Bug 14616

The previous code only worked becaues in most cases BIOS brought up the link already which isn't always the case.  Actually trying to bring the link up is what's taking time there, so I'm afraid we can't remove it because it adds 1 second delay.  If you have special configuration where 1 second delay isn't acceptable and you're sure link resetting is not necessary, please give a shot at libata.force=norst parameter.

Resolving as WILL_NOT_FIX for now.  Please feel free to add further comment or reopen.

Thanks.

Is it possible to check out if SATA link is already up, without trying to bring it up no matter what?

Because speaking frankly I hate adding any kernel boot strings as I see them as dirty hacks which in most cases mean the internal kernel logic doesn't work well or doesn't deal with the real world scenarios.

Besides, the original bug deals with resuming after any kind of suspend, that means we don't have to introduce this workaround for the initial boot sequence.

The standard behavior is to bring up link and reset unconditionally.  It was an obvious bug fix and taking a couple of secs there is acceptable behavior which will suit most users and safely cover most corner cases.  If for whatever reason you can't live with one sec delay there, your requirement is quite different from others, so the suggestion of boot parameter.  It's a normal latency necessary to initialize the hardware and I don't think tinkering with it to reduce 1 sec delay is a good idea.

Thanks.

Tejun,

There are no bug reports related to finding/initializing storage media _upon boot_, so why there's a need to introduce this "safe" behaviour even though everything worked before?

If you think this fix is a _MUST_ then feel free to close this bug as INVALID and I'll never bother writing anything here again.

Because that's how all other SATA drivers behave or the BIOS might choose not to bring up certain links if it's not occupied during BIOS detection and we will be missing hotplug events.  Also, there are cases where BIOS might not be the one right before this boot.  Resume is one case, kexec boot is another.  There are many corner cases so the baseline is to fully initialize the hardware to the best of the driver's knowledge.  It's basically how the whole thing is designed and supposed to behave.

As you can see from the other bug report, when we deviate from the baseline, corner cases start breaking.  Please note that when things break during resume, emergency kexec boot or whatnot, it's usually quite difficult to reproduce, painful to report and diagnose, so risking subtle breakages for 1sec delay reduction doesn't make whole lot of sense.

So, resolving as INVALID.

Hmmm.... one thing we can try is to avoid writing to SControl if the link already seems online but it will be a very risky change which might end up with long and thin regression distribution graph.  With parallel probing, I seriously don't think several secs for port probing matters at all tho.

Thanks.

libata.force=norst hack didn't help much.

Boot time was reduced by 0.3 seconds, which is still 0.7 seconds more/longer than without this "safe" patch.

Hey, that's 30% reduction.  :-)

Joke aside, I wonder what the difference is tho.  Ah, right, with hardreset ruled out, prereset thinks it needs to resume the link so link resuming will still happen.  Hmmm... Adding a parameter to inhibit unconditional link resuming isn't difficult at all but I'm still not convinced this matters at all.  Why does a 1 sec parallel delay matter so much?

Thanks.

Created attachment 23818 [details]
dmesg for 2.6.32-rc7 (not so good even with libata.force=norst)

Something booted in almost no time, i.e. perfectly, no it's not that perfect.

Do you see my point? :)

So, how can I get back "my" 0.7 seconds?

Frankly, I don't care.  Changing the default behvaior is too risky.  We can introduce a parameter to skip link resume if link already seems online but how many people are gonna use that?  If you need the .7s that much, just patch and roll your own kernel.

Thanks.