diff options
author | Mark Salyzyn <mark_salyzyn@xyratex.com> | 2012-01-17 17:52:24 +0100 |
---|---|---|
committer | James Bottomley <JBottomley@Parallels.com> | 2012-02-19 15:08:52 +0100 |
commit | 5954d7380f627371c4d8d7c59c08f9596aa2c674 (patch) | |
tree | 0908dd4a832ba0ef07c79237f9f6ab3f2ad20a36 /drivers/scsi/pm8001/pm8001_sas.h | |
parent | [SCSI] pm8001: Add FUNC_GET_EVENTS (diff) | |
download | linux-5954d7380f627371c4d8d7c59c08f9596aa2c674.tar.xz linux-5954d7380f627371c4d8d7c59c08f9596aa2c674.zip |
[SCSI] pm8001: deficient responses to IO_XFER_ERROR_BREAK and IO_XFER_OPEN_RETRY_TIMEOUT
IO_XFER_ERROR_BREAK and IO_XFER_OPEN_RETRY_TIMEOUT are deficient of the
required actions as outlined in the programming manual for the pm8001. Due to
the overlapping code requirements of these recovery responses, we found it
necessary to bundle them together into one patch.
When a break is received during the command phase (ssp_completion), this is a
result of a timeout or interruption on the bus. Logic suggests that we should
retry the command.
When a break is received during the data-phase (ssp_event), the task must be
aborted on the target or it will retain a data-phase lock turning the target
reticent to all future media commands yet will successfully respond to TUR,
INQUIRY and ABORT leading eventually to target failure through several
abort-cycle loops.
The open retry interval is exceedingly short resulting in occasional target
drop-off during expander resets or when targets push-back during bad-block
remapping. Increased effective timeout from 130ms to 1.5 seconds for each try
so as to trigger after the administrative inquiry/tur timeout in the scsi
subsystem to keep error-recovery harmonics to a minimum.
When an open retry timeout event is received, the action required by the
targets is to issue an abort for the outstanding command then logic suggests
we retry the command as this state is usually an indication of a credit block
or busy condition on the target.
We hijacked the pm8001_handle_event work queue handler so that it will handle
task as an argument instead of device for the workers in support of the
deferred handling outlined above.
Moderate to Heavy bad-path testing on a 2.6.32 vintage kernel, compile-testing
on scsi-misc-2.6 kernel ...
Signed-off-by: Mark Salyzyn <mark_salyzyn@xyratex.com>
Acked-by: Jack Wang <jack_wang@usish.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Diffstat (limited to 'drivers/scsi/pm8001/pm8001_sas.h')
-rw-r--r-- | drivers/scsi/pm8001/pm8001_sas.h | 5 |
1 files changed, 5 insertions, 0 deletions
diff --git a/drivers/scsi/pm8001/pm8001_sas.h b/drivers/scsi/pm8001/pm8001_sas.h index 83a48f3247b8..11008205aeb3 100644 --- a/drivers/scsi/pm8001/pm8001_sas.h +++ b/drivers/scsi/pm8001/pm8001_sas.h @@ -235,6 +235,7 @@ struct pm8001_ccb_info { struct pm8001_device *device; struct pm8001_prd buf_prd[PM8001_MAX_DMA_SG]; struct fw_control_ex *fw_control_context; + u8 open_retry; }; struct mpi_mem { @@ -484,6 +485,10 @@ void pm8001_dev_gone(struct domain_device *dev); int pm8001_lu_reset(struct domain_device *dev, u8 *lun); int pm8001_I_T_nexus_reset(struct domain_device *dev); int pm8001_query_task(struct sas_task *task); +void pm8001_open_reject_retry( + struct pm8001_hba_info *pm8001_ha, + struct sas_task *task_to_close, + struct pm8001_device *device_to_close); int pm8001_mem_alloc(struct pci_dev *pdev, void **virt_addr, dma_addr_t *pphys_addr, u32 *pphys_addr_hi, u32 *pphys_addr_lo, u32 mem_size, u32 align); |