certificates
and /etc/grid-security/vomsdir
respectively. Previously, dcap ignored the dcap.authn.capath
and
dcap.authn.vomsdir
properties. With this release, these properties
are honoured.
pool
If a stage request has been forwarded to the nearline subsystem and the file is deleted then the stage request is now cancelled.
Nearline storage configuration is now applied atomically rather than one attribute at a time.
The maximum number of concurrent nearline operations may be adjusted
using the hsm set
admin command. Previously, if the concurrency is
increased then additional activity is only started as more requests
are received, ignoring any existing backlog. This has been fixed so
increasing the concurrency limit will immediately start more activity
if there is a backlog of requests.
This release reduces the CPU processing required when starting a third-party HTTPS transfer.
When transferring files to some third-party service using HTTPS, the remote host must provide a trusted certificate. The trust material (CA certificates and CRLs) that are necessary for this are now refreshed without requiring the pool to be restarted.
poolmanager
Stage and pool-to-pool transfers, triggered by poolmanager, are monitored by poolmanager querying the source pool to check if the request is still active. In previous versions of dCache, for n transfers from a pool, this monitoring generated O(n*n) of network traffic. In particular, poolmanager’s monitoring of bulk staging requests can result in sufficient network traffic between the pool and poolmanager that other dCache activity is impacted. With this release, the monitoring traffic is reduced in general, most prominently for large concurrent activity, such as bulk stage requests.
The rc destroy
command has been removed. The implementation was
incomplete and the operation was dangerous.
If the door resubmits a request to poolmanager that triggers staging of a file, poolmanager uses additional memory while waiting for the stage to complete. For bulk staging, this additional memory usage can be significant, potentially resulting in poolmanager exhausting the available memory. This release updates poolmanager so a door resubmitting a stage request does not increase the memory footprint of poolmanager.
The poolmanager has an incomplete feature called ‘clumping’. This is where poolmanager can handle multiple select-pool-for-read requests for the same file as a single request. There is a hard-coded clumping limit that prevents too many files from being handled together; this allows poolmanager to choose an alternative pool when staging or replicating to make the file available. Once this limit is reached, subsequent open requests for the same file will fail in poolmanager. Doors handle this failure differently; the xrootd and webdav doors propagates this failure back to the client, while the dcap and ftp doors will resubmit the request to poolmanager. This release increases the clumping limit from 1 to 20; it also changes the behaviour of the doors so all will retry the pool-selection if the clumping limit is reached. This fixes the ‘request clumping limit reached’ failures.
This release fixes a race-condition between two clients opening the same file. Previously, if poolmanager starts processing the second open request as it is finishing processing the first then there is a tiny chance that the second request is lost.
srm
The srm service will check the user certificate (generally a
voms-proxy certificate) to extract the DN and primary FQAN. To do
this, the srm service needs to know where the trusted CAs and trusted
VOMS server identities are stored; these are typically
/etc/grid-security/certificates
and /etc/grid-security/vomsdir
respectively. Previously the SRM ignored the srm.authn.capath
property and there was no way to configure the trusted VOMS server
identity path. With this release, the srm.authn.capath
is
honoured and the srm.authn.vomsdir
property is introduced.
When the SRM server sends an asychronous response to the client, the client must make subsequent queries to check if the request has completed. Part of each response is a hint suggesting for how long the client should wait before querying for a progress update. Prior to dCache v2.9, the srm service provided this hint where subsequent responses suggested a longer wait time. With v2.9, this feature was lost, resulting in the SRM client hammering the srm service. This has been fixed.
When the SRM client wants to initialise a download (using the srmPrepareToGet call) for a file that is already online there is a very small risk that the srm service will hang and provide no reply. This has been fixed.
When processing a srmBringOnline request, the server checks there is at least one transfer protocol that both the client and server support. In previous versions, this was checked when processing each file in the srmBringOnline request, failing the file if no common transfer protocol is found. This has been updated so the transfer protocols are checked once for the entire request. This new behaviour matches how the transfer protocols are checked for other types of SRM requests.
webdav
Update door to use the webdav.root
configuration property instead of
webdavRootPath
; the latter continues to be supported for backwards
compatibility.
Allow more control over webdav door behaviour; in particular, thread
usage and connection handling. This allows sites to tune their door
to match hardware capabilities and client activity. The following
configuration properties are introduced webdav.limits.acceptors
,
webdav.limits.idle-time
, webdav.limits.idle-time.unit
,
webdav.limits.low-resource-idle-time
,
webdav.limits.low-resource-idle-time.unit
,
webdav.limits.handshake-time
, webdav.limits.handshake-time.unit
,
webdav.limits.backlog
, webdav.limits.threads.max
,
webdav.limits.threads.min
, webdav.limits.threads.idle-time
,
webdav.limits.threads.idle-time.unit
, webdav.limits.queue-length
,
webdav.limits.graceful-shutdown
,
webdav.limits.graceful-shutdown.unit
. See the webdav.properties
defaults file for more details.
If the client reads or writes through the door and the door detects a lack of progress then it will cancel the transfer. Previously, the door did not kill the mover, leaving it to time out. With this release, the door will kill the mover when it cancels the transfer.
The door is updated so, when a client initialises a third-party transfer, the distinguished name and primary FQAN comes from the gplazma login result rather than being generated locally. This change means the node hosting the webdav door does not require a vomsdir setup.
xrootd
Previously the xrootd batch service required the user to specify a non-empty pool queue for uploading. This requirement has been relaxed.
Changelog from 2.10.0 to 2.10.1
- b14aa7f
- [maven-release-plugin] prepare release 2.10.1
- 33a7ec7
- srm: srm,dcap: Use configured vomsdir and capath
- cc0c6d6
- build: add work-around for JDK bug and PowerMock
- 85e97f7
- pool: Fix concurrency settings for script nearline storage
- 63cbc4e
- poolmanager: Fix ‘request clumping limit reached’ failures
- ca2afba
- poolmanager: Avoid request leak
- e970b90
- srm: Fix race condition in srmPrepareToGet
- b85329b
- webdav: Fix certificate handling for HTTP third-party copy
- 9051f82
- poolmanager: Fix excessive p2p and stage alive checks
- 04db77b
- srm: Restore estimated wait time update
- f0951c1
- webdav: Kill mover in case transfer times out
- 87ad152
- webdav: Make threading and TCP connection limits configurable
- ce24ff6
- pool: Fix several minor issues in nearline storage subsystem
- 4a2b9b5
- srm,dcap: Use configured vomsdir and capath
- 3d3263f
- Don’t use legacy properties
- 14dbed7
- [maven-release-plugin] prepare for next development iteration
Release 2.10.0
Changes affecting several services
Libraries
Updated Spring to v4.0.5, AspectJ to 1.8.1, SMC to v6.3.0, Jetty to 8.1.15.v20140411, Wicket to v6.16.0, Guava to v17.0, HikariCP to v1.3.9, Netty to v3.9.2.Final, BerkleyDB v6.0.11, Apache commons-compress 1.8.1, Scala v2.11.1
Logging FTP client disconnects
The FTP protocol contains commands that a client uses to indicate that it is about to disconnect from the server. Not all clients use this; some simply disconnect the network connection, which is unfortunate as dCache cannot know whether or not this indicates a problem with the client.
dCache logs when the client disconnects unexpectly (without using the BYE command). Previous versions of dCache also logged this in pnfsmanager and spacemanager but, with v2.10, dCache no longer logs these events in either service; logging in the door continues.
Messaging
Most messages sent from one dCache cell to another are requesting some information or that some action should take place. These messages include an expiry date. The expiry date provides a hint when the response message will not be of interest. For example, when the srm sends a message to gplazma (to log the user in), the srm will fail the request if it doesn’t receive a response from gplazma within some timeout period (3 minutes, by default). Therefore, after 3 minutes, the srm is not interested in the reply from gplazma.
Services can make use of this information and reject messages where it cannot reply in time. This is most useful when dCache is recovering from being overloaded: dCache can recover faster because it can reject some of the backlog of work.
In previous versions of dCache, this behaviour was limited to pnfsmanager and the pool’s migraion module. With dCache v2.10, this behaviour applies to all services.
Certain messages request the current status of a service. Often, the result of these queries provides similar information. Currently, for certain pieces of information, objects are created afresh, despite there being relatively few distinct values. With 2.10, these same-valued Strings are represented by the same object, reducing the memory footprint of services, especially webdamin.
Kerberos
dCache will now use the kerberos configuration in /etc/krb5.conf
by
default. The existing configuration is still supported, but the
system configuration file is now taken as the default.
http third-party transfers
dCache v2.10 sees greatly improved support for HTTP third-party transfers. These are transfers between dCache and some other HTTP server, so that the traffic does not go through the client.
In keeping with gsiftp, http third-party transfers are initiated on the pool. dCache now supports both pull and push transfers. A pull transfer is where the pool fetches a file from the remote server using the HTTP GET method. A push request is where the pool uploads a file to the remote server using the HTTP PUT method.
A transfer can use either an unencrypted transport (http://...
) or
make use of SSL/TLS encryption (https://...
). If an SSL/TLS
transport is used then the pool can use an X.509 credential when
establishing the SSL/TLS connection.
Both transfer directions (push and pull) will attempt to check the integrity of the transferred data. For pull requests, all data integrity information comes from the headers supplied in the GET response. For push requests, the pool will make a subsequent HEAD request and use the supplied information to verify the data was sent correctly.
For push requests, if a PUT request is successful but the subsequent HEAD request reveals data corruption then the pool will attempt to delete the file with an HTTP DELETE request. If this DELETE fails then the error message is updated accordingly. For pull requests, no additional cleanup steps are needed.
File checksum values are discovered via the RFC 3230 extension to HTTP. Unfortunately, this is not widely supported by HTTP servers; therefore, dCache supports two levels of data integrity checking: weak and strong. Weak data integrity is satisfied when
the remote service and the dCache server agree on the content length,
none of the checksums supplied by the remote server (if any) disagree with the checksums known by dCache.
Strong data integrity is satisfied when, in addition to satisfying the weak conditions:
- there is at least one checksum supplied by the remote server that agrees with a dCache-local checksum.
Weak data integrity can be satified by any HTTP server whereas strong data integrity requires an HTTP server with RFC 3230 support. A transfer satisfying weak integrity might not check the checksum values match; strong integrity requires that checksums exist and match.
Third-party HTTP transfers also allow the client to specify zero or more (arbitrary) headers when requesting a file be transferred. This allows the client to customise the HTTP request the pool sends to the remote server. Such headers could include authorisation information or some other information to steer how the remote server handles the HTTP request.
Third-party transfers may be initiated by the webdav service and the srm service. See the following sections on those services for details on how to trigger third-party copying.
Changes to services
ssh1
The ssh v1 admin interface has been removed; ssh v2 is now the only way to connect to the admin interface.
webadmin
The active transfers page has been reimplemented. The new implementation focuses on faster rendering, faster issuing of commands, and a reduced memory footprint.
webdav
The WebDAV door uses the StringTemplate language for generating directory listings for web-browsers and for error pages. With previous versions of dCache, a missing or unreadable template file resulted in a NullPointerException. With dCache v2.10, there is a hard-coded “broken dCache” page that describes the problem and suggests how to fix it.
dCache v2.9 saw the introduction of an activity log that recorded SRM and FTP commands and their corresponding responses in the log file Domain.access. This allows easy investigation of client-specific problems. With v2.10, support has been extended to include HTTP/WebDAV requests. The access file now includes a single line for each HTTP/WebDAV request. Logged information includes the mapped user, the request method, the URL, the numerical response code and explanation, the client IP address, and the user-agent.
The default HTML rendering of directory listings and error messages has been updated to use the Bootstrap framework, which supports a rich set of features. This new default provides a deliberately toned-down presentation. The default page now contains two links for each file: one that hints the browser should show the content and the other that the browser should download the file. As before, admins may customise the rendering of directory listings and error messages. To support clients with poor or no Internet connection, all required javascript libraries and Bootstrap files for the default layout are included with dCache. This release also fixes a rendering bug in the breadcrumbs.
third-party transfers
The webdav door now supports requesting third-party file transfers.
This is an extension to the WebDAV COPY command, which is normally
limited to internal copies. A client may request a file be
transferred to a remote site by specifying the remote location as the
Destination
HTTP request header. Currently supported transports are
GridFTP (destination URI starts gsiftp://
) and HTTP. For HTTP both
plain (http://
) and with SSL/TLS (https://
) are supported.
The Credential
HTTP request header may be used by the client to
describe whether or not the pool should use a delegated credential
when transferring the file. The accepted values are none
and
gridsite
. If none
then no credential is to be used; if gridsite
then a delegated credential is to be used. If the Credential
header
isn’t specified then a transport-specific default is used: for gsiftp
and https transports the default is gridsite
; for http transport the
default is none
. Some combinations are not supported: gsiftp with
none
and http with gridsite
are not supported.
If a delegated credential is to be used, the client must delegate a
credential to dCache using the delegation service, which is part of
the srm service. If no useful certificate could be found, the webdav
door will redirect the client to itself and include the
X-Delegate-To
HTTP header in its response. This header contains a
space-separated list of delegation endpoints.
For ‘http’ and ‘https’ transfers, the client may choose whether to
require weak or strong verification; see above for a definition of
these. The RequireChecksumVerification
header controls this
behaviour; if this header has a value ‘true’, strong verification is
required. If ‘false’, weak verification is sufficient. If the header
is not specified, a configurable default value is used.
Again, for ‘http’ and ‘https’ transfers, the client may specify
additional or replacement HTTP headers that the pool should use when
making the transfer. These are specified as headers that start
TransferHeader
. This prefix is removed and the header used by the
pool. For example, to have the pool use basic authentication with
userid ‘Aladdin’ and password ‘open sesame’, the client would add the
following header to its request:
TransferHeaderAuthentication: Basic QWxhZGRpbjpvcGVuIHNlc2FtZQ==
Once the transfer is accepted, the client will receive periodic progress information. If the client closes the TCP connection, the transfer is cancelled.
alarms
The alarms system, through the logback library, has always supported sending emails if an alert is found. However, enabling this feature required editing an XML file. With dCache v2.10, the standard dCache configuration now includes properties that allow the admin to configure this behaviour through dCache’s normal configuration properties.
space-manager
If a space reservation is over-subscribed, the free space is shown as a negative number. This breaks the layout, with that line being badly formatted. This is fixed with dCache v2.10
dcap
The dcap door has been made more robust when processing unexpected replies from other dCache services.
admin
The prompt shown in the admin interface has changed from green to bold. This is in keeping with other dCache CLIs.
For many services, if the ‘info’ command is called after a cell is accepting commands but before initialising has completed, a NullPointer error is produced internally and no information is provided. This has been fixed.
billing
Billing messages are no longer sent to the pinboard, so the ‘show pinboard’ admin command is now limited to showing information about the service itself.
Some of the information about a transfer is written as a colon-separated list of items. In previous versions of dCache, the protocol version and client hostname were not separated, resulting in entries like:
{Http-1.1zitpcx6184.desy.de...}
With dCache v2.10, this has been fixed for the http and xrootd transfers, so entries now are correctly separated:
{Http-1.1:zitpcx6184.desy.de...}
info
The info service, which collects and collates information from various dCache components, now supports presenting the gathered data in different formats. Previous versions of dCache only supported a dCache-specific XML format. With dCache v2.10 two additional formats are available.
There are two ways the client can choose the format. The HTTP client
can specify the MIME type as the “Accept” request header;
JSON-specific clients may do this automatically. The alternative is
to append “?format=_some_format_” to the URI; e.g.,
http://httpd-service.example.org:2288/info/summary/pools?format=json
.
There are four supported MIME types (application/xml
, text/xml
,
application/json
and text/x-ascii-art
) and three supported formats
for ‘?format=’ (xml
, json
and pretty
). The default format
continues to be XML.
srm
dCache v2.10 sees a major overhaul in how the srm controls and limits its interaction with other dCache services. The following sections describe these changes.
Background
The SRM protocol has the concept of a client’s request not being processed immediately, with the request being queued prior to any work being conducted. This is most apparent for asynchronous requests, where the client reconnects periodically to discover what progress, if any, has taken place. The following also applies for synchronous requests (where the client waits to learn the result of the query); however, the client is oblivious of these details.
In general, a successful request will start out in a “queued” state. After dCache starts processing the request, the request will have an “in-progress” state. For transfer requests (GET or PUT), the request will be in a “ready” state when there is a TURL for the client to use, and in a “success” state after a successful transfer. For non-transfer requests (LS, BRING-ONLINE, …) the request will go into the “success” state once dCache finishes processing it, skipping the “ready” state.
To answer most client-issued requests, the srm service must trigger some activity in other dCache services. For the most part, the srm service simply issues requests to other services, collects the responses and converts them into a corresponding SRM reply.
The srm service has several execution pools to throttle processing of requests. These limits are applied to individual files, so, for example, from a single request to read 10 files, each of the files is scheduled as if there were 10 requests, each reading a single file.
Prior to dCache v2.10, the srm throttled itself by limiting the number of requests it can process concurrently. Once this limit is reached, additional requests are placed in a queue with a limited capacity. Once this queue limit is reached, subsequent requests are rejected. The effect is to throttle the frequency of messages being sent.
In practice, a single thread could easily saturate the rest of dCache, which meant that this was not a particularly effective way to throttle the srm service.
Furthermore, the queue limits would also affect requests that had already begun processing. Thus under high load, one could observe requests being failed even if they had already started processing. At least in theory, this could lead to a starvation problem, in which a lot of processing could take place while no requests would ever succeed.
New approach
Rather than rate-limiting the delivery of internal messages, dCache v2.10 provides the ability to limit the number of requests in each of the different states mentioned above (“queued”, “in-progress” and “ready”). This is taking into account “in flight” requests.
There are now three types of limit: max-requests, max-inprogress and max-transfers.
-
The max-requests limit controls the total number of requests that are queued or being actively worked on. Once this limit is reached, any more requests from an SRM client are rejected.
This limit allows the srm service to protect against running out of memory. Setting a value too low will result in the srm not fully utilising the available memory; setting it too high will risk running out of memory under heavy load.
-
The max-inprogress limit controls the concurrent activity within dCache; once this limit is reached, subsequent requests are either queued or rejected, depending on the max-requests limit.
This allows the srm service to stop itself from overloading core dCache services when processing SRM requests. The max-inprogress limit also controls the maximum number of staging requests for GET and BRING-ONLINE requests. For COPY requests, it sets the maximum number of current transfers.
Setting the value too low will result in the srm artificially limiting its performance, as requests will needlessly spend time queued. Setting the value too high will risk overloading core components when the srm service is under heavy load.
-
The max-transfers limit controls the number of TURLs handed out to clients. Once this limit is reached, the srm keeps subsequent requests that have a TURL in the “in-progress” state - at least from the point of view of the client. Internally, such requests are in the “ready queued” state and no longer count towards the max-inprogress limit. Once the SRM client completes transfers, those requests that have a TURL but are in the “ready queued” state can become “ready”.
This limit is mostly to allow dCache to protect pools by limiting the number of concurrent transfers. It will also provide some protection for the transfer doors (typically ftp) and poolmanager.
Since this limit type affects only transfers, setting the value too low will result in artificially poor transfer rates (with client requests spending a large amount of time in the “in-progress” state despite having a TURL) while other SRM requests are processed quickly. Setting the value too high risks overloading pools when dCache is under heavy load.
There are request-specific properties that may be configured for each
of these three types of limit. These have names like
srm.request.TYPE.{max-requests, max-inprogress, max-transfers}
,
where TYPE
is one of get
, put
, ls
, bring-online
, copy
or
reserve-space
.
For example:
srm.request.bring-online.max-inprogress
controls the maximum number of concurrent staging requests.srm.request.get.max-transfers
controls the maximum number of concurrent downloads (maximum number of TURLs handed out at any time).srm.request.ls.max-requests
controls how many file metadata and directory listing queries to allow (either actively being processed or queued) before rejecting new requests.
The max-requests limit properties (srm.request.get.max-requests
,
srm.requests.put.max-requests
, etc) all take a common default value
controlled by the srm.request.max-requests
property. The current
default value for srm.request.max-requests
is 10,000. Note that the
srm.request.max-requests
value applies to each SRM request type
(GET, PUT, BRING-ONLINE, …) independently. Therefore, the overall
maximum number of requests is six times the configured value.
The max-inprogress properties (srm.request.get.max-inprogress
,
srm.request.put.max-inprogress
, etc.) all take individual default
values. The current default values are to allow concurrent processing
of 10,000 BRING-ONLINE requests, 1,000 GET and COPY requests, 50 PUT
requests, 50 LS requests, and 10 RESERVE-SPACE requests.
The two max-transfers properties (srm.request.get.max-transfers
and
srm.request.put.max-transfers
) take pre–2.10 srm configuration
properties as default values.
As with prior versions of dCache, if a request suffers a transitory failure, the srm will retry the operation after waiting awhile. Prior to 2.10, such retries were treated specially. With 2.10, the retry is treated as if it were a fresh request, so is subject to the three limits described above.
srm protecting itself
The srm service continues to protect against trying to process too
many requests concurrently. As with pre–2.10 srm, this is achieved by
scheduling activity using a set of threadpools. These are controlled
by the srm.request.type.threads configuration property, where type
is one of {get, bring-online, put, copy, ls, reserve-space}. These
all take srm.request.threads
as a default value. The threadpool
sizes also accept pre–2.10 configuration (for example,
srm.request.put.threads takes the value of srmPutReqThreadPoolSize as
a default value).
Notes for sites upgrading
We have tried to reduce the impact of this change by continuing to support some pre–2.10 configuration. However, since this change reflects a new processing model, sites should expect that some configuration changes are needed. Therefore, admins are recommended to familiarise themselves with the new scheduler options and adjust their configuration accordingly.
The dcache check-config
command will describe which configuration
properties are no longer supported, and so may indicate the need for
alternative configuration.
HTTP third-party transfers
The srm service now supports http and https transfers in both pull (dCache pool makes an HTTP GET request) and pull (pool makes an HTTP PUT request) operations. The SRM client can steer the transfer by supplying ExtraInfo arguments. The ExtraInfo arguments are a set of arbitrary key-value pairs that the client can specify when making the request. The currently supported ExtraInfo keys are:
verified
If set, takes a value of eithertrue
orfalse
. This controls whether weak or strong verification is required. If set to false then weakly verified transfers are successful. If set to true then a transfer must be strongly verified to be successful.header-
. All ExtraInfo elements that start “header-” are converted into HTTP headers. The HTTP header key is the ExtraInfo key without the initial “header-”. The HTTP header value is the ExtraInfo value. These HTTP headers are used when the pool makes a request.
If the SRM client delegates a credential as part of the GSI connection, this credential is used to authenticate when the pool makes an SSL/TLS-based connection to the remote server. If the client doesn’t delegate then dCache will attempt the transfer without any client credential.
See the above section on HTTP/HTTPS third-party transfers for more details.
Other srm improvements
The dCache srm implementation places client requests into one of a number of internal states. Each of these internal states broadly correspond to a single SRM state, as described above (“queued”, “in-progress”, …). With dCache v2.10 this similarity has been made more explicit and, as a result, some incorrect mappings have been fixed.
Previously, the srm would log “illegal state transition” messages if a user aborts a request involving more than one file. This has been fixed.
If the srm service sends a message to some other service due to some client request (e.g., prepareToPut), and the response from this message takes too long, the srm service will fail the client request. It could happen that, after the client request has been failed, the srm service receives the reply message, stating the requested operation was successful. If this happens then the client will see dCache in an inconsistent state: the srm service told the SRM client its request failed, yet its request resulted in successful activity. In the case of preparing for uploading a file, pinning a file and creating a space-reservation, the srm service can safely “undo” the effect of the message by removing the temporary upload directory, unpinning the file, or releasing the space-reservation. This behaviour has been added with dCache v2.10.
The error message sent to the SRM client if the srm service receives certain malformed LS requests has been improved.
The output of the admin interface ‘ls’ command (which provides an overview of the current requests and allows more detailed information about any specific request) has been improved. The timestamps are easier to read; the history also shows the time spent in each internal state. Also, a lot of “noise” has been removed, making it easier to understand what is happening.
pool
Wake up sweeper if new entry is added.
Make pool-to-pool transfers more robust against two problems: busy networks during the initial phase of the internal transfer (when the mover is queued) and the source sending the URL for the transfer just as the destination was giving up.
The allocate-now (NFS-only) garbage-collection now also utilises all reclaimable space.
The classic dCache behaviour is to block when a client writes into a file and there is insufficient space to satisfy the request. For the NFS protocol, this blocking behaviour is seen by the client as the server being unresponsive and will trigger the client banning the pool and attempting to recover by sending subsequent IO to the door. For NFS v4.1 traffic, writing through the door is not supported, so the upload will fail. With dCache v2.10, when the pool runs out of space, this information is relayed back to the NFS client so it can behave correctly.
The number of files has been added to the output from the ‘info’ command.
poolmanager
The wrandom partition type has been removed. The same behaviour may be achieved with the WAAS partition type.
gplazma
The nis plugin will fail if no matching pincipal was found.
The ldap plugin default value has been updated.
The xacml plugin now supports the gums server returning GID values.
nfs
The nfs door may now be configured whether or not it is well-known.
The admin interface for the nfs door has been updated to include better on-line help.
The ‘pool reset id’ command has been added to the nfs door to reset pool’s device id. This command is used to recover from unknown failures: the Linux kernel may choose to ban a pool (due to some real or imagined protocol violation) and send IO through the NFS door. This command allows the door to simulate a pool being rebooted, triggering the client to reassess its opinion of the pool.
ftp
Additional, useful information has been added to the output of the info command for the cell handling a client’s command channel.
pnfsmanager
The error message reporting a problem when a user cancels an upload has been fixed.
When querying for a file’s checksum values, the chimera pnfsmanager plugin would try each supported checksum algorithm. In most cases, a file will have only a single checksum value, so many queries gave no additional information, but placed load on the PostgreSQL database and increased latency. Now such requests are handled with a single database query.
Scripts
The dcache-star
command has been updated so it will create a valid
(albeit empty) StAR record if there is a database connection failure.
The dcache billing
command has been extended to support the
--since
and --until
options. These allow reports to be generated
within a specific period. The output has been enhanced so search
results can be provided in JSON or YAML format, in addition to the
existing raw and files format. Please be aware that the JSON and YAML
formats are not finalised: future versions of dCache may make slight
alterations as experience is garnered.
Changelog from 2.9 to 2.10.0
- befd0f8
- [maven-release-plugin] prepare release 2.10.0
- f602324
- srm: fix NPE when checking delegated credential
- 4b2812d
- srm: fix deserialise of TExtraInfo
- efe4315
- srm: fix IllegalStateTransition for COPY requests
- 76a393b
- nfs4: use only address string in ProtocolInfo#toString() method
- d8fff55
- Add support for webdav 3rd-party transfer using http and https
- a150e36
- webdav: allow users to choose between showing file’s content inline and downloading it
- 30c515b
- Fix broken commit
- 62bb02a
- srm,pool: improve support for 3rd-party HTTP copy
- 7c8a5ea
- webdav: add support for remote copy
- ed5bb82
- chimera: fix liquibase changeset to fail peacefully
- ab65294
- fix .gitignore to work for Eclipse on MacOS X
- f354005
- libs: revet luquibase to 3.1.2
- f28f204
- Revert “Update to DataNucleus 4”
- 23c6d2b
- srm: make timestamps easier to read
- 27fa708
- srm: fix retry and failure messages for scheduled jobs
- ff2397f
- srm: fix status of all-failed container jobs
- 908ae5a
- srm-client: by default only delegate for srmcp with 3rd-party copies
- 201af9e
- alarms: Fix checksum alarm filter.
- 99e73b0
- src: kill more of obsolete code
- 3445ee3
- src: kill more obsolete code
- 87081b2
- src: more obsolete code
- caf476c
- src: remove obsolete code
- 9e4251e
- pool: Add file count to ‘info’ output
- efa2de6
- deb: Add python-psycopg2 as a recommended dependency
- 5e71072
- Update to DataNucleus 4
- 75d3718
- Update third party dependencies
- 38f00db
- nfsv4: get client from current session for v4.1
- cf399f9
- delegation-shell: generate delegated credential of correct type
- a8438a5
- pool: add soft/hard space allocation modes
- 2a4efa6
- pool: add Allocator#allocateNow()
- d22ad32
- pool: Account#allocateNow() should trigger sweeper to get some space
- c22bba3
- pool: Update repository javadoc
- 5da93bb
- Suppress nfs’ file protocol from being published
- ade506e
- pool: Fix race condition that leads to orphaned files
- 87bcfb2
- srm: improve error message
- 7fed889
- dcache-webadmin: fix pool queue plot grid placeholder image
- 3740796
- webadmin: fix regression in alarm deletion filter
- d7f67a7
- cells: remove System time dependency in unit-test
- 1722627
- webadmin: Reimplement active transfers page
- 52d2f1c
- Intern several strings on message deserialization
- 43ddb9f
- webadmin: Simplify titel initialization of base page
- 5026cf9
- pool: log with error if checksum scanner gets an IO error
- b756c66
- nfs: add a command to reset pool’s device id
- 4d14b40
- nfs: use new admin command interface
- b2575fa
- nfs: remove double locking
- 6a6fcfd
- ftp: Close passive port on failure
- 44ecaa3
- ftp: Fix indentation
- 0764f2a
- srm: Undo external operations if the request has expired
- 15fa2ee
- srm: Cancel upload path if request was aborted
- 8b7bc70
- ftp: fix LIST, NLST and MLSD cmd-channel output
- f1bee28
- pool: Fix regressions in new HSM subsystem
- e73977e
- Fix ArrayIndexOutOfBoundsException when processing multiple surls invoking srmget.
- d12513c
- Fix issue with incorrect determination of whether or not a string is encoded by replacing URLDecoder with import org.springframework.web.util.UriUtils as the former does not handle ‘+’ character.
- 40356b5
- gplazma-xacml: Add handling of GID returned by GUMS in gPlazma2 XACML plugin
- 97bd31a
- dcache alarms: add properties and fuller documentation for configuring SMTP appender
- 6f97e8d
- system-test: fix web font corruption
- cb46d8a
- config: do not set default values for kdc and realm
- 5c63d0e
- http: fix multipart response size
- 3fd5888
- dcache-star: create valid xml even on db connection failure
- 93e94be
- xrootd,nfs: Add support for space management
- ab39955
- pinmanager,srm,spacemanager,xrootd: Fix pool monitor expiration
- 3ac48ab
- srm: Fix lock congestion
- 3f40abe
- http: do not use depricated API
- b0afa4f
- chimera-provider: fix regressions introduced by API change
- e671d5c
- chimera: remove FileSystemProvider#getInodeChecksum
- 1a79834
- dcache-chimera: query for all checksums in one go
- 0cb9af2
- Update third party libs
- e85f9cf
- webdav: Bootstrap for directory listing
- 4137a00
- xrootd: Quote IPv6 addresses in kXR_redirect replies
- c2d5170
- nfs: remove unused argument in NFSv41Door#getPool()
- bd7a2fe
- cells: Fix hidden NPE during initialization
- 1933780
- xrootd: Fix error message for read patch check
- 39c2c0e
- dcache-webadmin: fix login redirect bug which misleadingly results in Access Denied Page
- cabce53
- xrootd: Upgrade to xrootd4j 1.3.3
- 722ea6c
- system-test: Make log configuration more similar to production config
- 8f2a7a0
- cells: Minor cleaning of CellPath and CellAddressCore
- 41a3b9b
- billing: Add date range and json/yaml output for billing search tool
- c3a83d9
- admin: Make prompt bold
- 4a838d2
- scripts: allow java8 runtime
- 24a244f
- gplazma: authzdb add test case
- 8f63805
- dcap: do not call toString() on error object
- 2f2badb
- dcap: ensure that we remove session on DoorTransferFinishedMessage
- 3f34bda
- info-provider: fix works-by-accent bug when importing dCache config
- bdc8f49
- webdav: switch activity logging to NetLogger format
- 005b7bf
- pom: update jglobus library
- 35ab3f8
- srm: Use srm.enable.space-reservation rather than dcache.enable.space-reservation
- 85d9e29
- Fix rendering of negative numbers in columm layout
- 4bd5947
- namespace: Improve error message in upload cancellation
- 58192a1
- srm: Fix NPE
- eeb9301
- ftp,dcap: Don’t export child doors
- eb9212c
- ftp: Add more information to info page
- a966d7a
- cleaner: Shut down timer thread
- afac6ab
- pool: Improve robustness of p2p
- ed1975e
- httpd: Fix a couple of bugs
- 39b2c11
- srm: Avoid duplicate status code update on status query
- 0eb1af0
- xrootd: split protocol name form the client IP in XrootdProtocolInfo
- ee90392
- chimera: Fix duplicate key supression
- 9bbb9dc
- Revert “srm: Drop fail fast for protocol checks”
- fd9d0c7
- srm: Fix request cancellation due to SURL deletion
- 4837397
- srm: Fix request expiration
- 45f658f
- nfs: make cell export configurable
- 2e3034d
- fix type in gplazma ldap default value
- 27fdd4d
- Drop messages when TTL is exceeded
- 360a230
- cells: Fix message counting in case of queue overflow
- 69fbf87
- pool: Improve how we deal with unexpected files in pools
- c34fa7e
- pool: Fix error handling in staging code
- c67f043
- system-test: Pick up DN of dCache user cert if available
- 3a22b10
- pool: Log P2P fatal errors with stack trace
- 6723151
- info: add JSON serialisation
- f6c4fc5
- srm: Fix transition ordering for restored requests
- 3a27973
- hsm scripts: fix command and options order
- 8364809
- billing: do not generate billing message if file based store is disabled
- 47852ba
- webdav: generate error page when template couldn’t be found
- c688083
- webdav: work-around race-condition in StringTemplate library
- 239a7f9
- systemtest: regenerate host credentials, add user credentials
- d728fc0
- httpd: log bugs with stacktrace
- eb82e9f
- core: remove obsolete code
- d90196a
- srm: Fix deadlock in pinning
- e0e9a51
- srm: Schedule resubmit of pin request to avoid deadlock
- 0f5a4d5
- srm: Drop fail fast for protocol checks
- d078fdb
- pnfsmanager: Fix expected size extraction and tag caching
- b30bc4a
- scripts: fix credentials script to work for non-Linux machines
- b97bec9
- srm: Minor cleanup of job scheduling
- 4b84cef
- srm: Simplify scheduler configuration
- 608b736
- ssh1: Remove service from dCache
- c1e6fed
- Silence log messages commonly associated with client disconnects
- 4b7d7a0
- fhs: Use gzip compression for deb package
- 654b6ce
- srm: Simplify unpin callbacks
- 7209d9f
- srm: Align SRM scheduler with SRM request states
- 4fc36a0
- srm: Report correct state for RQUEUED and TQUEUED
- 0194173
- statistics: Fix NPE
- a01653f
- pool: Remove legacy hsm initialization from default pool setup
- 418c8ff
- core: split protocol name form the client IP in HttpProtocolInfo
- 5cf8806
- srm: Split standalone configuration from generic configuration
- cbb5936
- srm: fix retry of copy-pull
- a4c7f34
- alarms, billing: fix liquibase character varying and cleanup.
- 460b9e0
- systemtest: add support for generating user credentials
- 7e6e2ef
- pool: Report correct client IP to billing for passive FTP transfers
- 942035a
- srm: Fix race condition in listing
- 6750fd4
- poolmanager: Fix infinite loop in lru partition type
- 3c6da51
- gplazma-nis: pulgin must fail if no matching principal found
- c34be64
- nfs: do not use CDC as try-with-resource
- dfc809c
- databases: update hikari minimum connection configuration
- 193cba9
- nfs4: do not log stacktrace on IO errors
- b937f41
- nfs: make cache time unit property name guideline compliant
- 80ade65
- poolmanager,pool: Fix precision underflow in WASS
- d56c1e0
- poolmanager: remove wrandom partition type
- dc0188b
- pool: small cleanups in sweeper
- f82ced7
- pool: wake-up sweeper if new entry is added
- 89e8063
- releases: update dCache version to v2.10