While testing live partition mobility, we have observed occasional crashes
of the Linux partition. What we've seen is that during the live migration,
for specific configurations with large amounts of memory, slow network
links, and workloads that are changing memory a lot, the partition can end
up being suspended for 30 seconds or longer. This resulted in the following
scenario:
CPU 0 CPU 1
------------------------------- ----------------------------------
scsi_queue_rq migration_store
-> blk_mq_start_request -> rtas_ibm_suspend_me
-> blk_add_timer -> on_each_cpu(rtas_percpu_suspend_me
_______________________________________V
|
V
-> IPI from CPU 1
-> rtas_percpu_suspend_me
-> __rtas_suspend_last_cpu
-- Linux partition suspended for > 30 seconds --
-> for_each_online_cpu(cpu)
plpar_hcall_norets(H_PROD
-> scsi_dispatch_cmd
-> scsi_times_out
-> scsi_abort_command
-> queue_delayed_work
-> ibmvfc_queuecommand_lck
-> ibmvfc_send_event
-> ibmvfc_send_crq
- returns H_CLOSED
<- returns SCSI_MLQUEUE_HOST_BUSY
-> __blk_mq_requeue_request
-> scmd_eh_abort_handler
-> scsi_try_to_abort_cmd
- returns SUCCESS
-> scsi_queue_insert
Normally, the SCMD_STATE_COMPLETE bit would protect against the command
completion and the timeout, but that doesn't work here, since we don't
check that at all in the SCSI_MLQUEUE_HOST_BUSY path.
In this case we end up calling scsi_queue_insert on a request that has
already been queued, or possibly even freed, and we crash.
The patch below simply increases the default I/O timeout to avoid this race
condition. This is also the timeout value that nearly all IBM SAN storage
recommends setting as the default value.
Link: https://lore.kernel.org/r/1610463998-19791-1-git-send-email-brking@linux.vnet.ibm.com
Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
764907293e