On a dual controller setup with multipath enabled, some MEDIUM ERRORs
caused both paths to be failed, thus I/O got queued/blocked since the
'queue_if_no_path' feature is enabled by default on IPR controllers.
This example disabled 'queue_if_no_path' so the I/O failure is seen at
the sg_dd program. Notice that after the sg_dd test-case, both paths
are in 'failed' state, and both path/priority groups are in 'enabled'
state (not 'active') -- which would block I/O with 'queue_if_no_path'.
# sg_dd if=/dev/dm-2 bs=4096 count=1 dio=1 verbose=4 blk_sgio=0
<...>
read(unix): count=4096, res=-1
sg_dd: reading, skip=0 : Input/output error
<...>
# dmesg
[...] sd 2:2:16:0: [sds] FAILED Result: hostbyte=DID_OK driverbyte=DRIVER_SENSE
[...] sd 2:2:16:0: [sds] Sense Key : Medium Error [current]
[...] sd 2:2:16:0: [sds] Add. Sense: Unrecovered read error - recommend rewrite the data
[...] sd 2:2:16:0: [sds] CDB: Read(10) 28 00 00 00 00 00 00 00 20 00
[...] blk_update_request: I/O error, dev sds, sector 0
[...] device-mapper: multipath: Failing path 65:32.
<...>
[...] device-mapper: multipath: Failing path 65:224.
# multipath -l
1IBM_IPR-0_59C2AE0000001F80 dm-2 IBM ,IPR-0
59C2AE00
size=5.2T features='0' hwhandler='1 alua' wp=rw
|-+- policy='service-time 0' prio=0 status=enabled
| `- 2:2:16:0 sds 65:32 failed undef running
`-+- policy='service-time 0' prio=0 status=enabled
`- 1:2:7:0 sdae 65:224 failed undef running
This is not the desired behavior. The dm-multipath explicitly checks
for the MEDIUM ERROR case (and a few others) so not to fail the path
(e.g., I/O to other sectors could potentially happen without problems).
See dm-mpath.c :: do_end_io_bio() -> noretry_error() !->! fail_path().
The problem trace is:
1) ipr_scsi_done() // SENSE KEY/CHECK CONDITION detected, go to..
2) ipr_erp_start() // ipr_is_gscsi() and masked_ioasc OK, go to..
3) ipr_gen_sense() // masked_ioasc is IPR_IOASC_MED_DO_NOT_REALLOC,
// so set DID_PASSTHROUGH.
4) scsi_decide_disposition() // check for DID_PASSTHROUGH and return
// early on, faking a DID_OK.. *instead*
// of reaching scsi_check_sense().
// Had it reached the latter, that would
// set host_byte to DID_MEDIUM_ERROR.
5) scsi_finish_command()
6) scsi_io_completion()
7) __scsi_error_from_host_byte() // That would be converted to -ENODATA
<...>
8) dm_softirq_done()
9) multipath_end_io()
10) do_end_io()
11) noretry_error() // And that is checked in dm-mpath :: noretry_error()
// which would cause fail_path() not to be called.
With this patch applied, the I/O is failed but the paths are not. This
multipath device continues accepting more I/O requests without blocking.
(and notice the different host byte/driver byte handling per SCSI layer).
# dmesg
[...] sd 2:2:7:0: [sdaf] Done: SUCCESS Result: hostbyte=0x13 driverbyte=DRIVER_OK
[...] sd 2:2:7:0: [sdaf] CDB: Read(10) 28 00 00 00 00 00 00 00 40 00
[...] sd 2:2:7:0: [sdaf] Sense Key : Medium Error [current]
[...] sd 2:2:7:0: [sdaf] Add. Sense: Unrecovered read error - recommend rewrite the data
[...] blk_update_request: critical medium error, dev sdaf, sector 0
[...] blk_update_request: critical medium error, dev dm-6, sector 0
[...] sd 2:2:7:0: [sdaf] Done: SUCCESS Result: hostbyte=0x13 driverbyte=DRIVER_OK
[...] sd 2:2:7:0: [sdaf] CDB: Read(10) 28 00 00 00 00 00 00 00 10 00
[...] sd 2:2:7:0: [sdaf] Sense Key : Medium Error [current]
[...] sd 2:2:7:0: [sdaf] Add. Sense: Unrecovered read error - recommend rewrite the data
[...] blk_update_request: critical medium error, dev sdaf, sector 0
[...] blk_update_request: critical medium error, dev dm-6, sector 0
[...] Buffer I/O error on dev dm-6, logical block 0, async page read
# multipath -l 1IBM_IPR-0_59C2AE0000001F80
1IBM_IPR-0_59C2AE0000001F80 dm-6 IBM ,IPR-0
59C2AE00
size=5.2T features='1 queue_if_no_path' hwhandler='1 alua' wp=rw
|-+- policy='service-time 0' prio=0 status=active
| `- 2:2:7:0 sdaf 65:240 active undef running
`-+- policy='service-time 0' prio=0 status=enabled
`- 1:2:7:0 sdh 8:112 active undef running
Signed-off-by: Mauricio Faria de Oliveira <mauricfo@linux.vnet.ibm.com>
Acked-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>