freebsd stable - Oct 2012 - mpt doesn't propagate read errors and dies on a single sector?

Hi,

I have a Sun X4540 with LSI C1068E based SAS controllers (FW version: 
1.27.02.00-IT).
My problem is if one drive starts to fail with read errors, the machine 
becomes completely unusable (running stable/9 with ZFS), because -it 
seems- ZFS can't see that there are read errors on a device, the mpt 
driver (controller, kernel?) wants to re-issue the operation endlessly.

Here is a verbose (dev.mpt.0.debug=7 level) dump:
mpt0: Address Reply:
SCSI IO Request Reply @ 0xffffff87ffcfdc00
         IOC Status    Success
         IOCLogInfo    0x00000000
         MsgLength     0x09
         MsgFlags      0x00
         MsgContext    0x000200eb
         Bus:          0
         TargetID      3
         CDBLength     10
         SCSI Status:  Check Condition
         SCSI State:   (0x00000001)AutoSense_Valid
         TransferCnt   0x20000
         SenseCnt      0x0012
         ResponseInfo  0x00000000
(da3:mpt0:0:3:0): READ(10). CDB: 28 0 3a 38 5d e 0 1 0 0
(da3:mpt0:0:3:0): CAM status: SCSI Status Error
(da3:mpt0:0:3:0): SCSI status: Check Condition
(da3:mpt0:0:3:0): SCSI sense: MEDIUM ERROR asc:11,0 (Unrecovered read error)
(da3:mpt0:0:3:0): Info: 0x3a385d1a
(da3:mpt0:0:3:0): Error 5, Unretryable error
SCSI IO Request @ 0xffffff80003046f0
         Chain Offset  0x00
         MsgFlags      0x00
         MsgContext    0x000200ea
         Bus:                0
         TargetID            3
         SenseBufferLength   32
         LUN:              0x0
         Control           0x02000000  READ  SIMPLEQ
         DataLength      0x00020000
         SenseBufAddr    0x0c65d5e0
         CDB[0:10]       28 00 3a 38 5e 0e 00 01 00 00
         SE64 0xffffff87ffd1c430: Addr=0x000000010e858000 
FlagsLength=0xd3020000
          64_BIT_ADDRESSING LAST_ELEMENT END_OF_BUFFER END_OF_LIST
mpt0: Address Reply:
SCSI IO Request Reply @ 0xffffff87ffcfdd00
         IOC Status    Success
         IOCLogInfo    0x00000000
         MsgLength     0x09
         MsgFlags      0x00
         MsgContext    0x000200ea
         Bus:          0
         TargetID      3
         CDBLength     10
         SCSI Status:  Check Condition
         SCSI State:   (0x00000001)AutoSense_Valid
         TransferCnt   0x20000
         SenseCnt      0x0012
         ResponseInfo  0x00000000

And I get these check condition SCSI errors endlessly. If ZFS is enabled 
at boot, the machine can't even start because of this (zpool import 
never finishes), if I boot without ZFS, and try to import, the zpool 
command stucks in the vdev_g state:
  1163 root          1  20    0 35440K  5200K vdev_g  6   0:01 0.10% zpool
procstat -k 1163
   PID    TID COMM             TDNAME KSTACK
  1163 100116 zpool            -                mi_switch 
sleepq_timedwait _sleep biowait vdev_geom_read_guid vdev_geom_open 
vdev_open vdev_open_children vdev_raidz_open vdev_open 
vdev_open_children vdev_root_open vdev_open spa_load spa_tryimport 
zfs_ioc_pool_tryimport zfsdev_ioctl devfs_ioctl_f

Could it be that GEOM/ZFS doesn't receive this read error and waits 
indefinitely for the command to complete?