Re: [git patches] libata updates for 3.3

[Jeff Garzik <jeff@garzik.org>]

On 01/14/2012 12:21 AM, Linus Torvalds wrote:
> On Sun, Jan 8, 2012 at 4:32 PM, Jeff Garzik<jeff@garzik.org>  wrote:
>>
>> Summary (very little excitement at all this time):
>>
>> 0) Will play around with git signed tags with the next update.
>>
>> 1) PM improvements, including runtime suspend/resume work
>
> Hmm. I don't know if this comes from the PM improvements or even this
> particular pull, but links that aren't connected are *really* slow.
>
> Annoyingly so.
>
> My Macbook Air that I finally can resume reliably again used to come
> back almost immediately from resume. No longer. And the reason seems
> to be this:
>
>   [  243.306149] ata_piix 0000:00:1f.2: setting latency timer to 64
>   [  243.306180] bcma: Found rev 6 PMU (capabilities 0x108C2606)
>   [  246.579648] ata1.01: failed to resume link (SControl 0)
>   [  246.735472] ata1.00: SATA link up 3.0 Gbps (SStatus 123 SControl 300)
>   [  246.735485] ata1.01: SATA link down (SStatus 0 SControl 0)
>   [  246.743632] ata1.00: ACPI cmd ef/03:46:00:00:00:a0 (SET FEATURES)
> filtered out
>   [  246.744353] ata1.00: configured for UDMA/100
>   [  246.744537] sd 0:0:0:0: [sda] Starting disk
>   [  247.769806] ata2.00: failed to resume link (SControl 0)
>   [  248.796207] ata2.01: failed to resume link (SControl 0)
>   [  248.807665] ata2.00: SATA link down (SStatus 4 SControl 0)
>   [  248.807681] ata2.01: SATA link down (SStatus 0 SControl 0)
>   [  248.808338] PM: resume of devices complete after 5511.027 msecs
>   [  248.882074] PM: Finishing wakeup.
>
> Notice the basically five-second timeout all basically for "failed to
> resume link: for things that didn't have anything connected to them
> anyway.
>
> This is a bog-standard Intel controller, there's nothing odd there.
>
> I'm pretty sure this used to be much faster, but I haven't bisected
> any of it (and with all the problems I had with resume both due to
> wireless and MCE, I really wouldn't want to even try).
>
> Taking 5.5 seconds to come back from suspend-to-ram really is too
> long. Not *all* of it is the SATA part, but a lot of it is.
>
> For ATA suspend/resume, could we perhaps only resume the ports that
> *used* to have something on them? And then, if somebody has plugged
> something into the others, not consider that a resume thing at all,
> but a hotplug thing that happens *after* the resume?
>
> If it takes five seconds to notice new hardware after a resume, nobody
> cares. But the disk we had before obviously needs to get resumed.. But
> it does seem like it's the "no link" part that takes long.

We definitely notice new hardware after a resume, but you're right -- it 
should not take that long to work through ports that are empty.

Will take a look tomorrow (kid->doctor+relatives today, uff) at the most 
recent PM push; my quick testing did not show any problems, but 
suspend/resume varies widely across hardware platforms.  I think I might 
even have a MacBook I can test.  Apple platforms test to be weird too...  ;)

	Jeff