All-Paths-Down (APD)
An All-Paths-Down (APD) situation occurs when all paths to a device are down. As there is no indication whether this is a permanent or temporary device loss, the ESXi host keeps reattempting to establish connectivity. APD style situations commonly occur when the LUN is incorrectly unpresented from the ESXi host. The ESXi host, still believing the device is available, retries all SCSI commands indefinitely. This has an impact on the management agents, as their commands are not responded to until the device is again accessible. This causes the ESXi host to become inaccessible or not-responding in the vCenter Server.
If PDL SCSI sense codes are not returned from a device (when unable to contact the storage array, or with a storage array that does not return the supported PDL SCSI codes), then the device is in an
All-Paths-Down (APD) state, and the ESXi host continues to send I/O requests until the host receives a response.
As the ESXi host is not able to determine if the device loss is permanent (PDL) or transient (APD), it indefinitely retries SCSI I/O, including:
- Userworld I/O (hostd management agent)
- Virtual machine guest I/O
Notes:
- If an I/O request is issued from a guest, the operating system should timeout and stop the I/O.
- Due to the nature of an APD situation, there is no clean way to recover.
- The APD situation needs to be resolved at the storage array/fabric layer to restore connectivity to the host.
- All affected ESXi hosts may require a reboot to remove any residual references to the affected devices that are in an APD state.
- Performing a vMotion migration of unaffected virtual machines is not possible, as the management agents may be affected by the APD condition, and the ESXi host may become unmanaged. As a result, a reboot of an affected ESXi host forces an outage to all non-affected virtual machines on that host.
- vSphere 6.0 and later have a powerful new feature as part of vSphere HA called VM Component Protection (VMCP). VMCP protects virtual machines from storage related events, specifically Permanent Device Loss (PDL) and All Paths Down (APD) incidents. For more information, see the VM Component Protection section in the vSphere Availability Guide.
Permanent Device Loss (PDL)
A clear distinction has been made between a device that is permanently lost (PDL) and a transient issue where all paths are down (APD) for an unknown reason.
For example, in the vmkernel logs, if a SCSI sense code of
H:0x0 D:0x2 P:0x0 Valid sense data: 0x5 0x25 0x0 or
Logical Unit Not Supported is logged by the storage device to the ESXi host, this indicates that the device is permanently inaccessible to the ESXi host, or is in a
Permanent Device Loss (PDL) state. The ESXi host no longer attempts to re-establish connectivity or issue commands to the device.
Devices that suffer a non-recoverable hardware error are also recognized as being in a Permanent Device Loss (PDL) state.
This table outlines possible SCSI sense codes that determine if a device is in a PDL state:
SCSI Sense Code | Description |
---|
H:0x0 D:0x2 P:0x0 Valid sense data: 0x__ /0x25/0x0 | *LOGICAL UNIT NOT SUPPORTED |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x__/0x68/0x0 | *LOGICAL UNIT NOT CONFIGURED |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x4/0x4c/0x0 | HARDWARE ERROR/LOGICAL UNIT FAILED SELF-CONFIGURATION |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x4/0x3e/0x3 | HARDWARE ERROR/LOGICAL UNIT FAILED SELF-TEST |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x4/0x3e/0x1 | HARDWARE ERROR/LOGICAL UNIT FAILURE |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x2/0x4c/0x0 | NOT READY/LOGICAL UNIT FAILED SELF-CONFIGURATION |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x2/0x3e/0x3 | NOT READY/LOGICAL UNIT FAILED SELF-TEST |
H:0x0 D:0x2 P:0x0 Valid sense data: 0x2/0x3e/0x1 | NOT READY/LOGICAL UNIT FAILURE |
*ESXi only checks ASC/ASCQ and if it happens to be 0x25/0x0 or 0x68/0x0, it marks device as PDL.
Note: Some iSCSI arrays map LUN-to-Target as a one-to-one relationship. That is, there is only ever a single LUN per Target. In this case, the iSCSI arrays do not return the appropriate SCSI sense code, so a PDL on these arrays types cannot be detected.
Planned versus unplanned PDL
A planned PDL occurs when there is an intent to remove a device presented to the ESXi host. The datastore must first be unmounted, then the device detached before the storage device can be unpresented at the storage array. For more information, see
How to detach a LUN device from ESXi hosts (2004605).
An unplanned PDL occurs when the storage device is unexpectedly unpresented from the storage array without the unmount and detach being executed on the ESXi host.
VMware provides a feature called Auto-remove for automatic removal of devices during an unplanned PDL. For more information, see see
Disabling PDL AutoRemove feature vSphere 6.x/7.x (2059622).
To clean up an unplanned PDL:
- All running virtual machines from the datastore must be powered off and unregistered from the vCenter Server.
- From the vSphere Client, navigate to the Configuration tab of the ESXi host, and click Storage.
- Right-click the datastore being removed, and click Unmount. The Confirm Datastore Unmount window displays.
- When the prerequisite criteria have been passed, the OK button appears.
Note: If you see this error when unmounting the LUN:
Call datastore refresh for object <name_of_LUN> on vCenter server <name_of_vCenter> failed.
You may have a snapshot LUN presented. To resolve this issue, remove that snapshot LUN on the array side. For more information, contact your storage vendor.
- Perform a rescan on all of the ESXi hosts that had visibility to the LUN.
Note: If there are active references to the device or pending I/O, the ESXi host still lists the device after the rescan. Check for virtual machines, templates, ISO images, floppy images, and raw device mappings which may still have an active reference to the device or datastore.
- If the LUN is still being used and available again, go to each host, right-click the LUN, and click Mount.
Note: One possible cause for an unplanned PDL is that the LUN ran out space causing it to become inaccessible.