IJ38923 |
Critical |
While updating the symlink target path on an AFM enabled fileset, the inode is not copied to the previous snapshot causing the assert.
(show details)
Symptom |
Crash |
Environment |
Linux |
Trigger |
AFM caching with symlinks and snapshots. |
Workaround |
None |
|
5.0.5.14 |
AFM |
IJ38924 |
Suggested |
Issuing io_uring IORING_OP_READ_FIXED requests to read data into preallocated buffers fails with an error.
(show details)
Symptom |
IO error |
Environment |
Linux |
Trigger |
No pre-conditions are necessary. |
Workaround |
When using io_uring, use IORING_OP_READ instead of IORING_OP_READ_FIXED. This would require changing the application issuing the requests and might come at a performance penalty. |
|
5.0.5.14 |
Core GPFS |
IJ38925 |
Suggested |
Today there is no command to bring an AFM Inactive fileset to active.
(show details)
Symptom |
AFM fileset moving to Inactive/dropped states. |
Environment |
All |
Trigger |
Fileset moving to Inactive state and needing recovery for any reason. |
Workaround |
Wait for an IO operation on the fileset or touch a file inside the fileset to simulate an incoming IO and trigger recovery on the fileset in question. |
|
5.0.5.14 |
AFM |
IJ38926 |
Suggested |
When the handler for AFM replication is created on the gateway node, the handler create time, the last replay time, and the last sync time are all initialized to now time. If for some reason the handler couldn't be mounted and replicate to home, this leads to AFM printing the last replay time as the same time as handler create time and gives a misconception that replication has actually happened.
(show details)
Symptom |
Error output |
Environment |
Linux |
Trigger |
Checking AFM replication handler for last replay and sync time, when there's a recovery pending and not happening on the fileset. |
Workaround |
None |
|
5.0.5.14 |
AFM |
IJ38927 |
High Importance
|
When running IO through KNFS and file audit logging enabled, an invalid pointer might be accessed.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger | Certain patterns of KNFS IO with file audit logging enabled |
Workaround |
None |
|
5.0.5.14 |
File audit logging |
IJ38928 |
Suggested |
32bit GPFS API library not available in default path on Ubuntu.
(show details)
Symptom |
Error output/message |
Environment |
Linux (x86_64) |
Trigger |
Build an application with 32bit GPFS API library on Ubuntu. |
Workaround |
Modify the build process of the application to search for the 32bit GPFS API library in a different directory. |
|
5.0.5.14 |
GPFS API |
IJ38949 |
High Importance
|
The SUID and SGID bits are not cleared after a successful write/truncate to a file by a non-owner.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger | Create a file with the SUID and SGID bits set. As a non-root user or a non-group member user, write to the file with the write() system call or truncate the file with the truncate() system call. |
Workaround |
Ensure that only owners can write to an executable binary file that has the SUID/SGID bit set. |
|
5.0.5.14 |
Core GPFS |
IJ36560 |
Suggested |
If a workload involves opening and creating lots of files concurrently under the same directory, some of the opens may suffer high open times.
(show details)
Symptom |
Performance Impact/Degradation. |
Environment |
Windows (x86_64) |
Trigger |
Workload that creates and opens many file concurrently in the same directory path. |
Workaround |
None |
|
5.0.5.14 |
Core GPFS |
IJ39127 |
High Importance
|
An error 22 is hit when trying to get the valid data blocks on a file in resync.
(show details)
Symptom |
Unexpected Behavior |
Environment |
Linux (AFM gateway nodes) |
Trigger | Running Resync with Uncached (possibly evicted) files at the SW cache site. |
Workaround |
None |
|
5.0.5.14 |
AFM |
IJ38848 |
Suggested |
NFS mount point is not getting killed if home fileset is unresponsive or hung. This is causing multiple NFS mount to be created for the same fileset.
(show details)
Symptom |
Too much memory consumption on the NFS mount point. |
Environment |
Linux |
Trigger |
Gateway node is getting more memory consumption on the NFS mount due to existing multiple mount points of the fileset. |
Workaround |
None |
|
5.0.5.14 |
AFM DR |
IJ39144 |
Critical |
DMAPI read event is generated on AFM deferred deletion files causing unnecessary recalls if there exists only AFM recovery snapshot.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
Deletion of migrated files on AFM fileset. |
Workaround |
None |
|
5.0.5.14 |
AFM |
IJ39384 |
Critical |
AFM fails to upload the object if the name starts with a '-' character.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
AFM+COS caching with special file names. |
Workaround |
None |
|
5.0.5.14 |
AFM |
IJ39388 |
Suggested |
If the system pool is also used for data, auto recovery mis-calculates avaiable metadata fg count and might trigger tsrestripefs -r wrongly.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
If the system pool is used for both data and metadata in a FPO cluster and if a disk/node failure causes the good failure group count to become less than the default metadata replication. |
Workaround |
Do not use system pool for data. |
|
5.0.5.14 |
FPO |
IJ39411 |
Suggested |
The IBM Spectrum Scale admin commands and handling of file system encryption keys require the use of more robust settings.
(show details)
Symptom |
None |
Environment |
All |
Trigger |
None |
Workaround |
None |
|
5.0.5.14 |
Admin commands |
IJ37872 |
High Importance
|
Missing sqlite-3 packages on IBM Spectrum Scale Erasure Code Edition environments can cause admin command hangs.
(show details)
Symptom |
Hang |
Environment |
All |
Trigger |
Problem occurs in an IBM Spectrum Scale Erasure Code Edition environment when the sqlite-3 package is installed on some nodes but not on others. |
Workaround |
None |
|
5.0.5.13 |
Admin commands |
IJ37873 |
High Importance
|
SGNotQuiesced assertion in dbshLockInode during file system quiesce
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Operations which do file system quiesce. |
Workaround |
None |
|
5.0.5.13 |
Snapshots |
IJ37875 |
Suggested |
An error message "Could not retrieve minReleaseVersion" is logged in the systemhealth monitor log file (mmsysmonitor.log).
(show details)
Symptom |
Error output/message |
Environment |
All (with performance monitoring installed) |
Trigger |
The error message is logged whenever an mmperfmon query is executed. |
Workaround |
None; The error message can be ignored. |
|
5.0.5.13 |
System health |
IJ36358 |
High Importance
|
mmap reads from lots of threads might cause a deadlock in DeclareResourceUsage.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
mmap reads from lots of threads. |
Workaround |
Disable mmap pagepoolresource usage declaration with the "mmchconfig mmapDeclarePageUsage=false" command. |
|
5.0.5.13 |
Core GPFS |
IJ37909 |
High Importance
|
When there are multiple threads trying to flush the same file and the file is large with many blocks, there could be mutex contention which can lead to performance degradation.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Multiple threads trying to flush the same large file. |
Workaround |
Reduce the number of worker threads. |
|
5.0.5.13 |
Core GPFS |
IJ37910 |
High Importance
|
SGNotQuiesced assertion in dbshLockInode during file system quiesce.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Operations which do file system quiesce. |
Workaround |
None |
|
5.0.5.13 |
Snapshots |
IJ37978 |
High Importance
|
Inodes are not reclaimed after the hardlinks are corrected during the AFM prefetch. This causes more inodes to be in-use than actual number of files present in the fileset.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
AFM prefetch |
Workaround |
None |
|
5.0.5.13 |
AFM |
IJ37107 |
High Importance
|
AFM fileset resync failed with EINVAL error (22)
(show details)
Symptom |
I/O error |
Environment |
Linux |
Trigger |
AFM fileset resync operation (mmafmctl command with resync subcommand) |
Workaround |
None |
|
5.0.5.13 |
AFM |
IJ37979 |
Suggested |
When SGPanic occurs, the dealloc queue subblocks count could be wrong and cause "(deallocHighSeqNum - deallocFlushedSeqNum) >= deallocQueueSubblocks" assertion failure.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
In rare cases, the block deallocation around SGPanic time might cause this assertion. |
Workaround |
None |
|
5.0.5.13 |
Core GPFS |
IJ38041 |
Suggested |
When a fileset is in chmodAndUpdateAcl permission change mode, creating a file with the open() system call under a parent directory with inherit entries and changing permission's of the newly created file via NFS results in duplicated and incorrect entries in the file's NFSv4 ACL.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Have a fileset in chmodAndUpdateAcl permission change mode and a parent directory with inherit entries. Using NFS, create a file with the open() system call and change the permissions of the file with chmod. |
Workaround |
Use chmodAndSetAcl permission change mode for filesets and avoid having inherit entries in the parent directory. |
|
5.0.5.13 |
NFS |
IJ38052 |
High Importance
|
Due to a change in procps output in Cygwin version 3.3, IBM Spectrum Scale fails to start.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Windows (x86_64) |
Trigger |
IBM Spectrum Scale startup |
Workaround |
Downgrade Cygwin. |
|
5.0.5.13 |
Admin commands |
IJ38077 |
Suggested |
mmvdisk recovery group conversion might conflict with settings for nsdRAIDSmallBufferSize from the previous deployment scripts. mmvdisk will apply a value of -1 to this setting, which conflicts with the original value of 256KiB. The result is that the Daemon will print a warning message on start up, warning the user that nsdRAIDSmallBufferSize has been reduced to a value of 4KiB. This might impact performance.
(show details)
Symptom |
Error output/message, Performance Impact/Degradation |
Environment |
Linux |
Trigger |
mmvdisk recovery group conversion from the pre-2020 server config settings. |
Workaround |
Delete the old nsdRAIDSmallBufferSize setting of 256K in SDRFS, or delete any -1 values that were part of the mmvdisk rg conversion override. |
|
5.0.5.13 |
ESS, GNR |
IJ38081 |
High Importance
|
AFM Prefetch with --dir-list-file option where the list contains encoded directory names is not being processed and queued.
(show details)
Symptom |
Unexpected Behavior |
Environment |
Linux (AFM gateway nodes) |
Trigger |
Running prefetch (with or without --metadata-only option) using a list of encoded directory names - like the one generated from checkUncached (during mchfileset command run). |
Workaround |
Decode the directory list by hand and feed it to prefetch. |
|
5.0.5.13 |
AFM |
IJ38148 |
High Importance
|
Given a parent directory with the SGID bit set, a file created with the SGID bit specified by a user who does not belong to the same group as the directory can still have the SGID bit set.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Create a file with the SGID bit specified as a non-member group user in a directory with the SGID bit set. |
Workaround |
Remove the SGID bit from the directory. |
|
5.0.5.13 |
Core GPFS |
IJ36547 |
High Importance
|
A newly mounting node either due to user mount or an expelled node rejoining the cluster can fail assert 'llfP->lockRangeNode != NodeAddr(-1U, 0, NodeAddr::naNormal)' if it happens in the middle of a, mmrestripefs, mmaddisk, mmdeldisk, or mmfsck operation.
(show details)
Symptom |
Node expel/Lost Membership |
Environment |
All |
Trigger |
Mounted node failure in the middle of an mmrestripefs operation |
Workaround |
None |
|
5.0.5.12 |
Core GPFS |
IJ34995 |
High Importance
|
mmlsquota is reporting wrong results with: 1. extra output lines with "no limits" for users or groups that don't have usage on the fileset 2. extra output lines, all showing "no limits" when no limits (quotas) are set for a user or group in the fileset
(show details)
Symptom |
The mmlsquota output for a user is showing "no limits" when there is quota usage but no limits set. |
Environment |
All |
Trigger |
Existing code error in mmlsquota when perfileset-quota is enabled. |
Workaround |
None |
|
5.0.5.12 |
Quotas |
IJ36553 |
Suggested |
When the last block of a file is not a full GPFS block, replica compare function could report false replica mismatch.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Running replica compare with mmrestripefs or mmrestripefile |
Workaround |
None |
|
5.0.5.12 |
Core GPFS |
IJ36556 |
Critical |
If there are node failures during burst of file create or delete activity, then it is possible for the cached free inode counters on the file system manager to become out of date.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Node failures in the middle of large number of file creates or deletes. |
Workaround |
Run 'mmfsadm test imapWork <fs> inodeManager' or 'mmchmgr <fs> <another node>' |
|
5.0.5.12 |
Core GPFS |
IJ36855 |
High Importance
|
The IBM Spectrum Scale HDFS Transparency connector version 3.1.0-9, 3.1.1.7 and 3.3.0-0 contain Apache Log4j libraries that are affected by the security vulnerabilities CVE-2019-17571 and CVE-2021-4104.
(show details)
Symptom |
NA |
Environment |
All |
Trigger |
The IBM Spectrum Scale HDFS Transparency connector is not vulnerable in default configurations. |
Workaround |
Manually patch the affected log4j libraries. |
|
5.0.5.12 |
HDFS Connector |
IJ36338 |
High Importance
|
"More than 22 minutes searching for a free buffer in the pagepool" assertion failure.
(show details)
Symptom |
Abend/Long Waiters |
Environment |
All |
Trigger |
This problem is more likely to occur in a cluster which has file systems with both large block size and small block size (compared to scatter buffer size). |
Workaround |
Change 'scatterBufferSize' config to a smaller size. |
|
5.0.5.12 |
Core GPFS |
IJ36861 |
Critical |
GPFS daemon could assert while running mmadddisk. This can only happen if a new storage pool is being created as a result of running mmadddisk and a storage pool had been deleted in the past by using mmdeldisk.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Creating a new storage pool by using mmadddisk command. |
Workaround |
Increase the number of disks being added by using mmadddisk command or avoid creating a new storage pool. |
|
5.0.5.12 |
Core GPFS |
IJ36862 |
High Importance
|
When split reads are spawned out to helper gateway nodes and there is a remote error at the home site causing the fileset to be evasively put to unmounted state - there is a window where putting the fileset to unmounted can conflict with an ongoing split read message.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
AFM fileset having to move to unmounted state when there are split reads in progress on the helper gateway node. |
Workaround |
None |
|
5.0.5.12 |
AFM |
IJ34184 |
High Importance
|
Daemon assert going off when generating DMAPI event: addr.isReserved() || addr.getClusterIdx() == clusterIdx in file cfgmgr.h, resulting in a daemon crash.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
DMAPI is enabled and a remote cluster is used while a DMAPI event is being generated after a remote client node left the cluster. |
Workaround |
None |
|
5.0.5.12 |
DMAPI |
IJ36967 |
Critical |
AFM gateway node crashes during the fileset recovery because invalid file handle are used to get inodes in the kernel.
(show details)
Symptom |
Crash |
Environment |
Linux |
Trigger |
AFM fileset recovery |
Workaround |
None |
|
5.0.5.12 |
AFM |
IJ36969 |
Suggested |
Certain characters such as newline (\n) or backslash (\), etc. were not escaped correctly resulting in invalid JSON. JSON parsers are not be able to read the event correctly.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Filenames, ACLs, or xattrs with escape characters |
Workaround |
You can programmatically escape existing events to create valid JSON before the parser tries to ingest the event. |
|
5.0.5.12 |
File audit logging, Watch folder |
IJ36970 |
High Importance
|
While trying to set extended attributes, SetXAttrHandlerThread could deadlock with itself trying to obtain a WW lock on the buffer while holding XW lock.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
Changing extended attributes on a file or directory |
Workaround |
None |
|
5.0.5.12 |
Core GPFS |
IJ36558 |
Suggested |
When running file audit logging, signal 11 is possible at FileMetadata::set_mtimeUpdate(unsigned int)
(show details)
Symptom |
Signal 11 |
Environment |
Linux |
Trigger |
Daemon crash |
Workaround |
None |
|
5.0.5.12 |
File audit logging |
IJ36557 |
Suggested |
If the number of quorum nodes in the cluster is not greater than the minQuorumNodes configure setting, the mmchconfig command fails without a clear message.
(show details)
Symptom |
Error message |
Environment |
All |
Trigger |
Problem arises when minQuorumNodes configuration value is greater than or equal to the number of quorum nodes in the cluster. |
Workaround |
If setting tiebreakerDisks parameter fails because the number of quorum nodes in the cluster is not great than minQuorumNodes, use the mmchconfig command to set minQuorumNodes to the default value or a value lower than the number of quorum nodes in the cluster. |
|
5.0.5.12 |
Admin commands |
IJ36536 |
High Importance
|
Running workloads with many lookups done to GPFS in a highly concurrent way have a performance impact.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux |
Trigger |
Lots of lookups to the file system. One known case is setting LD_LIBRARY_PATH to directories on a GPFS file system on zLinux. The zLinux dynamic linker issues a much higher number of lookups for each entry in LD_LIBRARY_PATH, making this scenario more likely to occur. |
Workaround |
Reduce the number of concurrent lookups. |
|
5.0.5.12 |
Core GPFS |
IJ36842 |
Critical |
GPFS daemon could assert while mounting the file system on a client node with code level prior to V5.1.1.0. This can only happen if a new storage pool is being created by mmadddisk and the storage pool had been deleted in the past by using mmdeldisk.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Creating a new storage pool via mmadddisk command. Then mount the fs from a client node with code level prior to V5.1.1.0. |
Workaround |
Increase the number of disks being added by using the mmadddisk command or avoid creating a new storage pool. |
|
5.0.5.12 |
Core GPFS |
IJ37027 |
High Importance
|
NULL pointer dereference in kxGanesha
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
It can happen on the Ganesha NFS server or the kernel NFS server over a GPFS file system. |
Workaround |
None |
|
5.0.5.12 |
NFS |
IJ34842 |
Suggested |
If the NFSv4 client holds a file lock for read/write operations, then client may report I/O error after CES-IP failover.
(show details)
Symptom |
I/O error |
Environment |
All |
Trigger |
If NFSv4 client holds a file lock for write operation, then CES-IP failover from current active NFS server(lets say protocol node1) to other server (protocol node2) may cause I/O failure on client. |
Workaround |
None |
|
5.0.5.12 |
NFS |
IJ34927 |
High Importance
|
logAssertFailed: exclLockWord == 0
(show details)
Symptom |
assert |
Environment |
POWER |
Trigger |
NA |
Workaround |
Disable the assert. |
|
5.0.5.11 |
Core GPFS |
IJ35248 |
High Importance
|
One thread is trying to initialize the AFM relationship with the remote site and in the meanwhile another thread has initiated an unlink on the same fileset. The relationship initialization thread doesn't give up causing unlink thread to wait forever, causing the deadlock.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Linux (AFM gateway nodes) |
Trigger |
Unlinking an AFM fileset while the AFM relationship for this fileset is being initialized. |
Workaround |
Make sure the fileset is in a Dirty/Active state when unlinking an AFM Fileset. |
|
5.0.5.11 |
AFM |
IJ35257 |
Suggested |
Eviction tries to create a pruned list file in the current working directory instead of putting it into internal IBM Spectrum Scale directories on /var. If the PWD is an RO directory, then the eviction command fails with a "Could not open file" error.
(show details)
Symptom |
Error output |
Environment |
AIX, Linux |
Trigger |
Run mmafmctl eviction subcommand with list-file option from inside a Read-Only current working directory. |
Workaround |
Keep the current working directory as some RW permission directory so that eviction can create the temporary eviction pruned list file in the PWD and proceed. |
|
5.0.5.11 |
AFM |
IJ35258 |
Suggested |
Reading an evicted file in a snapshot at the AFM cache site should return an EIO and fail always.
(show details)
Symptom |
Unexpected Behavior |
Environment |
AIX, Linux |
Trigger |
Evict a file from the fileset, Create a local snapshot, and then try to read the evicted file from the snapshot. |
Workaround |
Set the afmSnapUncachedRead config to yes at the cluster level and then read the file in the snapshot to get the EIO error. |
|
5.0.5.11 |
AFM |
IJ35259 |
High Importance
|
mmrepquota command failure with message size too big due to empty quota entries.
(show details)
Symptom |
mmrepquota fails with message 'size too big' |
Environment |
All |
Trigger |
The problem is triggered when there are too many un-used quota entries accumulated in the quota file. |
Workaround |
None |
|
5.0.5.11 |
Quotas |
IJ35285 |
Suggested |
Running mmfsck command with --status-report will fail with error "Option '-N' is incorrect.", if the cluster has the config parameter defaultHelperNodes set on the cluster.
(show details)
Symptom |
Error message |
Environment |
All |
Trigger |
Running mmfsck command with --status-report on cluster that has the config parameter defaultHelperNodes set on the cluster. |
Workaround |
1) Set mmchconfig defaultHelperNodes=DEFAULT and re-execute the command. 2) Run "mmfsadm dump fsck" on all nodes running mmfsck. |
|
5.0.5.11 |
FSCK |
IJ35440 |
High Importance
|
Drives on an ESS 3000 may not show up after a boot or reboot of a canister. You can detect these errors using lspci -s 0x87 | grep DpcSta | grep Trigger+ or lspci -s 0x3c | grep DpcSta | grep Trigger+
(show details)
Symptom |
Component Level Outage |
Environment |
Linux (x86_64) |
Trigger |
ESS 3000 boot or reboot canister (very rare) |
Workaround |
You can use the setpci utility to manually clear the DPC error flag of the 0x87 and 0x3c busses. This will force the devices to attempt to train. If the drive still does not train, there is some other issue. |
|
5.0.5.11 |
ESS, GNR |
IJ35441 |
Critical |
When offline fsck is run in repair mode (-y) on a file system having inode size less than 4K the following assert is possible: "logAssertFailed: (dacIb[i] == (*DiskAddr::invalidDiskAddrP)) || (dacIb[i] == (*DiskAddr::dittoDiskAddrP)) || (dacIb[i] == (*DiskAddr::cdittoDiskAddrP)) || (dacIb[i] == (*DiskAddr::invalidZDiskAddrP)) || (dacIb[i] == (*DiskAddr::brokenZDiskAddrP)) || (dacIb[i] == (*DiskAddr::brokenDiskAddrP))"
(show details)
Symptom |
Overwrite disk addresses/Node assert |
Environment |
All |
Trigger |
This issue will hit if offline fsck -y is run on a file system having inode size less than 4K (4096 bytes). |
Workaround |
Disable the patch queue feature by running the following command before running offline fsck -y: # mmdsh -N all mmfsadm test fsck usePatchQueue 0 |
|
5.0.5.11 |
FSCK |
IJ33526 |
Suggested |
A mechanism is needed that, when mmfsd takes on a fatal signal (sigsegv, etc), the signal handler can either complete in finite time, or the daemon needs to be forcibly terminated. Lack of such a mechanism can result in a deadlock.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
GPFS calls non-signal-safe function in signal handler |
Workaround |
None |
|
5.0.5.11 |
Core GPFS |
IJ35535 |
High Importance
|
Race condition during daemon shutdown could lead to kernel crash if DIO workload is running.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
GPFS daemon shutdown while DIO workload is still running |
Workaround |
Unmount all file systems before shutting down the GPFS daemon. |
|
5.0.5.11 |
Core GPFS |
IJ35687 |
High Importance
|
If a sufficiently large number of inodes are deleted on a node, then it is possible for the background deletion process to miss processing some of the inodes.
(show details)
Symptom |
Unexpected Behavior |
Environment |
All |
Trigger |
Large number of file deletions on a node |
Workaround |
Run 'mmfsadm test imapWork <fs> fullDeletion' or 'mmchmgr <fs> <another node>' |
|
5.0.5.11 |
Core GPFS |
IJ34384 |
Critical |
Online fsck can report lost blocks that are false positives. Repairing this can allow the block to be used by multiple files at the same time causing corruption.
(show details)
Symptom |
Operation failure due to FS corruption |
Environment |
All |
Trigger |
Multi-pass fsck scan during file operations that update inode metadata. |
Workaround |
Run offline fsck. |
|
5.0.5.11 |
Core GPFS |
IJ35688 |
Suggested |
Copying an inode block which contains a bad deleted inode could trigger SIGFPE signal, then crash the mmfsd daemon.
(show details)
Symptom |
daemon crash |
Environment |
All |
Trigger |
This issue only happens on a bad deleted inode when snapshot is being used. |
Workaround |
None |
|
5.0.5.11 |
Snapshots |
IJ35689 |
Suggested |
ACL changed when running AFM failover in SW.
(show details)
Symptom |
Fileset root ACL gets changed |
Environment |
Linux |
Trigger |
Some fileset is being recreated or the fileset root metadata is changed while running failover. |
Workaround |
None |
|
5.0.5.11 |
AFM |
IJ35919 |
Critical |
GPFS API calls from 32bit application fail on SLES15 SP3.
(show details)
Symptom |
Error output/message |
Environment |
Linux (x86_64 and s390x) |
Trigger |
Running on SLES15 SP3 and an application trying to issue 32bit GPFS API calls. |
Workaround |
1. Apply the fix manually by editing the file /usr/lpp/mmfs/src/gpl-linux/ss.c to remove the checks for HAVE_COMPAT_IOCTL 2. Run mmbuildgpl again. 3. Restart GPFS. |
|
5.0.5.11 |
GPFS API |
IJ35921 |
Suggested |
IBM Spectrum Scale ships several ilm samples. One of them is a mmfind tool and to use the tool, findUtil_processOutputFile.c needs to be compiled. But the compilation of findUtil_processOutputFile.c fails on some Linux distros.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Compiling mmfindUtil_processOutputFile.c |
Workaround |
Modify mmfindUtil_processOutputFile.c before compiling it. |
|
5.0.5.11 |
Admin commands |
IJ34331 |
Suggested |
When the mmchmgr command is used to assign a new file system manager, it could fail with "No log available" message after the current file system panics with "No log available" error. This can happen if file system is not externally mounted on any node.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Using the mmchmgr command to assign a new file system manager |
Workaround |
Mount the file system before issuing the mmchmgr command. |
|
5.0.5.10 |
Core GPFS |
IJ34346 |
High Importance
|
If FIPS is enabled, call home uploads fail; manual call home uploads crash with an error, mentioning FIPS.
(show details)
Symptom |
Component Level Outage |
Environment |
Linux |
Trigger |
Enabling FIPS |
Workaround |
Disable FIPS. |
|
5.0.5.10 |
Call home |
IJ34351 |
High Importance
|
On a cluster with two quorum nodes and tiebreaker disks, an unexpected quorum loss can be seen on the challenger node when the current cluster manager sees a mmshutdown (or node reboot).
(show details)
Symptom |
File System Outage (unexpected GPFS file system unmount for about 30 seconds) |
Environment |
All |
Trigger |
GPFS shutdown (mmshutdown) or node reboot of current cluster manager |
Workaround |
Move the cluster manager role by hand using the 'mmchmgr -c ' command. |
|
5.0.5.10 |
Cluster Manager |
IJ34354 |
High Importance
|
When an application reads with an IO size that is a multiple of the GPFS block, prefetching doesn't start until the application issue a second read request unless the read starts at the beginning of the file or prefetchAggressiveness is set to prefetchOnFirstAccess. This can cause slow read performance when read IO size is very large.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Application issue read with IO size that is much larger than the GPFS block size. |
Workaround |
Set prefetchAggressiveness configuration to prefetchOnFirstAccess or reduce the read IO size to the GPFS block size. |
|
5.0.5.10 |
Core GPFS |
IJ34355 |
High Importance
|
mmkeyserv client register, deregister or rkm change command will fail if the new RKM.conf contains expired certificates.
(show details)
Symptom |
Error output/message Unexpected Results/Behavior |
Environment |
Windows (x86_64), Linux |
Trigger |
This occurs when there is a client that is registered to multiple tenants and the certificate has expired or when there are multiple clients that are registered to at least one tenant and the their certificates have expired. |
Workaround |
Use the mmkeyserv client update to update the client certificates. Otherwise, shut down GPFS and retry the command. |
|
5.0.5.10 |
Admin commands, Encryption |
IJ34356 |
High Importance
|
When a file system has a high number of block allocation regions, the processing of the allocation manager RPC could be slower than expected.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
Running mmdf |
Workaround |
Avoid running mmdf. |
|
5.0.5.10 |
Core GPFS |
IJ34357 |
High Importance
|
With thousands of client nodes mounted in the file system, adding some more disks serviced by ESS 3000 nodes can cause long waiters trying to get NSD disk information on each client node.
(show details)
Symptom |
Stuck mmadddisk command |
Environment |
ESS 3000, ECE |
Trigger |
Create new NSD disks from an ESS 3000 or an ECE cluster and add them to a file system before starting GPFS service. |
Workaround |
Restart the GPFS service on the ESS 3000 or ECE nodes, or fail over the RG master from one node to the other one. |
|
5.0.5.10 |
Admin commands |
IJ34381 |
Suggested |
The timestamps displayed by "mmdiag --iohist" on Windows nodes may show incorrect values, especially for the decimal part of the seconds. This may also cause misreporting of the duration of the affected I/O operations.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Windows (x86_64) |
Trigger |
Running "mmdiag --iohist" on Windows nodes |
Workaround |
None |
|
5.0.5.10 |
Admin commands |
IJ34391 |
Critical |
Running online fsck in repair mode (-o -y) can cause it to detect and repair false positive lost blocks (i.e. blocks that are assigned to files) and mark it as free, and doing this can lead to duplicate block corruptions.
(show details)
Symptom |
Data corruption due to duplicate blocks |
Environment |
All |
Trigger |
Running online fsck in repair mode (-o -y) |
Workaround |
Use offline fsck to fix corruptions. |
|
5.0.5.10 |
Online FSCK |
IJ34783 |
Suggested |
Kernel assert: Signal 11 at SharedHashTab::htInit
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
A special combination of maxStatCache and maxFilesToCache configuration |
Workaround |
Change the maxStatCache/maxFilesToCache configurations. |
|
5.0.5.10 |
Core GPFS |
IJ34784 |
Suggested |
File data not synced after a recovery
(show details)
Symptom |
Recovery silently completed without syncing the data with the remote. |
Environment |
Linux |
Trigger |
When tspcachescan gets an error code of more than 255. |
Workaround |
None |
|
5.0.5.10 |
AFM |
IJ34785 |
Suggested |
File data not synced after a recovery
(show details)
Symptom |
Assert is logged. |
Environment |
Linux |
Trigger |
When write operation is happening in the NSD backend and checking the remotefs quiesce or not. |
Workaround |
None |
|
5.0.5.10 |
AFM |
IJ34805 |
High Importance
|
Assert: SGNotQuiesced sgmrpc.C
(show details)
Symptom |
Scale mmfsd daemon process crash |
Environment |
All |
Trigger |
Snapshot create or delete operations |
Workaround |
None |
|
5.0.5.10 |
Core GPFS |
IJ34609 |
High Importance
|
Deadlock while queueing PIO read if there is one active gateway node and home becomes unresponsive.
(show details)
Symptom |
Deadlock |
Environment |
Linux (AFM gateway nodes) |
Trigger |
PIO read with unresponsive home. |
Workaround |
Restart GPFS at the gateway node. |
|
5.0.5.10 |
AFM |
IJ34822 |
Suggested |
When multiple nodes are creating files in the same directory, creates can slow down during recovery.
(show details)
Symptom |
Long Waiters |
Environment |
All |
Trigger |
File system crash |
Workaround |
None |
|
5.0.5.10 |
Core GPFS |
IJ34813 |
Critical |
Hard lockup between 2 pemsmod kernel threads can panic the kernel. A kernel panic will mean system down time and maybe quorum loss for the customer. Stack trace at vmcore-dmesg.txt will have something like this: [88432.803601] CPU: 27 PID: 14563 Comm: pemsRollUpQueue Kdump: loaded Tainted: G
(show details)
Symptom |
Kernel crash |
Environment |
Linux (x86_64) |
Trigger |
System running heavy I/O workload can hit this issue. |
Workaround |
None |
|
5.0.5.10 |
ESS, GNR |
IJ34943 |
High Importance
|
AFM gateway node crashes if the home is not responding while mounting the fileset target path.
(show details)
Symptom |
Crash |
Environment |
Linux |
Trigger |
AFM caching with unresponsive home. |
Workaround |
None |
|
5.0.5.10 |
AFM |
IJ33370 |
High Importance
|
If the disks for a file system are not ready to be used yet and the command "mmfsadm dump deferreddeletions" is run at the same time, the command will fail with the side effect of causing a long waiter 'waiting for SG cleanup' when the file system is deleted and recreated.
(show details)
Symptom |
Long Waiters |
Environment |
All |
Trigger |
None |
Workaround |
None |
|
5.0.5.9 |
Core GPFS |
IJ33371 |
High Importance
|
GPFS allows the length of NSD names to be up to 255 characters and there are no rules that say it must contain an alpha. If there are NSD names with all digits and long enough, this can be a problem. With long digit names, two NSDs can incorrectly be identified as the same NSD.
(show details)
Symptom |
Error output/message Unexpected Results/Behavior |
Environment |
All |
Trigger |
Long NSD names of all digits |
Workaround |
Add an alphabetic character in to the NSD name. |
|
5.0.5.9 |
Admin commands |
IJ33372 |
Critical |
During reconnect in the middle of a write operation, the below error may be reported: 2021-03-30_12:59:35.050-0400: [W] Encountered first checksum error on network I/O from NSD Client 10.10.10.10
(show details)
Symptom |
IO error |
Environment |
Linux (s390x) |
Trigger |
Network is not good which can lead to TCP connections reconnect. |
Workaround |
None |
|
5.0.5.9 |
Core GPFS |
IJ33386 |
Suggested |
Command: err 46: tsunlinkfileset -f after mmunlinkfileset commands are invoked.
(show details)
Symptom |
Unable to unlink / delete the fileset which encountered this error |
Environment |
All |
Trigger |
Invoking the mmunlinkfileset command |
Workaround |
Reboot the node and retry the mmunlinkfileset command. |
|
5.0.5.9 |
Filesets |
IJ33392 |
High Importance
|
With the introduction of the 5-level page tables, supported by Intel's Ice Lake processor generation, user space memory gets expanded by a factor of 512. This resulted in the change of kernel base address and due to this GPFS asserts with message "logAssertFailed: (UIntPtr)(vmallocStart)" while validating kernel addresses.
(show details)
Symptom |
Assert |
Environment |
Linux (x86_64) |
Trigger |
Systems that attempt to install IBM Spectrum Scale on a newer Intel x86_64 processor with 5-level page. |
Workaround |
Disable 5-level page table setting by adding no5lvl to the kernel command line and then rebooting the node. Check the documentation of the Linux distribution used for details on how to apply this change. For example, on RHEL8:
# grubby --update-kernel=ALL --args="no5lvl" # cat /proc/cmdline BOOT_IMAGE=(hd0,msdos1)/vmlinuz-4.18.0-240.10.1.el8_3.x86_64 root=/dev/mapper/rhel-root ro crashkernel=auto resume=/dev/mapper/rhel-swap rd.lvm.lv=rhel/root rd.lvm.lv=
rhel/swap rhgb quiet net.ifnames=0 biosdevname=0 no5lvl |
|
5.0.5.9 |
Core GPFS |
IJ33393 |
High Importance
|
When users run the mmlsfileset command, it doesn't show the junction paths of some fileset randomly.
(show details)
Symptom |
Unexpected results |
Environment |
All |
Trigger |
One fileset's root directory has been corrupted for an unknown reason. |
Workaround |
None |
|
5.0.5.9 |
Filesets |
IJ33394 |
Suggested |
Assert "(verify == 0) || (ofP == __null) || (ofP->sgP == __null) || ofP->isRoSnap() || (ofP->metadata.getInodeStatus() != 1) || (!ofP->sgP->isFileIncludedInSnapshot(ofP->getInodeNum(), ofP->getSnapId(), getInodeStatus())) || (ofP->assertInodeWasCopiedToPrevSnapshot()) || (ofP->isBeingRestriped() || ofP->beenRestriped)".
(show details)
Symptom |
Daemon crash |
Environment |
All |
Trigger |
Operations triggering a statlite call on a node without sufficient stat file token |
Workaround |
Disable the statlite config parameter by "mmfschconfig statliteMaxAttrAge=0 -i". |
|
5.0.5.9 |
Core GPFS |
IJ33410 |
Suggested |
--safe-limit option is ignored when eviction is invoked manually using the "mmafmctl device evict" command.
(show details)
Symptom |
Unexpected Results |
Environment |
All |
Trigger |
AFM manual eviction with safe limit |
Workaround |
None |
|
5.0.5.9 |
AFM |
IJ33627 |
Suggested |
When a thread performing shutdown and a thread initiating startup run concurrently, it is possible that it could result in a kernel crash.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
Very small race window in GPFS cleanup process |
Workaround |
None |
|
5.0.5.9 |
Core GPFS |
IJ33680 |
Suggested |
If a Linux node is overloaded and the thread cannot be scheduled quickly could result in a kernel panic: RIP list_del_entry_valid.cold
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
mmshutdown on a busy Linux node |
Workaround |
None |
|
5.0.5.9 |
Core GPFS |
IJ33702 |
High Importance
|
cNFS does not work on RHEL8.x. This is due to a change in pidof commnand in RHEL8.
(show details)
Symptom |
Unexpected Results/Behavior Node Reboot |
Environment |
Red Hat Enterprise Linux 8.x |
Trigger |
Enable cNFS on RHEL8.x nodes. |
Workaround |
Downgrade or upgrade procps-ng to procps-ng-3.3.15-3.el8. |
|
5.0.5.9 |
cNFS |
IJ33704 |
Suggested |
Fileset path is taking the chars count to read the actual mount path from a given directory path. If the directory mount has the same chars till the count then prefetch starts processing successfully.
(show details)
Symptom |
dirpath is checked properly |
Environment |
All |
Trigger |
Prefetch starts working on an invalid path where the dir path doesn't belong to the same fileset. |
Workaround |
None |
|
5.0.5.9 |
AFM |
IJ33715 |
Suggested |
GPFS has fileset level permissions which can deny setting the mode or EAs on the fileset entities depending on which mode this targets. AFM doesn't consider this flag on the fileset and we end up getting E_PERM from the home which causes the queue to get stalled. Normal queue goes fine, but its mostly the recovery or resync queue that hits this.
(show details)
Symptom |
Unexpected Results |
Environment |
Linux |
Trigger |
Set the Same Fileset level permissions flag (setAclOnly, ChmodOnly, chmodAndUpdateAcl etc.) on both Cache/Primary and/or Home/Secondary sites and perform IO to the fileset and then run Recovery or Resync. |
Workaround |
Drop those operations that stall the queue when the Fileset level permissions are enabled only at one of the 2 sites. |
|
5.0.5.9 |
AFM |
IJ33740 |
High Importance
|
AFM Prefetch is not generating the prefetch end callback event registered through the afmPrepopEnd event.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Register for afmPrepopEnd callback event, and run AFM prefetch with list file or directory option. |
Workaround |
None |
|
5.0.5.9 |
AFM |
IJ33741 |
High Importance
|
The automatic restart of NFS (remedy action) is blocked by an open unmounted_fs_check event which is not relevant for NFS/SMB exports.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux (CES nodes running NFS) |
Trigger |
File systems with the automount flag and an unmounted file system |
Workaround |
Remove the "automount" flag from the "testFS" file system. |
|
5.0.5.9 |
System health |
IJ33759 |
High Importance
|
The Mellanox firmware manager was called frequently (around every minute) by the system health monitor. That caused a high CPU load.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
The Mellanox firmware check is executed too frequently by the system health monitor. There is no need for so much checking. |
Workaround |
None |
|
5.0.5.9 |
System health |
IJ33778 |
Suggested |
RAS event dir_sharedroot_perm_problem was received by mmhealth sometimes without a need and sometimes with a need, but the description of the event does not say what is wrong with the permissions and which permissions should be provided.
(show details)
Symptom |
Error output |
Environment |
Linux |
Trigger |
cesSharedRoot does not have permissions 'rx' for 'group' and 'others'. |
Workaround |
Provide the necessary permissions for cesSharedRoot ('rx' for 'group' and 'others'). |
|
5.0.5.9 |
System health |
IJ33860 |
High Importance
|
cNFS does not work on RHEL8.x. This is due to a change in pidof commnand in RHEL8.
(show details)
Symptom |
Unexpected Results/Behavior Node Reboot |
Environment |
Red Hat Enterprise Linux 8.x |
Trigger |
Enable cNFS on RHEL8.x nodes. |
Workaround |
Downgrade or upgrade procps-ng to procps-ng-3.3.15-3.el8. |
|
5.0.5.9 |
cNFS |
IJ33861 |
Suggested |
If a file system is set to maintenance mode then it is listed as 'SUSPENDED', but only a 'unmounted_fs_check' event is shown as the reason. It should say 'maintenance state' instead.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
The 'fs_maintenance_mode' event is only at info-level, since it is a user intended state. Info-level events are in general not reported by 'mmhealth node show' since they do not indicate an issue or error state. A code change was done to allow the 'fs_maintenance_mode' event to be listed as a reason. |
Workaround |
None |
|
5.0.5.9 |
System health |
IJ33862 |
Suggested |
Ganesha fails to open files when over 1 million files are open.
(show details)
Symptom |
Check for logs "Futility count exceeded. Client load is opening FDs faster than the LRU thread can close them." and values of current_open and former_open. |
Environment |
Linux |
Trigger |
Whenever a client opens more than 1 million files. |
Workaround |
None |
|
5.0.5.9 |
CES NFS |
IJ32972 |
Suggested |
Manual procedure to decommission a DataNode is not supported.
(show details)
Symptom |
NA |
Environment |
All |
Trigger |
NA |
Workaround |
None |
|
5.0.5.9 |
HDFS Transparency |
IJ31047 |
High Importance
|
Assertion (!OWNED_BY_CALLER(lockWordCopy, lockWordCopy) Failure at line 1275 in file dSynch.C when accessing snapshot files
(show details)
Symptom |
Daemon Crash |
Environment |
All |
Trigger |
Data prefetch is triggered after sequentially accessing snapshot files. |
Workaround |
None |
|
5.0.5.8 |
Snapshots |
IJ32501 |
High Importance
|
When getting the stats of a file, users could run into the assert: "Assert exp((verify == 0) || (ofP == __null) || (ofP->sgP == __null) || ofP->isRoSnap() || (ofP->metadata.getInodeStatus() != 1) || !ofP->sgP->isFileIncludedInSnapshot(ofP->getInodeNum(), ofP->getSnapId(), getInodeStatus())) || (ofP->assertInodeWasCopiedToPrevSnapshot()) || (ofP->isBeingRestriped() || ofP->beenRestriped)" if there are writes to the same file from other nodes.
(show details)
Symptom |
Daemon Crash |
Environment |
All |
Trigger |
Getting the lite stat of a file while writes are in progress from other nodes. |
Workaround |
Run the mmchconfig command to reset the configuration "statliteMaxAttrAge=0", which will disable the statlite and avoid this problem, but it may also impact the writes performance on the other nodes as well. |
|
5.0.5.8 |
gpfs_statlite API |
IJ32365 |
Critical |
AFM prefetch fails with "too many open files" error.
(show details)
Symptom |
Unexpected results |
Environment |
All |
Trigger |
AFM prefetch |
Workaround |
None |
|
5.0.5.8 |
AFM |
IJ32345 |
Suggested |
The systemhealth monitor did not detect all paths for RDMA support (libibverbs.so library) on Ubuntu Machines. Therefore, it reports a "ib_rdma_libs_wrong_path" issue.
(show details)
Symptom |
Error output/messages |
Environment |
Ubuntu Linux |
Trigger |
The issue shows up on Ubuntu machines with RDMA in use. |
Workaround |
None |
|
5.0.5.8 |
System health |
IJ32361 |
Suggested |
After convert legacy recovery group to mmvdisk managed recovery group, poor write performance observed from an application and the gpfs daemon did not come up because of OOM issue on the some nodes.
(show details)
Symptom |
Abend/Crash Performance Impact/Degradation |
Environment |
Linux |
Trigger |
When you convert legacy recovery group to mmvdisk managed recovery group by the following command: mmvdisk recoverygroup convert --recovery-group RgName[,RgName] --node-class NcName |
Workaround |
Use the following command to reset the pagepool to 60%: mmvdisk server change --node-class NcName --pagepool 60% --recycle one |
|
5.0.5.8 |
ESS, GNR |
IJ32375 |
Critical |
Application performance degradation while running on AFM filesets.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux |
Trigger |
AFM replication |
Workaround |
None |
|
5.0.5.8 |
AFM, AFM DR |
IJ32503 |
Critical |
GPFS daemon could assert with: Assert exp(start + offsetToRef(elen) <= dhP->hashTabRef) when operate on a corrupted directory block. The assert also prevent repairs using mmfsck.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Corruption in the directory block |
Workaround |
None |
|
5.0.5.8 |
Core GPFS |
IJ32504 |
HIPER |
AFM recovery may incorrectly delete the files at home or secondary if there is any network issues while doing the home readdir.
(show details)
Symptom |
Unexpected Results |
Environment |
Linux |
Trigger |
AFM recovery |
Workaround |
Resync the fileset if there are any missing files at home. |
|
5.0.5.8 |
AFM, AFM DR |
IJ32581 |
High Importance
|
When doing preallocation and writes (e.g., Spectrum Protect Plus copy restore), the block usage of the file system is more than the total data size of these files.
(show details)
Symptom |
More disk space usage than expected. |
Environment |
All |
Trigger |
Preallocate the data blocks of the file, and then write as much data as the file size. |
Workaround |
Use the following command: mmchattr --compact=fragment |
|
5.0.5.8 |
Disk space preallocation of files |
IJ32601 |
High Importance
|
mmhealth reports degraded network with reason "ib_rdma_nic_unrecognized" even though all RDMA ports are operational.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
Uninitialized data can lead to invalid input for mmhealth. |
Workaround |
None |
|
5.0.5.8 |
RDMA |
IJ32653 |
High Importance
|
AFM prefetch fails with error 238 if the prefetch list file contains symlinks and if their target paths do not exist as part of the same fileset.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
AFM prefetch |
Workaround |
None |
|
5.0.5.8 |
AFM |
IJ32796 |
High Importance
|
Operations requiring allocation of full metadata blocks. Examples: Expand number of allocated inode Create new independent fileset.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Operations requiring allocation of full metadata blocks. Examples: Expand number of allocated inode or create new independent fileset |
Workaround |
Add more disks to the system pool. |
|
5.0.5.8 |
Core GPFS |
IJ32797 |
Suggested |
The Linux fallocate(2) API doesn't work correctly on Spectrum Scale file systems when punching a hole beyond the end of file.
(show details)
Symptom |
Punching a hole beyond the end of a file fails with EINVAL(22) error. |
Environment |
Linux |
Trigger |
Punching a hole through Linux fallocate(2) API |
Workaround |
None |
|
5.0.5.8 |
fallocate(2) |
IJ32813 |
High Importance
|
Issuing a "mmchnode --daemon-interface" attempts to change the cluster configuration repository (CCR). When this mmchnode is issued from a Windows node, CCR gets committed with invalid IPv4 information, rendering the cluster in a non-working state.
(show details)
Symptom |
The mmchnode command fails with a message trail resembling: 'mmchnode: Unable to commit new changes.' 'mmchnode: [E] The command was unable to reach the CCR service on any quorum node. Ensure the CCR service (mmfsd or mmsdrserv daemon) is running on all quorum nodes and the communication port is not blocked by the firewall.' 'mmchnode: 6027-1271 Unexpected error from function setRunningCommand. Return code: 149' |
Environment |
Windows (x86_64) |
Trigger |
Issuing "mmchnode --daemon-interface" command on a Windows node specifying an alternate IPv4 address. |
Workaround |
None. A manual CCR restore (mmsdrrestore --ccr-repair) may be necessary to restore the cluster to a working state. |
|
5.0.5.8 |
CCR |
IJ32814 |
Suggested |
Offline fsck will not be able to repair all corruptions when using the option of applying patch file ((i.e mmfsck FSchk -v --patch-file path-towrite-patchfile --patch) to repair the corruptions When repairing corruption by applying patch file the fsck output would show the below messages indicating the issue: ---------------- Invalid BlockType Inode. Skipping patch. ---------------
(show details)
Symptom |
Error output/message and all corruptions not fixed |
Environment |
All |
Trigger |
Offline fsck repairing corruptions by applying patch file. |
Workaround |
Run fsck repair with the regular option -y to fix the corruptions. |
|
5.0.5.8 |
FSCK |
IJ32859 |
Suggested |
When the mmdf command is run from a directory where the current working directory has become stale (directory was deleted after going to it), the command states it was run from an invalid directory.
(show details)
Symptom |
Command states it was run from an invalid directory. But the command fails with various additional errors. |
Environment |
All |
Trigger |
Running the mmdf command from a directory that is stale (directory was deleted after going to it). |
Workaround |
Only use MM commands in a valid current working directory. Move to a directory that still exists within the nodes file systems. |
|
5.0.5.8 |
Core GPFS |
IJ33000 |
High Importance
|
In the current implementation of Eviction on a file, the eviction program acquires a DMAPI lock first on the file first and punches a hole on it. The program can be terminated at any point without the DMAPI lock to be released - causing a lock leak and hence later DMAPI lock acquire on the file can deadlock and the only way to come out of this is to bounce the mmfsd.
(show details)
Symptom |
Deadlock |
Environment |
AIX, Linux |
Trigger |
Try to evict a file or list of files, and the eviction getting killed midway through. |
Workaround |
None |
|
5.0.5.8 |
AFM |
IJ32892 |
High Importance
|
On an AIX node, in some occasions, including the /var file system becoming full, mmfsd is unable to run child processes, and that results in different failures, depending on the process which mmfsd attempts to run. Among the operations which have been seen to fail: - mmadddisk - mmauth Once the problem is triggered, it will remain until the mmfsd daemon is restarted. If the problem is initiated by the /var file system getting full, freeing up space on that file system is not
enough to solve the problem. An indication that problem is taking place is in the output of command /usr/lpp/mmfs/bin/tslsfs nonexistent_FS (that is, passing the name of a nonexistent file system as parameter to the command above). In a system where the problem is occurring, the output will be mmcommon getEFOptions nonexistent_FS failed. Return code 1. While on a system without the problem, the output will be mmcommon: File system nonexistent_FS is not known to the GPFS cluster.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
AIX |
Trigger |
A likely trigger for the problem is the /var file system being filled, possibly around the time an operation is taking place that results in information being produced to the mmfs.log file. |
Workaround |
Once the issue in /var is resolved, restarte mmfsd. |
|
5.0.5.8 |
Core GPFS |
IJ32893 |
High Importance
|
When a dependent fileset is created and linked under AFM independent fileset, ACLs form the home dependent fileset are not fetched and set at the cache dependent fileset. This happens only for the dependent fileset root path.
(show details)
Symptom |
Unexpected Results |
Environment |
Linux |
Trigger |
AFM caching with dependent filesets |
Workaround |
None |
|
5.0.5.8 |
AFM |
IJ32906 |
High Importance
|
DEADLOCK PROBECLUSTERTHREAD WAITING FOR SG CLEANUP
(show details)
Symptom |
Deadlock |
Environment |
Linux (PPC64) |
Trigger |
Generating file audit logging or watch folder events on a ppc64 node and there are x86 or ppc64le nodes also using the same file system. |
Workaround |
None |
|
5.0.5.8 |
File audit logging, Watch folder |
IJ32929 |
Suggested |
The mmfs.log may be filled with netstat: not found messages on systems running SLES 15. This is the result of running mmdiag --network command explicitly or through mmhealth monitoring service
which uses the netstat command.
(show details)
Symptom |
Error output/message |
Environment |
SLES 15 |
Trigger |
When the mmdiag command is invokes with the --network option on SLES15 systems directly on the command line or through system health monitoring. |
Workaround |
Install the deprecated net-tools package. |
|
5.0.5.8 |
GUI, System health |
IJ30160 |
High Importance
|
When mmbackup or tsapolicy is called to scan files, it could report "no such file or directory" for existing files.
(show details)
Symptom |
Unexpected behavior |
Environment |
All |
Trigger |
Running mmbackup or tsapolicy operation while there are file deletions in progress. |
Workaround |
Rerun the mmbackup or the tsapolicy command. |
|
5.0.5.7 |
mmbackup, tsapolicy, GPFS API, DMAPI, AFM |
IJ30308 |
Suggested |
mmcrfs fails to create file systems when the cluster is configured with minQuorumNodes greater than one and tiebreakerDisks are in use.
(show details)
Symptom |
Error Output/Message Unexpected Results/Behavior |
Environment |
All |
Trigger |
CCR Cluster minQuorumNodes parameter is set to more than one and tiebreakerDisks are in use. |
Workaround |
Either set minQuorumNodes back to the default or disable tiebreakerDisks. |
|
5.0.5.7 |
Admin commands |
IJ31134 |
High Importance
|
mmbackup does not honor --max-backup-size in a snapshot backup.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Running snapshot backup with the --max-backup-size option. |
Workaround |
None |
|
5.0.5.7 |
mmbackup |
IJ30673 |
High Importance
|
When aioSyncDelay config is enabled, the buffer steal and the aio writes that need to be done as buffered I/O may race with each other and causes log assert isSGPanicked in clearBuffer.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
Enabling aioSyncDelay config and do aio writes. |
Workaround |
Issue mmchconfig aioSyncDelay=0 to disable it. |
|
5.0.5.7 |
Core GPFS |
IJ30700 |
Suggested |
GPFS command reports incorrect default for nsdRAIDMaxRecoveryRetries.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Command output |
Workaround |
None |
|
5.0.5.7 |
Admin commands |
IJ31060 |
High Importance
|
State of Physical disk shown as "unknown" in mmhealth and GUI for ECE.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
ECE, this does not affect GNR/ESS |
Workaround |
None |
|
5.0.5.7 |
System health, GUI |
IJ31785 |
Suggested |
mmhealth has issues with high inode consumption.
(show details)
Symptom |
Error output/messages |
Environment |
All |
Trigger |
One fileset out of multiple filesets for the same file system ran out of inode space. |
Workaround |
If the file system contains multiple filesets (even the root file system) then check if any of them exceeds the defined limits and fix this issue. Use mmdf and mmlsfileset commands for more diagnostic. |
|
5.0.5.7 |
System health |
IJ31208 |
High Importance
|
Daemon assert: (ofP == NULL) || (getPseudoIbdP() == ibdP) || (ibdP->assignedIndDA.isALLOC()) in file Metadata.h, resulting in mmfsd daemon crash.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Concurrent random writes from multiple nodes with new disk space allocation, and some of these writing nodes fail. |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ30797 |
High Importance
|
GPFS daemon could fail with logAssertFailed: getDeEntType() == detUnlucky when reading a directory block that contain unexpected data due to corruption.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Unexpected data corruption in directory block |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ32008 |
High Importance
|
The free space stats report through the "df" command or statfs API is very out of date on client nodes.
(show details)
Symptom |
Stale free space report on client nodes |
Environment |
All |
Trigger |
NA |
Workaround |
Run a mmdf command on the problematic client node or issue the "df" command from the fs manager node. |
|
5.0.5.7 |
Core GPFS |
IJ29239 |
High Importance
|
logAssertFailed: ofP->mnodeStatusIs(0x4)
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Have many processes/threads on different nodes repeatedly writing the same data to the inode file. |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ31851 |
Suggested |
While running offline fsck the node asserts with signal 11 when checking log files.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Running offline fsck while deleting the system pool disk. |
Workaround |
Patch the SGdesc using tsdbfs and set numlogitems to 0: # tsdbfs fs patch desc numlogitems 0 |
|
5.0.5.7 |
FSCK |
IJ31852 |
High Importance
|
Assert exp(isValidSocket(sock)) in line 2722 of file thread.C
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
mmhealth daemon is down. |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ32009 |
High Importance
|
GNR rebalance unable to complete after many days.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Add an enclosure to an existing RG or MES upgrade. |
Workaround |
None |
|
5.0.5.7 |
ESS, ECE, GNR |
IJ31853 |
Suggested |
The provided improvements result in a more robust functionality of the MM command interface.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
NA |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ31880 |
High Importance
|
GPFS daemon assert: retryCount <= 300
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Running mmrestripefs command and disk related commands such as mmchdisk, mmdeldisk. |
Workaround |
Avoid running mmrestripefs and disk related commands such as mmchdisk, mmdeldisk at the same time. |
|
5.0.5.7 |
Core GPFS |
IJ29447 |
High Importance
|
When performing IO on a very large file, contention for InodeCacheObjMutex could occur as a number of buffers for the file increases. This is more likely to happen on file systems with smaller block sizes.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Running IO to a large file. |
Workaround |
None |
|
5.0.5.7 |
Core GPFS |
IJ31902 |
High Importance
|
There is a small possibility for both replica to be placed in the same failure group when there is a disk configuration change and one failure group is low on free space.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Disk configuration change via mmdeldisk, mmchdisk. |
Workaround |
Avoid delete disk when one failure group is low on free space. |
|
5.0.5.7 |
Core GPFS |
IJ31927 |
Critical |
GPFS daemon could fail unexpectedly with assert: commitRanges[i].valid || commitRanges[i].numBytes <= lfP->sgP->getWriteCacheThreshold(). This could happen after mmchfs command was issued to reduce write cache threshold while applications are actively writing to the file system.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Reducing the write cache threshold using the mmchfs command. |
Workaround |
Do not reduce write cache threshold while there is write workload on the file system. |
|
5.0.5.7 |
HAWC |
IJ32035 |
Critical |
Assert in mmfsd "Signal 6 at verbs::parseConfigVerbsPorts, at verbsInit.C:4407", resulting in a Spectrum Scale crash at start up.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
RDMA is active. |
Workaround |
Disable RDMA. |
|
5.0.5.7 |
RDMA |
IJ32038 |
High Importance
|
There appears to be an issues at the systemd layer that causes startup service to fail with connection time out during reboot. If auto load is set to yes, GPFS may not be able to start up or it may get stuck waiting for the environment to be initialized.
(show details)
Symptom |
GPFS does not start after a reboot. |
Environment |
Linux |
Trigger |
This issue affects clusters with auto load set to yes and hitting systemd connection time out during reboot. |
Workaround |
Manually restart GPFS. |
|
5.0.5.7 |
GPFS startup, CCR, systemd |
IJ30432 |
High Importance
|
When mmdelnode is issued against a node whose mmfsd daemon is still up, several of the nodes in the cluster can fail with messages such as the following: [E] Deleted node 169.28.113.36 (nodeX) is still up. [E] Node 169.28.113.36 (nodeX) has been deleted from the cluster
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Issuing mmdelnode against a node whose mmfsd daemon is still up on the target node, or against a node whose status cannot be determined. |
Workaround |
Ensure that the mmfsd is down on a given node before issuing a mmdelnode command against that node. If the status of the target node cannot be determined, ensure that the node gets powered down. |
|
5.0.5.6 |
Cluster Membership |
IJ30346 |
Suggested |
Some processes may not be woken up as they should during a cluster manager change. That might lead to potential deadlocks.
(show details)
Symptom |
Long Waiters |
Environment |
All |
Trigger |
Cluster manager change |
Workaround |
None |
|
5.0.5.6 |
Core GPFS |
IJ30393 |
High Importance
|
The GPFS daemon could fail with logAssertFailed: fromNode != regP->owner. This could occur when a file system's disk configuration is changed just as a new file system manager is taking over.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Disk configuration change by using commands such as mmadddisk, mmdeldisk and mmchdisk |
Workaround |
Avoid disk related commands right after a new file system has been appointed. |
|
5.0.5.6 |
Core GPFS |
IJ30402 |
High Importance
|
"Disk in use" error when using not partitioned DASD devices. DASD '/dev/dasdk' is in use. Unmount it first! mmcrnsd: Unexpected error from fdasd -a /dev/dasd. Return code: 1 mmcrnsd: [E] Unable to partition DASD device /dev/disk/by-path/ccw-0.0.0500 mmcrnsd: Failed while processing disk stanza on node node01.abc.de %nsd: device=/dev/disk/by-path/ccw-0.0.0500 nsd=scale_data01 servers=node01.abc.de usage=dataAndMetadata
(show details)
Symptom |
Upgrade/Install failure |
Environment |
Linux (s390x) |
Trigger |
Running the mmcrnsd command on not partitioned DASD devices. |
Workaround |
Before running the mmcrnsd command, partition the DASD devices with fdasd -a, refresh the partition table on all server nodes, and specify the partitions of the devices in the mmcrnsd stanza file. Then, run the mmcrnsd command. |
|
5.0.5.6 |
Installation toolkit |
IJ30408 |
Suggested |
AIO operations on encrypted files are handled as buffered IO, further decreasing the performance of the AIO operation in addition to the crytographic overhead introduced by the encryption of files in the file system.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
AIX (POWER) |
Trigger |
Using AIO |
Workaround |
None |
|
5.0.5.6 |
Encryption |
IJ30451 |
Suggested |
When a user starts the mmrestripefile command against a big file with the -b option, it could take a long time (For example, more than 20 minutes) to return but no data movement is seen between disks. This is because the big file is already balanced.
(show details)
Symptom |
Performance Impact |
Environment |
All |
Trigger |
Using mmrestripefile to rebalance big files. |
Workaround |
None |
|
5.0.5.6 |
mmrestripefile command |
IJ30409 |
Suggested |
Kernel v4.7 changes the inode ACLs cache mechanism, and GPFS (5.0.5.2+, 4.2.3.23+) does not adapt to the new kernel behaviors. The following two typical issues are observed: 1. normal user can access one file, and root removes the file access privilege from the user by chmod command => the user can still access the file 2. normal user cannot access one file, and root grants the file access privilege for the user by chmod command => the user cannot access the file either.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux with kernel v4.7 or later |
Trigger |
GPFS does not adapt to the new kernel behavior. |
Workaround |
Remount the file system. |
|
5.0.5.6 |
Core GPFS |
IJ30458 |
Suggested |
skipRecall config does not work.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
When skipRecall config is used. |
Workaround |
None |
|
5.0.5.6 |
DMAPI |
IJ30427 |
Suggested |
The mmfs.log shows several "sdrServ: Communication error" messages.
(show details)
Symptom |
Error output/messages |
Environment |
All |
Trigger |
Running the systemhealth monitor. |
Workaround |
Restart the systemhealth monitor (mmsysmoncontrol restart). |
|
5.0.5.6 |
System health |
IJ30461 |
High Importance |
mmbackup could backup files unnecessary after failure.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
mmbackup failing during IBM Spectrum Protect client and server backup phase |
Workaround |
None |
|
5.0.5.6 |
mmbackup |
IJ30462 |
High Importance |
Memory leak on file system manager node during quota revoke storm
(show details)
Symptom |
Cluster/File System Outage |
Environment |
All |
Trigger |
Quota usage reaches limit setting |
Workaround |
Restart GPFS. |
|
5.0.5.6 |
Quotas |
IJ30463 |
Medium Importance
|
While migrating a file to the cloud, the gpfs daemon might hit a signal in StripeGroup::decNumAccessRights()
(show details)
Symptom |
Daemon hits a signal during file migration to the cloud. |
Environment |
Linux |
Trigger |
Migrating a file to the cloud. |
Workaround |
None |
|
5.0.5.6 |
TCT |
IJ30429 |
High Importance |
mmfsd crashed due to signal 11 when verifying the file system descriptor.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Descriptor verification |
Workaround |
None |
|
5.0.5.6 |
Core GPFS |
IJ30466 |
Suggested |
mmsmb exportacl list doesn't show "@" of the SMB share name.
(show details)
Symptom |
Error output/messages |
Environment |
Linux |
Trigger |
Using an '@' character in SMB share name. |
Workaround |
Do not use an '@' character in SMB share names. |
|
5.0.5.6 |
SMB |
IJ30397 |
High Importance |
mmvdisk throws an exception for a list operation when the daemon node name is not identical to the admin node name.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
Daemon node name doesn't match admin node name. |
Workaround |
None |
|
5.0.5.6 |
GNR, ESS |
IJ30465 |
High Importance |
Assert failure luEnclosureSlotP == __null
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
A surprise disk pull |
Workaround |
None |
|
5.0.5.6 |
GNR, ESS |
IJ30352 |
Critical |
logAssertFailed (*respPP != __null) cacheops.C
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Writing to the file in AFM read-only mode to a fileset which is exported using a ganesha NFS server. |
Workaround |
Export the AFM read-only fileset path in read-only mode from the ganesha NFS server. |
|
5.0.5.6 |
AFM |
IJ30493 |
Suggested |
The administrator is unable to change the page pool setting on the GNR recovery group server. The problem is seen only on recovery groups not managed by mmvdisk. The mmchconfig command will fail, and the following error message is displayed: The --force-rg-server flag must be used to change the pagepool
(show details)
Symptom |
Error output/message Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
mmchconfig pagepool=newValue on a GNR server not administered by the mmvdisk command set. |
Workaround |
None |
|
5.0.5.6 |
GNR, ESS |
IJ30143 |
High Importance |
After dm_punch_hole call, a dm_get_allocinfo could return improper results for the information of the data blocks allocation.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Calling dm_punch_hole to create hole in a file. |
Workaround |
None |
|
5.0.5.6 |
DMAPI |
IJ30621 |
Suggested |
Running "mmces events list" stdout prints many trailing white characters (empty spaces), unnecessarily.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Running mmces events list. |
Workaround |
None |
|
5.0.5.6 |
CES |
IJ30634 |
High Importance |
An RPC message could be handled twice when TCP reconnect happens This could cause log assertion, FS struct error or be silently handled depending on the type of RPC.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Network is not good which leads to TCP connection reconnect. |
Workaround |
None |
|
5.0.5.6 |
Core GPFS |
IJ30785 |
Suggested |
If the mmvdisk command takes more than 60 seconds to complete, mmhealth reports all pdisks as vanished. On larger systems with many I/O nodes and pdisks, 60 second timeouts are not enough.
(show details)
Symptom |
Unnecessary events shown in mmhealth / GUI |
Environment |
Linux, ESS I/O nodes |
Trigger |
Running mmvdisk command on a system with many I/O nodes and pdisks. |
Workaround |
Manually increase the time out of the mmvdisk command execution in the mmhealth monitoring. |
|
5.0.5.6 |
System health, GUI |
IJ30864 |
High Importance |
In a mixed AIX/Linux cluster, the mmbackup command could fail with gskit/ssl errors after upgrading IBM Spectrum Protect code to 8.1.11, which introduced new rpm for gskit 8.0-55.17 that is not compatible with gpfs.gskit version.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
AIX and Linux mixed cluster |
Trigger |
Run mmbackup in a AIX/Linux mixed cluster with IBM Spectrum Protect 8.1.11. |
Workaround |
None |
|
5.0.5.6 |
mmbackup |
IJ29444 |
Suggested |
After Node-B successfully reestablishes a broken connection to Node-A, Node-A still shows the reconnect_start state (DEGRADED).
(show details)
Symptom |
Error output/messages |
Environment |
All |
Trigger |
Reconnecting broken connections |
Workaround |
Restart the systemhealth monitor (mmsysmoncontrol restart). |
|
5.0.5.6 |
System health |
IJ30878 |
HIPER |
AFM gateway node crashes if the home is not responding and multiple threads are trying to read the same file.
(show details)
Symptom |
Crash |
Environment |
Linux |
Trigger |
Reading the same file from multiple threads when the home is not responding. |
Workaround |
Use the undocumented config option "mmfsadm afm readbypass -1" on the gateway node. |
|
5.0.5.6 |
AFM |
IJ30973 |
HIPER |
AFM gateway node asserts if the home is not responding and multiple threads are trying to read the same file.
(show details)
Symptom |
Crash |
Environment |
Linux |
Trigger |
Reading the same file from multiple threads when the home is not responding. |
Workaround |
Use the undocumented config option "mmfsadm afm readbypass -1" on the gateway node. |
|
5.0.5.6 |
AFM |
IJ30675 |
Critical |
Daemon (AFM) assert goes off: getReqP->r_length <= ksP->r_bufSize
(show details)
Symptom |
Unexpected Results |
Environment |
Linux |
Trigger |
AFM caching mode and a read of an uncached file. |
Workaround |
Increase the afmReadBufferSize config value. |
|
5.0.5.6 |
AFM |
IJ30976 |
HIPER |
Revalidation on a AFM fileset fails on a RHEL 8.3 gateway node and home changes may not be detected causing the data or metadata mismatch between cache and home.
(show details)
Symptom |
Unexpected Results |
Environment |
Red Hat Enterprise Linux 8.3 |
Trigger |
AFM caching, resync and failover. |
Workaround |
None |
|
5.0.5.6 |
AFM, AFM DR |
IJ30778 |
Critical |
With async refresh enabled, file system quiesce is blocked during the remote operation and it might result in a deadlock if the remote is not responding.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
AFM caching mode with async refresh |
Workaround |
None |
|
5.0.5.6 |
AFM |
IJ29433 |
Medium Importance
|
The systemhealth monitor reported a gpfs_down event and triggered a failover even though the system was fine.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Unknown |
Workaround |
None |
|
5.0.5.5 |
System health |
IJ29517 |
Critical |
When an uncached file is renamed in the local-updates mode, the file is not copied to the previous snapshot causing the setInodeDirtyAndVerify assert.
(show details)
Symptom |
Crash |
Environment |
All |
Trigger |
Rename of a uncached file in local-updates mode fileset |
Workaround |
None |
|
5.0.5.5 |
AFM |
IJ29434 |
High Importance
|
While the GPFS daemon is shutting down, there is chance that a specific trace will be logged and it may crash the kernel.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
mmshutdown when trace is enabled and there is still a workload which accesses the GPFS file system |
Workaround |
Stop the workload before mmshutdown, or disable mmtrace before mmshutdown. |
|
5.0.5.5 |
Core GPFS |
IJ29530 |
Suggested |
IBM Spectrum Scale has core dump triggered in dAssignSharedBufferSpace() due to segmentation fault hit by the mmfsd or the lxtrace daemon.
(show details)
Symptom |
Daemon crash |
Environment |
Linux |
Trigger |
Enabling overwrite tracing and using internal lxtrace command to recycle the trace data. |
Workaround |
You should not directly use the lxtrace command. Instead, you should use mmtrace or mmtracectl command to recycle the trace data. |
|
5.0.5.5 |
Trace |
IJ29435 |
High Importance
|
On zLinux, while running an mmap workload with traceIoData configuration enabled, the trace code may trigger a page fault and cause the kernel to crash.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux (s390x only) |
Trigger |
Running mmap workload On zLinux with traceIoData configuration enabled. |
Workaround |
Do not enable traceIoData. |
|
5.0.5.5 |
mmap |
IJ25754 |
High Importance
|
Quota clients request quota shares based on the workload and most of the time the quota shares given to an active client is much larger than the previously pre-defined amount (e.g. 20 file system blocks). The unused or excess quota shares are returned to the quota manager periodically. At the quota manager side, when the quota usage exceeds the established soft quota limits, the grace period is triggered. At this event, the quota shares are reclaimed and the quota share distribution falls back to a more conservative fashion (based on predetermined amount). In certain workloads, when the partial quota shares are returned to the manager along with the usage updates and as a result it triggers the soft quota limit exceeded event, some amount of quota shares are lost due to mismanagement of quota shares between the client and the manager, leading to a permanent loss of quota shares correctable by using the mmcheckquota command.
(show details)
Symptom |
Quota shares loss, thus increasing the in-doubt values, caused by the soft quota exceeded events. The loss of shares can't be reclaimed without running the mmcheckquota command. |
Environment |
All |
Trigger |
Timing and workload specific caused by when the quota usage exceeds the soft limits. |
Workaround |
None. Establishing soft quota limits such that the usage is less likely to trigger repeatedly the "soft quota limit exceeded" events minimizes the timing window of hitting this issue. |
|
5.0.5.5 |
Quotas |
IJ29453 |
High Importance
|
In cases which have small pagepool size and large file system block size, GPFS may wait for page reservation unnecessarily because GPFS tends to reserve more pages than necessary.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Linux |
Trigger |
Workloads which access a file both in mmap and regular read/write method |
Workaround |
None |
|
5.0.5.5 |
mmap |
IJ29535 |
Critical |
After mmimgrestore, the mmfsd could assert when handling the mmlsfileset command for a dependent fileset: logAssertFailed: fsOfP->getDirLayoutP() != __null
(show details)
Symptom |
Operation failure due to file system corruption |
Environment |
All |
Trigger |
Running mmimgrestore |
Workaround |
None |
|
5.0.5.5 |
DMAPI |
IJ29490 |
High Importance
|
Under heavy workload (especially with file creation/deletion involved) with quota function enabled, some race issues are exposed such that the filesetId is not handled correctly, causing a GPFS daemon assert.
(show details)
Symptom |
Cluster/File System Outage |
Environment |
All |
Trigger |
Heavy workload with file creation/deletion involved |
Workaround |
None |
|
5.0.5.5 |
Quotas |
IJ29495 |
High Importance
|
mmchnode fails when more than the current number of quorum nodes becomes quorum nodes again.
(show details)
Symptom |
Cluster/File System Outage |
Environment |
All |
Trigger |
Running the 'mmchnode --quorum -N' command with number of new quorum nodes greater than the number of current quorum nodes. |
Workaround |
Restart the GPFS daemon on the nodes which should become quorum nodes before attempting the 'mmchnode --quorum ...' command. |
|
5.0.5.5 |
mmchnode --quorum, CCR |
IJ29502 |
Suggested |
If the cluster is configured with a separate daemon and admin interfaces, the -Y output of mmgetstate only shows the admin node name.
(show details)
Symptom |
Command output |
Environment |
All |
Trigger |
Running mmgetstate -Y in a cluster with separate daemon and admin interfaces. |
Workaround |
Get the daemon node name from other commands such as mmlscluster. |
|
5.0.5.5 |
Admin commands |
IJ29678 |
High Importance
|
logAssertFailed: !"Cleanup hit contended Fileset lock."
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
File system specific sync, fileset cleanup actions like unlink fileset, or unmount the file system on a given node. |
Workaround |
None |
|
5.0.5.5 |
Filesets |
IJ29679 |
Suggested |
The mmkeyserv command displays the latest expiration date from the KMIPT certificate chain. It should display the expiration date of the end-entity certificate.
(show details)
Symptom |
Error output/message |
Environment |
Linux, AIX |
Trigger |
mmkeyserv command "show" output |
Workaround |
Use openssl to display and view each certificate in the chain to determine the correct expiration date. |
|
5.0.5.5 |
Admin commands, Encryption |
IJ29682 |
Critical |
When truncating a migrated immutable file with DMAPI interfaces, the data of the file becomes zero, although the file is immutable.
(show details)
Symptom |
Data of the file becomes zero |
Environment |
All |
Trigger |
Truncate operation against a migrated immutable file |
Workaround |
None |
|
5.0.5.5 |
Immutable and append-only files |
IJ29514 |
Critical |
File system manager could assert with exp(isStoragePoolIdValid(poolId)) during log recovery if a node fails shortly after running mmdeldisk.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Node failure shortly after deleting a disk from the file system by using mmdeldisk. |
Workaround |
None |
|
5.0.5.5 |
Core GPFS |
IJ29515 |
Suggested |
Incorrect quota check result due to OpenFile reuse/updateShadowTab
(show details)
Symptom |
Unexpected Results |
Environment |
All |
Trigger |
Running mmcheckquota |
Workaround |
None |
|
5.0.5.5 |
Quotas |
IJ29686 |
High Importance
|
On clusters having minReleaseLevel at 5.0.1, with mixed version nodes available from 5.0.1.X through till 5.0.5.X nodes and where the gateway node is at level 5.0.5.X, the newer level of gateway nodes finds it hard to co-exist with the older level nodes causing a recovery failure repeatedly.
(show details)
Symptom |
Unexpected Behavior |
Environment |
Linux |
Trigger |
Triggering AFM Fileset recovery on a mixed version cluster with gateway node at the 5.0.5.X level and minReleaseLevel on the cluster is at the 5.0.1 level. |
Workaround |
None |
|
5.0.5.5 |
AFM |
IJ29719 |
Suggested |
While reading the file, the file can be evicted and its captured checksum shows inconsistency for this opened file.
(show details)
Symptom |
File gets evicted even though the file is opened for operations. |
Environment |
Linux |
Trigger |
Read/write file and run manual evict on the same file. |
Workaround |
None |
|
5.0.5.5 |
AFM |
IJ29533 |
Suggested |
AFM Prefetch doesn't print how many files/inodes it has completed processing for queueing. It would help understand the progress because the actual queueing happens when enough files/inodes have been accumulated and it takes longer to update any progress to the user.
(show details)
Symptom |
Unexpected Behavior |
Environment |
Linux |
Trigger |
Running an AFM listfile or directory prefetch sub-command from the mmafmctl command. |
Workaround |
None |
|
5.0.5.5 |
AFM |
IJ29683 |
Medium Importance
|
A NFS client might be blocked for a while after a failover before it continues I/O.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux (CES nodes with Ganesha/NFS installed) |
Trigger |
A Ganesha/NFS failover |
Workaround |
None |
|
5.0.5.5 |
System health |
IJ29685 |
Suggested |
I/O hangs with mirrored disk while recovery group resigns repetitively due to vdisk fault tolerance exceeded.
(show details)
Symptom |
I/O hang due to long waiters 'waiting for stateful NSD server error takeover (2)' |
Environment |
All |
Trigger |
Mirrored disks and a failed disk |
Workaround |
None |
|
5.0.5.5 |
GNR |
IJ29690 |
Suggested |
The systemhealth monitor reports file systems used with NFS/SMB exports as unmounted even when they are mounted and functional.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux (CES nodes with Ganesha/NFS or SMB installed) |
Trigger |
mmhealth running and file systems having NFS/SMB exports |
Workaround |
Not available. To suppress the misleading events altogether try the following: Edit file /var/mmfs/mmsysmon/mmsysmonitor.conf In section [events], add (if not exist) the following "ignore" entry: ignore =local_exported_fs_unavail,nfs_exports_down,smb_exports_down Change that file for all CES nodes and then restart mmsysmon (mmsysmoncontrol restart) |
|
5.0.5.5 |
System health |
IJ29542 |
Suggested |
Several RAS events had inconsistent values in their SEVERITY and STATE. For instance, the event "network_bond_degraded", which STATE=DEGRADED, has SEVERITY=INFO. As a result, related failures were not propagated properly.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Conditions for raising of one of the following RAS events are met: - bond_degraded - ces_bond_degraded - ces_ips_all_unassigned - gnr_pdisk_unknown - heartbeat_missing - scale_mgmt_server_failed - tct_cs_disabled exports |
Workaround |
None. You need to look up the RAS event definition each time to understand if it is a problem. |
|
5.0.5.5 |
System health, GUI |
IJ29910 |
Suggested |
In IBM Spectrum Scale Erasure Code Edition, it is possible for all of the server's pdisks (physical disks) to become missing, either due to network failure, node failure, or through a planned "node suspend" maintenance procedure. When this
happens, the system will continue to function if there is sufficient remaining fault tolerance. However, smaller configurations with less ECE nodes are exposed to a race condition where pdisk state changes can interrupt a system-wide descriptor update which causes the recovery group to resign. It is also possible to experience this problem with higher probability when using small ESS configurations, such as the GS1 or GS2 enclosures. For both ESS and ECE, a possible symptom may appear in the mmfs.log in this form when a pdisk state change is quickly followed by a resign message claiming VCD write failures before the system fault tolerance is exceeded: 2020-12-01_19:01:36.696-0400: [D] Pdisk n004p005 of RG rg1 state changed from ok/00000.180 to missing/ suspended/00050.180. 2020-12-01_19:01:36.697-0400: [E] Beginning to resign recovery group rg1 due to "VCD write failure", caller err 217 when "updating VCD: RGD" Note that a "VCD write failure" with err 217 is a generic message issued when fault tolerance is exceeded during critical system updates, but in this case the race condition resigns the system when only a handful of missing disks are found.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
pdisk state updates and system-wide descriptor updates on configurations with smaller recovery groups, such as a small number of nodes on ECE or a small Enclosure on ESS. |
Workaround |
Allow the system to resign and recover on its own. For smaller ECE configurations, avoid leaving a node suspended for long periods of time during maintenance tasks. |
|
5.0.5.5 |
GNR, ESS |
IJ29916 |
Suggested |
When the file system is in panic on a quota client node, the outstanding quota share is not relinquished. Quota share Indoubt value is reported and the shares can only be reclaimed by mmcheckquota.
(show details)
Symptom |
Unexpected Results |
Environment |
All |
Trigger |
File system is in panic |
Workaround |
None |
|
5.0.5.5 |
Quotas |
IJ29919 |
Suggested |
Incorrect quota check results on small files with fragments
(show details)
Symptom |
Unexpected Results |
Environment |
All |
Trigger |
Newly created file contains fragments during mmcheckquota execution |
Workaround |
None |
|
5.0.5.5 |
Quotas |
IJ28891 |
Suggested |
In a kernel >4.10 and file-sizes being a multiple of page sizes, a false error is returned once the read offset reaches file size.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Having a file size that is a multiple of the page size |
Workaround |
Do not use a file size that is a multiple of the page size. |
|
5.0.5.5 |
Core GPFS |
IJ29312 |
High Importance
|
Node crash if tremendous parallel access to file with NFS
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
File operation lease is broken. |
Workaround |
None |
|
5.0.5.5 |
kNFS |
IJ29826 |
Critical |
AFM gateway nodes runs out of memory during resync glibc is known to use as many arenas as 8 times the number of CPU threads a systems has. This makes a multi-threaded program like AFM which allocates memory for queues to use a lot more memory than actually needed.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
AFM resync under heavy workload |
Workaround |
None |
|
5.0.5.5 |
AFM |
IJ29939 |
Critical |
Assertion 'exp(lfVersion != other.lfVersion || lfVersion == 0 || tailLsn == other.tailLsn)' in line 99 of file openlog.C
(show details)
Symptom |
Abend/Crash/Lost Membership Data/metadata corruption |
Environment |
Linux |
Trigger |
New cluster manager election started by a quorum node other than the current CC Mgr. This cluster manager election gets started by a so-called challenge written to the tiebreaker disks which must be answered by the current CC manager. |
Workaround |
None |
|
5.0.5.5 |
File IO Cluster manager election |
IJ29763 |
Suggested |
The tsfindinode utility incorrectly reports file path as not found for valid inodes.
(show details)
Symptom |
Incorrect output |
Environment |
All |
Trigger |
There is no problem trigger as this is a tsfindinode issue. |
Workaround |
Use mmapplypolicy to get file paths using the following procedure: Create a policy rule file: RULE LIST 'paths' DIRECTORIES_PLUS WHERE INODE=inum1 OR INODE=inum2 OR ... Then, run the policy scan using the preceding rule: mmapplypolicy -f /tmp -P policy.rule -I defer -m 64 View the result in /tmp/list.files |
|
5.0.5.5 |
tsfindinode |
IJ29209 |
High Importance
|
logAssertFailed: ofP->isInodeValid() at mnUpdateInode when doing stat() or gpfs_statlite()
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
A file is actively written on one node and stat() or gpfs_statlite() are repeatedly called on the file on another node. |
Workaround |
None |
|
5.0.5.5 |
Core GPFS |
IJ30125 |
Suggested |
When there are many threads doing sync writes through the same file descriptor, a contention on inodeCacheOjbMutex between them could impact the performance of writes.
(show details)
Symptom |
Sync write performance becomes bad after upgrading to 5.0.x. |
Environment |
All |
Trigger |
Multiple threads are doing sync writes on the same opened file concurrently. |
Workaround |
Enable the config parameter "prefetchAggressisvenessWrite" to work around the problem if the workload is only doing sync writes. |
|
5.0.5.5 |
Sync writes |
IJ30076 |
Suggested |
Inodes not getting freed after user deleted them
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
A lot of inodes are deleted on a node which overflow background deletion queues. |
Workaround |
mmfsadm dump deferreddeletions mmfsadm test imapWork fs fullDeletion mmfsadm dump deferreddeletions |
|
5.0.5.5 |
Core GPFS |
IJ29780 |
Suggested |
mmvdisk --replace command results in message: Location XXX contains multiple disk devices.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
The problem occurs when hardware errors cause the lower logical block addresses on a disk to become unreadable. |
Workaround |
Contact IBM support for a workaround disk replacement procedure, or upgrade to a code level with the fix. |
|
5.0.5.5 |
GNR, ESS |
IJ29106 |
High Importance
|
The IBM Spectrum Scale HDFS Transparency connector version 3.1.0-6 contains 2 NullPointerExceptions in the HDFS NameNode service. The application accessing the data is not impacted, but these exceptions are seen in the NameNode log file.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Running the HDFS Transparency workload |
Workaround |
None |
|
5.0.5.5 |
HDFS Connector |
IJ29133 |
Suggested |
The IBM Spectrum Scale HDFS Transparency connector version 3.1.0-6 modified the label for the open operation when the configuration is set to "Scale" for the ranger.enabled parameter. When retrieving the JMX stats, the open is reported as GetBlockLocations.
(show details)
Symptom |
Open counts are reported under the label GetBlockLocations |
Environment |
All |
Trigger |
Retrieving JMX stats when ranger.enabled = scale |
Workaround |
Use GetBlockLocation values instead of Open. |
|
5.0.5.5 |
HDFS Connector |
IJ28470 |
Suggested |
The CLiC cryptographic engine by IBM Spectrum Scale has been sunset.
(show details)
Symptom |
None |
Environment |
Linux (PPC64), AIX |
Trigger |
None |
Workaround |
None |
|
5.0.5.4 |
Encryption |
IJ26212 |
Suggested |
mmdumpkthreads is stuck in zombie state.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Run gpfs.snap or the mmdumpkthreads command. |
Workaround |
The zombie process is harmless. Running the "mmfsadm dump thread" command can clear these mmdumpkthreads zombie processes. |
|
5.0.5.4 |
System health |
IJ27098 |
High Importance
|
mmfsd daemon asserting with Assert exp(cfg) in FSFlashDevice.C.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
On nodes with LROC-devices |
Workaround |
None |
|
5.0.5.4 |
LROC |
IJ27087 |
High Importance
|
When QoS throttling is in use, and an application
uses ionice with certain IO priorities, it is
possible for not only that application to experience
degraded performance due to the throttline, but also
for other file system operations to be delayed or to
fail.
(show details)
Symptom |
I/O hang |
Environment |
All |
Trigger |
Application runs with lower I/O priority when QoS is being used. |
Workaround |
None |
|
5.0.5.4 |
QoS |
IJ27923 |
Suggested |
When a user turns off the file system maintenance mode, the file system cannot be mounted.
(show details)
Symptom |
Can not mount the file system. |
Environment |
All |
Trigger |
Perform some file system operations while the maintenance mode is being turned off. |
Workaround |
Stop the attempts of file system operations while the maintenance mode is being turned off. |
|
5.0.5.4 |
Core GPFS |
IJ28498 |
High Importance
|
Long waiters
(show details)
Symptom |
Long waiters |
Environment |
All |
Trigger |
Large pagepool size with most of it used by a single file doing mixed buffered IO and DIO, with other buffered IOs needing to steal buffers. |
Workaround |
Use a smaller pagepool size. Try to avoid doing buffered IO on a file opened for DIO. |
|
5.0.5.4 |
Core GPFS |
IJ28505 |
Suggested |
Opening or closing parenthesis not being accepted in bind_dn password.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
Using '(' or ')' in the bind_dn password. |
Workaround |
Do not use '(' or ')' in the bind_dn password. |
|
5.0.5.4 |
Authentication |
IJ28604 |
Suggested |
Umount file system operation is not successful because of open files.
(show details)
Symptom |
Umount is not successful on the remote file system. |
Environment |
Linux |
Trigger |
Using open control files on remote. |
Workaround |
Use the mmumount -C all_remote command. |
|
5.0.5.4 |
AFM, AFM DR |
IJ28605 |
Suggested |
mmlsmount with the --report and -Y options may not take into account nodes which do not have the file system mounted.
(show details)
Symptom |
Unexpected results |
Environment |
All |
Trigger |
File system is mounted on some nodes and not on others. |
Workaround |
None |
|
5.0.5.4 |
Core GPFS |
IJ28606 |
Suggested |
mmhealth cluster show: faster heartbeat_missing
(show details)
Symptom |
Network outage |
Environment |
Linux, AIX |
Trigger |
Network outage |
Workaround |
Running "mmhealth config interval high" would have a similar effect on heartbeat events but at a cost of a lot of system resources. |
|
5.0.5.4 |
System health |
IJ28434 |
Suggested |
mmgetstate -Y format is showing a negative value
(show details)
Symptom |
Negative value in mmgetstate -Y |
Environment |
Linux |
Trigger |
mmgetstate -Y format |
Workaround |
None |
|
5.0.5.4 |
AFM, AFM DR |
IJ28611 |
HIPER |
AFM uses special control files to replicate ACLs, EAs and to check fileset mode at the home/secondary site. This special control file is not being recognized correctly thus affecting the EA and ACLs replication and fileset mode recognition.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
AFM and AFM DR replication. |
Workaround |
None |
|
5.0.5.4 |
AFM, AFM DR |
IJ28609 |
Suggested |
In mmcheckquota, when a quota entry is processed from a deleted fileset, the quota entry is correctly skipped, but this makes the mmcheckquota process exit with an error.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Error handling |
Workaround |
Rerun mmcheckquota. |
|
5.0.5.4 |
Core GPFS |
IJ28584 |
HIPER |
File and memory leaks in the kernel
(show details)
Symptom |
Memory leak |
Environment |
Linux |
Trigger |
AFM migration on newer kernel versions >= 3.13 |
Workaround |
None |
|
5.0.5.4 |
AFM |
IJ27905 |
Critical |
Failback performance issues
(show details)
Symptom |
Failback performance issues |
Environment |
Linux |
Trigger |
FIO 2MSeqwrite Node suspend/resume. AZ CLI vm stop/start, spontaneous vm reboots and error injects |
Workaround |
None |
|
5.0.5.4 |
CES |
IJ28184 |
High Importance
|
mmfsd daemon assert going off: Assert exp(rmsgP != __null) in file llcomm.C, resulting in a daemon crash.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux (CES nodes) |
Trigger |
The RPC component may hit this assert because of missing a memory barrier. |
Workaround |
None |
|
5.0.5.4 |
Core GPFS |
IJ27087 |
High Importance
|
Application runs with I/O priority mapping into a not supported QoS class, which does have IOPS limitation with 1 IOPS, thus leading to I/Os being queued to wait for enough tokens to service the I/O operation. This causes long waiters.
(show details)
Symptom |
I/O hang |
Environment |
Linux (CES nodes) |
Trigger |
Application runs with lower I/O priority when QoS is being used. |
Workaround |
None |
|
5.0.5.4 |
QoS |
IJ28608 |
High Importance
|
If call home data collection process was interrupted because of a the power loss, the following data collection of the same schedule will fail due to the directory already existing.
(show details)
Symptom |
Component level outage |
Environment |
Linux |
Trigger |
Power loss or restart of the Sysmonitor (mmsysmon.py) during a call home data collection |
Workaround |
Run "mmdsh -N all rm -rf /var/mmfs/tmp/gather/*/*". |
|
5.0.5.4 |
Call home |
IJ28626 |
Suggested |
Read-only offline fsck reports references to down disk as corrupt which is not correct behavior.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Read only offline fsck with a down disk |
Workaround |
Ignore fsck report about corrupt references to down disks. |
|
5.0.5.4 |
FSCK |
IJ27414 |
High Importance
|
mmfsck --estimate-only panics at the end of fsck and a new stripe group manager disallows new mounts.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Run multi-node mmfsck --estimate-only while file system is mounted. |
Workaround |
Start and abort an online fsck for all the nodes that had run the mmfsck --estimate-only command. |
|
5.0.5.4 |
FSCK |
IJ28631 |
High Importance
|
Given an IBM Spectrum Scale cluster with 'verbsRdmaCm' set to 'enable' and configured to use RDMA via RoCE, individual nodes will fail to establish a RDMA connection to other nodes when the IP addresses configured on the RDMA adapters include a non-link local IPv6 address.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux |
Trigger |
An IBM Spectrum Scale cluster with 'verbsRdmaCm' set to 'enable' configured to use RDMA via RoCE and RDMA adapters configured with non-link local IPv6 addresses. |
Workaround |
Remove all non-link local IPv6 addresses from all RDMA adapters which are configured to be used by IBM Spectrum Scale. |
|
5.0.5.4 |
RDMA |
IJ28610 |
Suggested |
While a node is tryiing to join a cluster, mmfsd start could encounter a null pointer reference and crash with a signal 11 with a backstack that looks like this: [D] #0: 0x0000559601506BCE RGMaster::getNode FullDomainName(NodeAddr, char**) + 0xAE at ??:0 [D] #1: 0x000055960150CAA2 RGMaster::rgListServers(int, unsigned int) + 0x212 at ??:0 [D] #2: 0x000055960145F21C runTSLsRecoveryGroupV2 (int, StripeGroup*, int, char**) + 0xA8C at ??:0 [D] #3: 0x0000559601460371 runTSLsRecoveryGroup (int, StripeGroup*, int, char**) + 0xB1 at ??:0
(show details)
Symptom |
Daemon crashes with signal 11 |
Environment |
Linux |
Trigger |
A node trying to join a cluster. |
Workaround |
None |
|
5.0.5.4 |
GNR |
IJ28848 |
Suggested |
Can not create SMB share using utf8 chars through CLI.
(show details)
Symptom |
The mmsmb share name is not created and the command fails. |
Environment |
Linux |
Trigger |
A user issuing the mmsmb create command on a non-CES node. |
Workaround |
You can create the SMB share name (using utf8 character set) on a CES node directly. |
|
5.0.5.4 |
SMB |
IJ28801 |
High Importance
|
A bad extended encryption attribute is in a snapshot file, then attempt to delete that snapshot.
(show details)
Symptom |
Unable to delete snapshot. |
Environment |
Linux |
Trigger |
A bad extended encryption attribute is in a snapshot file, then attempt to delete that snapshot. |
Workaround |
None |
|
5.0.5.4 |
Snapshots, Encryption |
IJ28849 |
Critical |
On file system with HAWC enabled, data written to disk could be lost after a node failure when using the system call fdatasync() or Ganesha.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Node failure such as file system panic, quorum loss, etc. with HAWC enabled file system |
Workaround |
Disable HAWC or avoid using the system call fdatasync() to flush the data to disk. |
|
5.0.5.4 |
HAWC |
IJ28877 |
High Importance
|
When file audit logging or watch folder is enabled on a file system, unmounting the file system might result in a waiter that will not clear. This may cause other commands to hang.
(show details)
Symptom |
A persistent waiter that causes commands to hang. |
Environment |
Linux |
Trigger |
Unmounting the file system that has file audit logging or watch folder enabled. |
Workaround |
Recycle the daemon. |
|
5.0.5.4 |
File audit logging, Watch folder |
IJ28889 |
Critical |
mmhealth does not work on AIX.
(show details)
Symptom |
Component level outage |
Environment |
AIX (POWER) |
Trigger |
All new IBM Spectrum Scale installations on AIX 7.1 and AIX 7.2 are generally affected in a major way (Sysmonitor cannot startup). Upgraded installations have some degraded functions. |
Workaround |
1. Touch mmsysmon.json 2. Run mmccr fput mmsysmon.json mmsysmon.json 3. Run mmdsh -N all mmsysmoncontrol restart After this workaround, Sysmonitor will be able to startup, but some of its features (e.g. hiding events) will stay broken. |
|
5.0.5.4 |
System health |
IJ29002 |
Critical |
If the default replication (-m or -r) setting for a file system is set to 1 and mmvdisk is used to add an additional vdisk set to the file system, an exception will be hit if the --failure-groups option is not used.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
File systems using vdisk based NSDs which have either data or metadata replication set to 1. |
Workaround |
In some cases, using the --failure-groups option avoids this issue. |
|
5.0.5.4 |
ESS, GNR |
IJ29004 |
Suggested |
The systemhealth monitor reports data and name nodes as down for the HadoopConService. In fact, both were running.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Running the systemhealth monitor and HadoopConServices |
Workaround |
None |
|
5.0.5.4 |
System health |
IJ28890 |
Critical |
AFM metadata prefetch does not handle hardlinks.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
AFM metadata prefetch |
Workaround |
None |
|
5.0.5.4 |
AFM |
IJ28897 |
Suggested |
When using --skip-inode-check option of offline fsck, it reports false positive extendedAcl corruption.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Offline fsck with --skip-inode-check option |
Workaround |
Do not use --skip-inode-check option of offline fsck. |
|
5.0.5.4 |
FSCK |
IJ29182 |
Critical |
The --metadata-only option hit the assert Assert exp(!"Assert on Structure Error") in prefetch.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
Using the --metadata-only option with a prefetch list-file |
Workaround |
None |
|
5.0.5.4 |
AFM, AFM DR |
IJ29210 |
High Importance
|
mmvdisk recovery group fails when creating log vdisks when creating a new recovery group in a cluster with preexisting recovery groups. An error message "Disk XXX is already registered for use by GPFS" will appear on the command console, and the recovery group creation will fail. Once the problem condition is hit, IBM support must be contacted to correct the conflicting cluster information.
(show details)
Symptom |
Upgrade/Install failure |
Environment |
Linux |
Trigger |
Users who have used mmchrecoverygroup --servers or mmvdisk recoverygroup to change --primary --backup on a preexisting recovery groups in a cluster may be affected when adding a new recovery group to that cluster. |
Workaround |
Avoid using the "--servers" option in mmchrecoverygroup or the "--primary" and "--backup" options in mmvdisk recovery group change to invoke planned temporary recovery group reassignment for maintenance operations. Instead, use the "--active" option in mmvdisk or mmchrecoverygroup to specify temporary recovery group reassignment. |
|
5.0.5.4 |
Admin commands, ESS, GNR |
IJ29313 |
Suggested |
Running prefetch stats is failing with err 22.
(show details)
Symptom |
Prefetch stats fail with err 22 (Invalid Argument) |
Environment |
Linux |
Trigger |
Prefetch stats are triggered on inactive cache fileset state. |
Workaround |
None |
|
5.0.5.4 |
AFM |
IJ29337 |
Critical |
GPFS maintains EA (Extended Attribute) registry to verify the EA priority. Due to incorrect EA registry addition without SG format version check, policy and inode scans might fail in the mixed node cluster environment. This problem could occur while running policy or inode scans in a mixed node environment running with 5.0.5.2, 5.0.5.3, and 5.1.0.0 and other old version nodes as the file system manager.
(show details)
Symptom |
Unexpected results |
Environment |
All |
Trigger |
Running policy or inode scans in a mixed node cluster environment running with 5.0.5.2, 5.0.5.3, and 5.1.0.0 along with older version nodes as the file sytsem manager. |
Workaround |
None |
|
5.0.5.4 |
AFM, Core GPFS |
IJ28321 |
Critical |
logAssertFailed:isValidSocket(sock) line 2661 of file thread.C
(show details)
Symptom |
mmfsd crash |
Environment |
All |
Trigger |
File system unmount operations and disk path rediscovery or NSD disk stats changes |
Workaround |
There is no stable workaround to avoid this, but stopping the unmount of the file system and a stable NSD server node with stable disks could help. |
|
5.0.5.3 |
Core GPFS |
IJ28305 |
HIPER |
When the fileset is in the stopped state (mmafmctl device stop -j fileset), modifying the file's metadata such as owner permissions and file times may not be replicated to the home/secondary site after the fileset is restarted (mmafmctl device start -j fileset).
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
AFM fileset is in stopped state and recovery is run later. |
Workaround |
None |
|
5.0.5.3 |
AFM, AFM DR |
IJ28314 |
Suggested |
In a distributed IBM Spectrum Scale environment in the presence of repetitive node failures can result in the declustered array becoming stuck in-transition with long waiters. Long waiters may occur, and file system operations may become stalled.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Linux |
Trigger |
Occurs in IBM Spectrum Scale RAID configurations involving multiple log groups, such as IBM Spectrum Scale Erasure Code Edition or ESS 3000, that experience repetitive node shutdowns or network partitioning affecting a small subset of nodes in the cluster. |
Workaround |
Restart the daemon. |
|
5.0.5.3 |
GNR |
IJ28316 |
Critical |
AFM list file prefetch hangs.
(show details)
Symptom |
Unresponsiveness |
Environment |
Linux |
Trigger |
AFM list file prefetch |
Workaround |
Split the list file and rerun prefetch. |
|
5.0.5.3 |
AFM |
IJ28162 |
HIPER |
AFM replicating extended attributes and ACLs can cause resync/recovery performance issues.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
Linux |
Trigger |
AFM resync and recovery; also occurs during the regular AFM replication. |
Workaround |
None |
|
5.0.5.3 |
AFM, AFM DR |
IJ28161 |
Suggested |
Changing the file system name via mmchfs -W option does not work if the file system manager is on another node that is not a Linux node. This is due to the new device name not being created in /dev at the time the daemon open the file system with the new device name.
(show details)
Symptom |
Command failure error message in mmfs.log |
Environment |
AIX on Power, Windows on x86_64 |
Trigger |
Change the device name of a file system when the file system manager is on another node and that node is not a Linux node. |
Workaround |
Ensure the next file system appointment is on a Linux node. This can be done by rearranging the list of quorum nodes. If there are no Linux nodes in the cluster, run the mmchfs command on the file system manager node. |
|
5.0.5.3 |
Core GPFS |
IJ27709 |
High Importance
|
When a node becomes a quorum node (via mmchnode command) and this node was a quorum node sometime in the past and a mmfsd/mmsdrserv restart has not occurred on this node since; In that case, the mmchnode command will fail with output like "initialize (1, 'node-11', ('192.168.1.1', 1191)) failed (err 73) mmchnode: 6027-1639 Command failed. Examine previous error messages to determine cause." Also an assertion of the following type occurs on the node which should become a quorum node (GPFS log): "ccrmmfsd.C:806: assertion 'nodeId == ccrNodeId'"
(show details)
Symptom |
mmchnode command fails GPFS daemon asserts on the node which should become a quorum node. |
Environment |
All |
Trigger |
Changing a node, which was a quorum node sometime in past and is now a nonquorum node to a quorum node again via the mmchnode command. During this entire time the GPFS daemon (mmfsd/mmsdrserv) hasn't seen a restart. |
Workaround |
Restart the mmfsd (if possible) on the node which should become a quorum node before the mmchnode command is attempted. |
|
5.0.5.3 |
mmchnode, CCR |
IJ27801 |
Suggested |
The GPFS kernel module exports an ioctl interface used by the mmfsd daemon and some of the mm* commands. The provided improvements result in a more robust functionality of the kernel module.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
NA |
Workaround |
None |
|
5.0.5.3 |
Core GPFS |
IJ27929 |
High Importance
|
When recycling the recovery group nodes, access to the recovery groups can be lost, even if the nodes are recycled one at a time.
(show details)
Symptom |
File system outage |
Environment |
Linux |
Trigger |
Recycling active recovery group servers |
Workaround |
Recycle the nodes manually. |
|
5.0.5.3 |
GNR |
IJ27922 |
Medium Importance
|
A race condition within GNR fast write logging mechanism could lead to a thread incorrectly considering itself as the candidate to start zone flushing operations, while there are still other threads that have not yet completed the fast (short) writing operation yet in this zone. The assert stopped the flushing operation.
(show details)
Symptom |
Assert will be triggered when this problem occurs: lbZoneP[zone].zHoldWord == 0 |
Environment |
Linux |
Trigger |
The problem normally occurs when multiple short write requests are in process at the same time. Some threads with bigger request size, while others with smaller request size. |
Workaround |
None |
|
5.0.5.3 |
GNR, ECE, ESS 3000 |
IJ27711 |
High Importance
|
When a metanode receives many RPC requests from the clients nodes, it is possible for mutex contention to occur which in turn can lead to high CPU usage. This could happen when many nodes share access to the same file/directory such as the root directory of the file system.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Many nodes share read access to same file and or directory |
Workaround |
Set the preferDesignatedMnode configuration variable to 1. Set the mnodeRemoteAccessThreadhold configuration variable to 0. |
|
5.0.5.3 |
Core GPFS |
IJ25651 |
High Importance
|
GPFS shuts down with the following message "logAssertFailed: holdCount > 0" found in the GPFS log file
(show details)
Symptom |
GPFS will abruptly shut down with the following message "logAssertFailed: holdCount > 0" found in the GPFS log file. |
Environment |
All |
Trigger |
Heavy pagepool activity, particularly with heavy page steal activity and also low memory configurations with compacted objects |
Workaround |
Reduce the pagepool contention by tuning the workload and increasing the pagepool size. |
|
5.0.5.3 |
Core GPFS |
IJ26952 |
High Importance
|
logAssertFailed: isValidSocket(sock): This assertion goes off when the sock file descriptor is too large
(show details)
Symptom |
Abend/Crash |
Environment |
AIX on Power, Linux |
Trigger |
Socket file descriptor leak |
Workaround |
None |
|
5.0.5.3 |
Core GPFS |
IJ26702 |
Suggested |
mmauth grant sets the local file system (which is remote for cache) access to Read-only access. So when you access the fileset over the control file, it returns a E_ROFS error from remote file system and it logs irrelevant information.
(show details)
Symptom |
Logged the error messsage "AFM is not enabled". |
Environment |
All |
Trigger |
When the access grant for the file system is Read-only at home. |
Workaround |
None |
|
5.0.5.3 |
AFM |
IJ27144 |
High Importance
|
When reconfiguring Object protocol authentication using the mmuserauth command, the command may occasionally hang while waiting for systemctl to shutdown a service. Looking at the process table may show systemctl waiting for the child process "pkttyagent" to complete.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Red Hat Enterprise Linux versions 7.7 and later |
Trigger |
This issue affects customers running IBM Spectrum Scale with Object protocol on Red Hat Enterprise Linux version 7 at versions 7.7 and later. |
Workaround |
Since the problem is intermittent, stopping the mmuserauth command and retrying it may result in the command successfully completing. |
|
5.0.5.3 |
Object |
IJ26697 |
High Importance
|
When multiple applications on a single node perform readdir and lookups on the same directory in a loop, it could lead to token starvation on other nodes trying to perform rename/create operation on the same directory. This will show up as slow application performance on affected nodes.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Running multiple application that repeatedly perform readdir/lookup (ls -l) on the same directory. |
Workaround |
None |
|
5.0.5.3 |
Core GPFS |
IJ26830 |
Critical |
If a file is unlinked after opening it, fallocate(2) on that fd will fail with ENOENT.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Unlinking a file after opening it, then call fallocate(2) on that fd. |
Workaround |
None |
|
5.0.5.3 |
Core GPFS |
IJ27150 |
Suggested |
Disable enableIpv6 by using mmchconfig enableIpv6=no does not work. The command treats the no value like yes value.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Run mmchconfig enableIpv6=no |
Workaround |
Instead of mmchconfig enableIpv6=no, issue mmchconfig enableIpv6=default |
|
5.0.5.3 |
Admin commands |
IJ27241 |
High Importance
|
On a Linux node defined to have the GPFS node role of snmp_collector, the GPFS control scripts might not properly determine if the GPFS Net-SNMP subagent is running. Consequently, multiple GPFS Net-SNMP subagent processes might be running at the same time. This might affect the ability of the snmp_collector node to respond to SNMP queries or send SNMP traps. Recurring error messages might be found in GPFS log files showing a GPFS Net-SNMP subagent is unable to register with the Net-SNMP agent (snmpd).
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
GPFS Net-SNMP subagent processes left over from shutdown. |
Workaround |
Manually kill all left over GPFS Net-SNMP subagent processes on shutdown. |
|
5.0.5.3 |
SNMP, Admin commands |
IJ27249 |
Suggested |
The ECE disk inventory utility to list out all the disk slots in the system could hit an exception when one of the LSI adapters doesn't have a disk. This may result in the missing slot locations for the pdisks. The slot location is important to identify the disk drive in disk replacement, so it may cause the disk replacement to fail. If multiple disks fail without being replaced, it may risk the data reliability of the ECE storage system.
(show details)
Symptom |
Error output/message Cluster/File System Outage |
Environment |
Red Hat Enterprise Linux |
Trigger |
NA |
Workaround |
When disk replacement fails due to this problem, contact IBM support to apply work around steps for disk replacement. |
|
5.0.5.3 |
GNR, ECE |
IJ26972 |
Suggested |
Grafana bridge returns "- ERROR - Metric mynode|GPFSNSDFS|exfld|gpfs_nsdfs_bytes_read cannot be found. Please check if the corresponding sensor is configured - "
(show details)
Symptom |
Error message |
Environment |
All |
Trigger |
Using grafana bridge |
Workaround |
None |
|
5.0.5.3 |
Core GPFS |
IJ27264 |
High Importance
|
The check for mismatched NSDs in a storage pool may fail if the regular NSD was created with a pool designation different than the one used in the mmadddisk stanza file.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Adding regular NSDs to a filesystem which contains vdisk-based NSDs into a new pool which does not match the pool used when the NSD was created. |
Workaround |
Use the --force-nsd-mismatch option. |
|
5.0.5.3 |
GNR, ESS |
IJ27236 |
High Importance
|
In certain conditions, the monitoring code will not work properly resulting in an erroneous state being shown.
(show details)
Symptom |
Error output/message Unexpected Results/Behavior |
Environment |
All |
Trigger |
CES policy is "node-affinity" and ces-groups are declared. |
Workaround |
Switch CES Policy to "even-coverage". |
|
5.0.5.3 |
System health |
IJ27285 |
Suggested |
Remote mount is not responsive and the control file failed(-1) to set. Due to this, it is returning E_STALE and the lookup gets requeued with E_RESTART every time.
(show details)
Symptom |
ls command hung and failed to return. |
Environment |
All |
Trigger |
ls command is executed on the fileset root. |
Workaround |
None |
|
5.0.5.3 |
AFM |
IJ27339 |
Suggested |
Notification messages for TLS socket disconnect as a result of a peer's idle disconnect were printed in the mmfs log file and the sys log, Creating confusion whether there is a real problem or not.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Expiration of the idleSocketTimemout timer |
Workaround |
None |
|
5.0.5.3 |
Authentication |
IJ27366 |
Critical |
With parallel IO enabled, it is possible that all the afmMaxWorkerThreads are used and there are no threads to handle the parallel IO responses.
(show details)
Symptom |
Deadlock |
Environment |
All |
Trigger |
AFM with parallel IO under heavy workload |
Workaround |
Disable parallel IO. |
|
5.0.5.3 |
AFM, AFM DR |
IJ27375 |
High Importance
|
Windows offline bit is set on the directory after AFM replication.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
AFM revalidation in modes such as IW/LU/RO |
Workaround |
None |
|
5.0.5.3 |
AFM |
IJ27003 |
Critical |
AFM sets incorrect creation time on symlinks during the migration.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
AFM caching and migration |
Workaround |
None |
|
5.0.5.3 |
AFM |
IJ27038 |
Critical |
AFM incorrectly sets secondary mode fileset permissions to primary mode fileset permissions during a resync operation.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Running changeSecondary command when secondary and primary fileset have different permissions. |
Workaround |
Change the secondary fileset permissions to the primary fileset permissions before running the changeSecondary command. |
|
5.0.5.3 |
AFM |
IJ27646 |
Suggested |
mmadquery fails with error - size limit exceeded
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
AD server having large number of entries (> 1000) |
Workaround |
None |
|
5.0.5.3 |
Authentication |
IJ27683 |
Critical |
With parallel IO enabled, the PIO response handler thread on MDS might deadlock due to an EIO error from the helper gateway node.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
AFM with parallel IO under heavy workload |
Workaround |
Disable parallel IO. |
|
5.0.5.3 |
AFM, AFM DR |
IJ16663 |
High Importance
|
When multiple applicatons on a single node perform readdir and lookups on the same directory in a loop, it could lead token starvation on other nodes trying to operate on the same directory. This will result in as slow application performance on affected nodes.
(show details)
Symptom |
Performance Impact/Degradation |
Environment |
All |
Trigger |
Running multiple application that repeatedly perform readdir/lookup (ls -l) on the same directory. |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ22152 |
Suggested |
There is no automatic method to generate the GPL installable package for customized RHEL release.
(show details)
Symptom |
NA |
Environment |
RHEL |
Trigger |
NA |
Workaround |
Follow the steps in /usr/lpp/mmfs/src/README. |
|
5.0.5.2 |
Core GPFS |
IJ26349 |
Critical |
For a DMAPI enabled file system, migrating files into an external storage pool may cause problems with snapshot files. In some cases, it might assert "okToIncreaseIndLevel". In another case when DMAPI is not enabled, adding extended attributes to a sparse file may trigger the same assert.
(show details)
Symptom |
Daemon crash with assert "okToIncreaseIndLevel" |
Environment |
All |
Trigger |
Using snapshots with a DMAPI enabled file system and migrating files to external storage pools. |
Workaround |
None |
|
5.0.5.2 |
Snapshots, DMAPI |
IJ24387 |
Suggested |
For mmces related log files, set mode bits to 622.
(show details)
Symptom |
Non root user can accidentally delete these files. |
Environment |
Linux |
Trigger |
NA |
Workaround |
Manually change the file access. |
|
5.0.5.2 |
CES |
IJ25803 |
Critical |
open(2) with O_NOATIME flag may not work as expected.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
open(2) with O_NOATIME flag |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ26355 |
Suggested |
Inode operations ->set_acl and ->get_acl are not supported in GPFS, and starting with kernel v3.14 some commands such as nfs4_setfacl may fail.
(show details)
Symptom |
Error output/message |
Environment |
Linux with kernel version 3.14 or later |
Trigger |
Missing functions in the GPFS kernel portability layer |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ26510 |
Critical |
GPFS daemon crashes and file system gets unmounted. The GPFS daemon crashes because of daemon code hit an assert that indicates "the allocated disk address is not expected".
(show details)
Symptom |
GPFS daemon crashes with logAssertFailed: dataBlockNum == lastDataBlock || !newDA.isALLOC(), file system unmounted |
Environment |
All |
Trigger |
The assert went off during file sync as it detected that a middle data block of a file has a fragment block address (DA) - it's illegal; only the last data block can be a fragment. |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ26341 |
High Importance
|
GPFS Access Control Lists (ACL) can only store limited types of Access Control Entries (ACE), specifically plain Access-Allowed-ACE, Access-Denied-ACE, System-Audit-ACE, and System-Alarm-ACE. GPFS does not support storing of any of the Object-specific-ACEs corresponding to ACL_REVISION_DS. An attempt to set an ACL (containing the unsupported ACE types, such as the Object-specific-ACEs), can result in a kernel bugcheck.
(show details)
Symptom |
Abend/Crash |
Environment |
Windows |
Trigger |
The ACL being set on a GPFS file or directory contains an unsupported ACE type, such as Object-specific-ACE. |
Workaround |
None |
|
5.0.5.2 |
Windows, ACLs |
IJ26342 |
Critical |
If incorrect arguments are passed to tschcarrier, a cleanup routine trys to remove some in-memory objects which are not created and this leads to a segmentation fault.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Customers running tschcarrier directly |
Workaround |
None |
|
5.0.5.2 |
GNR |
IJ26356 |
Suggested |
When a data management application receives a DMAPI postrename event, it fails to get the file handle for the renamed file with a "no such file" error. This is because Spectrum Scale is delivering a DMAPI postrename event before the Linux kernel updates its directory lookup cache for the file being renamed.
(show details)
Symptom |
The data management application gets a ENOENT error if it is trying to get the file handle for the renamed file immediately after receiving a DMAPI postrename event. |
Environment |
All |
Trigger |
DMAPI enabled Spectrum Scale file system and a DMAPI postrename event |
Workaround |
Let your data management application take a short sleep and retry to get the file handle of the renamed file name. |
|
5.0.5.2 |
DMAPI |
IJ26022 |
High Importance
|
After upgrading the Spectrum Scale version to 5.0.4.4 or later, hit this assert (logAssertFailed: status == ASumWriting) on the upgraded node when it takes over the file system manager role.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Upgrade to 5.0.4.4 or later and move the file system manager to the newly upgraded node. |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ26358 |
Suggested |
Eviction is failing to execute if there is a space character in the path name.
(show details)
Symptom |
Data is not evicted from the site. |
Environment |
All |
Trigger |
A space character in the path name |
Workaround |
None |
|
5.0.5.2 |
AFM |
IJ26423 |
High Importance
|
When adding a vdisk set which contains multiple node classes to a file system, some node classes may be omitted.
(show details)
Symptom |
Unexpected result |
Environment |
Linux |
Trigger |
Vdisk sets which contain multiple node classes. |
Workaround |
Run the mmvdisk fs add command again. |
|
5.0.5.2 |
GNR |
IJ26438 |
High Importance
|
For pre-4.1 file systems, after the mmquotaoff command deactivates user/group/fileset quota, the old quota file will be deinstalled and converted to a normal file. If the system pool cannot contain the data, the old quota file will need to be moved from the system pool to a data pool. If the file system has DMAPI enabled (-z yes), the deinstallation process will encounter assert exp(context != unknownOp) in moveDataBetweenPools.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
mmquotaoff, a pre-4.1 file system, DMAPI enabled |
Workaround |
Use mmchfs -z no to disable DMAPI first before running mmquotaoff command. |
|
5.0.5.2 |
Quotas |
IJ26436 |
Suggested |
The Spectrum Scale mmbackup command translates include/exclude options into policy rules for backup candidate file lists. If the path name specified in the include/exclude option contains any white space, mmbackup translates it incorrectly because space is used as the default delimiter in "dsmc query inclexcl" output.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Using path name with white space in Spectrum Protect include/exclude options |
Workaround |
Use customized policy rule. |
|
5.0.5.2 |
mmbackup |
IJ26348 |
High Importance
|
A race condition between the RG master resign/recovery and the mdi operation on the worker side could lead to a bug in the RG master recovery and cause the the working index (WI) entry to be stuck in the Assigned state. This would further cause the integrity manager thread to block and cause long waiter "wait for working index entry to be committed" on the RG master node. Under this state, it could lead to data integrity issues when the next RG master resign/recovery event occurs. Note that this could only happen in ECE/ESS 3000 environment, and not in legacy ESS environment.
(show details)
Symptom |
The immediate effect will be that on the RG (root owner node) a long waiter will be observed: "wait for working index entry to be committed". Under this state, if the metadata block related to this working index is updated, RG recovery could cause metadata roll back and cause data integrity issues, or the other possible outcome is recovery failure which renders a failure message as "MDI VCD recovery failure in build used metaslot bitmaps: 16". Note that this long waiter could be caused by other situations, and not all cases will result data integrity issue. |
Environment |
Linux |
Trigger |
This is normally exposed by a race condition when the RG master resign/recovery event happens in the middle of the worker side mdi operation, specifically when worker's RPC which modifies mdi state, interleaves with RG master's RPC which inquires about the mdi state. |
Workaround |
The immediate bug does not cause a problem right away, but it does expose a condition that additional events could lead to a recovery failure. When the recovery failure is detected, data has already been corrupted. It will require special steps to allow recovery to complete but in the meantime a certain amount of data will be lost. |
|
5.0.5.2 |
GNR, ECE, ESS 3000 |
IJ26789 |
Critical |
This is a deadlock issue between the file deletion thread and another thread that is finding the inode from the Linux kernel inode hash list while holding the scale file lock. The other thread is mostly doing DMAPI calls.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux with kernel version 4 and later |
Trigger |
Deleting files managed by DMAPI |
Workaround |
None |
|
5.0.5.2 |
DMAPI |
IJ26835 |
High Importance
|
Calling dm_path_to_handle API from DMAPI RENAME event handler could run into a deadlock on the i_mutex lock that is obtained when the file's parent is renamed on V3.0.x Linux kernel versions.
(show details)
Symptom |
Deadlock |
Environment |
Linux |
Trigger |
Calling dm_path_to_handle against a file being renamed from a DMAPI RENAME event handler |
Workaround |
None |
|
5.0.5.2 |
DMAPI |
IJ26679 |
Suggested |
Rare deadlock causing mmrestripefs to hang under high load
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
Running mmrestripefs under high load. |
Workaround |
Run mmrestripefs during off-hours. |
|
5.0.5.2 |
Policy, ILM restripe, Rebalance |
IJ25547 |
Critical |
GPFS daemon crashes, file system gets unmounted on this node. The GPFS daemon crashes because of daemon code hit an assert which indicates the allocated disk address is not
expected.
(show details)
Symptom |
GPFS daemon crashes with logAssertFailed: dataBlockNum == lastDataBlock || !newDA.isALLOC(), file system unmounted. |
Environment |
All |
Trigger |
The assert went off during file sync as it detected that a middle data block of the file has a fragment block address (DA) - it's illegal, we only can have the last data block be a fragment. |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ25843 |
High Importance
|
If a multi-threaded program reads or writes to a file in regular or mmap mixed mode, it may assert with "logAssertFailed: TokenPermits(get_token_mode(), lm_want)"
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Multi-threaded program reads or writes to a file in regular or mmap mixed mode |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ26434 |
High Importance
|
The EA overflow block is a metadata block that should be read using a continuous buffer, but due to a code error, it is considered to be a data block, so a scatter buffer is used which causes a log assert failure.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
If the file's EA entries cannot be placed in the inode, GPFS will place it in a per file EA overflow block. The size of the EA overflow block depends on how many EA entries the file has. If the scatterBufferSize value is smaller than the EA overflow block size, GPFS may hit log assert when the user accesses the file. |
Workaround |
If the size of scatter buffer is bigger than or equal to the EA overflow block size, the log assert can be avoid. Before GPFS version 5.0.0, the default scatterBufferSize value is 32KB that is small and more likely to hit this log assert. Changing the scatterBufferSize value to 256KB could be a work around. |
|
5.0.5.2 |
Core GPFS |
IJ26921 |
Critical |
When deleting an old snapshot or accessing snapshot files from an old snapshot, the operations could run into a out-of-stack space error if a snapshot file contains an overflow block. This causes the mmfsd process to crash with a memory fault error.
(show details)
Symptom |
Daemon crash |
Environment |
All |
Trigger |
Accessing snapshot files from older snapshots |
Workaround |
None |
|
5.0.5.2 |
Snapshots |
IJ25905 |
High Importance
|
When a new fileset is created, the in-doubt quota shares for that fileset should start out at 0. In some situations, a fileset can be created and start out with a non-zero in-doubt value. This can occur when a fileset had been deleted previously, and then the new fileset re-uses the same ID that the deleted fileset had used.
(show details)
Symptom |
NA |
Environment |
All |
Trigger |
NA |
Workaround |
Running mmcheckquota will reset all of the in-doubt values. |
|
5.0.5.2 |
Quotas |
IJ26301 |
High Importance
|
mmadddisk hits assert exp(isStoragePoolIdValid(poolId)) when trying to open a disk due to stripe group descriptor update.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Creating and or deleting storage pools |
Workaround |
None |
|
5.0.5.2 |
Core GPFS |
IJ25906 |
Critical |
After an upgrade from 5.0.4.3 to 5.0.5.0, none of the CES IP addresses can be assigned.
(show details)
Symptom |
Hang of "mmces address add". Other internal CES IP processing could also hang. |
Environment |
Linux (protocol nodes) |
Trigger |
In 5.0.4.3: 2020-04-10_20:44:52.462-0400: [I] mmcesop: run: /usr/sbin/ip a add 192.168.100.254/24 dev enp11s0f1.2500:0 rc=0 In 5.0.5.0: 2020-06-09_17:36:34.014-0400: [I] mmcesop: run: /usr/sbin/ip a add 192.168.100.254/24 dev enp11s0f1.2500:0 label enp11s0f1.2500:0 RTNETLINK answers: Numerical result out of range rc=2 |
Workaround |
None |
|
5.0.5.2 |
CES |
IJ34824 |
High Importance
|
The Ganesha process crashes or the Ganesha work pool threads hang. The crash occurs when nfs4_acl_release_entry() calls hashtable_getlatch(). See the Symptom section of this APAR to view the stack traces for both scenarios.
(show details)
Symptom |
Abend/Crash
The Ganesha process crashed with this stack trace:
(gdb) bt
#0 raise (sig=11)
#1 0x0000000000443529 in crash_handler (signo=11, info=0x7f7a4cdc7570, ctx=0x7f7a4cdc7440)
#2
#3 0x000000000052ac40 in UNALIGNED_LOAD64
#4 0x000000000052ac7b in Fetch64
#5 0x000000000052b46c in CityHash64 (s=0x0, len=20225737280)
#6 0x00000000004fe949 in fsal_acl_hash_both (hparam=0x1c81700, key=0x7f7a4cdca2e0, index=0x7f7a4cdca264, rbthash=0x7f7a4cdca250)
#7 0x00000000004f1f5a in compute (ht=0x1c81700, key=0x7f7a4cdca2e0, index=0x7f7a4cdca264, rbt_hash=0x7f7a4cdca250)
#8 0x00000000004f2946 in hashtable_getlatch (ht=0x1c81700, key=0x7f7a4cdca2e0, val=0x7f7a4cdca2c0, may_write=true, latch=0x7f7a4cdca2a0)
#9 0x00000000004ff415 in nfs4_acl_release_entry (acl=0x7f7e3c005600)
#10 0x000000000053cda2 in fsal_release_attrs (attrs=0x7f7dec285798)
#11 0x000000000053e2fa in mdcache_lru_clean (entry=0x7f7dec2856d0)
#12 0x00000000005424d0 in mdcache_lru_get (sub_handle=0x7f7b9c9e2b10)
#13 0x0000000000551884 in _mdcache_alloc_handle (export=0x1ceebc0, sub_handle=0x7f7b9c9e2b10, fs=0x1cd9120, reason=MDC_REASON_DEFAULT,
func=0x5c2cb0 <__func__.24082> "mdcache_new_entry", line=709)
#14 0x0000000000553563 in mdcache_new_entry (export=0x1ceebc0, sub_handle=0x7f7b9c9e2b10, attrs_in=0x7f7a4cdcaec0, attrs_out=0x0, new_directory=false,
entry=0x7f7a4cdcae38, state=0x0, reason=MDC_REASON_DEFAULT)
#15 0x0000000000547c6f in mdcache_alloc_and_check_handle (export=0x1ceebc0, sub_handle=0x7f7b9c9e2b10, new_obj=0x7f7a4cdcafd0, new_directory=false,
attrs_in=0x7f7a4cdcaec0, attrs_out=0x0, tag=0x5c1a4c "lookup ", parent=0x7f7e3ccb4f80, name=0x7f7cf47b03c0 "ndjsmsdxscm05_2106250501_NfsStats.1.gz",
invalidate=0x7f7a4cdcaebf, state=0x0)
#16 0x000000000055572f in mdc_lookup_uncached (mdc_parent=0x7f7e3ccb4f80, name=0x7f7cf47b03c0 "ndjsmsdxscm05_2106250501_NfsStats.1.gz",
new_entry=0x7f7a4cdcb140, attrs_out=0x0)
#17 0x000000000055bdfb in mdcache_readdir_chunked (directory=0x7f7e3ccb4f80, whence=1853889278, dir_state=0x7f7a4cdcb2f0, cb=0x434608 ,
attrmask=122830, eod_met=0x7f7a4cdcb51f)
#18 0x000000000054981f in mdcache_readdir (dir_hdl=0x7f7e3ccb4fb8, whence=0x7f7a4cdcb2d0, dir_state=0x7f7a4cdcb2f0, cb=0x434608 ,
attrmask=122830, eod_met=0x7f7a4cdcb51f)
#19 0x0000000000434f1b in fsal_readdir (directory=0x7f7e3ccb4fb8, cookie=1853889278, nbfound=0x7f7a4cdcb510, eod_met=0x7f7a4cdcb51f, attrmask=122830,
cb=0x47c32a , opaque=0x7f7a4cdcb4b0)
#20 0x000000000047d9cb in nfs4_op_readdir (op=0x7f7b9c14ea10, data=0x7f7a4cdcbdd0, resp=0x7f7b9cb64440)
#21 0x00000000004614f2 in nfs4_Compound (arg=0x7f7b9c74f868, req=0x7f7b9c74f160, res=0x7f7b9c1426f0)
#22 0x000000000045e355 in nfs_rpc_process_request (reqdata=0x7f7b9c74f160)
(gdb)
The Ganesha work pool threads hang with this stack trace:
(gdb) bt
#0 futex_abstimed_wait (private=0, abstime=0x0, expected=3, futex_word=0x7feca00186dc)
#1 __pthread_rwlock_wrlock_full (abstime=0x0, rwlock=0x7feca00186d0)
#2 __GI___pthread_rwlock_wrlock (rwlock=0x7feca00186d0)
#3 0x00000000004ff09e in nfs4_acl_release_entry (acl=0x7feca00186c0)
#4 0x000000000053cda2 in fsal_release_attrs (attrs=0x7feb5064b688)
#5 0x000000000053e2fa in mdcache_lru_clean (entry=0x7feb5064b5c0)
#6 0x00000000005424d0 in mdcache_lru_get (sub_handle=0x7fea4c54d860)
#7 0x0000000000551884 in _mdcache_alloc_handle (export=0xede340, sub_handle=0x7fea4c54d860, fs=0xeccc00, reason=MDC_REASON_SCAN,
func=0x5c2cb0 <__func__.24082> "mdcache_new_entry", line=709)
#8 0x0000000000553563 in mdcache_new_entry (export=0xede340, sub_handle=0x7fea4c54d860, attrs_in=0x7fe8c8cc9940, attrs_out=0x0, new_directory=false,
entry=0x7fe8c8cc9818, state=0x0, reason=MDC_REASON_SCAN)
#9 0x000000000055840b in mdc_readdir_chunk_object (name=0x7fe8c8cc9ae3 "jbonura1.core.21884.gz", sub_handle=0x7fea4c54d860, attrs_in=0x7fe8c8cc9940,
dir_state=0x7fe8c8cc9ef0, cookie=30487614)
#10 0x0000000000559530 in mdc_readdir_chunked_cb (name=0x7fe8c8cc9ae3 "jbonura1.core.21884.gz", sub_handle=0x7fea4c54d860, attrs=0x7fe8c8cc9940,
dir_state=0x7fe8c8cc9ef0, cookie=30487614)
#11 0x00007fece215d203 in read_dirents ()
#12 0x0000000000559d2f in mdcache_populate_dir_chunk (directory=0xec76f0, whence=28136498, dirent=0x7fe8c8cca160, prev_chunk=0x0, eod_met=0x7fe8c8cca15f)
#13 0x000000000055b10b in mdcache_readdir_chunked (directory=0xec76f0, whence=28008299, dir_state=0x7fe8c8cca2f0, cb=0x434608 ,
attrmask=122830, eod_met=0x7fe8c8cca51f)
#14 0x000000000054981f in mdcache_readdir (dir_hdl=0xec7728, whence=0x7fe8c8cca2d0, dir_state=0x7fe8c8cca2f0, cb=0x434608 ,
attrmask=122830, eod_met=0x7fe8c8cca51f)
#15 0x0000000000434f1b in fsal_readdir (directory=0xec7728, cookie=28008299, nbfound=0x7fe8c8cca510, eod_met=0x7fe8c8cca51f, attrmask=122830,
cb=0x47c32a , opaque=0x7fe8c8cca4b0)
#16 0x000000000047d9cb in nfs4_op_readdir (op=0x7fea4c38d780, data=0x7fe8c8ccadd0, resp=0x7fea4cf0b6a0)
#17 0x00000000004614f2 in nfs4_Compound (arg=0x7fea4cd3a338, req=0x7fea4cd39c30, res=0x7fea4cb18cf0)
#18 0x000000000045e355 in nfs_rpc_process_request (reqdata=0x7fea4cd39c30) |
Environment |
Linux |
Trigger |
The issue mostly occurs when the system is loaded with heavy workloads and mdcache hits high watermark. |
Workaround |
None |
|
5.0.5.2 |
NFS |
IJ25321 |
High Importance
|
Signal 11 occurs while mmdiag --threads is running. For example: [E] Signal 11 at location 0x55F6FF53A47D in process 7840
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Running mmdiag --threads |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25146 |
High Importance
|
A race condition between disks having errors and recovery groups or log groups resigning could lead to a bug in the GNR log: 'vtrack recovery failed to scrub and repair the stale data on the disk'. It can further lead to data corruption if all good copies of the mirrored data are lost.
(show details)
Symptom |
This problem could occur either to the fast write log or the VCD log, and the impact will be failure to complete recovery. The log will indicate:
2020-05-29_16:25:21.693-0700: [E] Recovery failure: at least 3400 sectors of the fast-write log for LG root of RG RG1 could not be recovered. 2020-05-29_16:25:21.693-0700: [E] Beginning to resign log group root in recovery group RG1 due to "recovery failure", caller err 333 when "recovering log group worker" or 2020-04-29_09:04:42.567-0400: [E] Recovery failure: at least 41616 sectors of the VCD log for LG LG001 of RG rg1 could not be recovered. 2020-04-29_09:04:42.567-0400: [E] Beginning to resign log group LG001 in recovery group rg1 due to "recovery failure", caller err 333 when "recovering log group worker" |
Environment |
NA |
Trigger |
This is normally exposed by a race condition between a number of disks gradually hitting errors causing recovery groups or log groups to resign. Some strips of the log vtrack might not be marked as stale, and this could lead to the other strips copying the stale data by mistake. |
Workaround |
None |
|
5.0.5.1 |
GNR |
IJ25463 |
High Importance
|
GPFS daemon assert: exp(this->mutexMagic == MUTEX_MAGIC_VALID) dSynch.C. This could occur during file system unmount.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Unmounting a file system |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25468 |
High Importance
|
GPFS daemon assert: ofP->mnodeStatusIs(0x4) || ofP->mnodeStatusIs(0x2) && indAccessLock.isLockedExclusive() in sync.C
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Multiple nodes accessing or updating the same file |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25714 |
High Importance
|
Potential deadlock when a file is accessed concurrently via mmap and regular file access methods.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
On a system with insufficient free memory, a multi-threaded application accesses the same file in both mmap and regular read/write methods |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25469 |
Suggested |
GPFS daemon failed to start with "Cannot allocate memory" error when prefetchThreads is set to less than 3.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Setting prefetchThreads to less than 3 |
Workaround |
Set prefetchThreads to 3 or higher. |
|
5.0.5.1 |
Core GPFS |
IJ25478 |
High Importance
|
GPFS daemon assert: !"Log file migrate check failed: need" in sgmdata.C. This could happen during mmrestripefs/mmdeldisk/mmrpldisk command.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
File system panic on a client node while running mmrestripefs/mmdeldisk/mmrpldisk command |
Workaround |
Rerun the command. |
|
5.0.5.1 |
Core GPFS |
IJ25511 |
High Importance
|
Under rare circumstances, all quorum nodes could be expelled in case the current cluster manager is expelled due to an error on network level on the current cluster manager node, which results in a cluster wide quorum loss. This applies only in the case RDMA has been activated and all GPFS RPCs are going over RDMA (verbsPorts,verbsRdma and verbsRdmaSend must be set). The current cluster manager will be expelled due to the network error (as expected) and the new elected cluster manager cannot make progress during its following group protocol, because it waits for a 10 seconds linger timeout down in the CCR, when a cached socket connection to the former cluster manager gets closed. This way all quorum nodes will be expelled.
(show details)
Symptom |
Node expel/Lost Membership/Quorum loss |
Environment |
Linux x86_64 |
Trigger |
Error inject on network level (daemon IP address) of the current cluster manager node when RDMA is used for all GPFS RPCs |
Workaround |
None |
|
5.0.5.1 |
RDMA, Cluster Membership, Cluster Manager |
IJ22372 |
High Importance
|
If the gpfsready script fails during GPFS startup, GPFS goes down automatically. This causes GPFS shutdown to hang for 5 minutes.
(show details)
Symptom |
Hang during mmfsd startup |
Environment |
Linux, AIX |
Trigger |
mmfsd startup and 'verifyGpfsReady' is set to 'yes'
|
Workaround |
None |
|
5.0.5.1 |
mmfsd startup |
IJ25555 |
Suggested |
Input file with carriage return causes mmfileid to fail with arithmetic syntax error.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Carriage return characters in an input file |
Workaround |
Manually ensure that all input files are free of carriage return characters. |
|
5.0.5.1 |
Core GPFS |
IJ24043 |
High Importance
|
The mmunlinkfileset command hangs and long waiter "waiting to quiesce" appears. A thread is hung and waiting inside the gpfs_s_delete_inode kernel extension routine.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
Run the mmunlinkfileset command (or commands which need file system quiesce like mmcrsnapshot and mmdelsnapshot) and delete files. |
Workaround |
None |
|
5.0.5.1 |
Filesets, Snapshots |
IJ25557 |
High Importance
|
fallocate(2) will set the wrong file size if the file has fragments and the end position of the fallocate range fits in the last block.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
fallocate(2) on a file which has fragments and the end position of the fallocate range fits in the last block |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25587 |
Suggested |
File system could panic with an error code 2 during the unmount process. This could happen if mmdelsnapshot command is running at the time of the unmount.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Unmounting the file system while the mmdelsnapshot command is still running |
Workaround |
Avoid unmounting the file system while the mmdelsnapshot command is still running. |
|
5.0.5.1 |
Snapshots |
IJ24352 |
Suggested |
LogAssert in file vnodeops.C (oldCount >= 0)
(show details)
Symptom |
Abend/Crash |
Environment |
AIX |
Trigger |
Unclear |
Workaround |
None |
|
5.0.5.1 |
Snapshots |
IJ25591 |
Suggested |
The time of the certificateExpiration field of mmkeyserv -Y output is not correct.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
mmkeyserv -Y option |
Workaround |
None |
|
5.0.5.1 |
Admin commands |
IJ24499 |
High Importance
|
When attempting to move or rename a file from a source to a destination, and the destination file already exists and has permissions set such that its deletion is not allowed, the move/rename operation wrongly ends up overwriting the destination file.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
The destination of the move/rename operation must already exist and have ACLs set such that its deletion is not allowed. |
Workaround |
None |
|
5.0.5.1 |
ACLs |
IJ24742 |
Suggested |
If root squash is enabled on a remote file system, the root user is remapped to a user specified UID and AFM fails to access the remote file system.
(show details)
Symptom |
Unexcepted results |
Environment |
Linux |
Trigger |
Accessing a remote file system using AFM while root squash is enabled |
Workaround |
None |
|
5.0.5.1 |
AFM, AFM DR |
IJ25388 |
Suggested |
A configuration variable "sharedTmpDir" is to instruct mmapplypolicy where to create temporary files shared among nodes during a policy scan. The internal configuration variable table missed containing this variable, which caused an error during GPFS daemon initialization but it does not result in any functional problem.
(show details)
Symptom |
Error message |
Environment |
All |
Trigger |
Restarting the GPFS daemon after configuring "sharedTmpDir" |
Workaround |
None |
|
5.0.5.1 |
Policy, ILM |
IJ25592 |
High Importance
|
The timeout option provided by the mmkeyserv command allows a system admin to fine tune the communication with the key server by providing a timeout value. This timeout value does not apply to the communication with the key server occurring during the execution of the mmkeyserv command, creating the potential of the command returning timeout errors.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Network congestion or resource starvation |
Workaround |
Ensure that the host where the mmkeyserv command is invoked has good connectivity to the key server. |
|
5.0.5.1 |
Admin commands |
IJ25600 |
High Importance
|
The GPFS kernel module's cryptographic functionality used block chaining ciphers that were previously deprecated. Newer versions of the Linux distro removed the deprecated ciphers and, consequently, the block chaining ciphers are not available any longer. On those distros, the GPFS kernel module is updated to use symmetric key ciphers.
(show details)
Symptom |
IO error |
Environment |
Linux |
Trigger |
Use of asynchronous direct I/O with kernels that do not provide block chain ciphers. You can look at /proc/crypto and see if there are any cbc(aes) ciphers of blkcipher type. If so, that will be used. If cbc(aes) is of type skcipher, that will be used. On zLinux, both bkcipher and skcipher alg types are available for cbc(aes); however, the blkcipher alg will be used as it is hardware accelerated. |
Workaround |
None |
|
5.0.5.1 |
Encryption |
IJ25601 |
Medium Importance
|
If a file is migrated to a cloud using TCT, accessing the file in snapshot will not show any contents.
(show details)
Symptom |
No data seen on read |
Environment |
Linux |
Trigger |
Accessing a TCT migrated file thru one of the snapshots |
Workaround |
None |
|
5.0.5.1 |
TCT |
IJ25613 |
Suggested |
In the unlikely scenario that the GPFS configuration file (mmfs.cfg) becomes corrupted, the mmfsd daemon may be affected.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
A corrupted mmfs.cfg file |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25617 |
Suggested |
Even though the tenant contains no keys, it cannot be deleted when there are other clients registered to it. On the same cluster, the client should have already deregistered. So, it is likely that the registered client is from another cluster.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
Deleting the tenant when there are other clients registered to it |
Workaround |
Deregister all registered key clients from the key server before deleting the tenant. |
|
5.0.5.1 |
Admin commands, Encryption |
IJ24558 |
Suggested |
When a fileset is in the deleteRequired state, a blank character is missing between the "latest:" string and the snapshot name if it is too long, thus leading to parsing issues on the snapshot name from the output of mmlsfileset.
(show details)
Symptom |
Error message |
Environment |
All |
Trigger |
A fileset is in deleteRequired state and a global snapshot contains this fileset with a long snapshot name |
Workaround |
None |
|
5.0.5.1 |
mmlsfileset command |
IJ25620 |
High Importance
|
AFM doesn't allow tuning the AFM tunables per node in the cluster. All of them seem to be only for the whole cluster level. A few of them such as afmHardMemThreshold and afmMaxParallelRecoveries need to be tuned at each gateway node.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux, AIX |
Trigger |
Try to set afmHardMemThreshold and afmMaxParallelRecoveries using mmchnode command with -N option. |
Workaround |
None |
|
5.0.5.1 |
AFM |
IJ25624 |
Suggested |
When a GPFS command is being blocked by another command, an informational message will be displayed to remind the user of the blocked command that the command will be resumed after the running conflicted command completes. This is not true for some long running commands, like mmlssnapshot and mmdelsnapshot.
(show details)
Symptom |
Unclear error message |
Environment |
All |
Trigger |
Any two conflicted GPFS commands, such as mmlssnapshot and mmdelsnasphot |
Workaround |
None |
|
5.0.5.1 |
Conflicting GPFS commands |
IJ25369 |
High Importance
|
Prefetch enhancements in 5.0.2 introduced a minor internal checksuch that if the list file is present in an NFS mount commonacross the application and gateway nodes, AFM skips copying this list file from app to the gateway node and uses the list fileas is on the gateway node. But sometimes the same path and file name can exist andhave 2 entirely unrelated files.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux gateway nodes, Linux and AIX application nodes |
Trigger |
Have same named list file on the gateway node and the application node. These files should have a different list of files. Try running prefetch with this list file from the application node. |
Workaround |
Rename the list file on the application node to have a unique name such that it doesn't match any similar file on the gateway node. Or, remove or rename similar named list files on the gateway node so that AFM prefetch doesn't detect the same named list file from the application node that is present on the gateway node. |
|
5.0.5.1 |
AFM |
IJ25652 |
High Importance
|
Cannot commit some log records successfully. But it might have updated the VtrackMap. Due to the resign, update to the memory metadata may not have happened. In such a scenario, if another thread is trying to do a VtrackMap flush operation, it successfully writes the metadata block, then the metadata version updated by the VtrackMap entry and metadata block will be the same, although the log will contain the latest version of the record.
(show details)
Symptom |
Error message |
Environment |
Linux |
Trigger |
A RG resign during a track write operation |
Workaround |
None |
|
5.0.5.1 |
GNR |
IJ25656 |
High Importance
|
AFM builds up temporary files in the /var/ directory for recovery procedures of the AFM fileset. These files are not deleted until the next recovery on the same fileset.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Failure of the AFM normal queue, and recovery getting triggered on next operation on the fileset |
Workaround |
User might have to remove the /var/mmfs/afm/-/recovery files manually, making sure that recovery has completed flushing the queue. |
|
5.0.5.1 |
AFM |
IJ25578 |
Suggested |
mmces event list command does not accept all options described in the man page.
(show details)
Symptom |
Code does not provide all functionality described in the man page. |
Environment |
Linux |
Trigger |
Execute the command: mmces events list HDFS |
Workaround |
Use mmhealth commands for the requested information. |
|
5.0.5.1 |
CES |
IJ25657 |
High Importance
|
AFM is sending lookups on non existent or newly created files to the old system even though readdir was performed on the directory. This is causing too many lookups being sent to old system leading to performance degradtion.
(show details)
Symptom |
Performance degradation |
Environment |
All |
Trigger |
AFM migration |
Workaround |
None |
|
5.0.5.1 |
AFM |
IJ25547 |
Critical |
GPFS daemon crashes and the file system gets unmounted. The GPFS daemon crashes because the daemon code hit an assert which indicates the disk address is not expected.
(show details)
Symptom |
GPFS daemon crashes with logAssertFailed: !"oldDiskAddrFound.compAddr(*oldDiskAddrP)", and the file system gets unmounted. |
Environment |
All |
Trigger |
Failed writes and seeks within a file |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25660 |
Suggested |
Applications that have lots of GPFS system calls may fail with a SIGSEGV.
(show details)
Symptom |
Application encountered SIGSEGV on first GPFS system call; backtrace shows a crash before entering the kernel |
Environment |
All (Primarily on x86_64 and secondarily on Z Linux) |
Trigger |
Extremely high concurrency with threads immediately calling gpfs_stat() (or any GPFS system call) |
Workaround |
Paced or metered thread startup |
|
5.0.5.1 |
GPFS system call library |
IJ25661 |
High Importance
|
RPC message sending thread hung, such as the following: Waiting 34581.7807 sec since 10:04:23, monitored, thread 13813 EEWatchDogThread: on ThCond 0x3FFCDC00E418 (MsgRecordCondvar), reason 'RPC wait'
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
All |
Trigger |
Network is not good which leads to reconnect. |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |
IJ25663 |
Critical |
A CES-IP was declared in /etc/hosts with just the IPand without a host name. This causes hangs in process.
(show details)
Symptom |
The mmces address add command hangs. Other internal CES IP processing could also hang. |
Environment |
Linux (protocol nodes) |
Trigger |
/etc/hosts is not populated with the 3 expected parameters. |
Workaround |
Fix the declarations in /etc/hosts. |
|
5.0.5.1 |
CES |
IJ25664 |
High Importance
|
In a rare case, mmap(2)/munmap(2) system call may block file system quiesce and cause quiesce timeout.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Linux |
Trigger |
Commands like mmdelfileset, mmcrsnapshot, mmdelsnapshot start file system quiesce. At the same time an application calls mmap(2)/munmap(2). |
Workaround |
None |
|
5.0.5.1 |
Filesets, Snapshots |
IJ25472 |
Medium Importance
|
tsbuhelper checkprotection might not work correctly if the filename contains two or more spaces.
(show details)
Symptom |
Spaces being removed from the input file names |
Environment |
Linux |
Trigger |
Using tsbuhelper checkprotection when file name contains more than two spaces |
Workaround |
None |
|
5.0.5.1 |
tsbuhelper checkprotection subcommand |
IJ25685 |
High Importance
|
On systems that are booted in FIPS, the ssh client produces extra messages on stdout. The message "FIPS mode initialized" causes GPFS command to fail. GPFS requires that the shell command produces no extraneous messags.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
Command failure on systems that booted in FIPS mode |
Workaround |
Disable FIPS at boot or write a wrapper script that will remove the extraneous message from ssh. The wrapper script can be added to the GPFS remote shell command. |
|
5.0.5.1 |
Admin commands |
IJ25686 |
Suggested |
If files are already cached and not any file is queued for prefetch then it returns an error.
(show details)
Symptom |
NA |
Environment |
All |
Trigger |
NA |
Workaround |
NA |
|
5.0.5.1 |
AFM |
IJ25579 |
High Importance
|
Cygwin version 3.1.5 released on June 1, 2020, has changed its implementation of symlinks. Cygwin symlinks are now Windows reparse points instead of the older-style system file with header. Due to this change, GPFS on Windows fails to interpret the new Cygwin symlinks. This results in errors during the GPFS daemon startup, specifically in its attempt to load the authorized public key.
(show details)
Symptom |
Abend/Crash |
Environment |
Windows |
Trigger |
Upgrade to Cygwin version 3.1.5 (or later). |
Workaround |
Revert to older level of Cygwin (version 3.1.4 or earlier). |
|
5.0.5.1 |
Windows |
IJ25687 |
High Importance
|
When a snapshot gets deleted, some of the blocks of the snapshot are copied to the previous snapshot to maintain data and metadata consistency. If due to some reason when the deletion of the snapshot is interrupted, there can be a scenario where the blocks are copied to the previous snapshot but the block disk addresses are not yet removed from the snapshot being deleted, the status of such snapshots change to DeleteRequired. This condition is expected for DeletedRequired snapshots and will be handled properly when the deletion of the snapshot is retried. But if before deleting the DeleteRequired snapshot, if an offline fsck is run on the file system then fsck will falsely report such blocks as duplicate address between the DeleteRequired snapshot and its previous snapshots.
(show details)
Symptom |
Incorrect report of corruptions by fsck |
Environment |
All |
Trigger |
This issue can occur when the delete of a snapshot has been interrupted due to some reason causing the snapshot to now show status of DeleteRequired and then offline fsck is run on the file system before deleting the DeleteRequired snapshot. |
Workaround |
Complete deletion of all the DeleteRequired status snapshots before running offline fsck. |
|
5.0.5.1 |
FSCK |
IJ25689 |
Suggested |
Fsck reports a bad DA in a to-be-deleted inode even though such inodes will be cleaned up during normal GPFS operations.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Corrupt disk address in a to-be-deleted inode |
Workaround |
None needed; it is not harmful to allow fsck to repair the corruption. |
|
5.0.5.1 |
FSCK |
IJ25692 |
Suggested |
Running an administrative command within a few seconds after running a read-only offline fsck can lead to an assert.
(show details)
Symptom |
Error Abend/Crash |
Environment |
All |
Trigger |
Administrative command run right after offline fsck |
Workaround |
None |
|
5.0.5.1 |
FSCK |
IJ25698 |
Suggested |
Offline fsck reports false positive replica mismatch if the NSD goes down midway.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
NSD going down while offline fsck is running |
Workaround |
None |
|
5.0.5.1 |
FSCK |
IJ25791 |
High Importance
|
Due to a delayed file close in the VFS layer and a context mismatch, closing the file after the replication does not wait for the file system quiesce causing the remote log assert.
(show details)
Symptom |
Assert/Crash |
Environment |
Linux |
Trigger |
AFM replication using NSD backend with file system quiesce |
Workaround |
None |
|
5.0.5.1 |
AFM, AFM DR |
IJ25792 |
High Importance
|
AFM does not replicate the file the correct amount of times when the time is set using the gpfs_set_times_path and gpfs_set_times API.
(show details)
Symptom |
Unexpected behaviour |
Environment |
Linux |
Trigger |
Using non-POSIX API to set file time on AFM fileset |
Workaround |
None |
|
5.0.5.1 |
AFM, AFM DR |
IJ25856 |
Suggested |
Offline fsck requires a certain amount of pagepool memory for it to run with a single inode scan pass. If the needed amount of pagepool memory is not available, it will display a warning message before starting the fsck scan indicating the number of inode scan passes it will take with the current available pagepool memory. It also shows the amount of pagepool memory it would need to run a complete single inode scan pass. For example, following is the message displayed by fsck if there is insufficient pagepool memory available for fsck to run with a single inode scan pass:
---------------- Available pagepool memory will require 3 inode scan passes by mmfsck.
To scan inodes in a single pass, total pagepool memory of 11767119872 bytes is needed.
The currently available total memory f or use by mmfsck is 8604614656 bytes. Continue fsck with multiple inode scan passes? n
---------------- Now the problem is that in some cases it will display an incorrect value of pagepool memory needed. Another side effect of this problem is that in some cases fsck might not show the above message and instead shows the below incorrect message:
--------------- There is not enough free memory available for use by mmfsck in . Continue fsck with multiple inode scan passes? n
(show details)
Symptom |
Incorrect value in message |
Environment |
All |
Trigger |
This issue is most likely to trigger when offline fsck is run on a large file system but where the nodes do not have small pagepool memory. |
Workaround |
While there is no specific workaround here, you can either choose to continue running fsck with multiple inode scan passes or else try to increase the pagepool with some random amount and keep incrementing as much as possible till it makes fsck run with a single inode scan pass. |
|
5.0.5.1 |
FSCK |
IJ25665 |
HIPER |
mmap may expand the file size to a page boundary if the file is a sparse file and the file size is not a multiple of page size.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
At the mmap page fault time GPFS will allocate a disk block if it's not allocated yet. A mmap read/write beyond file size will cause SIGBUS. But if the file size is not a multiple of page size, the page covers eof is accessible, and a mmap write to this page should keep current file size. |
Workaround |
None |
|
5.0.5.1 |
Core GPFS |