On 2021-04-22 12:52 p.m., Bart Van Assche wrote:
On 4/22/21 8:56 AM, Douglas Gilbert wrote:
In driver manuals Seagate often list the PRE-FETCH command as optional. As
in: pay us some extra money and we will put it in. That suggests to me some
big OEM likes PRE-FETCH. Where I found it supported in WDC manuals, they
didn't support the IMMED bit which sort of defeats the purpose of it IMO.
Since the sd driver does not submit the PRE-FETCH command, how about
moving support for CONDITION MET into the sg code and treating CONDITION
MET as an error inside the sd, sr and st drivers? I think that would
allow to simplify scsi_status_is_good(). The current definition of that
function is as follows:
static inline int scsi_status_is_good(int status)
{
/*
* FIXME: bit0 is listed as reserved in SCSI-2, but is
* significant in SCSI-3. For now, we follow the SCSI-2
* behaviour and ignore reserved bits.
*/
status &= 0xfe;
return ((status == SAM_STAT_GOOD) ||
(status == SAM_STAT_CONDITION_MET) ||
/* Next two "intermediate" statuses are obsolete in*/
/* SAM-4 */
(status == SAM_STAT_INTERMEDIATE) ||
(status == SAM_STAT_INTERMEDIATE_CONDITION_MET) ||
/* FIXME: this is obsolete in SAM-3 */
(status == SAM_STAT_COMMAND_TERMINATED));
}
The whole stack needs to treat SAM_STAT_CONDITION_MET as a non-error.
However the complex multi-layer return values are represented,
reducing them to a comparison with zero, spread all over the
stack just seems bad software engineering. IMO a predicate function
(i.e. returning bool) is needed.
I would argue that in the right circumstances, the sd driver should
indeed by using PRE-FETCH. It would need help from the upper layers.
It is essentially "read-ahead" in the case where the next LBA
does not follow the last read LBA. A smarter read-ahead ...
Doug Gilbert
