When upgrading to VMware vSphere ESXi 8.0 Update 3, administrators utilizing QLogic Fibre Channel over Ethernet (FCoE) adapters may experience severe storage (HPE Synergy) disruptions. This issue manifests as database virtual machines hanging, overall performance degradation, and applications going offline due to datastore disconnects. The ESXi host loses access to volumes, and disks appear degraded within the vSphere Client. These connectivity issues are tied directly to scatter-gather (SG) element alignment failures within the qedf driver. This issue has not been observed on hosts using Emulex adapters utilizing the lpfc driver, or QLogic adapters utilizing the qlnativefc driver
Database VMs hang and overall host storage performance degrades on after upgrading version 8.0.Update 3 - Build 24022510.
Datastore disconnects cause hosted applications to fail.
Volumes and disks appear degraded.
Reviewing the /var/run/log/vmkernel.log reveals DMA mapping and IO errors:
qedf:vmhba#:qedfc_eh_abort:####:Info: IO not found
Unable to map SG array on path vmhba##:C#:T#:L#. Status: DMA mapping could not be completed DMA Error: Can't meet SG element alignment.
/var/run/log/hostd.log reveals lost access to volumes and VMFS heartbeat timeouts:Lost access to volume ########-########-####-########## (datastore) due to connectivity issues. Recovery attempt is in progress and outcome will be reported shortly.,Information,,esx.problem.vmfs.heartbeat.timedout,,Hostname,MM/DD/YYYY, HH:MM:SS"
VMware ESXi 8.0 Update 3 - Build 24022510
QLogic FastLinQ QL41xxx Series/QLogic FastLinQ QL45xxx 10/25 GbE Controller (FCoE)
qedf driver
This issue is caused by a conflict between the DMA constraints exposed by the qedf driver and the Pluggable Storage Architecture (PSA) logic. When I/O requests with large block sizes are issued, they may be split incorrectly based on these constraints, leading to DMA mapping failures and subsequent loss of volume access.
On ESXi 8.0.3 , DMA allocation optimization for large blocks controlled by the ESXi adv configure "ScsiPathSplitUseSimpleCloneBuf
The issue is fixed on ESXi version 8.0 U3h - Build 25067014 & higher as per VMware ESXi 8.0 Update 3h Release Notes
See Download Broadcom products, patches and software to download this release.
To workaround this issue, modify the SCSI path split configuration on the affected ESXi hosts.
ScsiPathSplitUseSimpleCloneBuffer setting (Default is 1): esxcli system settings advanced list -o /Scsi/ScsiPathSplitUseSimpleCloneBuffer Path: /Scsi/ScsiPathSplitUseSimpleCloneBuffer Type: integer Int Value: 1 <<<<==== The default value of this config option Default Int Value: 1 Min Value: 0 Max Value: 1 String Value: Default String Value: Valid Characters: Description: Use simple clone buffers to split IOs at path layer whenever possible. Host Specific: false Impact: noneesxcli system settings advanced set -o /Scsi/ScsiPathSplitUseSimpleCloneBuffer -i 0esxcli system settings advanced list -o /Scsi/ScsiPathSplitUseSimpleCloneBuffer Path: /Scsi/ScsiPathSplitUseSimpleCloneBuffer Type: integer Int Value: 0 <<<<==== This value is reset to 0 after step #2 Default Int Value: 1 Min Value: 0 Max Value: 1 String Value: Default String Value: Valid Characters: Description: Use simple clone buffers to split IOs at path layer whenever possible. Host Specific: false Impact: noneNo reboot is required and changes will be persistent through a reboot
This issue has not been observed on hosts using Emulex adapters (lpfc driver) or QLogic adapters utilizing the qlnativefc driver.
Subscribe to this knowledge article to get updates on this issue.