tbd Cluster File Systems, Inc. * version 1.6.1 * CONFIGURATION CHANGE. This version of Lustre WILL NOT INTEROPERATE with 1.4.x versions automatically. In many cases a special upgrade step is needed. Please read the user documentation before upgrading any part of a 1.4.x system. * WARNING: Lustre configuration and startup changes are required with 1.6.x releases. See https://mail.clusterfs.com/wikis/lustre/MountConf for details. * Support for kernels: 2.6.9-42.0.10.EL (RHEL 4) 2.6.5-7.283 (SLES 9) 2.4.21-47.0.1.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) 2.6.16.27-0.9 (SLES 10) * Client support for unpatched kernels: (see https://mail.clusterfs.com/wikis/lustre/PatchlessClient) 2.6.16 - 2.6.19 vanilla (kernel.org) 2.6.9-42.0.8EL (RHEL 4) * Recommended e2fsprogs version: 1.39.cfs5 * bug fixes * Note that reiserfs quotas are temporarily disabled on SLES 10 in this kernel. Severity : enhancement Bugzilla : 10997 Description: lfs setstripe use optional parameters instead of postional parameters. -------------------------------------------------------------------------------- tbd Cluster File Systems, Inc. * version 1.6.0 * CONFIGURATION CHANGE. This version of Lustre WILL NOT INTEROPERATE with older versions automatically. In many cases a special upgrade step is needed. Please read the user documentation before upgrading any part of a 1.4.x system. * WARNING: Lustre configuration and startup changes are required with this release. See https://mail.clusterfs.com/wikis/lustre/MountConf for details. * Support for kernels: 2.6.9-42.0.10.EL (RHEL 4) 2.6.5-7.283 (SLES 9) 2.4.21-47.0.1.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) 2.6.16.27-0.9 (SLES 10) * Client support for unpatched kernels: (see https://mail.clusterfs.com/wikis/lustre/PatchlessClient) 2.6.16 - 2.6.19 vanilla (kernel.org) 2.6.9-42.0.8EL (RHEL 4) * Recommended e2fsprogs version: 1.39.cfs5 * bug fixes * Note that reiserfs quotas are temporarily disabled on SLES 10 in this kernel. Severity : enhancement Bugzilla : 8007 Description: MountConf Details : Lustre configuration is now managed via mkfs and mount commands instead of lmc and lconf. New obd types (MGS, MGC) are added for dynamic configuration management. See https://mail.clusterfs.com/wikis/lustre/MountConf for details. Severity : enhancement Bugzilla : 4482 Description: dynamic OST addition Details : OSTs can now be added to a live filesystem Severity : enhancement Bugzilla : 9851 Description: startup order invariance Details : MDTs and OSTs can be started in any order. Clients only require the MDT to complete startup. Severity : enhancement Bugzilla : 4899 Description: parallel, asynchronous orphan cleanup Details : orphan cleanup is now performed in separate threads for each OST, allowing parallel non-blocking operation. Severity : enhancement Bugzilla : 9862 Description: optimized stripe assignment Details : stripe assignments are now made based on ost space available, ost previous usage, and OSS previous usage, in order to try to optimize storage space and networking resources. Severity : enhancement Bugzilla : 4226 Description: Permanently set tunables Details : All writable /proc/fs/lustre tunables can now be permanently set on a per-server basis, at mkfs time or on a live system. Severity : enhancement Bugzilla : 10547 Description: Lustre message v2 Details : Add lustre message format v2. Severity : enhancement Bugzilla : 9866 Description: client OST exclusion list Details : Clients can be started with a list of OSTs that should be declared "inactive" for known non-responsive OSTs. Severity : minor Frequency : SFS test only (otherwise harmless) Bugzilla : 6062 Description: SPEC SFS validation failure on NFS v2 over lustre. Details : Changes the blocksize for regular files to be 2x RPC size, and not depend on stripe size. Severity : enhancement Bugzilla : 9293 Description: Multiple MD RPCs in flight. Details : Further unserialise some read-only MDT RPCs - learn about intents. To avoid overly-overloading MDT, introduce a limit on number of MDT RPCs in flight for a single client and add /proc controls to adjust this limit. Severity : enhancement Bugzilla : 22484 Description: client read/write statistics Details : Add client read/write call usage stats for performance analysis of user processes. /proc/fs/lustre/llite/*/offset_stats shows non-sequential file access. extents_stats shows chunk size distribution. extents_stats_per_process show chunk size distribution per user process. Severity : enhancement Bugzilla : 22485 Description: per-client statistics on server Details : Add ldlm and operations statistics for each client in /proc/fs/lustre/mds|obdfilter/*/exports/ Severity : enhancement Bugzilla : 22486 Description: improved MDT statistics Details : Add detailed MDT operations statistics in /proc/fs/lustre/mds/*/stats Severity : enhancement Bugzilla : 10968 Description: VFS operations stats Details : Add client VFS call stats, trackable by pid, ppid, or gid /proc/fs/lustre/llite/*/vfs_ops_stats /proc/fs/lustre/llite/*/vfs_track_[pid|ppid|gid] Severity : minor Frequency : always Bugzilla : 6380 Description: Fix client-side osc byte counters Details : The osc read/write byte counters in /proc/fs/lustre/osc/*/stats are now working Severity : minor Frequency : always as root on SLES Bugzilla : 10667 Description: Failure of copying files with lustre special EAs. Details : Client side always return success for setxattr call for lustre special xattr (currently only "trusted.lov"). Severity : minor Frequency : always Bugzilla : 10345 Description: Refcount LNET uuids Details : The global LNET uuid list grew linearly with every startup; refcount repeated list entries instead of always adding to the list. Severity : enhancement Bugzilla : 2258 Description: Dynamic service threads Details : Within a small range, start extra service threads automatically when the request queue builds up. Severity : major Frequency : mixed-endian client/server environments Bugzilla : 11214 Description: mixed-endian crashes Details : The new msg_v2 system had some failures in mixed-endian environments. Severity : enhancement Bugzilla : 11229 Description: Easy OST removal Details : OSTs can be permanently deactivated with e.g. 'lctl conf_param lustre-OST0001.osc.active=0' Severity : enhancement Bugzilla : 11335 Description: MGS proc entries Details : Added basic proc entries for the MGS showing what filesystems are served. Severity : enhancement Bugzilla : 10998 Description: provide MGS failover Details : Added config lock reacquisition after MGS server failover. Severity : enhancement Bugzilla : 11461 Description: add Linux 2.4 support Details : Added support for RHEL 2.4.21 kernel for 1.6 servers and clients Severity : normal Bugzilla : 11330 Description: a large application tries to do I/O to the same resource and dies in the middle of it. Details : Check the req->rq_arrival time after the call to ost_brw_lock_get(), but before we do anything about processing it & sending the BULK transfer request. This should help move old stale pending locks off the queue as quickly as obd_timeout. Severity : major Frequency : when an incorrect nid is specified during startup Bugzilla : 10734 Description: ptlrpc connect to non-existant node causes kernel crash Details : LNET can't be re-entered from an event callback, which happened when we expire a message after the export has been cleaned up. Instead, hand the zombie cleanup off to another thread. Severity : enhancement Bugzilla : 10902 Description: plain/inodebits lock performance improvement Details : Grouping plain/inodebits in granted list by their request modes and bits policy, thus improving the performance of search through the granted list. Severity : minor Frequency : only for kernels with patches from Lustre below 1.4.3 Bugzilla : 11248 Description: Remove old rdonly API Details : Remove old rdonly API which unsed from at least lustre 1.4.3 Severity : major Frequency : only for devices with external journals Bugzilla : 10719 Description: Set external device read-only also Details : During a commanded failover stop, we set the disk device read-only while the server shuts down. We now also set any external journal device read-only at the same time. Severity : minor Frequency : when upgrading from 1.4 while trying to change parameters Bugzilla : 11692 Description: The wrong (new) MDC name was used when setting parameters for upgraded MDT's. Also allows changing of OSC (and MDC) parameters if --writeconf is specified at tunefs upgrade time. Severity : major Frequency : when setting specific ost indicies Bugzilla : 11149 Description: QOS code breaks on skipped indicies Details : Add checks for missing OST indicies in the QOS code, so OSTs created with --index need not be sequential. ------------------------------------------------------------------------------ TBD Cluster File Systems, Inc. * version 1.4.10 * Support for kernels: 2.6.16.21-0.8 (SLES10) 2.6.9-42.0.8EL (RHEL 4) 2.6.5-7.276 (SLES 9) 2.4.21-47.0.1.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) * Recommended e2fsprogs version: 1.39.cfs2-0 Severity : major Frequency : liblustre (e.g. catamount) on a large cluster with >= 8 OSTs/OSS Bugzilla : 11684 Description: System hang on startup Details : This bug allowed the liblustre (e.g. catamount) client to return to the app before handling all startup RPCs. This could leave the node unresponsive to lustre network traffic and manifested as a server ptllnd timeout. Severity : enhancement Bugzilla : 11667 Description: Add "/proc/sys/lustre/debug_peer_on_timeout" (liblustre envirable: LIBLUSTRE_DEBUG_PEER_ON_TIMEOUT) boolean to control whether to print peer debug info when a client's RPC times out. Severity : normal Frequency : always Bugzilla : 10214 Description: make O_SYNC working on 2.6 kernels Details : 2.6 kernels use different method for mark pages for write, so need add a code to lustre for O_SYNC work. Severity : minor Frequency : always Bugzilla : 11110 Description: Failure to close file and release space on NFS Details : Put inode details into lock acquired in ll_intent_file_open. Use mdc_intent_lock in ll_intent_open to properly detect all kind of errors unhandled by mdc_enqueue Severity : major Frequency : rare Bugzilla : 10866 Description: proc file read during shutdown sometimes raced obd removal, causing node crash Details : Add lock to prevent obd access after proc file removal. Severity : normal Frequency : Only for files larger than 4GB on 32-bit clients. Bugzilla : 11237 Description: improperly doing page alignment of locks Details : Modify lustre core code to use CFS_PAGE_* defines instead of PAGE_*. Make CFS_PAGE_MASK a 64-bit mask. Severity : normal Frequency : rarely Bugzilla : 11203 Description: RPCs being resent when they shouldn't be Details : Some RPCs that should not be resent are being resent. This can cause inconsistencies in the RPC state machine. Do not resend such requests. Severity : normal Frequency : rare, only with NFS export Bugzilla : 11669 Description: Crash on NFS re-export node Details : under very unusual load conditions an assertion is hit in ll_intent_file_open() Severity : major Frequency : only if OST filesystem is corrupted Bugzilla : 9829 Description: client incorrectly hits assertion in ptlrpc_replay_req() Details : for a short time RPCs with bulk IO are in the replay list, but replay of bulk IOs is unimplemented. If the OST filesystem is corrupted due to disk cache incoherency and then replay is started it is possible to trip an assertion. Avoid putting committed RPCs into the replay list at all to avoid this issue. Severity : normal Frequency : always Bugzilla : 10901 Description: large O_DIRECT requests fail under memory pressure/fragmentation Details : Large single O_DIRECT read and write calls can fail to allocate a sufficiently large buffer to process the request. In case of allocation failure the allocation is retried with a smaller buffer and broken into smaller requests. Severity : normal Frequency : always Bugzilla : 3244 Description: Addition of EXT3_FEATURE_RO_COMPAT_DIR_NLINKS flag for > 32000 subdirectories Details : Add EXT3_FEATURE_RO_COMPAT_DIR_NLINK flag to EXT3_FEATURE_RO_COMPAT_SUPP. This flag will be set whenever subdirectory count crosses 32000. This will aid e2fsck to correctly handle more than 32000 subdirectories. Severity : normal Frequency : always Bugzilla : 11090 Description: versioning check is incomplete Details : Checking the version difference of client vs. server, report error if the gap is too big. Severity : major Bugzilla : 11710 Frequency : always Description: add support PG_writeback bit Details : add support for PG_writeback bit for Lustre, for more carefull work with page cache in 2.6 kernel. This also fix some deadlocks and remove hack for work O_SYNC with 2.6 kernel. ------------------------------------------------------------------------------ 2006-02-09 Cluster File Systems, Inc. * version 1.4.9 * Support for kernels: 2.6.16.21-0.8 (SLES10) 2.6.9-42.0.3EL (RHEL 4) 2.6.5-7.276 (SLES 9) 2.4.21-47.0.1.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) * bug fixes * The backwards-compatible /proc/sys/portals symlink has been removed in this release. Before upgrading, please ensure that you change any configuration scripts or /etc/sysctl.conf files that access /proc/sys/portals/* or sysctl portals.* to use the corresponding entry in /proc/sys/lnet or sysctl lnet.*. This change can be made in advance of the upgrade on any system running Lustre 1.4.6 or newer, since /proc/sys/lnet was added in that version. * Note that reiserfs quotas are temporarily disabled on SLES 10 in this kernel. Severity : critical Frequency : MDS failover only, very rarely Bugzilla : 11125 Description: "went back in time" messages on mds failover Details : The greatest transno may be lost when the current operation finishes with an error (transno==0) and the client's last_rcvd record is over-written. Save the greatest transno in the mds_last_transno for this case. Severity : minor Frequency : always for specific kernels and striping counts Bugzilla : 11042 Description: client may get "Matching packet too big" without ACL support Details : Clients compiled without CONFIG_FS_POSIX_ACL get an error message when trying to access files in certain configurations. The clients should in fact be denied when mounting because they do not understand ACLs. Severity : major Frequency : Cray XT3 with more than 4000 clients and multiple jobs Bugzilla : 10906 Description: many clients connecting with IO in progress causes connect timeouts Details : Avoid synchronous journal commits to avoid delays caused by many clients connecting/disconnecting when bulk IO is in progress. Queue liblustre connect requests on OST_REQUEST_PORTAL instead of OST_IO_PORTAL to avoid delays behind potentially many pending slow IO requests. Severity : normal Frequency : occasionally with multiple writers to a single file Bugzilla : 11081 Description: shared writes to file may result in wrong size reported by stat() Details : Allow growing of kms when extent lock is cancelled Severity : minor Frequency : always with random mmap IO to multi-striped file Bugzilla : 10919 Description: mmap write might be lost if we are writing to a 'hole' in stripe Details : Only if the hole is at the end of OST object so that kms is too small. Fix is to increase kms accordingly in ll_nopage. Severity : normal Frequency : rare, only if OST filesystem is inconsistent with MDS filesystem Bugzilla : 11211 Description: writes to a missing object would leak memory on the OST Details : If there is an inconsistency between the MDS and OST filesystems, such that the MDS references an object that doesn't exist, writes to that object will leak memory due to incorrect cleanup in the error handling path, eventually running out of memory on the OST. Severity : minor Frequency : rare Bugzilla : 11040 Description: Creating too long symlink causes lustre errors Details : Check symlink and name lengths before sending requests to MDS. Severity : normal Frequency : only if flock is enabled (not on by default) Bugzilla : 11415 Description: posix locks not released on fd closure on 2.6.9+ Details : We failed to add posix locks to list of inode locks on 2.6.9+ kernels, this caused such locks not to be released on fd close and then assertions on fs unmount about still used locks. Severity : minor Frequency : MDS failover only, very rarely Bugzilla : 11277 Description: clients may get ASSERTION(granted_lock != NULL) Details : When request was taking a long time, and a client was resending a getattr by name lock request. The were multiple lock requests with the same client lock handle and mds_getattr_name->fixup_handle_for_resent_request found one of the lock handles but later failed with ASSERTION(granted_lock != NULL). Severity : major Frequency : rare Bugzilla : 10891 Description: handle->h_buffer_credits > 0, assertion failure Details : h_buffer_credits is zero after truncate, causing assertion failure. This patch extends the transaction or creates a new one after truncate. Severity : normal Frequency : NFS re-export or patchless client Bugzilla : 11179, 10796 Description: Crash on NFS re-export node (__d_move) Details : We do not want to hash the dentry if we don't have a lock. But if this dentry is later used in d_move, we'd hit uninitialised list head d_hash, so we just do this to init d_hash field but leave dentry unhashed. Severity : normal Frequency : NFS re-export or patchless client Bugzilla : 11135 Description: NFS exports has problem with symbolic link Details : lustre client didn't properly install dentry when re-exported to NFS or running patchless client. Severity : normal Frequency : NFS re-export or patchless client Bugzilla : 10796 Description: Various nfs/patchless fixes. Details : fixes reuse disconected alias for lookup process - this fixes warning "find_exported_dentry: npd != pd", fix permission error with open files at nfs. fix apply umaks when do revalidate. Severity : normal Frequency : occasional Bugzilla : 11191 Description: Crash on NFS re-export node Details : calling clear_page() on the wrong pointer triggered oops in generic_mapping_read(). Severity : normal Frequency : rarely, using O_DIRECT IO Bugzilla : 10903 Description: unaligned directio crashes client with LASSERT Details : check for unaligned buffers before trying any requests. Severity : major Frequency : rarely, using CFS RAID5 patches in non-standard kernel series Bugzilla : 11313 Description: stale data returned from RAID cache Details : If only a small amount of IO is done to the RAID device before reading it again it is possible to get stale data from the RAID cache instead of reading it from disk. Severity : major Frequency : depends on arch, kernel and compiler version, always on sles10 kernel and x86_64 Bugzilla : 11562 Description: recursive or deep enough symlinks cause stack overflow Details : getting rid of large stack-allocated variable in __vfs_follow_link Severity : minor Frequency : depends on hardware Bugzilla : 11540 Description: lustre write performance loss in the SLES10 kernel Details : the performance loss is caused by using of write barriers in the ext3 code. The SLES10 kernel turns barrier support on by default. The fix is to undo that change for ldiskfs. ------------------------------------------------------------------------------ 2006-12-09 Cluster File Systems, Inc. * version 1.4.8 * Support for kernels: 2.6.9-42.0.3EL (RHEL 4) 2.6.5-7.276 (SLES 9) 2.4.21-47.0.1.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) * bug fixes Severity : major Frequency : quota enabled and large files being deleted Bugzilla : 10707 Description: releasing more than 4GB of quota at once hangs OST Details : If a user deletes more than 4GB of files on a single OST it will cause the OST to spin in an infinite loop. Release quota in < 4GB chunks, or use a 64-bit value for 1.4.7.1+. Severity : minor Frequency : rare Bugzilla : 10845 Description: statfs data retrieved from /proc may be stale or zero Details : When reading per-device statfs data from /proc, in the {kbytes,files}_{total,free,avail} files, it may appear as zero or be out of date. Severity : minor Frequency : systems with MD RAID1 external journal devices Bugzilla : 10832 Description: lconf's call to blkid is confused by RAID1 journal devices Details : Use the "blkid -l" flag to locate the MD RAID device instead of returning all block devices that match the journal UUID. Severity : normal Frequency : always, for aggregate stripe size over 4GB Bugzilla : 10725 Description: "lfs setstripe" fails assertion when setting 4GB+ stripe width Details : Using "lfs setstripe" to set stripe size * stripe count over 4GB will fail the kernel with "ASSERTION(lsm->lsm_xfersize != 0)" Severity : minor Frequency : always if "lfs find" used on a local file/directory Bugzilla : 10864 Description: "lfs find" segfaults if used on a local file/directory Details : The case where a directory component was not specified wasn't handled correctly. Handle this properly. Severity : normal Frequency : always on ppc64 Bugzilla : 10634 Description: the write to an ext3 filesystem mounted with mballoc got stuck Details : ext3_mb_generate_buddy() uses find_next_bit() which does not perform endianness conversion. Severity : major Frequency : rarely (truncate to non-zero file size after write under load) Bugzilla : 10730, 10687 Description: Files padded with zeros to next 4K multiple Details : With filesystems mounted using the "extents" option (2.6 kernels) it is possible that files that are truncated to a non-zero size immediately after being written are filled with zero bytes beyond the truncated size. No file data is lost. Severity : enhancement Bugzilla : 10452 Description: Allow recovery/failover for liblustre clients. Details : liblustre clients were unaware of failover configurations until now. Severity : enhancement Bugzilla : 10743 Description: user file locks should fail when not mounting with flock option Details : Set up an error-returning stub in ll_file_operations.lock field to prevent incorrect behaviour when client is mounted without flock option. Also, set up properly f_op->flock field for RHEL4 kernels. Severity : minor Frequency : always on ia64 Bugzilla : 10905 Description: "lfs df" loops on printing out MDS statfs information Details : The obd_ioctl_data was not initialized and in some systems this caused a failure during the ioctl that did not return an error. Initialize the struct and return an error on failure. Severity : minor Frequency : SLES 9 only Bugzilla : 10667 Description: Error of copying files with lustre special EAs as root Details : Client side always return success for setxattr call for lustre special xattr (currently only "trusted.lov"). Severity : normal Frequency : rarely on clusters with both ia64+i386 clients Bugzilla : 10672 Description: ia64+i686 clients doing shared IO on the same file may LBUG Details : In rare cases when both ia64+i686 (or other mixed-PAGE_SIZE) clients are doing concurrent writes to the same file it is possible that the ia64 clients may LASSERT because the OST extent locks are not PAGE_SIZE aligned. Ensure that grown locks are always aligned on the request boundary. Severity : normal Frequency : specific use, occasional Bugzilla : 7040 Description: Overwriting in use executable truncates on-disk binary image Details : If one node attempts to overwrite an executable in use by another node, we now correctly return ETXTBSY instead of truncating the file. Severity : normal Frequency : rare Bugzilla : 2707 Description: chmod on Lustre root is propagated to other clients Details : Re-validate root's dentry in ll_lookup_it to avoid having it invalid by the follow_mount time. Severity : minor Frequency : liblustre clients only Bugzilla : 10883 Description: Race in 'instant cancel' lock handling could lead to such locks never to be granted in case of SMP MDS Details : Do not destroy not yet granted but cbpending locks in handle_enqueue Severity : minor Frequency : replay/resend of open Bugzilla : 10991 Description: non null lock assetion failure in mds_intent_policy Details : Trying to replay/resend lockless open requests resulted in mds_open() returning 0 with no lock. Now it sets a flag if a lock is going to be returned. Severity : enhancement Bugzilla : 10889 Description: Checksum enhancements Details : New checksum enhancements allow for resending RPCs that failed checksum checks. Severity : enhancement Bugzilla : 7376 Description: Tunables on number of dirty pages in cacche Details : Allow to set limit on number of dirty pages cached. Severity : normal Frequency : rare Bugzilla : 10643 Description: client crash on unmount - lock still has references Details : In some error handling cases it was possible to leak a lock reference on a client while accessing a file. This was not harmful to the client during operation, but would cause the client to crash when the filesystem is unmounted. Severity : normal Frequency : specific case, rare Bugzilla : 10921 Description: ETXTBSY on mds though file not in use by client Details : ETXTBSY is no longer incorrectly returned when attempting to chmod or chown a directory that the user previously tried to execute or a currently-executing binary. Severity : major Frequency : extremely rare except on liblustre-based clients Bugzilla : 10480 Description: Lustre space not freed when files are deleted Details : Clean up open-unlinked files after client eviction. Previously the unlink was skipped and the files remained as orphans. Severity : normal Frequency : rare Bugzilla : 10999 Description: OST failure "would be an LBUG" in waiting_locks_callback() Details : In some cases it was possible to send a blocking callback to a client doing a glimpse, even though that client didn't get a lock granted. When the glimpse lock is cancelled on the OST the freed lock is left on the waiting list and corrupted the list. Severity : major Frequency : all core dumps Bugzilla : 11103 Description: Broke core dumps to lustre Details : Negative dentry may be unhashed if parent does not have UPDATE lock, but some callers, e.g. do_coredump, expect dentry to be hashed after successful create, hash it in ll_create_it. ------------------------------------------------------------------------------ 2006-09-13 Cluster File Systems, Inc. * version 1.4.7.1 * Support for kernels: 2.6.9-42.0.2.EL (RHEL 4) 2.6.5-7.276 (SLES 9) 2.4.21-40.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) * bug fix Severity : major Frequency : always on RHEL 3 Bugzilla : 10867 Description: Number of open files grows over time Details : The number of open files grows over time, whether or not Lustre is started. This was due to a filp leak introduced by one of our kernel patches. ------------------------------------------------------------------------------ 2006-08-20 Cluster File Systems, Inc. * version 1.4.7 * Support for kernels: 2.6.9-42.EL (RHEL 4) 2.6.5-7.267 (SLES 9) 2.4.21-40.EL (RHEL 3) 2.6.12.6 vanilla (kernel.org) * bug fixes Severity : major Frequency : rare Bugzilla : 5719, 9635, 9792, 9684 Description: OST (or MDS) trips assertions in (re)connection under heavy load Details : If a server is under heavy load and cannot reply to new connection requests before the client resends the (re)connect, the connection handling code can behave badly if two service threads are concurrently handing separate (re)connections from the same client. Add better locking to the connection handling code, and ensure that only a single connection will be processed for a given client UUID, even if the lock is dropped. Severity : enhancement Bugzilla : 3627 Description: add TCP zero-copy support to kernel Details : Add support to the kernel TCP stack to allow zero-copy bulk sends if the hardware supports scatter-gather and checksumming. This allows socklnd to do client-write and server-read more efficiently and reduce CPU utilization from skbuf copying. Severity : minor Frequency : only if NFS exporting from client Bugzilla : 10258 Description: NULL pointer deref in ll_iocontrol() if chattr mknod file Details : If setting attributes on a file created under NFS that had never been opened it would be possible to oops the client if the file had no objects. Severity : major Frequency : rare Bugzilla : 9326, 10402, 10897 Description: client crash in ptlrpcd_wake() thread when sending async RPC Details : It is possible that ptlrpcd_wake() dereferences a freed async RPC. In rare cases the ptlrpcd thread alread processed the RPC before ptlrpcd_wake() was called and the request was freed. Severity : minor Frequency : always for liblustre Bugzilla : 10290 Description: liblustre client does MDS+OSTs setattr RPC for each write Details : When doing a write from a liblustre client, the client incorrectly issued an RPC to the MDS and each OST the file was striped over in order to update the timestamps. When writing with small chunks and many clients this could overwhelm the MDS with RPCs. In all cases it would slow down the write because these RPCs are unnecessary. Severity : enhancement Bugzilla : 9340 Description: allow number of MDS service threads to be changed at module load Details : It is now possible to change the number of MDS service threads running. Adding "options mds mds_num_threads={N}" to the MDS's /etc/modprobe.conf will set the number of threads for the next time Lustre is restarted (assuming the "mds" module is also reloaded at that time). The default number of threads will stay the same, 32 for most systems. Severity : major Frequency : rare Bugzilla : 10300 Description: OST crash if filesystem is unformatted or corrupt Details : If an OST is started on a device that has never been formatted or if the filesystem is corrupt and cannot even mount then the error handling cleanup routines would dereference a NULL pointer. Severity : normal Frequency : rare Bugzilla : 10047 Description: NULL pointer deref in llap_from_page. Details : get_cache_page_nowait can return a page with NULL (or otherwise incorrect) mapping if the page was truncated/reclaimed while it was searched for. Check for this condition and skip such pages when doing readahead. Introduce extra check to llap_from_page() to verify page->mapping->host is non-NULL (so page is not anonymous). Severity : minor Frequency : Sometimes when using sys_sendfile Bugzilla : 7020 Description: "page not covered by a lock" warnings from ll_readpage Details : sendfile called ll_readpage without right page locks present. Now we introduced ll_file_sendfile that does necessary locking around call to generic_file_sendfile() much like we do in ll_file_read(). Severity : normal Frequency : with certain MDS communication failures at client mount time Bugzilla : 10268 Description: NULL pointer deref after failed client mount Details : a client connection request may delayed by the network layer and not be sent until after the PTLRPC layer has timed out the request. If the client fails the mount immediately it will try to clean up before the network times out the request. Add a reference from the request import to the obd device and delay the cleanup until the network drops the request. Severity : normal Frequency : occasionally during client (re)connect Bugzilla : 9387 Description: assertion failure during client (re)connect Details : processing a client connection request may be delayed by the client or server longer than the client connect timeout. This causes the client to resend the connection request. If the original connection request is replied in this interval, the client may trip an assertion failure in ptlrpc_connect_interpret() which thought it would be the only running connect process. Severity : normal Frequency : only with obd_echo servers and clients that are rebooted Bugzilla : 10140 Description: kernel BUG accessing uninitialized data structure Details : When running an obd_echo server it did not start the ping_evictor thread, and when a client was evicted an uninitialized data structure was accessed. Start the ping_evictor in the RPC service startup instead of the OBD startup. Severity : enhancement Bugzilla : 10193 (patchless) Description: Remove dependency on various unexported kernel interfaces. Details : No longer need reparent_to_init, exit_mm, exit_files, sock_getsockopt, filemap_populate, FMODE_EXEC, put_filp. Severity : minor Frequency : rare (only users of deprecated and unsupported LDAP config) Bugzilla : 9337 Description: write_conf for zeroconf mount queried LDAP incorrectly for client Details : LDAP apparently contains 'lustreName' attributes instead of 'name'. A simple remapping of the name is sufficient. Severity : major Frequency : rare (only with non-default dump_on_timeout debug enabled) Bugzilla : 10397 Description: waiting_locks_callback trips kernel BUG if client is evicted Details : Running with the dump_on_timeout debug flag turned on makes it possible that the waiting_locks_callback() can try to dump the Lustre kernel debug logs from an interrupt handler. Defer this log dumping to the expired_lock_main() thread. Severity : enhancement Bugzilla : 10420 Description: Support NFS exporting on 2.6 kernels. Details : Implement non-rawops metadata methods for NFS server to use without changing NFS server code. Severity : normal Frequency : very rare (synthetic metadata workload only) Bugzilla : 9974 Description: two racing renames might cause an MDS thread to deadlock Details : Running the "racer" program may cause one MDS thread to rename a file from being the source of a rename to being the target of a rename at exactly the same time that another thread is doing so, and the second thread has already enqueued these locks after doing a lookup of the target and is trying to relock them in order. Ensure that we don't try to re-lock the same resource. Severity : major Frequency : only very large systems with liblustre clients Bugzilla : 7304 Description: slow eviction of liblustre clients with the "evict_by_nid" RPC Details : Use asynchronous set_info RPCs to send the "evict_by_nid" to all OSTs in parallel. This allows the eviction of stale liblustre clients to proceed much faster than if they were done in series, and also offers similar improvements for other set_info RPCs. Severity : minor Frequency : common Bugzilla : 10265 Description: excessive CPU usage during initial read phase on client Details : During the initial read phase on a client, it would agressively retry readahead on the file, consuming too much CPU and impacting performance (since 1.4.5.8). Improve the readahead algorithm to avoid this, and also improve some other common cases (read of small files in particular, where "small" is files smaller than /proc/fs/lustre/llite/*/max_read_ahead_whole_mb, 2MB by default). Severity : minor Frequency : rare Bugzilla : 10450 Description: MDS crash when receiving packet with unknown intent. Details : Do not LBUG in unknown intent case, just return -EFAULT Severity : enhancement Bugzilla : 9293, 9385 Description: MDS RPCs are serialised on client. This is unnecessary for some. Details : Do not serialize getattr (non-intent version) and statfs. Severity : minor Frequency : occasional, when OST network is overloaded/intermittent Bugzilla : 10416 Description: client evicted by OST after bulk IO timeout Details : If a client sends a bulk IO request (read or write) the OST may evict the client if it is unresposive to its data GET/PUT request. This is incorrect if the network is overloaded (takes too long to transfer the RPC data) or dropped the OST GET/PUT request. There is no need to evict the client at all, since the pinger and/or lock callbacks will handle this, and the client can restart the bulk request. Severity : minor Frequency : Always when mmapping file with no objects Bugzilla : 10438 Description: client crashes when mmapping file with no objects Details : Check that we actually have objects in a file before doing any operations on objects in ll_vm_open, ll_vm_close and ll_glimpse_size. Severity : minor Frequency : Rare Bugzilla : 10484 Description: Request leak when working with deleted CWD Details : Introduce advanced request refcount tracking for requests referenced from lustre intent. Severity : Enhancement Bugzilla : 10482 Description: Cache open file handles on client. Details : MDS now will return special lock along with openhandle, if requested and client is allowed to hold openhandle, even if unused, until such a lock is revoked. Helps NFS a lot, since NFS is opening closing files for every read/write openration. Severity : Enhancement Bugzilla : 9291 Description: Cache open negative dentries on client when possible. Details : Guard negative dentries with UPDATE lock on parent dir, drop negative dentries on lock revocation. Severity : minor Frequency : Always Bugzilla : 10510 Description: Remounting a client read-only wasn't possible with a zconf mount Details : It wasn't possible to remount a client read-only with llmount. Severity : enhancement Description: Include MPICH 1.2.6 Lustre ADIO interface patch Details : In lustre/contrib/ or /usr/share/lustre in RPM a patch for MPICH is included to add Lustre-specific ADIO interfaces. This is based closely on the UFS ADIO layer and only differs in file creation, in order to allow the OST striping to be set. This is user-contributed code and not supported by CFS. Severity : minor Frequency : Always Bugzilla : 9486 Description: extended inode attributes (immutable, append-only) work improperly when 2.4 and 2.6 kernels are used on client/server or vice versa Details : Introduce kernel-independent values for these flags. Severity : enhancement Frequency : Always Bugzilla : 10248 Description: Allow fractional MB tunings for lustre in /proc/ filesystem. Details : Many of the /proc/ tunables can only be tuned at a megabyte granularity. Now, Fractional MB granularity is be supported, this is very useful for low memory system. Severity : enhancement Bugzilla : 9292 Description: Getattr by fid Details : Getting a file attributes by its fid, obtaining UPDATE|LOOKUP locks, avoids extra getattr rpc requests to MDS, allows '/' to have locks and avoids getattr rpc requests for it on every stat. Severity : major Frequency : Always, for filesystems larger than 2TB Bugzilla : 6191 Description: ldiskfs crash at mount for filesystem larger than 2TB with mballoc Details : Kenrel kmalloc limits allocations to 128kB and this prevents filesystems larger than 2TB to be mounted with mballoc enabled. Severity : critical Frequency : Always, for 32-bit kernel without CONFIG_LBD and filesystem > 2TB Bugzilla : 6191 Description: filesystem corruption for non-standard kernels and very large OSTs Details : If a 32-bit kernel is compiled without CONFIG_LBD enabled and a filesystems larger than 2TB is mounted then the kernel will silently corrupt the start of the filesystem. CONFIG_LBD is enabled for all CFS-supported kernels, but the possibility of this happening with a modified kernel config exists. Severity : enhancement Bugzilla : 10462 Description: add client O_DIRECT support for 2.6 kernels Details : It is now possible to do O_DIRECT reads and writes to files in the Lustre client mountpoint on 2.6 kernel clients. Severity : enhancement Bugzilla : 10446 Description: parallel glimpse, setattr, statfs, punch, destroy requests Details : Sends glimpse, setattr, statfs, punch, destroy requests to OSTs in parallel, not waiting for response from every OST before sending a rpc to the next OST. Severity : minor Frequency : rare Bugzilla : 10150 Description: setattr vs write race when updating file timestamps Details : Client processes that update a file timestamp into the past right after writing to the file (e.g. tar) it is possible that the updated file modification time can be reset to the current time due to a race between processing the setattr and write RPC. Severity : enhancement Bugzilla : 10318 Description: Bring 'lfs find' closer in line with regular Linux find. Details : lfs find util supports -atime, -mtime, -ctime, -maxdepth, -print, -print0 options and obtains all the needed info through the lustre ioctls. Severity : enhancement Bugzilla : 6221 Description: support up to 1024 configured devices on one node Details : change obd_dev array from statically allocated to dynamically allocated structs as they are first used to reduce memory usage Severity : minor Frequency : rare Bugzilla : 10437 Description: Flush dirty partially truncated pages during truncate Details : Immediatelly flush partially truncated pages in filter_setattr, this way we completely avoid having any pages in page cache on OST and can retire ugly workarounds during writes to flush such pages. Severity : minor Frequency : rare Bugzilla : 10409 Description: i_sem vs transaction deadlock in mds_obd_destroy during unlink. Details : protect inode from truncation within vfs_unlink() context just take a reference before calling vfs_unlink() and release it when parent's i_sem is free. Severity : minor Frequency : always, if extents are used on OSTs Bugzilla : 10703 Description: index ei_leaf_hi (48-bit extension) is not zeroed in extent index Details : OSTs using the extents format would not zero the high 16 bits of the index physical block number. This is not a problem for any OST filesystems smaller than 16TB, and no kernels support ext3 filesystems larger than 16TB yet. This is fixed in 1.4.7 (all new/modified files) and can be fixed for existing filesystems with e2fsprogs-1.39-cfs1. Severity : minor Frequency : rare Bugzilla : 9387 Description: import connection selection may be incorrect if timer wraps Details : Using a 32-bit jiffies timer with HZ=1000 may cause backup import connections to be ignored if the 32-bit jiffies counter wraps. Use a 64-bit jiffies counter. Severity : major Frequency : during server recovery Bugzilla : 10479 Description: crash after server is denying duplicate export Details : If clients are resending connect requests to the server, the server refuses to allow a client to connect multiple times. Fixed a bug in the handling of this case. Severity : minor Frequency : very large clusters immediately after boot Bugzilla : 10083 Description: LNET request buffers exhausted under heavy short-term load Details : If a large number of client requests are generated on a service that has previously never seen so many requests it is possible that the request buffer growth cannot keep up with the spike in demand. Instead of dropping incoming requests, they are held in the LND until the RPC service can accept more requests. Severity : minor Frequency : Sometimes during replay Bugzilla : 9314 Description: Assertion failure in ll_local_open after replay. Details : If replay happened on an open request reply before we were able to set replay handler, reply will become not swabbed tripping the assertion in ll_local_open. Now we set the handler right after recognising of open request Severity : minor Frequency : very rare Bugzilla : 10584 Description: kernel reports "badness in vsnprintf" Details : Reading from the "recovery_status" /proc file in small chunks may cause a negative length in lprocfs_obd_rd_recovery_status() call to vsnprintf() (which is otherwise harmless). Exit early if there is no more space in the output buffer. Severity : enhancement Bugzilla : 2259 Description: clear OBD RPC statistics by writing to them Details : It is now possible to clear the OBD RPC statistics by writing to the "stats" file. Severity : minor Frequency : rare Bugzilla : 10641 Description: Client mtime is not the same on different clients after utimes Details : In some cases, the client was using the utimes() syscall on a file cached on another node. The clients now validate the ctime from the MDS + OSTs to determine which one is right. Severity : minor Frequency : always Bugzilla : 10611 Description: Inability to activate failout mode Details : lconf script incorrectly assumed that in python string's numeric value is used in comparisons. Severity : minor Frequency : always with multiple stripes per file Bugzilla : 10671 Description: Inefficient object allocation for mutli-stripe files Details : When selecting which OSTs to stripe files over, for files with a stripe count that divides evenly into the number of OSTs, the MDS is always picking the same starting OST for each file. Return the OST selection heuristic to the original design. Severity : minor Frequency : rare Bugzilla : 10673 Description: mount failures may take full timeout to return an error Details : Under some heavy load conditions it is possible that a failed mount can wait for the full obd_timeout interval, possibly several minutes, before reporting an error. Instead return an error as soon as the status is known. ------------------------------------------------------------------------------ 2006-02-14 Cluster File Systems, Inc. * version 1.4.6 * WIRE PROTOCOL CHANGE. This version of Lustre networking WILL NOT INTEROPERATE with older versions automatically. Please read the user documentation before upgrading any part of a live system. * WARNING: Lustre networking configuration changes are required with this release. See https://bugzilla.clusterfs.com/show_bug.cgi?id=10052 for details. * bug fixes * Support for kernels: 2.6.9-22.0.2.EL (RHEL 4) 2.6.5-7.244 (SLES 9) 2.6.12.6 vanilla (kernel.org) Severity : enhancement Bugzilla : 7981/8208 Description: Introduced Lustre Networking (LNET) Details : LNET is new networking infrastructure for Lustre, it includes a reorganized network configuration mode (see the user documentation for full details) as well as support for routing between different network fabrics. Lustre Networking Devices (LNDs) for the supported network fabrics have also been created for this new infrastructure. Severity : enhancement Description: Introduced Access control lists Details : clients can set ACLs on files and directories in order to have more fine-grained permissions than the standard Unix UGO+RWX. The MDS must be started with the "-o acl" mount option. Severity : enhancement Description: Introduced filesystem quotas Details : Administrators may now establish per-user quotas on the filesystem. Severity : enhancement Bugzilla : 7982 Description: Configuration change for the XT3 The PTLLND is now used to run Lustre over Portals on the XT3 The configure option(s) --with-cray-portals are no longer used. Rather --with-portals= is used to enable building on the XT3. In addition to enable XT3 specific features the option --enable-cray-xt3 must be used. Severity : major Frequency : rare Bugzilla : 7407 Description: Running on many-way SMP OSTs can trigger oops in llcd_send() Details : A race between allocating a new llcd and re-getting the llcd_lock allowed another thread to grab newly-allocated llcd. Severity : enhancement Bugzilla : 7116 Description: 2.6 OST async journal commit and locking fix to improve performance Details : The filter_direct_io()+filter_commitrw_write() journal commits for 2.6 kernels are now async as they already were in 2.4 kernels so that they can commit concurrently with the network bulk transfer. For block-allocated files the filter allocation semaphore is held to avoid filesystem fragmentation during allocation. BKL lock removed for 2.6 xattr operations where it is no longer needed. Severity : minor Frequency : rare Bugzilla : 8320 Description: lconf incorrectly determined whether two IP networks could talk Details : In some more complicated routing and multiple-network configurations, lconf will avoid trying to make a network connection to a disjoint part of the IP space. It was doing the math incorrectly for one set of cases. Severity : major Frequency : rare Bugzilla : 7359 Description: Fix for potential infinite loop processing records in an llog. Details : If an llog record is corrupted/zeroed, it is possible to loop forever in llog_process(). Validate the llog record length and skip the remainder of the block on error. Severity : minor Frequency : occasional (liblustre only) Bugzilla : 6363 Description: liblustre could not open files whose last component is a symlink Details : sysio_path_walk() would incorrectly pass the open intent to intermediate path components. Severity : minor Frequency : rare (liblustre only with non-standard tuning) Bugzilla : 7201 (7350) Description: Tuning the MDC DLM LRU size to zero triggers client LASSERT Details : llu_lookup_finish_locks() tries to set lock data on a lock after it has been released, only do this for referenced locks Severity : enhancement Bugzilla : 7328 Description: specifying an (invalid) directory default stripe_size of -1 would reset the directory default striping Details : stripe_size -1 was used internally to signal directory stripe removal, now use "all default" to signal dir stripe removal as a directory striping of "all default" is not useful Severity : minor Frequency : common for large clusters running liblustre clients Bugzilla : 7198 Description: doing an ls when liblustre clients are running is slow Details : sending a glimpse AST to a liblustre client waits for every AST to time out, as liblustre clients will not respond. Since they cannot cache data we refresh the OST lock LVB from disk instead. Severity : enhancement Bugzilla : 7198 Description: doing an ls at the same time as file IO can be slow Details : enqueue and other "small" requests can be blocked behind many large IO requests. Create a new OST IO portal for non-IO requests so they can be processed faster. Severity : minor Frequency : rare (only HPUX clients mounting unsupported re-exported NFS vol) Bugzilla : 5781 Description: an HPUX NFS client would get -EACCESS when ftruncate()ing a newly created file with mode 000 Details : the Linux NFS server relies on an MDS_OPEN_OWNEROVERRIDE hack to allow an ftruncate() as a non-root user to a file with mode 000. Lustre now respects this flag to disable mode checks when truncating a file owned by the user Severity : minor Frequency : liblustre-only, when liblustre client dies unexpectedly or becomes busy Bugzilla : 7313 Description: Revoking locks from clients that went dead or catatonic might take a lot of time. Details : New lock flags FL_CANCEL_ON_BLOCK used by liblustre makes cancellation of such locks instant on servers without waiting for any reply from clients. Clients drops these locks when cancel notification from server is received without replying. Severity : minor Frequency : liblustre-only, when liblustre client dies or becomes busy Bugzilla : 7311 Description: Doing ls on Linux clients can take a long time with active liblustre clients Details : Liblustre client cannot handle ASTs in timely manner, so avoid granting such locks to it in the first place if possible. Locks are taken by proxy on the OST during the read or write and dropped immediately afterward. Add connect flags handling, do not grant locks to liblustre clients for glimpse ASTs. Severity : enhancement Bugzilla : 6252 Description: Improve read-ahead algorithm to avoid excessive IO for random reads Details : Existing read-ahead algorithm is tuned for the case of streamlined sequential reads and behaves badly with applications doing random reads. Improve it by reading ahead at least read region, and avoiding excessive large RPC for small reads. Severity : enhancement Bugzilla : 8330 Description: Creating more than 1000 files for a single job may cause a load imbalance on the OSTs if there are also a large number of OSTs. Details : qos_prep_create() uses an OST index reseed value that is an even multiple of the number of available OSTs so that if the reseed happens in the middle of the object allocation it will still utilize the OSTs as uniformly as possible. Severity : major Frequency : rare Bugzilla : 8322 Description: OST or MDS may oops in ping_evictor_main() Details : ping_evictor_main() drops obd_dev_lock if deleting a stale export but doesn't restart at beginning of obd_exports_timed list afterward. Severity : enhancement Bugzilla : 7304 Description: improve by-nid export eviction on the MDS and OST Details : allow multiple exports with the same NID to be evicted at one time without re-searching the exports list. Severity : major Frequency : rare, only with supplementary groups enabled on SMP 2.6 kernels Bugzilla : 7273 Description: MDS may oops in groups_free() Details : in rare race conditions a newly allocated group_info struct is freed again, and this can be NULL. The 2.4 compatibility code for groups_free() checked for a NULL pointer, but 2.6 did not. Severity : minor Frequency : common for liblustre clients doing little filesystem IO Bugzilla : 9352, 7313 Description: server may evict liblustre clients accessing contended locks Details : if a client is granted a lock or receives a completion AST with a blocking AST already set it would not reply to the AST for LDLM_FL_CANCEL_ON_BLOCK locks. It now replies to such ASTs. Severity : minor Frequency : lfs setstripe, only systems with more than 160 OSTs Bugzilla : 9440 Description: unable to set striping with a starting offset beyond OST 160 Details : llapi_create_file() incorrectly limited the starting stripe index to the maximum single-file stripe count. Severity : minor Frequency : LDAP users only Bugzilla : 6163 Description: lconf did not handle in-kernel recovery with LDAP properly Details : lconf/LustreDB get_refs() is searching the wrong namespace Severity : enhancement Bugzilla : 7342 Description: bind OST threads to NUMA nodes to improve performance Details : all OST threads are uniformly bound to CPUs on a single NUMA node and do their allocations there to localize memory access Severity : enhancement Bugzilla : 7979 Description: llmount can determine client NID directly from Myrinet (GM) Details : the client NID code from gmnalnid was moved directly into llmount, removing the need to use this or specifying the client NID explicitly when mounting GM clients with zeroconf Severity : minor Frequency : if client is started with down MDS Bugzilla : 7184 Description: if client is started with down MDS mount hangs in ptlrpc_queue_wait Details : Having an LWI_INTR() wait event (interruptible, but no timeout) will wait indefinitely in ptlrpc_queue_wait->l_wait_event() after ptlrpc_import_delayed_req() because we didn't check if the request was interrupted, and we also didn't break out of the event loop if there was no timeout Severity : major Frequency : rare Bugzilla : 5047 Description: data loss during non-page-aligned writes to a single file from both multiple nodes and multiple threads on one node at same time Details : updates to KMS and lsm weren't protected by common lock. Resulting inconsistency led to false short-reads, that were cached and later used by ->prepare_write() to fill in partially written page, leading to data loss. Severity : minor Frequency : always, if lconf --abort_recovery used Bugzilla : 7047 Description: lconf --abort_recovery fails with 'Operation not supported' Details : lconf was attempting to abort recovery on the MDT device and not the MDS device Severity : enhancement Bugzilla : 9445 Description: remove cleanup logs Details : replace lconf-generated cleanup logs with lustre internal cleanup routines. Eliminates the need for client-cleanup and mds-cleanup logs. Severity : enhancement Bugzilla : 8592 Description: add support for EAs (user and system) on lustre filesystems Details : it is now possible to store extended attributes in the Lustre client filesystem, and with the user_xattr mount option it is possible to allow users to store EAs on their files also Severity : enhancement Bugzilla : 7293 Description: Add possibility (config option) to show minimal available OST free space. Details : When compiled with --enable-mindf configure option, statfs(2) (and so, df) will return least minimal free space available from all OSTs as amount of free space on FS, instead of summary of free spaces of all OSTs. Severity : enhancement Bugzilla : 7311 Description: do not expand extent locks acquired on OST-side Details : Modify ldlm_extent_policy() to not expand local locks, acquired by server: they are not cached anyway. Severity : major Frequency : when mmap is used/binaries executed from Lustre Bugzilla : 9482 Description: Unmmap pages before throwing them away from read cache. Details : llap_shrink cache now attempts to unmap pages before discarding them (if unmapping failed - do not discard). SLES9 kernel has extra checks that trigger if this unmapping is not done first. Severity : minor Frequency : rare Bugzilla : 6034 Description: lconf didn't resolve symlinks before checking to see whether a given mountpoint was already in use Severity : minor Frequency : when migrating failover services Bugzilla : 6395, 9514 Description: When migrating a subset of services from a node (e.g. failback from a failover service node) the remaining services would time out and evict clients. Details : lconf --force (implied by --failover) sets the global obd_timeout to 5 seconds in order to quickly disconnect, but this caused other RPCs to time out too quickly. Do not change the global obd_timeout for force cleanup, only set it for DISCONNECT RPCs. Severity : enhancement Frequency : if MDS is started with down OST Bugzilla : 9439,5706 Description: Allow startup/shutdown of an MDS without depending on the availability of the OSTs. Details : Asynchronously call mds_lov_synchronize during MDS startup. Add appropriate locking and lov-osc refcounts for safe cleaning. Add osc abort_inflight calls in case the synchronize never started. Severity : minor Frequency : occasional (Cray XT3 only) Bugzilla : 7305 Description: root not authorized to access files in CRAY_PORTALS environment Details : The client process capabilities were not honoured on the MDS in a CRAY_PORTALS/CRAY_XT3 environment. If the file had previously been accessed by an authorized user then root was able to access the file on the local client also. The root user capabilities are now allowed on the MDS, as this environment has secure UID. Severity : minor Frequency : occasional Bugzilla : 6449 Description: ldiskfs "too long searching" message happens too often Details : A debugging message (otherwise harmless) prints too often on the OST console. This has been reduced to only happen when there are fragmentation problems on the filesystem. Severity : minor Frequency : rare Bugzilla : 9598 Description: Division by zero in statfs when all OSCs are inactive Details : lov_get_stripecnt() returns zero due to incorrect order of checks, lov_statfs divides by value returned by lov_get_stripecnt(). Severity : minor Frequency : common Bugzilla : 9489, 3273 Description: First write from each client to each OST was only 4kB in size, to initialize client writeback cache, which caused sub-optimal RPCs and poor layout on disk for the first writen file. Details : Clients now request an initial cache grant at (re)connect time and so that they can start streaming writes to the cache right away and always do full-sized RPCs if there is enough data. If the OST is rebooted the client also re-establishes its grant so that client cached writes will be honoured under the grant. Severity : minor Frequency : common Bugzilla : 7198 Description: Slow ls (and stat(2) syscall) on files residing on IO-loaded OSTs Details : Now I/O RPCs go to different portal number and (presumably) fast lock requests (and glimses) and other RPCs get their own service threads pool that should be able to service those RPCs immediatelly. Severity : enhancement Bugzilla : 7417 Description: Ability to exchange lustre version between client and servers and issue warnings at client side if client is too old. Also for liblustre clients there is ability to refuse connection of too old clients. Details : New 'version' field is added to connect data structure that is filled with version info. That info is later checked by server and by client. Severity : minor Frequency : rare, liblustre only. Bugzilla : 9296, 9581 Description: Two simultaneous writes from liblustre at offset within same page might proceed at the same time overwriting eachother with stale data. Details : I/O lock withing llu_file_prwv was released too early, before data actually was hitting the wire. Extended lock-holding time until server acknowledges receiving data. Severity : minor Frequency : extremely rare. Never observed in practice. Bugzilla : 9652 Description: avoid generating lustre_handle cookie of 0. Details : class_handle_hash() generates handle cookies by incrementing global counter, and can hit 0 occasionaly (this is unlikely, but not impossible, because initial value of cookie counter is selected randonly). Value of 0 is used as a sentinel meaning "unassigned handle" --- avoid it. Also coalesce two critical sections in this function into one. Severity : enhancement Bugzilla : 9528 Description: allow liblustre clients to delegate truncate locking to OST Details : To avoid overhead of locking, liblustre client instructs OST to take extent lock in ost_punch() on client's behalf. New connection flag is added to handle backward compatibility. Severity : enhancement Bugzilla : 4928, 7341, 9758 Description: allow number of OST service threads to be specified Details : a module parameter allows the number of OST service threads to be specified via "options ost ost_num_threads={N}" in the OSS's /etc/modules.conf or /etc/modprobe.conf. Severity : major Frequency : rare Bugzilla : 6146, 9635, 9895 Description: servers crash with bad pointer in target_handle_connect() Details : In rare cases when a client is reconnecting it was possible that the connection request was the last reference for that export. We would temporarily drop the export reference and get a new one, but this may have been the last reference and the export was just destroyed. Get new reference before dropping old one. Severity : enhancement Frequency : if client is started with failover MDS Bugzilla : 9818 Description: Allow multiple MDS hostnames in the mount command Details : Try to read the configuration from all specified MDS hostnames during a client mount in case the "primary" MDS is down. Severity : enhancement Bugzilla : 9297 Description: Stop sending data to evicted clients as soon as possible. Details : Check if the client we are about to send or are sending data to was evicted already. (Check is done every second of waiting, for which l_wait_event interface was extended to allow checking of exit condition at specified intervals). Severity : minor Frequency : rare, normally only when NFS exporting is done from client Bugzilla : 9301 Description: 'bad disk LOV MAGIC: 0x00000000' error when chown'ing files without objects Details : Make mds_get_md() recognise empty md case and set lmm size to 0. Severity : minor Frequency : always, if srand() is called before liblustre initialization Bugzilla : 9794 Description: Liblustre uses system PRNG disturbing its usage by user application Details : Introduce internal to lustre fast and high-quality PRNG for lustre usage and make liblustre and some other places in generic lustre code to use it. Severity : enhancement Bugzilla : 9477, 9557, 9870 Description: Verify that the MDS configuration logs are updated when xml is Details : Check if the .xml configuration logs are newer than the config logs stored on the MDS and report an error if this is the case. Request --write-conf, or allow starting with --old_conf. Severity : enhancement Bugzilla : 6034 Description: Handle symlinks in the path when checking if Lustre is mounted. Details : Resolve intermediate symlinks when checking if a client has mounted a filesystem to avoid duplicate client mounts. Severity : minor Frequency : rare Bugzilla : 9309 Description: lconf can hit an error exception but still return success. Details : The lconf command catches the Command error exception at the top level script context and will exit with the associated exit status, but doesn't ensure that this exit status is non-zero. Severity : minor Frequency : rare Bugzilla : 9493 Description: failure of ptlrpc thread startup can cause oops Details : Starting a ptlrpc service thread can fail if there are a large number of threads or the server memory is very fragmented. Handle this without oopsing. Severity : minor Frequency : always, only if liblustre and non-default acceptor port was used Bugzilla : 9933 Description: liblustre cannot connect to servers with non-default acceptor port Details : tcpnal_set_default_params() was not called and was therefore ignoring the environment varaible TCPNAL_PORT, as well as other TCPNAL_ environment variables Severity : minor Frequency : rare Bugzilla : 9923 Description: two objects could be created on the same OST for a single file Details : If an OST is down, in some cases it was possible to create two objects on a single OST for a single file. No problems other than potential performance impact and spurious error messages. Severity : minor Frequency : rare Bugzilla : 5681, 9562 Description: Client may oops in ll_unhash_aliases Details : Client dcache may become inconsistent in race condition. In some cases "getcwd" can fail if the current directory is modified. Severity : minor Frequency : always Bugzilla : 9942 Description: Inode refcounting problems in NFS export code Details : link_raw functions used to call d_instantiate without obtaining extra inode reference first. Severity : minor Frequency : rare Bugzilla : 9942, 9903 Description: Referencing freed requests leading to crash, memleaks with NFS. Details : We used to require that call to ll_revalidate_it was always followed by ll_lookup_it. Also with revalidate_special() it is possible to call ll_revalidate_it() twice for the same dentry even if first occurence returned success. This fix changes semantic between DISP_ENQ_COMPLETE disposition flag to mean there is extra reference on a request referred from the intent. ll_intent_release() then releases such a request. Severity : minor Frequency : rare, normally benchmark loads only Bugzilla : 1443 Description: unlinked inodes were kept in memory on the client Details : If a client is repeatedly creating and unlinking files it can accumulate a lot of stale inodes in the inode slab cache. If there is no other client load running this can cause the client node to run out of memory. Instead flush old inodes from client cache that have the same inode number as a new inode. Severity : minor Frequency : SLES9 2.6.5 kernel and long filenames only Bugzilla : 9969, 10379 Description: utime reports stale NFS file handle Details : SLES9 uses out-of-dentry names in some cases, which confused the lustre dentry revalidation. Change it to always use the in-dentry qstr. Severity : major Frequency : rare, unless heavy write-truncate concurrency is continuous Bugzilla : 4180, 6984, 7171, 9963, 9331 Description: OST becomes very slow and/or deadlocked during object unlink Details : filter_destroy() was holding onto the parent directory lock while truncating+unlinking objects. For very large objects this may block other threads for a long time and slow overall OST responsiveness. It may also be possible to get a lock ordering deadlock in this case, or run out of journal credits because of the combined truncate+unlink. Solution is to do object truncate first in one transaction without parent lock, and then do the final unlink in a new transaction with the parent lock. This reduces the lock hold time dramatically. Severity : major Frequency : rare, 2.4 kernels only Bugzilla : 9967 Description: MDS or OST cleanup may trip kernel BUG when dropping kernel lock Details : mds_cleanup() and filter_cleanup() need to drop the kernel lock before unmounting their filesystem in order to avoid deadlock. The kernel_locked() function in 2.4 kernels only checks whether the kernel lock is held, not whether it is this process that is holding it as 2.6 kernels do. Severity : major Frequency : rare Bugzilla : 9635 Description: MDS or OST may oops/LBUG if a client is connecting multiple times Details : The client ptlrpc code may be trying to reconnect to a down server before a previous connection attempt has timed out. Increase the reconnect interval to be longer than the connection timeout interval to avoid sending duplicate connections to servers. Severity : minor Frequency : echo_client brw_test command Bugzilla : 9919 Description: fix echo_client to work with OST preallocated code Details : OST preallocation code (5137) didn't take echo_client IO path into account: echo_client calls filter methods outside of any OST thread and, hence, there is no per-thread preallocated pages and buffers to use. Solution: hijack pga pages for IO. As a byproduct, this avoids unnecessary data copying. Severity : minor Frequency : rare Bugzilla : 3555, 5962, 6025, 6155, 6296, 9574 Description: Client can oops in mdc_commit_close() after open replay Details : It was possible for the MDS to return an open request with no transaction number in mds_finish_transno() if the client was evicted, but without actually returning an error. Clients would later try to replay that open and may trip an assertion Simplify the client close codepath, and always return an error from the MDS in case the open is not successful. Severity : major Frequency : rare, 2.6 OSTs only Bugzilla : 10076 Description: OST may deadlock under high load on fragmented files Details : If there was a heavy load and highly-fragmented OST filesystems it was possible to have all the OST threads deadlock waiting on allocation of biovecs, because the biovecs were not released until the entire RPC IO was completed. Instead, release biovecs as soon as they are complete to ensure forward IO progress. Severity : enhancement Bugzilla : 9578 Description: Support for specifying external journal device at mount Details : If an OST or MDS device is formatted with an external journal device, this device major/minor is stored in the ext3 superblock and may not be valid for failover. Allow detecting and specifying the external journal at mount time. Severity : major Frequency : rare Bugzilla : 10235 Description: Mounting an MDS with pending unlinked files may cause oops Details : target_finish_recovery() calls mds_postrecov() which returned the number of orphans unlinked. mds_lov_connect->mds_postsetup() considers this an error and immediately begins cleaning up the lov, just after starting the mds_lov process Severity : enhancement Bugzilla : 9461 Description: Implement 'lfs df' to report actual free space on per-OST basis Details : Add sub-command 'df' on 'lfs' to report the disk space usage of MDS/OSDs. Usage: lfs df [-i][-h]. Command Options: '-i' to report usage of objects; '-h' to report in human readable format. ------------------------------------------------------------------------------ 2005-08-26 Cluster File Systems, Inc. * version 1.4.5 * bug fixes Severity : major Frequency : rare Bugzilla : 7264 Description: Mounting an ldiskfs file system with mballoc may crash OST node. Details : ldiskfs mballoc code may reference an uninitialized buddy struct at startup during orphan unlinking. Instead, skip buddy update before setup, as it will be regenerated after recovery is complete. Severity : minor Frequency : rare Bugzilla : 7039 Description: If an OST is inactive, its locks might reference stale inodes. Details : lov_change_cbdata() must iterate over all namespaces, even if they are inactive to clear inode references from the lock. Severity : enhancement Frequency : occasional, if non-standard max_dirty_mb used Bugzilla : 7138 Description: Client will block write RPCs if not enough grant Details : If a client has max_dirty_mb smaller than max_rpcs_in_flight, then the client will block writes while waiting for another RPC to complete instead of consuming its dirty limit. With change we get improved performance when max_dirty_mb is small. Severity : enhancement Bugzilla : 3389, 6253 Description: Add support for supplementary groups on the MDS. Details : The MDS has an upcall /proc/fs/lustre/mds/{mds}/group_upcall (set to /usr/sbin/l_getgroups if enabled) which will do MDS-side lookups for user supplementary groups into a cache. Severity : minor Bugzilla : 7278 Description: O_CREAT|O_EXCL open flags in liblustre always return -EEXIST Details : Make libsysio to not enforce O_EXCL by clearing the flag, for liblustre O_EXCL is enforced by MDS. Severity : minor Bugzilla : 6455 Description: readdir never returns NULL in liblustre. Details : Corrected llu_iop_getdirentries logic, to return offset of next dentry in struct dirent. Severity : minor Bugzilla : 7137 Frequency : liblustre only, depends on application IO pattern Description: liblustre clients evicted if not contacting servers Details : Don't put liblustre clients into the ping_evictor list, so they will not be evicted by the pinger ever. Severity : enhancement Bugzilla : 6902 Description: Add ability to evict clients by NID from MDS. Details : By echoing "nid:$NID" string into /proc/fs/lustre/mds/.../evict_client client with nid that equals to $NID would be instantly evicted from this MDS and from all active OSTs connected to it. Severity : minor Bugzilla : 7198 Description: Do not query file size twice, somewhat slowing stat(2) calls. Details : lookup_it_finish() used to query file size from OSTs that was not needed. Severity : minor Bugzilla : 6237 Description: service threads change working directory to that of init Details : Starting lustre service threads may pin the working directory of the parent thread, making that filesystem busy. Threads now change to the working directory of init to avoid this. Severity : minor Bugzilla : 6827 Frequency : during shutdown only Description: shutdown with a failed MDS or OST can cause unmount to hang Details : Don't resend DISCONNECT messages in ptlrpc_disconnect_import() if server is down. Severity : minor Bugzilla : 7331 Frequency : 2.6 only Description: chmod/chown may include an extra supplementary group Details : ll{,u}_mdc_pack_op_data() does not properly initialize the supplementary group and if none is specified this is used. Severity : minor Bugzilla : 5479 (6816) Frequency : rare Description: Racing open + rm can assert client in mdc_set_open_replay_data() Details : If lookup is in progress on a file that is unlinked we might try to revalidate the inode and fail in revalidate after lookup is complete and ll_file_open() enqueues the open again but it_open_error() was not checking DISP_OPEN_OPEN errors correctly. Severity : minor Frequency : always, if lconf --abort_recovery used Bugzilla : 7047 Description: lconf --abort_recovery fails with 'Operation not supported' Details : lconf was attempting to abort recovery on the MDT device and not the MDS device ------------------------------------------------------------------------------ 2005-08-08 Cluster File Systems, Inc. * version 1.4.4 * bug fixes Severity : major Frequency : rare (only unsupported configurations with a node running as an OST and a client) Bugzilla : 6514, 5137 Description: Mounting a Lustre file system on a node running as an OST could lead to deadlocks Details : OSTs now preallocates memory needed to write out data at startup, instead of when needed, to avoid having to allocate memory in possibly low memory situations. Specifically, if the file system is mounted on on OST, memory pressure could force it to try to write out data, which it needed to allocate memory to do. Due to the low memory, it would be unable to do so and the node would become unresponsive. Severity : enhancement Bugzilla : 7015 Description: Addition of lconf --service command line option Details : lconf now accepts a '--service ' option, which is shorthand for 'lconf --group --select =' Severity : enhancement Bugzilla : 6101 Description: Failover mode is now the default for OSTs. Details : By default, OSTs will now run in failover mode. To return to the old behaviour, add '--failout' to the lmc line for OSTs. Severity : enhancement Bugzilla : 1693 Description: Health checks are now provided for MDS and OSTs Details : Additional detailed health check information on MSD and OSTs is now provided through the procfs health_check value. Severity : minor Frequency : occasional, depends on IO load Bugzilla : 4466 Description: Disk fragmentation on the OSTs could eventually cause slowdowns after numerous create/delete cycles Details : The ext3 inode allocation policy would not allocate new inodes very well on the OSTs because there are no new directories being created. Instead we look for groups with free space if the parent directories are nearly full. Severity : major Bugzilla : 6302 Frequency : rare Description: Network or server problems during mount may cause partially mounted clients instead of returning an error. Details : The config llog parsing code may overwrite the error return code during mount error handling, returning success instead of an error. Severity : minor Bugzilla : 6422 Frequency : rare Description: MDS can fail to allocate large reply buffers Details : After long uptimes the MDS can fail to allocate large reply buffers (e.g. zconf client mount config records) due to memory fragmentation or consumption by the buffer cache. Preallocate some large reply buffers so that these replies can be sent even under memory pressure. Severity : minor Bugzilla : 6266 Frequency : rare (liblustre) Description: fsx running with liblustre complained that using truncate() to extend the file doesn't work. This patch corrects that issue. Details : This is the liblustre equivalent of the fix for bug 6196. Fixes ATTR_SIZE and lsm use in llu_setattr_raw. Severity : critical Bugzilla : 6866 Frequency : rare, only 2.6 kernels Description: Unusual file access patterns on the MDS may result in inode data being lost in very rare circumstances. Details : Bad interaction between the ea-in-inode patch and the "no-read" code in the 2.6 kernel caused the inode and/or EA data not to be read from disk, causing single-file corruption. Severity : critical Bugzilla : 6998 Frequency : rare, only 2.6 filesystems using extents Description: Heavy concurrent write and delete load may cause data corruption. Details : It was possible under high-load situations to have an extent metadata block in the block device cache from a just-unlinked file overwrite a newly-allocated data block. We now unmap any metadata buffers that alias just-allocated data blocks. Severity : minor Bugzilla : 7241 Frequency : filesystems with default stripe_count larger than 77 Description: lconf+mke2fs fail when formatting filesystem with > 77 stripes Details : lconf specifies an inode size of 4096 bytes when the default stripe_count is larger than 77. This conflicts with the default inode density of 1 per 4096 bytes. Allocate smaller inodes in this case to avoid pinning too much memory for large EAs. ------------------------------------------------------------------------------ 2005-07-07 Cluster File Systems, Inc. * version 1.4.3 * bug fixes Severity : minor Frequency : rare (extremely heavy IO load with hundreds of clients) Bugzilla : 6172 Description: Client is evicted, gets IO error writing to file Details : lock ordering changes for bug 5492 reintroduced bug 3267 and caused clients to be evicted for AST timeouts. The fixes in bug 5192 mean we no longer need to have such short AST timeouts so ldlm_timeout has been increased. Severity : major Frequency : occasional during --force or --failover shutdown under load Bugzilla : 5949, 4834 Description: Server oops/LBUG if stopped with --force or --failover under load Details : a collection of import/export refcount and cleanup ordering issues fixed for safer force cleanup Severity : major Frequency : only filesystems larger than 120 OSTs Bugzilla : 5990, 6223 Description: lfs getstripe would oops on a very large filesystem Details : lov_getconfig used kfree on vmalloc'd memory Severity : minor Frequency : only filesystems exporting via NFS to Solaris 10 clients Bugzilla : 6242, 6243 Description: reading from files that had been truncated to a non-zero size but never opened returned no data Details : ll_file_read() reads zeros from no-object files to EOF Severity : major Frequency : rare Bugzilla : 6200 Description: A bug in MDS/OSS recovery could cause the OSS to fail an assertion Details : There's little harm in aborting MDS/OSS recovery and letting it try again, so I removed the LASSERT and return an error instead. Severity : enhancement Bugzilla : 5902 Description: New debugging infrastructure for tracking down data corruption Details : The I/O checksum code was replaced to: (a) control it at runtime, (b) cover more of the client-side code path, and (c) try to narrow down where problems occurred Severity : major Frequency : rare Bugzilla : 3819, 4364, 4397, 6313 Description: Racing close and eviction MDS could cause assertion in mds_close Details : It was possible to get multiple mfd references during close and client eviction, leading to one thread referencing a freed mfd. Severity: : enhancement Bugzilla : 3262, 6359 Description: Attempts to reconnect to servers are now more aggressive. Details : This builds on the enhanced upcall-less recovery that was added in 1.4.2. When trying to reconnect to servers, clients will now try each server in the failover group every 10 seconds. By default, clients would previously try one server every 25 seconds. Severity : major Frequency : rare Bugzilla : 6371 Description: After recovery, certain operations trigger a failed assertion on a client. Details : Failing over an mds, using lconf -d --failover, while a client was doing a readdir() call would cause the client to LBUG after recovery completed and the readdir() was resent. Severity : enhancement Bugzilla : 6296 Description: Default groups are now added by lconf Details : You can now run lconf --group without having to manually add groups with lmc. Severity : major Frequency : occasional Bugzilla : 6412 Description: Nodes with an elan id of 0 trigger a failed assertion Severity : minor Frequency : always when accessing e.g. tty/console device nodes Bugzilla : 3790 Description: tty and some other devices nodes cannot be used on lustre Details : file's private_data field is used by device data and lustre values in there got lost. New field was added to struct file to store fs-specific private data. Severity : minor Frequency : when exporting Lustre via NFS Bugzilla : 5275 Description: NFSD failed occasionally when looking up a path component Details : NFSD is looking up ".." which was broken in ext3 directories that had grown large enough to become hashed. Severity : minor Frequency : Clusters with multiple interfaces not on the same subnet Bugzilla : 5541 Description: Nodes will repeatedly try to reconnect to an interface which it cannot reach and report an error to the log. Details : Extra peer list entries will be created by lconf with some peers unreachable. lconf now validates the peer before adding it. Severity : major Frequency : Only if a default stripe is set on the filesystem root. Bugzilla : 6367 Description: Setting a default stripe on the filesystem root prevented the filesystem from being remounted. Details : The client was sending extra request flags in the root getattr request and did not allocate a reply buffer for the dir EA. Severity : major Frequency : occasional, higher if lots of files are accessed by one client Bugzilla : 6159, 6097 Description: Client trips assertion regarding lsm mismatch/magic Details : While revalidating inodes the VFS looks up inodes with ifind() and in rare cases can find an inode that is being freed. The ll_test_inode() code will free the lsm during ifind() when it finds an existing inode and then the VFS later attaches this free lsm to a new inode. Severity : major Frequency : rare Bugzilla : 6422, 7030 Description: MDS deadlock between mkdir and client eviction Details : Creating a new file via mkdir or mknod (starting a transaction and getting the ns lock) can deadlock with client eviction (gets ns lock and trying to finish a synchronous transaction). Severity : minor Frequency : occasional Description: While starting a server, the fsfilt_ext3 module could not be loaded. Details : CFS's improved ext3 filesystem is named ldiskfs for 2.6 kernels. Previously, lconf would still use the ext3 name when trying to load modules. Now, it will correctly use ext3 on 2.4 and ldiskfs on 2.6. Severity : enhancement Description: The default stripe count has been changed to 1 Details : The interpretation of the default stripe count (0, to lfs or lmc) has been changed to mean striping across a single OST, rather than all available. For general usage we have found a stripe count of 1 or 2 works best. Severity : enhancement Description: Add support for compiling against Cray portals. Details : Conditional compiling for some areas that are different on Cray Portals. Severity : major Frequency : occasional Bugzilla : 6409, 6834 Description: Creating files with an explicit stripe count may lead to a failed assertion on the MDS Details : If some OSTs are full or unavailable, creating files may trigger a failed assertion on the MDS. Now, Lustre will try to use other servers or return an error to the client. Severity : minor Frequency : occasional Bugzilla : 6469 Description: Multiple concurrent overlapping read+write on multiple SMP nodes caused lock timeout during readahead (since 1.4.2). Details : Processes doing readahead might match a lock that hasn't been granted yet if there are overlapping and conflicting lock requests. The readahead process waits on ungranted lock (original lock is CBPENDING), while OST waits for that process to cancel CBPENDING read lock and eventually evicts client. Severity : enhancement Bugzilla : 6931 Description: Initial enabling of flock support for clients Details : Implements fcntl advisory locking and file status functions. This feature is provided as an optional mount flag (default off), and is NOT CURRENTLY SUPPORTED. Not all types of record locking are implemented yet, and those that are are not guaranteed to be completely correct in production environments. mount -t lustre -o [flock|noflock] ... Severity : major Frequency : occasional Bugzilla : 6198 Description: OSTs running 2.4 kernels but with extents enabled might trip an assertion in the ext3 JBD (journaling) layer. Details : The b_committed_data struct is protected by the big kernel lock in 2.4 kernels, serializing journal_commit_transaction() and ext3_get_block_handle->ext3_new_block->find_next_usable_block() access to this struct. In 2.6 kernels there is finer grained locking to improve SMP performance of the JBD layer. Severity : minor Bugzilla : 6147 Description: Changes the "SCSI I/O Stats" kernel patch to default to "enabled" ----------------------------------------------------------------------------- 2005-05-05 Cluster File Systems, Inc. * version 1.4.2 NOTE: Lustre 1.4.2 uses an incompatible network protocol than previous versions of Lustre. Please update all servers and clients to version 1.4.2 or later at the same time. You must also run "lconf --write-conf {config}.xml" on the MDS while it is stopped to update the configuration logs. * bug fixes - fix for HPUX NFS client breakage when NFS exporting Lustre (5781) - mdc_enqueue does not need max_mds_easize request buffer on send (5707) - swab llog records of type '0' so we get proper header size/idx (5861) - send llog cancel req to DLM cancel portal instead of cb portal (5515) - fix rename of one directory over another leaking an inode (5953) - avoid SetPageDirty on 2.6 (5981) - don't re-add just-being-destroyed locks to the waiting list (5653) - when creating new directories, inherit the parent's custom striping settings if present parent (3048) - flush buffers from cache before direct IO in 2.6 obdfilter (4982) - don't hold i_size_sem in ll_nopage() and ll_ap_refresh_count (6077) - don't hold client locks on temporary worklist from l_lru (5666) - handle IO errors in 2.6 obdfilter bio completion routine (6046) - automatically evict dead clients (5921) - Update file size properly in create+truncate+fstat case (6196) - Do not unhash mountpoint dentries, do not allow removal of mountpoints (5907) - Avoid lock ordering deadlock issue with write/truncate (6203,5654) - reserve enough journal credits in fsfilt_start_log for setattr (4554) - ldlm_enqueue freed-export error path would always LBUG (6149,6184) - don't reference lr_lvb_data until after we hold lr_lvb_sem (6170) - don't overwrite last_rcvd if there is a *_client_add() error (6086) - Correctly handle reads of files with no objects (6243) - lctl recover will also mark a device active if deactivate used (5933) * miscellania - by default create 1 inode per 4kB space on MDS, per 16kB on OSTs - allow --write-conf on an MDS with different nettype than client (5619) - don't write config llogs to MDS for mounts not from that MDS (5617) - lconf should create multiple TCP connections from a client (5201) - init scripts are now turned off by default; run chkconfig --on lustre and chkconfig --on lustrefs to use them - upcalls are no longer needed for clients to recover to failover servers (3262) - add --abort-recovery option to lconf to abort recovery on device startup (6017) - add support for an arbitrary number of OSTs (3026) - Quota support protocol changes. - forward compatibility changes to wire structs (6007) - rmmod NALs that might be loaded because of /etc/modules.conf (6133) - support for mountfsoptions and clientoptions to the Lustre LDAP (5873) - improved "lustre status" script - initialize blocksize for non-regular files (6062) - added --disable-server and --disable-client configure options (5782) - introduce a lookup cache for lconf to avoid repeated DB scans (6204) - Vanilla 2.4.29 support - increase maximum number of obd devices to 520 (6242) - remove the tcp-zero-copy patch from the suse-2.4 series (5902) - Quadrics Elan drivers are now included for the RHEL 3 2.4.21 and SLES 9 2.6.5 kernels - limit stripes per file to 160 (the maximum EA size) (6093) 2005-03-22 Cluster File Systems, Inc. * version 1.4.1 * bug fixes - don't LASSERT in ll_release on NULL lld with NFS export (4655, 5760) - hold NS lock when calling handle_ast_error->del_waiting_lock (5746) - fix setattr mtime regression from lovcleanup merge (4829, 5669) - workaround for 2.6 crash in ll_unhash_aliases (5687, 5210) - small ext3 extents cleanups and fixes (5733) - improved mballoc code, several small races and bugs fixed (5733, 5638) - kernel version 43 - fix remove_suid bugs in both 2.4 and 2.6 (5695) - avoid needless client->OST connect, fix handle mismatch (5317) - fix DLM error path that led to out-of-sync client, long delays (5779) - support common vfs-enforced mount options (nodev,nosuid,noexec) (5637) - fix several locking issues related to i_size (5492,5624,5654,5672) - don't move pending lock onto export if it is already evicted (5683) - fix kernel oops when creating .foo in unlinked directory (5548) - fix deadlock in obdfilter statistics vs. object create (5811) - use time_{before,after} to avoid timer jiffies wrap (5882) - shutdown --force/--failover stability (3607,3651,4797,5203,4834) - Do not leak request if server was not able to process it (5154) - If mds_open unable to find parent dir, make that negative lookup(5154) - don't create new directories with extent-mapping (5909, 5936) * miscellania - fix lustre/lustrefs init scripts for SuSE (patch from Scali, 5702) - don't hold the pinger_sem in ptlrpc_pinger_sending_on_import - change obd_increase_kms to obd_adjust_kms (up or down) (5654) - lconf, lmc search both /usr/lib and /usr/lib64 for Python libs (5800) - support for RHEL4 kernel on i686 (5773) - provide error messages when incompatible logs are encountered (5898) 2005-02-18 Cluster File Systems, Inc. * version 1.4.0.10 (1.4.1 release candidate 1) * bug fixes - don't keep a lock reference when lock is not granted (4238) - unsafe list practices (rarely) led to infinite eviction loop (4908) - add per-fs limit of Lustre pages in page cache, avoid OOM (4699) - drop import inflight refcount on signal_completed_replay error (5255) - unlock page after async write error during send (3677) - handle missing objects in filter_preprw_read properly (5265) - no transno return for symlink open, don't save no-trasno open (3440) - don't try to complete elan receive that already failed (4012) - free RPC server reply state on error (5406) - clean up thread from ptlrpc_start_thread() on error (5160) - readahead could read extra page into cache that wasn't ejected (5388) - prevent races in class_attach/setup/cleanup/detach (5260) - don't dereference de->d_inode after l_dput of de (5458) - use "int" for stripe value returned from lock_to_stripe (5544) - mballoc allocation and error-checking fixes in 2.6 (5504) - block device patches to fix I/O request sizes in 2.6 (5482) - look up hostnames for IB nals (5602) - 2.6 changed lock ordering of 2 semaphores, caused deadlock (5654) - don't start multiple acceptors for the same port (5277) - fix incorrect LASSERT in mds_getattr_name (5635) - export a proc file for general "ping" checking (5628) - fix "lfs check" to not block when the MDS is down (5628) * miscellania - service request history (4965) - put {ll,lov,osc}_async_page structs in a single slab (4699) - create an "evict_client" /proc entry on OSTs, like the MDS has - fix mount usage message, return errors per mount(8) (5168) - change grep [] to grep "[]" in tests so they work in more UMLs - fix ppc64/x86_64 spec to use %{_libdir} instead of /usr/lib (5389) - remove ancient LOV_MAGIC_V0 EA support (5047) - add "disk I/Os in flight" and "I/O req time" stats in obdfilter - align r/w RPCs to PTLRPC_MAX_BRW_SIZE boundary for performance (3451) - allow readahead allocations to fail when low on memory (5383) - mmap locking landed again, after considerable improvement (2828) - add get_hostaddr() to lustreDB.py for LDAP support (5459) 2004-11-23 Cluster File Systems, Inc. * version 1.4.0 * bug fixes - send OST transaction number in read/write reply to free req (4966) - don't ASSERT in ptl_send_rpc() if we run out of memory (5119) - lock /proc/sys/portals/routes internal state, avoiding oops (4827) - the watchdog thread now runs as interruptible (5246) - flock/lockf fixes (but it's still disabled, pending 5135) - don't use EXT3 constants in llite code (5094) - memory shortage at startup could cause assertion (5176) * miscellania - reorganization of lov code - single portals codebase - Infiniband NAL - add extents/mballoc support (5025) - direct I/O reads in the obdfilter (4048) - kernel patches from LNXI for 2.6 (bluesmoke, perfctr, mtd, kexec) tbd Cluster File Systems, Inc. * version 1.2.9 * bug fixes - send OST transaction number in read/write reply to free req (4966) - don't ASSERT in ptl_send_rpc() if we run out of memory (5119) - lock /proc/sys/portals/routes internal state, avoiding oops (4827) - the watchdog thread now runs as interruptible (5246) - handle missing objects in filter_preprw_read properly (5265) - unsafe list practices (rarely) led to infinite eviction loop (4908) - drop import inflight refcount on signal_completed_replay error (5255) - unlock page after async write error during send (3677) - return original error code on reconstructed replies (3761) - no transno return for symlink open, don't save no-trasno open (3440) * miscellania - add pid to ldlm debugging output (4922) - bump the watchdog timeouts -- we can't handle 30sec yet - extra debugging for orphan dentry/inode bug (5259) 2004-11-16 Cluster File Systems, Inc. * version 1.2.8 * bug fixes - fix TCP_NODELAY bug, which caused extreme perf regression (5134) - allocate qswnal tx descriptors singly to avoid fragmentation (4504) - don't LBUG on obdo_alloc() failure, use OBD_SLAB_ALLOC() (4800) - fix NULL dereference in /proc/sys/portals/routes (4827) - allow failed mdc_close() operations to be interrupted (4561) - stop precreate on OST before MDS would time out on it (4778) - don't send partial-page writes before EOF from client (4410) - discard client grant for sub-page writes on large-page clients (4520) - don't free dentries not owned by NFS code, check generation (4806) - fix lsm leak if mds_create_objects() fails (4801) - limit debug_daemon file size, always print CERROR messages (4789) - use transno after validating reply (3892) - process timed out requests if import state changes (3754) - update mtime on OST during writes, return in glimpse (4829) - add mkfsoptions to LDAP (4679) - use ->max_readahead method instead of zapping global ra (5039) - don't interrupt __l_wait_event() during strace * miscellania - add software watchdogs to catch hung threads quickly (4941) - make lustrefs init script start after nfs is mounted - fix CWARN/ERROR duplication (4930) - return async write errors to application if possible (2248) - add /proc/sys/portal/memused (bytes allocated by PORTALS_ALLOC) - print NAL number in %x format (4645) - update barely-supported suse-2.4.21-171 series (4842) - support for sles 9 %post scripts - support for building 2.6 kernel-source packages - support for sles km_* packages 2004-10-07 Cluster File Systems, Inc. * version 1.2.7 * bug fixes - ignore -ENOENT errors in osc_destroy (3639) - notify osc create thread that OSC is being cleaned up (4600) - add nettype argument for llmount in #5d in conf-sanity.sh (3936) - reconstruct ost_handle() like mds_handle() (4657) - create a new thread to do import eviction to avoid deadlock (3969) - let lconf resolve symlinked-to devices (4629) - don't unlink "objects" from directory with default EA (4554) - hold socknal file ref over connect in case target is down (4394) - allow more than 32000 subdirectories in a single directory (3244) - fix blocks count for O_DIRECT writes (3751) - OST returns ENOSPC from object create when no space left (4539) - don't send truncate RPC if file size isn't changing (4410) - limit OSC precreate to 1/2 of value OST considers bogus (4778) - bind to privileged port in socknal and tcpnal (3689) * miscellania - rate limit CERROR/CWARN console message to avoid overload (4519) - GETFILEINFO dir ioctl returns LOV EA + MDS stat in 1 call (3327) - basic mmap support (3918) - kernel patch series update from b1_4 (4711) 2004-09-16 Cluster File Systems, Inc. * version 1.2.6 * bug fixes - avoid crash during MDS cleanup with OST shut down (2775) - fix loi_list_lock/oig_lock inversion on interrupted IO (4136) - don't use bad inodes on the MDS (3744) - dynamic object preallocation to improve recovery speed (4236) - don't hold spinlock over lock dumping or change debug flags (4401) - don't zero obd_dev when it is force cleaned (3651) - print grants to console if they go negative (4431) - "lctl deactivate" will stop automatic recovery attempts (3406) - look for existing locks in ldlm_handle_enqueue() (3764) - don't resolve lock handle twice in recovery avoiding race (4401) - revalidate should check working dir is a directory (4134) * miscellania - don't always mark "slow" obdfilter messages as errors (4418) 2004-08-24 Cluster File Systems, Inc. * version 1.2.5 * bug fixes - don't close LustreDB during write_conf until it is done (3860) - fix typo in lconf for_each_profile (3821) - allow dumping logs from multiple threads at one time (3820) - don't allow multiple threads in OSC recovery (3812) - fix debug_size parameters (3864) - fix mds_postrecov to initialize import for llog ctxt (3121) - replace config semaphore with spinlock (3306) - be sure to send a reply for a CANCEL rpc with bad export (3863) - don't allow enqueue to complete on a destroyed export (3822) - down write_lock before checking llog header bitmap (3825) - recover from lock replay timeout (3764) - up llog sem before sending rpc (3652) - reduce ns lock hold times when setting kms (3267) - change a dlm LBUG to LASSERTF, to maybe learn something (4228) - fix NULL deref and obd_dev leak on setup error (3312) - replace some LBUG about llog ops with error handling (3841) - don't match INVALID dentries from d_lookup and spin (3784) - hold dcache_lock while marking dentries INVALID and hashing (4255) - fix invalid assertion in ptlrpc_set_wait (3880) * miscellania - add libwrap support for the TCP acceptor (3996) - add /proc/sys/portals/routes for non-root route listing (3994) - allow setting MDS UUID in .xml (2580) - print the stack of a process that LBUGs (4228) 2004-07-14 Cluster File Systems, Inc. * version 1.2.4 * bug fixes - don't cleanup request in ll_file_open() on failed MDS open (3430) - make sure to unset replay flag from failed open requests (3440) - if default stripe count is 0, use OST count for inode size (3636) - update parent mtime/ctime on client for create/unlink (2611) - drop dentry ref in ext3_add_link from open_connect_dentry (3266) - free recovery state on server during a forced cleanup (3571) - unregister_reply for resent reqs (3063) - loop back devices mounting and status check on 2.6 (3563) - fix resource-creation race that can provoke i_size == 0 (3513) - don't try to use bad inodes returned from MDS/OST fs lookup (3688) - more debugging for page-accounting assertion (3746) - return -ENOENT instead of asserting if ost getattr+unlink race (3558) - avoid deadlock after precreation failure (3758) - fix race and lock order deadlock in orphan handling (3450, 3750) - add validity checks when grabbing inodes from l_ast_data (3599) * miscellania - add /proc/.../recovery_status to obdfilter (3428) - lightweight CDEBUG infrastructure, debug daemon (3668) - change default OSC RPC parameters to be better on small clusters - turn off OST read cache for files smaller than 32MB - install man pages and include them in rpms (3100) - add new init script for (un)mounting lustre filesystems (2593) - run chkconfig in %post for init scripts (3701) - drop scimac NAL (unmaintained) 2004-06-17 Cluster File Systems, Inc. * version 1.2.3 * bug fixes - clean kiobufs before and after use (3485) - strip trailing '/'s before comparing paths with /proc/mounts (3486) - remove assertions to work around "in-flight rpcs" recovery bug (3063) - change init script to fail more clearly if not run as root (1528) - allow clients to reconnect during replay (1742) - fix ns_lock/i_sem lock ordering deadlock for kms update (3477) - don't do DNS lookups on NIDs too small for IP addresses (3442) - re-awaken ptlrpcd if new requests arrive during check_set (3554) - fix cond_resched (3554) - only evict unfinished clients after recovery (3515) - allow bulk resend, prevent data loss (3570) - dynamic ptlrpc request buffer allocation (2102) - don't allow unlinking open directory if it isn't empty (2904) - set MDS/OST threads to umask 0 to not clobber client modes (3359) - remove extraneous obd dereference causing LASSERT failure (3334) - don't use get_cycles() when creating temp. files on the mds (3156) - hold i_sem when setting i_size in ll_extent_lock() (3564) - handle EEXIST for set-stripe, set proper directory name (3336) * miscellania - servers can dump a log evicting a client - lustre.dump_on_timeout=1 - fix ksocknal_fmb_callback() error messages (2918) 2004-05-27 Cluster File Systems, Inc. * version 1.2.2 * bug fixes - don't copy lvb into (possibly NULL) reply on error (2983) - don't deref dentry after dput, don't free lvb on error (2922) - use the kms to determine writeback rpc length (2947) - increment oti_logcookies when osc is inactive (2948) - update client's i_blocks count via lvb messages (2543) - handle intent open/close of special files properly (1557) - mount MDS with errors=remount-ro, like obdfilter (2009) - initialize lock handle to avoid ASSERT on error cleanup (3057) - don't use cancelling-locks' kms values (2947) - use highest lock extent for kms, not last one (2925) - don't dereference ERR_PTR() dentry in error handling path (3107) - fix thread race in portals_debug_dumplog() (3122) - create lprocfs device entries at setup instead of at attach (1519) - common AST error handler, don't evict client on completion race (3145) - zero nameidata in detach_mnt in 2.6 (3118) - verify d_inode after revalidate_special is valid in 2.6 (3116) - use lustre_put_super() to handle zconf unmounts in 2.6 (3064) - initialize RPC timeout timer earlier for 2.6 (3219) - don't dereference NULL reply buffer if mdc_close was never sent (2410) - print nal/nid for unknown nid (3258) - additional checks for oscc recovery before doing precreate (3284) - fix ll_extent_lock() error return code for 64-bit systems (3043) - don't crash in mdc_close for bad permissions on open (3285) - zero i_rdev for non-device files (3147) - clear page->private before handing to FS, better assertion (3119) - tune the read pipeline (3236) - fix incorrect decref of invalidated dentry (2350) - provide read-ahead stats and refine rpc in flight stats (3328) - don't hold journal transaction open across create RPC (3313) - update atime on MDS at close time (3265) - close LDAP connection when recovering to avoid server load (3315) - update iopen-2.6 patch with fixes from 2399,2517,2904 (3301) - don't leak open file on MDS after open resend (3325) - serialize filter_precreate and filter_destroy_precreated (3329) - loop device shouldn't call sync_dev() for nul device (3092) - clear page cache after eviction (2766) - resynchronize MDS->OST in background (2824) - refuse to mount the same filesystem twice on same mountpoint (3394) - allow llmount to create routes for mounting behind routers (3320) - push lock cancellation to blocking thread for glimpse ASTs (3409) - don't call osc_set_data_with_check() for TEST_LOCK matches (3159) - fix rare problem with rename on htree directories (3417) * miscellania - allow default OST striping configuration per directory (1414) - fix compilation for qswnal for 2.6 kernels (3125) - increase maximum number of MDS request buffers for large systems - change liblustreapi to be useful for external progs like lfsck (3098) - increase local configuration timeout for slow disks (3353) - allow configuring ldlm AST timeout - lustre.ldlm_timeout= 2004-03-22 Cluster File Systems, Inc. * version 1.2.1 * bug fixes - fixes for glimpse AST timeouts / incorrectly 0-sized files (2818) - don't overwrite extent policy data in reply if lock was blocked (2901) - drop filter export grants atomically with removal from device (2663) - del obd_self_export from work_list in class_disconnect_exports (2908) - don't LBUG if MDS recovery times out during orphan cleanup (2530) - swab reply message in mdc_close, other PPC fixes (2464) - fix destroying of named logs (2325) - overwrite old logs when running lconf --write_conf (2264) - bump LLOG_CHUNKSIZE to 8k to allow for larger clusters (2306) - fix race in target_handle_connect (2898) - mds_reint_create() should take same inode create lock (2926) - correct journal credits calculated for CANCEL_UNLINK_LOG (2931) - don't close files for self_export to avoid uninitialized obd (2936) - allow MDS with the same name as client node (2939) - hold dentry reference for closed log files for unlink (2325) - reserve space for all logs during transactions (2059) - don't evict page beyond end of stripe extent (2925) - don't oops on a deleted current working directory (2399) - handle hard links to targets without a parent properly (2517) - don't dereference NULL lock when racing during eviction (2867) - don't grow lock extents when lots of conflicting locks (2919) 2004-03-04 Cluster File Systems, Inc. * version 1.2.0 * bug fixes - account for cache space usage on clients to avoid data loss (974) - lfsck support in lustre kernel code (2349) - reduce journal credits needed for BRW writes (2370) - orphan handling to avoid losing space on client/server crashes - ptlrpcd can be blocked, stopping ALL progress (2477) - use lock value blocks to assist in proper KMS, faster stat (1021) - takes i_sem instead of DLM locks internally on obdfilter (2720) - recovery for initial connections (2355) - fixes for mds_cleanup_orphans (1934) - abort_recovery crashes MDS in b_eq (mds_unlink_orphan) (2584) - block all file creations until orphan recovery completes (1901) - client remove rq_connection from request struct (2423) - conf-sanity test_5, proper cleanup in umount log not availale (2640) - recovery timer race (2670) - mdc_close recovey bug (2532) - ptlrpc cleanup bug (2710) - mds timeout on local locks (2588) - namespace lock held during RPCs (2431) - handle interrupted sync write properly (2503) - don't try to handle a message that hasn't been replied to (2699) - client assert failure during cleanup after abort recovery (2701) - leak mdc device after failed mount (2712) - ptlrpc_check_set allows timedout requests to complete (2714) - wait for inflight reqs when ptlrpcd finishes (2710) - make sure unregistered services are removed from the srv_list - reset bulk XID's when resending them (caught by 1138 test) - unregister_bulk after timeout - fix lconf error (2694) - handle write after unfinished setstripe, stripe-only getstripe (2388) - readahead locks pages, leaves pending causing memory pressure (2673) - increase OST request buffers to 4096 on large machines (2729) - fix up permission of existing directories in simple_mkdir (2661) - init deleted item, add assertions ptlrpc_abort_inflight() (2725) - don't assign transno to errored transactions (2742) - don't delete objects on OST if given a bogus objid from MDS (2751) - handle large client PAGE_SIZE readdir on small PAGE_SIZE MDS (2777) - if rq_no_resend, then timeout request after recovery (2432) - fix MDS llog_logid record size, 64-bit array alignment (2733) - don't call usermode_helper from ptlrpcd, DEFAULT upcall (2773) - put magic in mount.lustre data, check for bad/NULL mount data (2529) - MDS recovery shouldn't delete objects that it has given out (2730) - if enqueue arrives after completion, don't clobber LVB (2819) - don't unlock pages twice when trigger_group_io returns error (2814) - don't deref NULL rq_repmsg if ldlm_handle_enqueue failed (2822) - don't write pages to disk if there was an error (1450) - don't ping imports that have recovery disabled (2676) - take buffered bytes into account when balancing socknal conn (2817) - hold a DLM lock over readdir always, use truncate_inode_pages (2706) - reconnect unlink llog connection after MDS reconnects to OST (2816) - remove little-endian swabbing of llog records (1987) - set/limit i_blksize to LL_MAX_BLKSIZE on client (2884) - retry reposting request buffers if they fail (1191) - grow extent at grant time to avoid granting a revoked lock (2809) - lock revoke doesn't evict page if covered by a second lock (2765) - disable VM readahead to avoid reading outside lock extents (2805) * miscellania - return LL_SUPER_MAGIC from statfs for the filesystem type (1972) - updated kernel patches for hp-2.4.20 kernel (2681) 2004-02-07 Cluster File Systems, Inc. * version 1.0.4 * kernel patches - fix truncated write corruption (2366) - fix for failed assertion in iopen_connect_dentry (1792,2517) * bug fixes - don't flag the ptlrpcd thread with PF_MEMALLOC (2636) - ensure len(uuid) < 37 in lmc (1171) - fix ia64 OOPS in llog_test (2255) - zero end of page at obdfilter for partial page writes (2648) - don't leave stale dentries around after renames (bug 2428) - fix timeouts when evicting a client with a single lock held (2642) - set deadline for the initial HELLO message to drain (2634) - print out dotted-quad IP addresses in the socknal (2302) * miscellania - additional debugging for MDS client eviction problem (2443) - fix mkfsoptions support for osts (2603, 2604) 2004-01-27 Cluster File Systems, Inc. * version 1.0.3 * kernel patches - add series for the vanilla 2.6.0 kernel - add series for the vanilla 2.4.24 kernel - add series for a cray x86/64 UL kernel drop - fix xattr patches for the vanilla 2.4.19 series * bug fixes - generate true UUIDs in lmc (1171) - have portals stack dumping break in UML (2466) - avoid bad dchild deref; avoid inum lock w/o creation (2362) - allocate with _NOFS in ldlm to avoid deadlock (1933) - wake callback waiting threads on client eviction (2460) - Add --ptldebug and --subsystem to lmc (1719) - update assertion to allow safe interrupt allocation - set rq_no_resend for cancel requests (2432) - recalculate ptlrpcd timeout after resend (2494) - call vfs_rmdir when removing pending directories (2368) - fix renaming a file to itself (2429) - lmc creates a default one-stripe lov (2454) - expand procfs space to handle large clusters (2326) - increase UML stack to avoid overflow - update lconf's list of debug and subsystem masks - fix lfs find --obd (2510) - /proc tunable for disabling filter read caching (2591) - stop rpm packages from altering slapd.conf (2301) - disable nagle in the socknal under 0conf (2578) - choose mds inode size based on stripe count (2572) - fix kernel-source rpm problems (2516) * miscellania - add --disable-doc to avoid pdf generation (2421) - update documentation, tests, type-os, comments - avoid format warnings on ia64 - remove the TOE NAL - tiny code cleanups by removing unused fields 2004-01-07 Cluster File Systems, Inc. * version 1.0.2 * bug fixes - fix obvious semaphore misuse in as-yet-unused setattr path (2348) - remove the most blatant lies from BUILDING file (2371) - change default debug level to reasonable production setting - reduce client side cache size to reduce cache flush time - reduce max RPCs in flight to avoid unnecessary file fragmentation - make TCP zerocopy and pinger support enabled by default (2476) - sync writes completed after process exits caused crashes (2319) - maintain correct mount count on the MDS (2356) - backout 1557, because 2316 wasn't really fixed - better file I/O statistics gathering in /proc - don't take unnecessary, deadlock-inducing bug in readpage (2383) - another kernel patch to fix zero-copy TCP function export - don't take duplicate lock when processing re-sent getattr (2420) - lctl uses obd_self_export instead of creating new conn (2353) - MDS/OST recovery case which requires object creation asserted (2425) - move lfs from /usr/sbin to /usr/bin in packages - fix race between mds_client_add and mds_client_free (2417) - use kmalloc instead of slabs in portals (2430) - don't create duplicate records when a failover MDS is present (2442) - remove unnecessary mount age check (2332) - don't remove directory inodes from locks prematurely (2451) - don't break if MDS service name is the same as hostname (2103) - fix races in client write RPC generation when cache full (2482) 2003-12-13 Cluster File Systems, Inc. * version 1.0.1 * bug fixes - remove now-unused request->rq_obd (278) - if an allocation fails, print out how much memory we've used (1933) - use PORTAL_SLAB_ALLOC for structures, to get GFP_MEMALLOC (1933) - add the "configurable stack size" patch to most series files (1256) - ability to write large log records, for 100+ OST configs (2306) - fix NULL deref when filter_prep fails (2314) - fix operator precedence error in filter_sync - dynamic allocation of socknal TX descriptors (2315) - fix a missed case in the GFP_MEMALLOC patch, can cause deadlock (2310) - fix gcc 2.96 compilation problem in xattr kernel patch (2294) - ensure that CWARN messages in Portals always get to the syslog - __init/__exit are not for prototype decls (ldlm_init/exit) - x86-64 compile warning fixes - fix gateway LMC keyword conflict (2318) - fix MDS lock inversions in getattr/reint paths (1844) - fix a rare lock re-ordering bug, which caused deadlock (2322) - fix i_sem/journal inversion in fsfilt_ext3_write_record (2306) - DLM race condition prevented some lock evictions (2328) - ENOMEM detection and retry on socknal sends (2230) - use GFP_NOFS throughout Lustre, to combat ENOMEM (2230) - move osc_rpcd into ptlrpc, for use in MDC and others (2329) - protect MDS inode fsdata with stronger locking; fixes assertion (2313) - better error messages when a client is rejected during recovery (1505) - avoid cancelling locks which were never granted, after failure (2330) - fix i_sem/journal inversion in mds_client_add (2333) - fix truncate/getattr lock cycle deadlock (2334) - use rpcd to send close; allows resend after timeout, avoid leak (1897) - fix two rare exit paths which could leak an l_lock() ref (2321) - fencepost error in MDS/OST orphan recovery (2226) - make log record alignment 8 bytes (1988) - lstripe now fails when requested offset > ost_count (2237) - ensure that all kernel series have a complete list.h (1607) - fix crashes in special-file operations (2316) - lctl create/brw OID mismatch, caused by obsolete filter loop (2339) * miscellania - allow configurable automake binary, for testing new versions - small update to the lfs documentation 2003-12-03 Cluster File Systems, Inc. * version 1.0.0 * fix negative export reference count in fsfilt_sync (2312) 2003-12-01 Cluster File Systems, Inc. * release candidate 0.9.1 * bug fixes - orphans are moved into the PENDING directory for possible recovery - replayed opens now open by fid for orphan/rename safety (1042) - last close of an orphan inode generates a transno (683) - chdir() and mount() now pin the directory entry (1020) - avoid CERROR in normal ll_setattr_raw() error case (1500) - discard very old requests without processing them (1502) - remove some common, well-understood CERRORs (1505) - require O_DIRECT I/O to be page-sized to workaround IA64 crash (1609) - clear "grant" flags in OST replies until OST grant code lands (1644) - fix read performance by not clobbering i_blksize on client (1598) - fix __ldlm_handle2lock oops by not dereferencing lock after PUT (1625) - make LRU size a /proc tunable, clears locks when reduced (707) - fix some lprocfs rot that prevented ptlbd from loading (1732) - server locks take references on exports now (1558) - build fixes for 2.4.20-rh trees (1663) - return an error from lov_create if all OSCs are inactive (1751) - fix import levels when a reconnect happens without a timeout (1597) - exit early from mds_open if we get a lookup error (1749) - partial page read at EOF wouldn't wait for disk before sending (1642) - avoid NULL deref in obdfilter when reading page past EOF (1592) - avoid LASSERT in ll_intent_lock if server failed very early (1090) - fix LBUG in ll_it_open_error with rc = -2 (1861) - write/truncate lock inversion (1639) - Don't auto-load obdclass, portals modules during cleanup (1495) - fix timestamps from jumping to "now" (1763) - extra journal assertions (1648) - add an extra multiunlink test (1771) - fix read_record/write_record API (1776) - fix leak of offset_extent, possible incorrect i_size later (1772) - fix lasserts in mis-matched transnos during open-unlink testing (1541) - Debugging for the kqswnal_get_idle_tx problems (1820) - Allow recovery to be attempted multiple times (1536) - Write out MDS last_rcvd file after it is first created (1600) - Fix tx_descriptor leak in failed transmit situations (1827) - ext3 journaling fixes for assertion failure after IO error (1871) - class_export_put() on freed export after completion AST error (1896) - Fix revalidate looping in VFS (1322) - Don't access a freed export during MDS_REINT timeout (1521) - Add open-unlink recovery support on the MDS (1673,1764) - Return an error if no MDS data was read from last_rcvd (1946) - Fix for lookup "." or ".." crash on error (1932,1931,1935) - Don't setup a disk device that doesn't match exported UUID (317) - Reduce bulk RPC timeout to avoid cascading client/OST failures (1845) - avoid committing NULL handle in force close - local.sh is now a one-stripe LOV configuration - POSIX utime.4 -EPERM on FIFO not owned by user (56) - fix ext3 htree duplicate directory entry corruption (1516) - POSIX creat.13, fstat.1, open.18, stat.3 new file atime/mtime (2020) - update to new LOV EA format (2097) - interoperability for different PAGE_SIZE/wordsize (686,1821,1343,2042) 2003-06-15 Phil Schwan * version v0_7 * bug fixes - imports and exports cleanup too early, need refcounts (349, 879, 1045) - per-import/export recovery handling (958, 931, 959) - multiple last-rcvd slots, for serving multiple FSes (949) - connections are again shared between multiple imp/exports (963, 964) - "umount -f" would hang if any requests needed to be sent (393, 978) - avoid pinning large req buffer by copying for queued messages (989) - add "uuid" to "lctl device" command to help upcalls (991) - "open" RPCs with transnos would confuse recovery counters (1037) - do proper endian conversion of all wire messages (288, 340, 891) - remove OST bulk get LBUGs, fix ost_brw_write cleanup (1126) - call waiting locks callback from LDLM recovery thread (1127, 1151) - fix ptlrpc_connection leak in target_handle_connect (1174) - fix import refcounting bug in OST and MDS cleanup (1134) - if an invalid-at-open-time OSC returned before close(), LBUG (1150) - fix very unlikely obd_types race condition (501) - remove osc_open hack for echo_client (1187) - we leaked exports/dlmimps for forcibly disconnected clients (1143) - a failure in read_inode2 leads to deadlock (1139) - cancel ack-locks as soon as transaction is committed (1072) - fix major leaks and crashes in the bulk I/O path (937, 1057) - make sure to commitrw after any preprw to avoid deadlock (1162) - failing to execute a file in a lustre FS would lock inode (1203) - small DEBUG_REQ fix to avoid dereferencing a NULL (1227) - don't ASSERT while cleaning up an incompletely-setup obd (1248) - obd_uuid2tgt would walk off the end of the list (1255) - on IA64 the osc would give portals incorrect bulk size (1258) - fix debug daemon ioctl interface; allows daemon on ia64 (1274) - fix lock inversion caused by new llite matching code (1282) - limit the number of dirty pages on a client to 10MB (1286) - timed out locks were not being corrected cancelled (1289) - fix O_DIRECT above 4GB on IA-32 (1292) * major user-visible changes - fail out/fail over policy now controlled by the upcall (993) * protocol changes - add OBD_PING to check server availability and failure (954) - lustre messages are now sent in sending host order (288, 340, 891) - add eadatalen to MDS getattr reply (340) - OST read replies may contain second buffer, with per-page status (593) 2003-03-11 Phil Schwan * version v0_6 * bug fixes - LDLM_DEBUG macro fix, for gcc 3.2 (850) - failed open()s could cause deadlock; fixed (867, 869) - stop cancelling OST locks when files are closed (481) - overlapping XID spaces caused network corruption (851, 853) - fix unsafe fsfilt counter arithmetic; change to atomic_t - setattr_raw added, to do single-RPC, server-side setattrs - lmc/lconf syntax change for OST UUIDs - fix crashy race condition between ptlrpc_free_req and osc_close - don't use request in mdc_enqueue if we hit a timeout (889) - don't set the inode i_size for regular files from the MDS (896) - handle out of order completion AST (842) - don't LBUG if a lock request times out after receiving AST (913) - avoid d_rehash race in ll_find_alias by rehashing inside dcache_lock - if a bad lock AST arrives, send an error instead of dropping entirely - return 0 from revalidate2 if ll_intent_lock returns -EINTR (912) - fix leak in bulk IO when only partially completed (899, 900, 926) - fix O_DIRECT for ia64 (55) - (almost) eliminate Lustre-kernel-thread effects on load average (722) - C-z after timeout could hang a process forever; fixed (977) * Features - client-side I/O cache (678, 924, 929, 941, 970) * protocol changes - READPAGE and SETATTRs which don't take server-side locks get their own portal 2003-02-11 Phil Schwan * version v0_5_20 * bug fixes - Fix ldlm_lock_match on the MDS to avoid matching remote locks (592) - Fix fsfilt_extN_readpage() to read a full page of directory entries, or fake the remainder if PAGE_SIZE != blocksize (500) - Avoid extra mdc_getattr() in ll_intent_lock when possible (534, 604) - Fix imbalanced LOV object allocation and out-of-bound access (469) - Most intent operations were removed, in favour of a new RPC mode that does a single RPC to the server and bypasses most of the VFS - All LDLM resource ID arrays were removed in favour of ldlm_res_id - Aggressively cancel local locks on DLM servers - mds_reint_unlink sends EA to the client if it's the last nlink. client uses that EA to unlink OST objects. - mds_reint_{rename,unlink,link} were rewritten to take ordered locks - recursive symlinks were fixed (439) - fixed NULL deref in DEBUG_REQ - filter_update_lastobjid no longer calls sync, which annoyed extN - fixed multi-client small-writes to a single file problem (445) - fixed mtime updates during file writes (607) - fixed vector writes on obdfilter causing problems when ENOSPC (670) - fixed bug in obd_brw_read/write() (under guise of testing 367) - fixed Linux OST size reporting problem (444, 656) - OST now updates object mtime with writes or setattr (607, 619) - client verifies file size before zeroing page past EOF (445) - OST now writes last allocated objid to disk with allocation (108) - LOV on echo now works (409) * protocol changes - mds_reint_unlink sends a new buffer, with the EA included. this buffer is only valid if body->valid & OBD_MD_FLEASIZE, which is only set if a regular file was being unlinked, and it was the last link - use PtlGet from the target for bulk writes (315) - OST now updates object mtime with writes or setattr (607, 619) - LDLM now has a grant-time callback to revalidate locked items, if necessary (604) - Many MDS operations were reorganized to combat race conditions * other changes - Merge b_intel branch (updated lprocfs code) - now at /proc/fs/lustre - configure check to avoid gcc version 2.96 20000731-2.96-98) (606) 2003-01-06 Andreas Dilger * version v0_5_19 * bug fixes - Fully reactivate OST imports after reconnection (512, others) - Make sure client sees our -ENOTCONN from mds_handle (513 - partial) - More graceful error handling for truncating on dead OST (515) - Don't error out unless we're actually accessing dead stripes (474) - Fix garbage sizes when stripes are missing (410) - LRU counters were broken, causing constant lock purge (433, 432) - garbage on read from stripes with failed OSTs (441) - mark OSCs as active before reconnecting during recovery (438) - lov_enqueue and lov_cancel need to handle inactive OSTs (403) - lfind did not preserve OST order in output (443) - symlinks cause hung clients, incorrect data (439) - stop dereferencing request after dropping refcount (457) - don't LASSERT(spin_is_locked) on non-SMP (455) - fixes for many rename() bugs - fstat didn't correctly synchronize attributes (399) - server must handle lock cancellation during blocking AST prep (487) - bulk descriptors were free()d too soon (511) - fix paths in lconf, which would load incorrect modules (451, 507) - fix confusing lconf 'host not found' error message (386) - fix lock order deadlock on OST (O/R i_sem before journal ops, 478) - fix race condition in mdc_blocking_ast() for inode access (526) - fix lov_unpackmd() unpacking wrong number of stripes (537) - fix lov_set_osc_active() marking wrong OSC inactive (440) - fix bad lstripe lov_unpackmd() assertion (fix layering too) (527) - fix multiple writes of stripe MD to MDS (358, maybe 519) - fix lstripe in several ways (kernel side) (527) - fix request leak in ldlm_cli_enqueue (262) - incorrect OSC was marked inactive after OST failure - call mds_fs_cleanup before unmounting filesystem (524) - fix races between taking ns_lock and ldlm_lock_change_resource - fix races updating LOV export open file list - fix lov_enqueue error path, avoid decref-ing bad lock handle - fix recovery NULL deref in ldlm_cli_cancel_unused - fix some DLM races by using new hash table for lock handles (419) - permit the client to specify desired inodes, at replay - duplicate requests when we queue them for replay reintegration - fix last_rcvd offset calculation - sync after each recovered transaction, so we always make progress - never, not always, ERESTART requests without transnos - store the lov_desc in the MDS, so we don't depend on getlovinfo to set it - skip replay if the MDS says that the client is already connected - don't check for a recovery-enabled export to match lctl's UUID - don't INC_USE_COUNT for phantom exports - don't crash when cleaning up phantom exports (567) - don't double-finish or set replay data for errored mdc_open requests - abort requests when they time out, so we don't get old replies - send/receive replies for AST messages again - if the client says that it doesn't have the lock, cancel it on the server - if we timeout during I/O, don't try to cancel an in-use lock; instead mark it as destroyed, it will all work out when decref is called - fix module use counts (22, 581) * protocol changes - ASTs now expect a reply (server cancels lock on error reply) 2002-12-02 Andreas Dilger * version v0_5_18 * bug fixes - fix many simultaneous client startup (392) - fix dentry->d_it clobbering - credentials weren't being shipped for readdir/getattr operations - remove invalid assertions triggered during some concurrent MD updates - proper Lustre versions added (336, 389) - fix memory leak for create error case (398) - fix LOV locking bug that would get cli/srv out of sync - fix echo client over LOV (409) - fix dbench 2, extN refcount problem (170, 258, 356, 418) - fix double-O_EXCL intent crash (424) - avoid sending multiple lock CANCELs (352) * Features - MDS can do multi-client recovery (modulo bugs in new code) * Documentation - many updates, edits, cleanups 2002-11-18 Phil Schwan * version v0_5_17 * bug fixes - fix null d_it dereference (346) - fix full OST/dbench hang (333) - fix permission problem with file removal (286) - fix removal of OSCs from LOV when they fail - fix NULL deref during bulk timeout (214) - fix problems related to multiple filesystems on one MDS (241) - fixed serious subtle metadata locking bugs - free locks on clients when inodes are removed due to memory pressure (201) - fix inode pointer in lock data (285) - partial support for multiple MDS on a single host (241) - data locks weren't cancelled at clear_inode time (290, 311) - intent locks could lead to unbounded lock growth (205) - added a maximum lock count, an LRU list, and a flusher - fix multiple rename (365) - properly abstracted the echo client - OSC locked 1 byte too many; fixed - rewrote brw callback code: - fixed recovery bugs related to LOVs (306) - fixed too-many-pages-in-one-write crash (191) - fixed (again) crash in sync_io_timeout (214) - probably fixed callback-related race (385) * protocol change - Add capability to MDS protocol - LDLM cancellations and callbacks on different portals 2002-10-28 Andreas Dilger * version v0_5_16 * bug fixes: - limit client IOV size to PTL_MD_MAX_IOV (611336, 191) - defer open object destruction to close time (601981, 138) - open/close OST file handle in obdo (OBD_MD_FLHANDLE) (601981, 138) - move LDLM_ENQUEUE/CONVERT back to MDS portal (625069) - abstract ll_lookup2, fix ll_revalidate2 to use abstraction (256) - don't call obd_setattr in ll_file_release for destroyed objects * protocol change to lustre_msg: move |version| and add |flags| * protocol change to osc_punch: "start" in "o_size", "end" in "o_blocks" * lock replay: for LDLM_FL_REPLAY trust client to do right thing * added replay of create, unlink, link and rename operations during MDS failover; recovery should be much more robust now * remove failed OSCs from LOVs (only lov_create uses this so far) * the lustre-HOWTO was brought (more) up to date (582544) 2002-10-23 Phil Schwan * version v0_5_15 * bug fixes: - in-use dentries weren't being reused properly (617851) - prevent multiple LDLM setup (599178) - fix LOV size calculations for truncate (617853) - fix client handling of MDS intent errors (POSIX) - fix permission bug in lovstripe.c test (624321) - fix MDS thread deadlock - move LDLM handler to DLM portal (625069) - truncate past end of file could corrupt data - proper cleanup after timeouts, crashes, etc (592524, 550815) - a race in recovery could return ETIMEDOUT to apps (623947) - building outside the source directory was fixed * the lustre-HOWTO was brought (more) up to date (582544) * major progress was made on recovery functionality 2002-10-10 Phil Schwan * version v0_5_14 * bug fixes: - recovery deadlock fix - rm -rf causes LBUG fix (617817) - file open by multiple tasks fix (618962) - directory permissions bugs (602707 and 620007) - journal_stop fixed with locking (611313) - O_APPEND failures resolved (618273, perhaps 614459) - lconf PATH fix (619770) - IA64 build fix (621450) - RPC buffer sizes scale with amount of memory 2002-10-01 Phil Schwan * version v0_5_13 * bug fixes: - locks would be cancelled without throwing away data pages, resulting in inconsistent data (605627) - inode attributes were not always being refreshed (605627, 612449) - lconf now continues to cleanup after lctl reports an error - MDS now enforces user permissions (602707) - lprocfs cleanup fixed, but not yet enabled (614157) - fixed infinite server hang, should a client not respond to an AST - avoid going into recovery if user calls readlink() with a buffer that's too small (613941) - AST RPCs no longer require replies (614867) -- this may be changed - don't crash server if client sends an IOV that's too big (611336) - fixed lock conversion deadlock (611892) - fixed the following of symlinks (614622) * recovery: the server can remove locks from a client that dies, other clients can make progress * more extN patch fixes * compile-time configurable ptlrpc buffer allocations * documentation - collaborative read cache document - Lustre Lite Performance CDR document-in-progress 2002-09-20 Andreas Dilger * version v0_5_12 * bug fix - fix typo in patch-2.4.18 2002-09-20 Andreas Dilger * version v0_5_11 * bug fixes - clear ptlrpc request each time in handle_incoming_request() - unlink of files now destroys the object on the OST 2002-09-19 Peter Braam * version 0_5_10 * add hard link support * change obdfile creation method * kernel patch changed 2002-09-19 Peter Braam * version 0_5_9 * bug fix - stack overflow bug in extN fixed 2002-09-18 Andreas Dilger * version 0_5_8 * documentation updates - add man pages for config tools - update tests/README to describe testing with new config tools - finish metadata API descriptions * bug fixes and cleanups - statfs workaround for 16TB limit - LOV stripe allocation improved, can stripe on subset of OSTs - LOV file size/IO offset was wrong for files > 4GB in size - object EA data was being dropped, caused files to be unreadable - memory overflow with non-LOV OST caused memory corruption - fixed regression tests to work with new config tools, obdfilter - fixed bug when directory size became larger than 1 block - fixed bug (for single client case) when PWD was deleted - invalidate local directory pages when doing intent-based ops - avoid LDLM oops when lock callback contained bad data 2002-09-09 Andreas Dilger * version 0_5_7 * documentation updates * bug fixes and cleanups - configuration tools - LOV - imports/exports - 64-bit compile warnings - 64-bit internal statfs data - many more * test_brw on persistent OST devices * MDS recovery * lprocfs (disabled) 2002-09-04 Andreas Dilger * version 0_5_6 * documentation updates * bug fixes and cleanups * configuration tools 2002-08-30 Peter J. Braam * version v0_5_5 * many small fixes to 0_5_4 * io/network handling * thinkos in MDS operations 2002-08-24 Peter J. Braam * version v0_5_4 * crucial basic fixes to 0.5.3 * IOR, Iozone work over Elan * EOF locks added 2002-08-07 Phil Schwan * version 0_5_3, our first alpha * we use the new Portals iovs * documentation updates * bug fixes and cleanups * small changes in the DLM wire protocol 2002-07-25 Peter J. Braam * version 0_5_1 with some initial stability, * locking on MD and file I/O. * documentation updates * several bug fixes since 0.5.0 * small changes in wire protocol 2002-07-18 Phil Schwan * version v0_4_5 * delivered as Lustre Light Alpha * fixed a crash after handling invalid MDS requests * fixed directory pages for architectures with non-4k pages sizes 2002-07-11 Andreas Dilger * release version v0_4_4 * Moves TCP acceptor to be on port 2432 (unused Coda port) instead of 1234. * Fixes a number of interruption problems with OST operations. * Update documentation for portals header changes * Move all wire protocol structs/defines to lustre_idl.h * Fixes symlink length bug. * Add tcpdump to repository. 2002-07-05 Andreas Dilger * release version v0_4_3 * Fixes statfs for inodes on extN. * Fixes bug in runtests which would delete /etc/hosts. * Use 64-bit object IDs wherever possible (not into VFS though) Remove ost_get_info, which is unused by lustre, and out of date. 2002-07-03 Peter Braam * release version v0_4_2 Fixes a lookup error (type not passed) * move forward to head of Portals * move forward to latest Lustre kernel 2002-06-25 Peter Braam * release version v0_4_1. Hopefully stable on single node use.