IJ42140 |
Suggested |
GPFS daemon dumps errors of type "NUMA mbind failed for pagepool address ..." in the mmfs log.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
Being on a system with more than one numa node. |
Workaround |
set verbsNumaAffinity=disable |
|
5.1.5.1 |
RDMA |
IJ41553 |
Suggested |
Part of GPFS are kernel modules that are loaded upon startup and used by other components. Usage counters were not used correctly in the tracedev module, which can lead to the module being unloaded while still in use, resulting in a kernel crash.
One case where this is possible is running the "mmvdisk server configure" and "mmvdisk server unconfigure" commands with the --recycle option.
(show details)
Symptom |
Abend/Crash |
Environment |
Linux |
Trigger |
Run GPFS shutdown and startup. This is a rare problem, so running this or the mentioned "mmvdisk server" command in a loop will be necessary to trigger the problem. |
Workaround |
Avoid stopping GPFS immediately after starting up. |
|
5.1.5.1 |
Core GPFS |
IJ42141 |
Suggested |
mmafmctl prefetch -Y hits segfault
(show details)
Symptom |
segfault. |
Environment |
Linux |
Trigger |
mmafmctl prefetch command with -Y option |
Workaround |
None |
|
5.1.5.1 |
AFM |
IJ42221 |
High Importance
|
When using mmcrcluster (or likely mmaddnode) with a hostname that resolves to over 64 characters in length, various failures can occur during creation.
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Long hostnames (e.g. from /etc/hosts) during node creation. |
Workaround |
Force hostnames to be under 64 characters long. |
|
5.1.5.1 |
CCR |
IJ42222 |
Suggested |
Node designation for pmcollector is not removed after a pod is rescheduled to another node
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
All |
Trigger |
pmcollector pod gets rescheduled to another node. |
Workaround |
Restart system health monitoring on the affected core pod with 'mmsysmoncontrol restart'. |
|
5.1.5.1 |
System Health |
IJ42283 |
High Importance
|
A race between recovery group recovery and pdisk state update broadcasting may make pdisk appear as missing which will prevent log group recovery and block I/O.
(show details)
Symptom |
Stuck IO |
Environment |
Linux |
Trigger |
Recovery group and log group failure from too many missing pdisks that could be caused by bad disks, nodes, or network. |
Workaround |
Restart the gpfs daemons running on the GNR nodes that manage the affected recovery group. |
|
5.1.5.1 |
ESS/GNR |
IJ42294 |
High Importance
|
AFM recovery fails with error 80 due to incorrect checks for the inode attributes. This error causes the replication to be stuck.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
AFM recovery |
Workaround |
None |
|
5.1.5.1 |
AFM |
IJ42360 |
Suggested |
When one CES node gets rebooted, NFS client lock requests might fail with a "NLM_DENIED" error.
(show details)
Symptom |
Lock request will fail (NLM_DENIED or NLM_BLOCKED error can be seen in tcpdump reply frame of LOCK Request). |
Environment |
All |
Trigger |
When one of the protocol nodes of the cluster gets rebooted or a failover happens and a lock request is attempted on the same file. |
Workaround |
None in NFSv3. Issue not present in NFSv4. |
|
5.1.5.1 |
NFS-Ganesha |
IJ42361 |
High Importance
|
AFM gateway node deadlocks during the read operation if both prefetch and application tries to read the same file simultaneously.
(show details)
Symptom |
Deadlock |
Environment |
All |
Trigger |
Read operation on AFM uncached file. |
Workaround |
None |
|
5.1.5.1 |
AFM |
IJ42422 |
Suggested |
If an upgrade is being performed to 5.1.3+ with msgqueue enabled and the IBM Spectrum Scale Knowledge Center instructions are not followed correctly, one can get into a state where 'mmmsgqueue config --remove-msgqueue' can no longer be run to finish the migration off of the msgqueue.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Upgrading cluster to 5.1.3+ without disabling and/or migrating off of the msgqueue beforehand. |
Workaround |
None |
|
5.1.5.1 |
Watch Folder / File audit logging |
IJ42423 |
High Importance
|
When one node in an ESS building block is down, its recovery group will attempt to fail over its recovery group to the remaining server in the building block. To ensure consistency of logged writes, it is necessary for the surviving node to read from the logtip backup device, which is an SSD that sits in the external storage enclosure.
If there are problems with the SAS fabric, and there are frequent transient I/O errors over the SAS fabric, a bug in the ESS log tip logic will not retry the read request and will fail to recover the recovery group.
This means that the fail over operation described above will not complete, resulting in a potential outage. The following message will occur (but is not a sufficient condition) in this situation. [E] Unable to read logTipBackup vdisk RG002LOGTIPBACKUP track 0 due to fatal pdisk IO errors!.
(show details)
Symptom |
Component Level Outage |
Environment |
Linux |
Trigger |
Failing over a recovery group when the SAS fabric is unstable. |
Workaround |
None |
|
5.1.5.1 |
ESS/GNR |
IJ42485 |
High Importance
|
When an NVMe device is becoming active, it is necessary for ESS to poll the device to determine if it is ready for I/O. It does this by polling the final LBA of the device to see if reads are allowed.
This is because the devices become visible to the OS prior to becoming ready to handle read/write requests. The original implementation, however, would incorrectly claim that media errors on the final LBA mean that the device isn't ready.
As a result, it is possible that legitimate media problems on the final LBA of an NVMe will induce ESS to claim that the entire device is not available. This problem can be identified by an NVMe pdisk going missing after seeing unrecovered read errors in the Spectrum Scale RAID recovery group event log (mmvdisk recoverygroup list --events).
(show details)
Symptom |
Component Level Outage |
Environment |
Linux |
Trigger |
Corrupted physical block mapped to the final logical block within an NVMe namespace. |
Workaround |
None |
|
5.1.5.1 |
ESS/GNR |
IJ42496 |
High Importance
|
Below assert going off: logAssertFailed: totalReceived == scatteredP->scattered_total_len || (totalReceived == 0 && scatteredIndex == scatteredP->scattered_count)
(show details)
Symptom |
Abend/Crash |
Environment |
All |
Trigger |
Network is not good which leads to TCP connection reconnect. |
Workaround |
None |
|
5.1.5.1 |
Core GPFS |
IJ42501 |
High Importance
|
pmsensor GPFSVFSX output 0 read and write stats but there are read/write operations, the problem here is that the format of data provided by mmpmon is not expected by Zimon, which caused the output to be wrong.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
Read GPFSVFSX stats |
Workaround |
None |
|
5.1.5.1 |
perfmon (Zimon) |
IJ42615 |
High Importance
|
Some smbtorture tests in vfs.fruit module are failing with errors such as NT_STATUS_OBJECT_NAME_COLLISION, NT_STATUS_INVALID_PARAMETER, bad name. VFS module may unintentionally use filesystem permissions instead of ACL from xattr.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
This is applicable only for fruit module, i.e. Mac OS clients. Problem can occur when accessing streams in VFS module where vfs_acl_xattr subroutine gets used. |
Workaround |
None |
|
5.1.5.1 |
SMB |
IJ42724 |
Suggested |
GEMS trace doesn't print ledState correctly, which might make trouble shooting more difficult.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
None |
Workaround |
None |
|
5.1.5.1 |
GNR |
IJ40858 |
High Importance
|
ECE can control LED for NVMe drives on some systems where LED control can be performed via sysfs. However, some system might be shipped with mixed drives with and without LED control. ECE disk inventory will break out with error intead of handle them properly.
(show details)
Symptom |
Unexpected Results/Behavior |
Environment |
Linux |
Trigger |
Run ECE on an NVMe system with and without LED control. |
Workaround |
None |
|
5.1.5.1 |
GNR |
IJ42745 |
High Importance
|
mmbuildgpl fails on SLES 15.3, new kernel 5.3.18-150300.59.90-default or later with error as below: ?No rule to make target 'vmlinux', needed by '/usr/lpp/mmfs/src/gpl-linux/kdump-kern-dummy.ko
(show details)
Symptom |
mmbuildgpl will fail on SLES 15.3 kernel version 5.3.18-150300.59.90-default or later. |
Environment |
Linux |
Trigger |
mmbuildgpl will fail on SLES 15.3 when kernel is upgraded to 5.3.18-150300.59.90-default. |
Workaround |
lear KBUILD_BUILTIN macro inside
/usr/lpp/mmfs/src/gpl-linux/Kbuild
KBUILD_BUILTIN :=
This can be done after below surrounding code
#For s390x: -pg and -fomit-frame-pointer are incompatible
ifeq ($(ARCH),s390)
ifdef CONFIG_FUNCTION_TRACER
ORIG_CFLAGS := $(KBUILD_CFLAGS)
KBUILD_CFLAGS = $(subst -pg,,$(ORIG_CFLAGS))
endif
endif
KBUILD_BUILTIN := |
|
5.1.5.1 |
Build |
IJ42746 |
High Importance
|
logAssertFailed: fileId.inodeNum > 0 when running AFM Recovery or Resync
(show details)
Symptom |
Lost Membership |
Environment |
Linux |
Trigger |
Role Reversal to make old Primary as Secondary and the old Secondary being promoted to Primary. |
Workaround |
None |
|
5.1.5.1 |
AFM |
IJ42724 |
Suggested |
GEMS trace doesn't print ledState correctly, which might make trouble shooting more difficult.
(show details)
Symptom |
Error output/message |
Environment |
Linux |
Trigger |
None |
Workaround |
None |
|
5.1.5.1 |
GNR |
IJ43115 |
High Importance
|
smb2.streams.names2 test in smbtorture failed with error NT_STATUS_CONNECTION_DISCONNECTED
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
This is applicable only for streams module. |
Workaround |
None |
|
5.1.5.1 |
SMB |
IJ43164 |
Critical |
If verbsRdmaSend configuration is enabled, and the verbs connection is disconnected and reconnected due to any error other than node shutdown or node failure, it may cause some RPC reply messages to be left in the internal table unintentionally.
These messages will remain in the internal table forever, as none of ack messages can clean them up. Deadlock will not occur immediately, because these RPC messages have been processed correctly.
However, the problem may occur when the 32-bit message IDs are wrapped and reused. Some new messages may be recognized as duplicated RPCs and be rejected by the destination node. These new messages will stay in 'pending' state and result in deadlock.
(show details)
Symptom |
Hang/Deadlock/Unresponsiveness/Long Waiters |
Environment |
Linux |
Trigger |
For a cluster which has the verbsRdmaSend configuration enabled, this problem may occur if the verbs connection is disconnected and reconnected due to any error other than node shutdown or node failure (for example because of network issue). |
Workaround |
Recycle GPFS daemon. |
|
5.1.5.1 |
RDMA |
IJ43166 |
Suggested |
The systemhealth monitor for GDS (GPUDirect Storage) does not warn if RDMA is not enabled. An enabled RDMA (mmchconfig verbsRdma=enable) is needed as prerequisite to use GDS.
(show details)
Symptom |
Unexpected Results/Behavior. GDS does not run when RDMA is not enabled. |
Environment |
All |
Trigger |
GDS monitoring enabled, but verbsRdma not enabled. |
Workaround |
None |
|
5.1.5.1 |
RDMA |
IJ43349 |
High Importance
|
The policy rule generated by mmbackup contains an EXTERNAL LIST rule that instructs mmapplypolicy to call BAexecScript which is also generated by mmbackup. One of the options passed to the script was -auditlogname which is a deprecated function.
(show details)
Symptom |
Component Level Outage |
Environment |
All |
Trigger |
This problem occurs if user-defined policy is used with -P option, and the policy contains -auditlogname option, and the Scale version is 5.1.4.0, 5.1.4.1, or 5.1.5.0. |
Workaround |
Remove "-auditlogname=" option from EXTERNAL LIST rule. |
|
5.1.5.1 |
mmbackup with -P option |
IJ43362 |
High Importance
|
Samba was failing to contact domain controllers with DNS SRV requests. Also Configuring permissions at the SMB share level (with mmsmb exportacl) appears to have no effect on file system operations over SMB.
(show details)
Symptom |
Error output/message |
Environment |
All |
Trigger |
This can happen rarely. |
Workaround |
None |
|
5.1.5.1 |
SMB |
IJ43416 |
High Importance
|
Incorrect permissions are set when the objects are downloaded using the mmafmcosctl command.
(show details)
Symptom |
Unexpected results |
Environment |
Linux |
Trigger |
Object download with mmafmcosctl command. |
Workaround |
None |
|
5.1.5.1 |
AFM |
IJ34399 |
Medium Importance |
logAssertFailed (lhsP->size == rhsP->size && lhsP->size == resultP->size) When the command mmrestripefs with -b option is running for rebalancing, when adding more dataOnly or metadataOnly disks, the rebalancing threads could hit this disks bitmap operation assert because of the disks bitmap size is increased due to adding disks.
(show details)
Symptom |
mmfsd daemon process died. |
Environment |
All |
Trigger |
Don't add dataOnly or metadataOnly disks to the file system while mmrestripefs -b option command is in progress. |
Workaround |
Don't add dataOnly or metadataOnly disks to the file system while mmrestripefs -b option command is in progress. |
|
5.1.5.0 |
mmrestripefs command with -b option |
IJ40965 |
Critical |
When AFM fileset snapshot is being created or deleted, either through manual snapshot commands or AFM DR periodic snapshot operations, user's file operations in AFM fileset could proceed without interlocking with snapshot operations, then trigger such assert went off.
(show details)
Symptom |
Daemon crash on AFM gateway node and file system manager node |
Environment |
All |
Trigger |
Doing file operations in AFM fileset, and create or delete snapshots for the same AFM fileset at the same time.
|
Workaround |
No, unless disable or stop snapshot operations in AFM fileset. |
|
5.1.5.0 |
AFM |
IJ41620 |
High Importance
|
Running mmbuildgpl on x86_64 with Linux kernels that include fixes for the retbleed vulnerability (CVE-2022-29900) results in an error. As a result, GPFS is not usable with these kernel versions. Specifically, this problem is hit with:
-
SLES 15 SP3 kernel update 5.3.18-150300.59.87.1 or higher
-
SLES 15 SP4 kernel update 5.14.21-150400.24.11.1
-
Ubuntu 22.04 kernel update 5.15.0-45.48
It is expected that the same changes will also be
backported to RHEL, but no RHEL kernel updates with retbleed
fixes have been released yet. The same applies to
Ubuntu 20.04; no kernel updates have been released yet
with this changes, but this should happen eventually.
The information provided by the
Linux distributions are useful references:
https://www.suse.com/security/cve/CVE-2022-29900.html
https://ubuntu.com/security/CVE-2022-29900
https://access.redhat.com/security/cve/CVE-2022-29900
(show details)
Symptom |
Component Level Outage (GPFS will be unusable on the node). |
Environment |
Linux (x86_64) |
Trigger |
This problem occurs when updating the Linux kernel to a version with retbleed patches included. |
Workaround |
The required change can also be applied manually:
-
Edit the file /usr/lpp/mmfs/src/gpl-linux/Kbuild
- Around line 100 there is a line:
$(KBHOSTPROGS) := lxtrace
- Before that line, add a new one with:
CFLAGS_kdump-kern.o += -mfunction-return=keep
- Save the file and run mmbuildgpl again.
|
|
5.1.5.0 |
Core GPFS |