[edit] WARNING: This is the _old_ Lustre wiki, and it is in the process of being retired. The information found here is all likely to be out of date. Please search the new wiki for more up to date information.

Change Log 1.8

From Obsolete Lustre Wiki
Revision as of 03:29, 23 June 2009 by Stinson1947 (Talk | contribs)
Jump to: navigation, search

Changes from v1.8.0 to v1.8.0.1

Support for networks:
socklnd - any kernel supported by Lustre
qswlnd - Qsnet kernel modules 5.20 and later
openiblnd - IbGold 1.8.2
o2iblnd - OFED 1.1, 1.2.0, 1.2.5, 1.3 and 1.4.1
viblnd - Voltaire ibhost 3.4.5 and later
ciblnd - Topspin 3.2.0
iiblnd - Infiniserv 3.3 + PathBits patch
gmlnd - GM 2.1.22 and later
mxlnd - MX 1.2.1 or later
ptllnd - Portals 3.3 / UNICOS/lc 1.5.x, 2.0.x

Support for kernels:
2.6.16.60-0.37 (SLES 10)
2.6.18-128.1.6.el5 (RHEL 5)
2.6.22.14 vanilla (kernel.org)

Client support for unpatched kernels: (see Patchless_Client)
2.6.16 - 2.6.22 vanilla (kernel.org)

Recommended e2fsprogs version: 1.40.11-sun1

File join has been disabled in this release, refer to bugzilla 16929

A new Lustre ADIO driver is available for MPICH2-1.0.7.

NFS export disabled when stack size < 8192. Since the NFSv4 export of Lustre filesystem with 4K stack may cause a stack overflow. For more information, please refer to bugzilla 17630

Severity: major
Description: Handle new CM events in OFED 1.4

Severity: enhancement
Description: Update OFED release to 1.4.1 RC4

Severity: enhancement
Description: Update kernel to SLES10 SP2 2.6.16.60-0.37.

Severity: enhancement
Description: Update to RHEL5.3 kernel-2.6.18-128.1.6.el5.

Severity: enhancement
Description: Add support for OFED 1.4.1.

Severity: enhancement
Description: build ofed 1.4.1 with mlx4_en (Mellanox ConnectX drivers in 10GbE mode) enabled

Severity: major (SLES10/OFED 1.4.1 only)
Description: BUG: soft lockup - CPU#7 stuck for 10s! [ll_imp_inval:18451]
Details: ll_imp_inval can sleep on waiting for a semaphore while holding a spinlock. Convert lco_lock to a semaphore to address the problem.

Severity: major, only with big OST
Description: Very poor metadata performance on Infiniband lustre configuration
Details: OST object precreation becomes very slow on big OSTs. This is due to the ialloc patch spending too much time scanning groups.

Severity: normal
Frequency: during recovery
Description: don't mix llog inodes with normal.
Details: allocate inodes for log in last inode group

Severity: major
Frequency: rare
Description: fix lqs' reference which won't be put in some situations
Details: This patch fixes:

    1. In quota_check_common(), this function will check quota
       for user and group, but only send one return via "pending".
       In most cases, the pendings should be same. But that is not
       always the case.
    2. If quotaoff runs between lquota_chkquota() and
       lquota_pending_commit(), the same thing will happen too.
       That is why it comes:
       -        if (!ll_sb_any_quota_active(qctxt->lqc_sb))
       -                RETURN(0);

Severity: enhancement
Description: improve lctl set/get_param
Details: handle the bad options, support more than one arguments, add '-F' option to append the indicator to the parameters.

Changes from v1.6.7.1 to v1.8.0

Support for networks:
socklnd - any kernel supported by Lustre
qswlnd - Qsnet kernel modules 5.20 and later
openiblnd - IbGold 1.8.2
o2iblnd - OFED 1.1, 1.2.0, 1.2.5, and 1.3.1
viblnd - Voltaire ibhost 3.4.5 and later
ciblnd - Topspin 3.2.0
iiblnd - Infiniserv 3.3 + PathBits patch
gmlnd - GM 2.1.22 and later
mxlnd - MX 1.2.1 or later
ptllnd - Portals 3.3 / UNICOS/lc 1.5.x, 2.0.x

Support for kernels:
2.6.16.60-0.31 (SLES 10)
2.6.18-92.1.17.el5 (RHEL 5)
2.6.22.14 vanilla (kernel.org)

Client support for unpatched kernels: (see Patchless_Client)
2.6.16 - 2.6.22 vanilla (kernel.org)

Recommended e2fsprogs version: 1.40.11-sun1

File join has been disabled in this release, refer to bugzilla 16929

A new Lustre ADIO driver is available for MPICH2-1.0.7.

NFS export disabled when stack size < 8192. Since the NFSv4 export of Lustre filesystem with 4K stack may cause a stack overflow. For more information, please refer to bugzilla 17630

Severity: minor
Description: minor fixes and cleanups
Details: use EXT_UNSET_BLOCK to avoid confusion with EXT_MAX_BLOCK. Initialize 'ix' variable in extents patch to stop compiler warning.

Severity: feature
Description: update FIEMAP ioctl to match upstream kernel version
Details: the FIEMAP block-mapping ioctl had a prototype version in ldiskfs 3.0.7 but this release updates it to match the interface in the upstream kernel, with a new ioctl number.

Severity: normal
Frequency: only if MMP is active and detects filesystem is in use
Description: if MMP startup fails, an oops is triggered
Details: if ldiskfs mounting doesn't succeed the error handling doesn't clean up the MMP data correctly, causing an oops.

Severity: enhancement
Description: Caching OSS
Details: introduce data caching on the OSS. The OSS now relies on the linux kernel page cache to keep recently accessed data in memory. It is worth noting that all write requests are still flushed synchronously as in lustre 1.6.

Severity: enhancement
Description: version based recovery
Details: introduce finer grained recovery able to detect transaction dependencies and can deal with transaction gaps caused by clients failing at the same time as the server.

Severity: enhancement
Description: Enable adaptive timeouts by default
Details: The Lustre timeout value in /proc/sys/lustre/timeout is now managed dynamically based on server load and should not need to be tuned manually based on cluster size. This allows Lustre to work under a wider variety of system sizes and loads, without unnecessarily causing lengthy recovery times.

Severity: enhancement
Description: Add OST Pools support
Details: File striping can now be set to use an arbitrary pool of OSTs

Severity: enhancement
Description: add lazystatfs mount option to allow statfs(2) to skip down OSTs
Details: allow skip disconnected ost for send statfs request and hide error in this case.

Severity: normal
Frequency: rare, on llog test 6
Description: don't allow connect to already connected import
Details: allowing connect to already connected import is hide connecting problem.

Severity: normal
Frequency: rare, connect and disconnect target at same time
Description: ASSERTION(atomic_read(&imp->imp_inflight) == 0
Details: don't call obd_disconnect under lov_lock. this long time operation and can block ptlrpcd which answer to connect request.

Severity: normal
Frequency: rare, on failed llog setup
Description: don't leak obd reference on failed llog setup
Details: for failed llog setup - mgc forget call class_destroy_import for client import, move destroy import to more generic place.

Severity: normal
Frequency: rare
Description: allow kill process which wait statahead result
Details: for some reasons 'ls' can stick in waiting result from statahead, in this case need way for kill this process.

Severity: normal
Frequency: rare
Description: don't lose wakeup for imp_recovery_waitq
Details: recover_import_no_retry or invalidate_import and import_close can both sleep on imp_recovery_waitq, but we was send only one wakeup to sleep queue.

Severity: normal
Frequency: rare, at shutdown
Description: panic at umount
Details: llap_shrinker can be raced with killing super block from list and this produce panic with access to already freeded pointer

Severity: normal
Frequency: rare
Description: panic in mds_open
Details: don't confuse mds_finish_transno() with PTR_ERR(-ENOENT)

Severity: normal
Frequency: rare
Description: stuck in cache_remove_extent() or panic with accessing to already freed look.
Details: release lock refernce only after add page to pages list.

Severity: normal
Frequency: start MDS on uncleanly shutdowned MDS device
Description: ll_sync thread stay in waiting mds<>ost recovery finished
Details: stay in waiting mds<>ost recovery finished produce random bugs due race between two ll_sync thread for one lov target. send ACTIVATE event only if connect realy finished and import have FULL state.

Severity: normal
Frequency: always with long access acl
Description: mds can't pack reply with long acl.
Details: mds don't control size of acl but they limited by reint/getattr reply buffer.

Severity: normal
Frequency: start MDS on uncleanly shutdowned MDS device
Description: aborting recovery hang on MDS
Details: don't throttle destroy RPCs for the MDT.

Severity: major
Frequency: on remount
Description: external journal device not working after the remount
Details: clear dev_rdonly flag for external journal devices in blkdev_put()

Severity: minor
Frequency: rare
Description: shutdown vs evict race
Details: client_disconnect_export vs connect request race. if client will evicted at this time - we start invalidate thread without referece to import and import can be freed at same time.

Severity: minor
Frequency: always
Description: shrink LOV EAs before replying
Details: correctly adjust LOV EA buffer for reply.

Severity: normal
Frequency: rare
Description: don't skip ost target if they assigned to file
Details: Drop slow OSCs if we can, but not for requested start idx. This means "if OSC is slow and it is not the requested start OST, then it can be skipped, otherwise skip it only if it is inactive/recovering/out-of-space.

Severity: enhancement
Description: Update to RHEL5 kernel-2.6.18-92.1.17.el5.

Severity: enhancement
Description: Update to SLES10 SP2 kernel-2.6.16.60-0.31.

Severity: normal
Frequency: rare, need acl's on inode.
Description: client can't handle ost additional correctly
Details: if ost was added after client connected to mds client can have hit lnet_try_match_md ... to big messages to wide striped files. in this case need teach client to handle config events about add lov target and update client max ea size at that event.

Severity: normal
Frequency: Create a symlink file with a very long name
Description: ldlm_cancel_pack()) ASSERTION(max >= dlm->lock_count + count)
Details: If there is no extra space in the request for early cancels, ldlm_req_handles_avail() returns 0 instead of a negative value.

Severity: major
Frequency: rare
Description: mds is deadlocked
Details: in rare cases, inode in catalog can have i_no less than have parent i_no, this produce wrong order for locking during open, and parallel unlink can be lock open. this need teach mds_open to grab locks in resource id order, not at parent -> child order.

Severity: enhancement
Description: Add /proc entry for import status
Details: The mdc, osc, and mgc import directories now have an import directory that contains useful import data for debugging connection problems.

Severity: enhancement
Description: Re-disable certain /proc logging
Details: Enable and disable client's offset_stats, extents_stats and extents_stats_per_process stats logging on the fly.

Severity: major
Frequency: Only on FC kernels 2.6.22+
Description: oops in statahead
Details: Do not drop reference count for the dentry from VFS when lookup, VFS will do that by itself.

Severity: enhancement
Description: Generic /proc file permissions
Details: Set /Proc file permissions in a more generic way to enable non-root users operate on some /proc files.

Severity: major
Description: Hitting mdc_commit_close() ASSERTION
Details: Properly handle request reference release in ll_release_openhandle().

Severity: normal
Description: only patchless client
Details: add workaround for race between add/remove dentry from hash

Severity: enhancement
Description: Allow OST glimpses to return PW locks

Severity: minor
Description: LBUG when llog conf file is full
Details: When llog bitmap is full, ENOSPC should be returned for plain log.

Severity: normal
Description: Prevent import from entering FULL state when server in recovery

Severity: major
Description: service mount cannot take device name with ":"
Details: Only when device name contains ":/" will mount treat it as client mount.

Severity: normal
Frequency: rare
Description: replace ptlrpcd with the statahead thread to interpret the async statahead RPC callback

Severity: normal
Frequency: on recovery
Description: I/O failures after umount during fail back
Details: if client reconnected to restarted server we need join to recovery instead of find server handler is changed and process self eviction with cancel all locks.

Severity: normal
Description: Kernel BUG tries to release flock
Details: Lustre does not destroy flock lock before last reference goes away. So always drop flock locks when client is evicted and perform unlock regardless of successfulness of speaking to MDS.

Severity: enhancement
Description: Upcall on Lustre log has been dumped
Details: Allow for a user mode script to be called once a Lustre log has been dumped. It passes the filename of the dumped log to the script, the location of the script can be specified via /proc/sys/lnet/debug_log_upcall.

Severity: minor
Frequency: rare
Description: avoid messages about idr_remove called for id that is not allocated
Details: Move assigment s_dev for clustered nfs to end of initialization, for avoid problem with error handling.

Severity: minor
Frequency: rare
Description: avoid Already found the key in hash [CONN_UNUSED_HASH] messages
Details: When connection is reused this not moved from CONN_UNUSED_HASH into CONN_USED_HASH and this prodice warning when put connection again in unused hash.

Severity: normal
Frequency: rare
Description: avoid ASSERTION(client_stat->nid_exp_ref_count == 0) failed
Details: release reference to stats when client disconnected, not when export destroyed for avoid races when client destroyed after main ost export.

Severity: normal
Description: more cleanup in mds_lov
Details: add workaround for get valid ost count for avoid warnings about drop too big messages, not init llog cat under semphore which can be blocked on reconnect and break normal replay, fix access to wrong pointer.

Severity: enhancement
Description: Export bytes_read/bytes_write count on OSC/OST.

Severity: normal
Description: Early reply size mismatch, MGC loses connection
Details: Apply the MGS_CONNECT_SUPPORTED mask at reconnect time so the connect flags are properly negotiated.

Severity: normal
Description: Properly propagate oinfo flags from lov to osc for statfs
Details: restore missing copy oi_flags to lov requests.

Severity: normal
Description: exports in /proc are broken
Details: recreate /proc entries for clients when they reconnect.

Severity: enhancement
Description: Add man pages for llobdstat(8), llstat(8), plot-llstat(8), l_getgroups(8), lst(8), routerstat(8)
Details: included man pages for llobdstat(8), llstat(8), plot-llstat(8), l_getgroups(8), lst(8), routerstat(8)

Severity: enhancement
Description: Implement lustre ll_show_options method.

Severity: normal
Description: exports in /proc are broken
Details: recreate /proc entries for clients when they reconnect.

Severity: normal
Description: don't fail open with -ERANGE
Details: if client connected until mds will be know about real ost count get LOV EA can be fail because mds not allocate enougth buffer for LOV EA.

Severity: normal
Description: Resolve device initialization race
Details: Prevent proc handler from accessing devices added to the obd_devs array but yet be intialized.

Severity: enhancement
Description: configure's --enable-quota should check the kernel .config for CONFIG_QUOTA
Details: configure is terminated if --enable-quota is passed but no quota support is in kernel

Severity: normal
Frequency: rare, on PPC clients
Description: don't swab ost objects in response about directory, because this not exist.
Details: bug similar bug 14856, but in different function.

Severity: enhancement
Description: lfs quota tool enhancement
Details: added units specifiers support for setquota, default to current uid/gid for quota report, short quota stats by default, nonpositional parameters for setquota, added llapi_quotactl manual page.

Severity: enhancement
Description: *optional* service tags registration
Details: if the "service tags" package is installed on a Lustre node When the filesystem is mounted, a local-node service tag will be created. See http://inventory.sun.com/ for more information about the Service Tags asset management system.

Severity: normal
Description: Client runs out of low memory
Details: Consider only lowmem when counting initial number of llap pages

Severity: normal
Frequency: occasional
Description: add refcount for osc callbacks, so avoid panic on shutdown

Severity: normal
Frequency: testing only
Description: sanity test 65a fails if stripecount of -1 is set
Details: handle -1 striping on filesystem in ll_dirstripe_verify

Severity: normal
Frequency: only in unusual configurations
Description: Kernel panic with find ost index.
Details: lov_obd have panic if some OST's have sparse indexes.

Severity: major
Frequency: rarely, if filesystem is mounted with -o flock
Description: do not process already freed flock
Details: flock can possibly be freed by another thread before it reaches to ldlm_flock_completion_ast.

Severity: normal
Frequency: rarely, if filesystem is mounted with -o flock
Description: LBUG during stress test
Details: Need properly lock accesses the flock deadlock detection list.

Severity: minor
Frequency: rarely, if binaries are being run from Lustre
Description: oops in page fault handler
Details: kernel page fault handler can return two special 'pages' in error case, don't try dereference NOPAGE_SIGBUS and NOPAGE_OMM.

Severity: minor
Frequency: rarely, during shutdown
Description: timeout with invalidate import.
Details: ptlrpcd_check call obd_zombie_impexp_cull and wait request which should be handled by ptlrpcd. This produce long age waiting and -ETIMEOUT ptlrpc_invalidate_import and as result LASSERT.

Severity: normal
Frequency: rarely
Description: ASSERTION(CheckWriteback(page,cmd)) failed
Details: badly clear PG_Writeback bit in ll_ap_completion can produce false positive assertion.

Severity: normal
Frequency: only with broken builds/installations
Description: no LBUG if lquota.ko and fsfilt_ldiskfs.ko are different versions
Details: just return an error to a user, put a console error message

Severity: enhancement
Description: enable MGS and MDT services start separately
Details: add a 'nomgs' option in mount.lustre to enable start a MDT with a co-located MGS without starting the MGS, which is a complement to 'nosvc' mount option.

Severity: normal
Frequency: always, on big-endian systems
Description: cleanup in ptlrpc code, related to PPC platform
Details: store magic in native order avoid panic's in recovery on PPC node and forbid from this error in future. Also fix possibly of twice swab data. Fix get lov striping to userpace.

Severity: normal
Frequency: rarely, if replay get lost on server
Description: server incorrectly drop resent replays lead to recovery failure.
Details: do not drop replay according to msg flags, instead we check the per-export recovery request queue for duplication of transno.

Severity: normal
Frequency: after recovery
Description: precreate to many object's after del orphan.
Details: del orphan st in oscc last_id == next_id and this triger growing count of precreated objects. Set flag LOW to skip increase count of precreated objects.

Severity: normal
Frequency: after recovery
Description: precreate to many object's after del orphan.
Details: del orphan st in oscc last_id == next_id and this triger growing count of precreated objects. Set flag LOW to skip increase count of precreated objects.

Severity: normal
Frequency: rare, on clear nid stats
Description: ASSERTION(client_stat->nid_exp_ref_count == 0)
Details: when clean nid stats sometimes try destroy live entry, and this produce panic in free.

Severity: major
Frequency: occasionally since 1.6.4
Description: Stack overflow during MDS log replay
Details: ease stack pressure by using a thread dealing llog_process.

Severity: minor
Frequency: very rare
Description: MDT cannot be unmounted, reporting "Mount still busy"
Details: Mountpoint references were being leaked during open reply reconstruction after an MDS restart. Drop mountpoint reference in reconstruct_open() and free dentry reference also.

Severity: normal
Frequency: rare
Description: wait until IO finished before start new when do lock cancel.
Details: VM protocol want old IO finished before start new, in this case need wait until PG_writeback is cleared until check dirty flag and call writepages in lock cancel callback.

Severity: normal
Frequency: rare
Description: mds_mfd_close() ASSERTION(rc == 0)
Details: In mds_mfd_close(), we need protect inode's writecount change within its orphan write semaphore to prevent possible races.

Severity: minor
Frequency: rare, on shutdown ost
Description: don't hit live lock with umount ost.
Details: shrink_dcache_parent can be in long loop with destroy dentries, use shrink_dcache_sb instead.

Severity: minor
Frequency: only when echo_client is used
Description: don't panic with use echo_client
Details: echo client pass NULL as client nid pointer and this produce NULL pointer dereference.

Severity: normal
Frequency: Always on 32-bit PowerPC systems
Description: fix build on PPC32
Details: compile code with -m64 flag produce wrong object file for PPC32.

Severity: normal
Frequency: rare
Description: MDS LBUG: ASSERTION(!IS_ERR(dchild))
Details: In reconstruct_* functions, LASSERTs on both the data supplied by a client, and the data on disk are dangerous and incorrect. Change them with client eviction.

Severity: enhancement
Description: skiplist implementation simplification
Details: skiplists are used to group compatible locks on granted list that was implemented as tracking first and last lock of each lock group the patch changes that to using doubly linked lists

Severity: normal
Description: delete compatibility for 32bit qdata
Details: as planned, when lustre is beyond b1_8, lquota won't support 32bit qunit. That means servers of b1_4 and servers of b1_8 can't be used together if users want to use quota.

Severity: normal
Frequency: only with administrator action
Description: mount failure if config log has invalid conf_param setting
Details: If administrator specified an incorrect configuration parameter with "lctl conf_param" this would cause an error during future client mounts. Instead, ignore the bad configuration parameter.

Severity: normal
Frequency: blocks per group < blocksize*8 and uninit_groups is enabled
Description: ldiskfs error: XXX blocks in bitmap, YYY in gd
Details: If blocks per group is less than blocksize*8, set rest of the bitmap to 1.

Severity: major
Frequency: Application do stride read on lustre
Description: The read performance will drop a lot if the application does stride read.
Details: Because the stride_start_offset are missing in stride read-ahead, it will cause clients read a lot of unused pages in read-ahead, then the read-performance drops.

Severity: normal
Description: more ldlm soft lockups
Details: In ldlm_resource_add_lock(), call to ldlm_resource_dump() starve other threads from the resource lock for a long time in case of long waiting queue, so change the debug level from D_OTHER to the less frequently used D_INFO.

Severity: enhancement
Description: add -gid, -group, -uid, -user options to lfs find

Severity: enhancement
Description: ll_recover_lost_found_objs - recover objects in lost+found
Details: OST corruption and subsequent e2fsck can leave objects in the lost+found directory. Using the "ll_recover_lost_found_objs" tool, these objects can be retrieved and data can be salvaged by using the object ID saved in the fid EA on each object.

Severity: minor
Frequency: rare
Description: this bug _only_ happens when inode quota limitation is very low (less than 12), so that inode quota unit is 1 at initialization.
Details: if remaining quota equates 1, it is a sign to demonstate that quota is effective now. So least quota qunit should be 2.

Severity: normal
Description: Hung threads in invalidate_inode_pages2_range
Details: The direct IO path doesn't call check_rpcs to submit a new RPC once one is completed. As a result, some RPCs are stuck in the queue and are never sent.

Severity: normal
Description: Procfs and llog threads access destoryed import sometimes.
Details: Sync the import destoryed process with procfs and llog threads by the import refcount and semaphore.

Severity: major
Description: mds fails to respond, threads stuck in ldlm_completion_ast
Details: Sort source/child resource pair after updating child resource.

Severity: major
Frequency: rare
Description: kernel BUG at ldiskfs2_ext_new_extent_cb
Details: If insertion of an extent fails, then discard the inode preallocation and free data blocks else it can lead to duplicate blocks.

Severity: normal
Description: don't always update ctime in ext3_xattr_set_handle()
Details: Current xattr code updates inode ctime in ext3_xattr_set_handle() In some cases the ctime should not be updated, for example for 2.0->1.8 compatibility it is necessary to delete an xattr and it should not update the ctime.

Severity: normal
Description: add quota statistics
Details: 1. sort out quota proc entries and proc code. 2. add quota statistics

Severity: normal
Frequency: often
Description: quotas are not honored with O_DIRECT
Details: all writes with the flag O_DIRECT will use grants which leads to this problem. Now using OBD_BRW_SYNC to guard this.

Severity: major
Frequency: rare
Description: Assertion in iopen_connect_dentry in 1.6.3
Details: looking up an inode via iopen with the wrong generation number can populate the dcache with a disconneced dentry while the inode number is in the process of being reallocated. This causes an assertion failure in iopen since the inode's dentry list contains both a connected and disconnected dentry.

Severity: normal
Description: assertion failure in ldlm_handle2lock()
Details: fix a race between class_handle_unhash() and class_handle2object() introduced in lustre 1.6.5 by bug 13622.

Severity: enhancement
Description: superblock lock contention with many SMP cores on one client
Details: several client filesystem locks were highly contended on SMP NUMA systems with 8 or more cores. Per-CPU datastructure and more efficient locking implemented to reduce contention.

Severity: minor
Frequency: rare
Description: Kernel BUG: sd_iostats_bump: unexpected disk index
Details: remove the limit of 256 scsi disks in the sd_iostat patch

Severity: minor
Frequency: rare
Description: oops in sd_iostats_seq_show()
Details: unloading/reloading the scsi low level driver triggers a kernel bug when trying to access the sd iostat file.

Severity: major
Frequency: rare
Description: Kernel panics during QLogic driver reload
Details: REQ_BLOCK_PC requests are not handled properly in the sd iostat patch, causing memory corruption.

Severity: minor
Frequency: rare
Description: journal_dev option does not work in b1_6
Details: pass mount option during pre-mount.

Severity: enhancement
Frequency:
Description: Add a FIEMAP(FIle Extent MAP) ioctl for ldiskfs
Details: FIEMAP ioctl will allow an application to efficiently fetch the extent information of a file. It can be used to map logical blocks in a file to physical blocks in the block device.

Severity: normal
Frequency: only with adaptive timeout enabled
Description: DEBUG_REQ() bad paging request
Details: ptlrpc_at_recv_early_reply() should not modify req->rq_repmsg because it can be accessed by reply_in_callback() without the rq_lock held.

Severity: normal
Frequency: only on Cray X2
Description: X2 build failures
Details: fix build failures on Cray X2.

Severity: normal
Description: xid & resent requests
Details: Initialize RPC XID from clock at startup (randomly if clock is bad).

Severity: major
Description: quota recovery deadlock during mds failover
Details: This patch includes att18982, att18236, att18237 in bz14840. Solve the problems: 1. fix osts hang when mds does failover with quotaon 2. prevent watchdog storm when osts threads wait for the recovery of mds

Severity: normal
Description: kernel panic on racer
Details: Do not access dchild->d_inode when IS_ERR(dchild) is true.

Severity: enhancement
Description: Add lustre_start utility to start or stop multiple Lustre servers from a CSV file.

Severity: major
Description: Lustre GPF in {:ptlrpc:ptlrpc_server_free_request+373}
Details: In case of memory pressure, list_del() can be called twice on req->rq_history_list, causing a kernel oops.

Severity: normal
Description: kptllnd_peer_check_sends()) ASSERTION(!in_interrupt()) failed
Details: fix stack overflow in the distributed lock manager by defering export eviction after a failed ast to the elt thread instead of handling it in the dlm interpret routine.

Severity: enhancement
Description: More exported tunables for mballoc
Details: Add support for tunable preallocation window and new tunables for large/small requests

Severity: normal
Description: Detect corruption of block bitmap and checking for preallocations
Details: Checks validity of on-disk block bitmap. Also it does better checking of number of applied preallocations. When corruption is found, it turns filesystem readonly to prevent further corruptions.

Severity: normal
Frequency: only for big-endian servers
Description: Check if big-endian system while mounting fs with extents feature
Details: Mounting a filesystem with extents feature will fail on big-endian systems since ext3-based ldiskfs is not supported on big-endian systems. Can be overridden with "bigendian_extents" mount option.

Severity: normal
Description: Excessive recovery window
Details: With AT enabled, the recovery window can be excessively long (6000+ seconds). To address this problem, we no longer use OBD_RECOVERY_FACTOR when extending the recovery window (the connect timeout no longer depends on the service time, it is set to INITIAL_CONNECT_TIMEOUT now) and clients report the old service time via pb_service_time.

Severity: normal
Description: Watchdog triggered on MDS failover
Details: enable OBD_CONNECT_MDT flag when connecting from the MDS so that the OSTs know that the MDS "UUID" can be reused for the same export from a different NID, so we do not need to wait for the export to be evicted.

Severity: enhancement
Description: Don't sync journal after every i/o
Details: Implement write RPC replay to allow server replies for write RPCs before data is on disk. However, this feature is disabled by default since some issues leading to data corruptions have been found during recovery (e.g. bug 19128). This feature can be enabled by running the following command on the OSSs: lctl set_param obdfilter.*.sync_journal=0

Severity: low
Description: Slow reads beyond 8Tb offsets.
Details: Page index integer overflow in ll_read_ahead_page

Severity: major
Frequency: rare, only if using MMP with Linux RAID
Description: MMP doesn't work with Linux RAID
Details: While using HA for Lustre servers with Linux RAID, it is possible that MMP will not detect multiple mounts. To make this work we need to unplug the device queue in RAID when the MMP block is being written. Also while reading the MMP block, we should read it from disk and not the cached one.

Severity: minor
Frequency: rare, during recovery
Description: Assertion failure in ldlm_lock_put
Details: Do not put cancelled locks into replay list, hold references on locks in replay list

Severity: critical
Description: Lustre detected file system corruption with inode out of bounds
Details: don't update i_size on MDS_CLOSE for directories. This causes directory corruptions on the MDT.

Severity: normal
Description: client doesn't try to reconnect
Details: correctly skip time estimate if in recovery

Personal tools
Navigation