Re: Disk spin down issue on shut down/suspend to disk

[Mark Lord <liml@rtr.ca>]

Mark Lord wrote:
> Tejun Heo wrote:
>> Mark Lord wrote:
..
>>>FWIW, Tejun, with 2.6.22, my new Seagate 160GB SATA drive (notebook)
>>>increments the "Power-Off_Retract_Count" on each suspend-to-RAM operation.
>>>It does not do any double spin-up/spin-down things though.
>>
>> Hmmm.. It shouldn't.  libata now issues STANDBYNOW prior to entering
>> STR.  Can you instrument code a bit and see whether it actually gets 
>> issued?
> 
> Heh.. I haven't instrumented it yet, but I did discover a bit more about it:
> 
> The Power-Off_Retract_Count incrmenents *only* when there's data in the
> on-drive write-cache.  So if I haven't written anything significantly large
> before suspending, then it often does NOT increment the retract counter.
> 
> But if I copy a couple of multi-MB files around and then suspend (to RAM),
> the retract count gets incremented.
> 
> So I've now just stuck "hdparm -F /dev/sda" into my suspend script,
> and that cures the problem completely for me.  "-F" does a FLUSH_CACHE,
> and requires a recent copy of hdparm.
> 
> Perhaps libata should also do a FLUSH_CACHE before any STANDBYNOW command,
> prior to entering STR, which is what my script is currently now doing..
> 
> I'll instrument libata and see what the current sequence is.

Okay, instrumented it and all of that now.

Yes, sd + libata-scsi send the proper FLUSH_CACHE_EXT and STANDBY1 commands
immediately prior to the suspend-to-RAM, and there are no other commands
sneaking through at the same time.  Good.

But the retract count still increments if I've recently written a signficant
amount of data.

My suspend script now has this little chunk of code at the point
where it actually does the suspend-to-RAM:

    sync; sync
    hdparm -F /dev/sda   ## flush drive write cache
    sleep 1              ## allow time for the flush to complete
    echo mem > /sys/power/state  ## suspend-to-RAM
 
Without the "sleep 1", it doesn't always eliminate the extra Retract,
so I hypothesize that the FLUSH_CACHE_EXT command is implemented in
an asynchronous fashion by the drive:  it returns immediately before
it has actually completed writing cached data to disk.  The "sleep 1"
seems to give it enough time to finish up, at least for me.

Cheers