On 14/07/2017 12:46, Gionatan Danti wrote:> Hi, so a
premature/preventive drive detachment is not a silver bullet,
and I buy it. However, I would at least expect this behavior to be
configurable. Maybe it is, and I am missing something?
Anyway, what really surprise me is *not* the drive to not be detached,
rather permitting that corruption make its way into real data. I naively
expect that when a WRITE_QUEUED or CACHE_FLUSH command aborts/fails
(which *will* cause data corruption if not properly handled) the I/O
layer has the following possibilities:
a) retry the write/flush. You don't want to retry indefinitely, so the
kernel need some type of counter/threshold; when the counter is reached,
continue with b). This would mask out sporadic errors, while propagating
recurring ones;
b) notify the upper layer that a write error happened. For synchronized
and direct writes it can notify that by simply returning the correct
exit code to the calling function. In this case, the block layer should
return an error to the MD driver, which must act accordlying: for
example, dropping the disk from the array.
c) do nothing. This seems to me by far the worst choice.
If b) is correcly implemented, it should prevent corruption to
accumulate on the drives.
Please also note the *type* of corrupted data: not only user data, but
filesystem journal and metadata also. The latter should be protected by
the using of write barriers / FUAs, so they should be able to stop
themselves *before* corruption.
So I have some very important questions:
- how does MD behave when flushing data to disk?
- does it propagate write barriers?
- when a write barrier fails, is the error propagated to the upper layers?
Thanks you all.
Hi all,
having some free time, I conducted some new tests and I am now able to
100% replicate the problem. To recap: a filesystem on a RAID1 array was
corrupted due to SATA WRITEs failing but *no* I/O error being reported
to higher layer (ie: mdraid/mdadm). I already submitted my findings on
the linux-scsi mailing list, but I want to share them here because they
can be useful to others.
On the affected machine, /var/log/messages shown some "failed command:
WRITE FPDMA QUEUED" entries, but *no* action (ie: kick off disk) was
taken by MDRAID. I tracked down the problem to an instable power supply
(switching power rail/connector solved the problem).
In the latest day I had some spare time and I am now able to regularly
replicate the problem. Basically, when a short powerloss happens, the
scsi midlayer logs some failed operations, but does *not* pass these
errors to higher layer. In other words, no I/O error is returned to the
calling application. This is the reason why MDRAID did not kick off the
instable disk on the machine with corrupted filesystem.
To replicated the problem, I wrote a large random file on a small MD
RAID1 array, pulling off the power of one disk from about 2 seconds. The
file write operation stopped for some seconds, than recovered. Running
an array check resulted in a high number of mismatch_cnt sectors. Dmesg
logged the following lines:
Aug 16 16:04:02 blackhole kernel: ata6.00: exception Emask 0x50 SAct
0x7fffffff SErr 0x90a00 action 0xe frozen
Aug 16 16:04:02 blackhole kernel: ata6.00: irq_stat 0x00400000, PHY RDY
changed
Aug 16 16:04:02 blackhole kernel: ata6: SError: { Persist HostInt
PHYRdyChg 10B8B }
Aug 16 16:04:02 blackhole kernel: ata6.00: failed command: WRITE FPDMA
QUEUED
Aug 16 16:04:02 blackhole kernel: ata6.00: cmd
61/00:00:10:82:09/04:00:00:00:00/40 tag 0 ncq 524288 out#012 res
40/00:d8:10:72:09/00:00:00:00:00/40 Emask 0x50 (ATA bus error)
Aug 16 16:04:02 blackhole kernel: ata6.00: status: { DRDY }
...
Aug 16 16:04:02 blackhole kernel: ata6.00: failed command: WRITE FPDMA
QUEUED
Aug 16 16:04:02 blackhole kernel: ata6.00: cmd
61/00:f0:10:7e:09/04:00:00:00:00/40 tag 30 ncq 524288 out#012
res 40/00:d8:10:72:09/00:00:00:00:00/40 Emask 0x50 (ATA bus error)
Aug 16 16:04:02 blackhole kernel: ata6.00: status: { DRDY }
Aug 16 16:04:02 blackhole kernel: ata6: hard resetting link
Aug 16 16:04:03 blackhole kernel: ata6: SATA link down (SStatus 0
SControl 310)
Aug 16 16:04:04 blackhole kernel: ata6: hard resetting link
Aug 16 16:04:14 blackhole kernel: ata6: softreset failed (device not ready)
Aug 16 16:04:14 blackhole kernel: ata6: hard resetting link
Aug 16 16:04:24 blackhole kernel: ata6: softreset failed (device not ready)
Aug 16 16:04:24 blackhole kernel: ata6: hard resetting link
Aug 16 16:04:35 blackhole kernel: ata6: link is slow to respond, please
be patient (ready=0)
Aug 16 16:04:42 blackhole kernel: ata6: SATA link down (SStatus 0
SControl 310)
Aug 16 16:04:46 blackhole kernel: ata6: hard resetting link
Aug 16 16:04:46 blackhole kernel: ata3: exception Emask 0x10 SAct 0x0
SErr 0x40d0202 action 0xe frozen
Aug 16 16:04:46 blackhole kernel: ata3: irq_stat 0x00400000, PHY RDY changed
Aug 16 16:04:46 blackhole kernel: ata3: SError: { RecovComm Persist
PHYRdyChg CommWake 10B8B DevExch }
Aug 16 16:04:46 blackhole kernel: ata3: hard resetting link
Aug 16 16:04:51 blackhole kernel: ata3: softreset failed (device not ready)
Aug 16 16:04:51 blackhole kernel: ata3: applying PMP SRST workaround and
retrying
Aug 16 16:04:51 blackhole kernel: ata3: SATA link up 3.0 Gbps (SStatus
123 SControl 300)
Aug 16 16:04:51 blackhole kernel: ata3.00: configured for UDMA/133
Aug 16 16:04:51 blackhole kernel: ata3: EH complete
Aug 16 16:04:52 blackhole kernel: ata6: softreset failed (device not ready)
Aug 16 16:04:52 blackhole kernel: ata6: applying PMP SRST workaround and
retrying
Aug 16 16:04:52 blackhole kernel: ata6: SATA link up 1.5 Gbps (SStatus
113 SControl 310)
Aug 16 16:04:52 blackhole kernel: ata6.00: configured for UDMA/133
Aug 16 16:04:52 blackhole kernel: ata6: EH complete
As you can see, while failed SATA operation were logged in dmesg (and
/var/log/messages), no I/O errors where returned to the upper layer
(MDRAID) or the calling application. I had to say that I *fully expect*
some inconsistencies: after all, removing the power wipes the volatile
disk's DRAM cache, which means data loss. However, I really expected
some I/O errors to be thrown to the higher layers, causing visible
reactions (ie: a disks pushed out the array). With no I/O errors
returned, the higher layer application are effectively blind.
Moreover, both disks show the *same RAID event number*, so the MDRAID
layer can not automatically offline/don't read the corrupted disk. Here
is the relevant output:
[root@blackhole storage]# mdadm -E /dev/sd[bc]1
/dev/sdb1:
Magic : a92b4efc
Version : 1.2
Feature Map : 0x0
Array UUID : 8d06acd3:95920a78:069a7fc3:a526ca8a
Name : blackhole.assyoma.it:200 (local to host
blackhole.assyoma.it)
Creation Time : Wed Aug 16 15:11:14 2017
Raid Level : raid1
Raid Devices : 2
Avail Dev Size : 2095104 (1023.00 MiB 1072.69 MB)
Array Size : 1047552 (1023.00 MiB 1072.69 MB)
Data Offset : 2048 sectors
Super Offset : 8 sectors
Unused Space : before=1960 sectors, after=0 sectors
State : clean
Device UUID : 97bfbe06:89016508:2cb250c9:937a5c2e
Update Time : Thu Aug 17 10:09:28 2017
Bad Block Log : 512 entries available at offset 72 sectors
Checksum : 52670329 - correct
Events : 759
Device Role : Active device 0
Array State : AA ('A' == active, '.' == missing, 'R' == replacing)
/dev/sdc1:
Magic : a92b4efc
Version : 1.2
Feature Map : 0x0
Array UUID : 8d06acd3:95920a78:069a7fc3:a526ca8a
Name : blackhole.assyoma.it:200 (local to host
blackhole.assyoma.it)
Creation Time : Wed Aug 16 15:11:14 2017
Raid Level : raid1
Raid Devices : 2
Avail Dev Size : 2095104 (1023.00 MiB 1072.69 MB)
Array Size : 1047552 (1023.00 MiB 1072.69 MB)
Data Offset : 2048 sectors
Super Offset : 8 sectors
Unused Space : before=1960 sectors, after=0 sectors
State : clean
Device UUID : bf660182:701430fd:55f5fde9:6ded709e
Update Time : Thu Aug 17 10:09:28 2017
Bad Block Log : 512 entries available at offset 72 sectors
Checksum : 5336733f - correct
Events : 759
Device Role : Active device 1
Array State : AA ('A' == active, '.' == missing, 'R' == replacing)
More concerning is the fact that these undetected errors can make their
way even when the higher application consistently calls sync() and/or
fsync. In other words, it seems than even acknowledged writes can fail
in this manner (and this is consistent with the first machine corrupting
its filesystem due to journal trashing - XFS journal surely uses sync()
where appropriate). The mechanism seems the following:
- an higher layer application issue sync();
- a write barrier is generated;
- a first FLUSH CACHE command is sent to the disk;
- data are written to the disk's DRAM cache;
- power is lost! The volatile cache lose its content;
- power is re-established and the disk become responsive again;
- a second FLUSH CACHE command is sent to the disk;
- the disk acks each SATA command, but real data are lost.
As a side note, when the power loss or SATA cable disconnection is
relatively long (over 10 seconds, as by eh timeout), the SATA disks
become disconnected (and the MD layer acts accordlying):
Aug 16 16:12:20 blackhole kernel: ata6.00: exception Emask 0x50 SAct
0x7fffffff SErr 0x490a00 action 0xe frozen
Aug 16 16:12:20 blackhole kernel: ata6.00: irq_stat 0x08000000,
interface fatal error
Aug 16 16:12:20 blackhole kernel: ata6: SError: { Persist HostInt
PHYRdyChg 10B8B Handshk }
Aug 16 16:12:20 blackhole kernel: ata6.00: failed command: WRITE FPDMA
QUEUED
Aug 16 16:12:20 blackhole kernel: ata6.00: cmd
61/00:00:38:88:09/04:00:00:00:00/40 tag 0 ncq 524288 out#012 res
40/00:d8:38:f4:09/00:00:00:00:00/40 Emask 0x50 (ATA bus error)
Aug 16 16:12:20 blackhole kernel: ata6.00: status: { DRDY }
...
Aug 16 16:12:32 blackhole kernel: sd 5:0:0:0: [sdf] FAILED Result:
hostbyte=DID_OK driverbyte=DRIVER_SENSE
Aug 16 16:12:32 blackhole kernel: sd 5:0:0:0: [sdf] Sense Key : Illegal
Request [current] [descriptor]
Aug 16 16:12:32 blackhole kernel: sd 5:0:0:0: [sdf] Add. Sense:
Unaligned write command
Aug 16 16:12:32 blackhole kernel: sd 5:0:0:0: [sdf] CDB: Write(10) 2a 00
00 09 88 38 00 04 00 00
Aug 16 16:12:32 blackhole kernel: blk_update_request: 23 callbacks
suppressed
Aug 16 16:12:32 blackhole kernel: blk_update_request: I/O error, dev
sdf, sector 624696
Now, I have few questions:
- is the above explanation plausible, or I am (horribly) missing something?
- why the scsi midlevel does not respond to a power loss event by
immediately offlining the disks?
- is the scsi midlevel behavior configurable (I know I can lower eh
timeout, but is this the right solution)?
- how to deal with this problem (other than being 100% sure power is
never lost by any disks)?
Thank you all,
regards.
--
Danti Gionatan
Supporto Tecnico
Assyoma S.r.l. - www.assyoma.it
email: g.danti@xxxxxxxxxx - info@xxxxxxxxxx
GPG public key ID: FF5F32A8
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
