On 9/14/23 15:53, Geert Uytterhoeven wrote:
> Hi Damien,
>
> On Thu, Sep 14, 2023 at 12:03 AM Damien Le Moal <dlemoal@xxxxxxxxxx> wrote:
>> On 9/13/23 19:21, Geert Uytterhoeven wrote:
>>>> Thanks for the report. The delay for the first data access from user space right
>>>> after resume is 100% expected, with or without this patch. The reason is that
>>>> waking up the drive (spinning it up) is done asynchronously from the PM resume
>>>> context, so when you get "PM suspend exit" message signaling that the system is
>>>> resumed, the drive may not yet be spinning. Any access will wait for that to
>>>> happen before proceeding. Depending on the drive that can take up to 10s or so.
>>>
>>> That does not match with what I am seeing: before this patch, there
>>> was no delay on first data access from user space, as the drive is fully
>>> spun up when system resume returns.
>>
>> Yes, that is a possibility, but not by design. Some users have complained about
>> the long resume times which causes laptop screens to be "black" until disks spin
>> up. That did not happen before 5.16, when the change to scsi using the PM async
>> ops to do suspend/resume created all the issues with suspend/resume on ata side.
>> I am going to run 5.15 again to check.
>>
>> The patch "do not issue START STOP UNIT on resume" was a botch attempt to remove
>> this delay. But it is a bad one because the ATA specs define that a drive can
>> get out of standby (spun down) power state only with a media access command (a
>> VERIFY command is used to spin up the disk). And furthermore, the specs also
>> says that even a reset shall not change the device power state. So issuing a
>> VERIFY command to spin up the drive is required per specs. Note that I do see
>> many of my drives (I have hundreds in the lab) spinning up on reset, which is
>> against the specs. But not all of them. So with the patch "do not issue START
>> STOP UNIT on resume", one risks not seeing the drive resuming correctly (timeout
>> errors on IDENTIFY command issued on resume will get the drive removed).
>>
>>> With this patch, system resume returns earlier, and the drive is only
>>> spun up when user space starts accessing data.
>>
>> Yes, but "when user space starts accessing data" -> "user space accesses are
>> processed only after the drive completes spinning up" would be more accurate.
>
> Sure, I wrote it down in terms of what the user is experiencing, which may
> not be identical to what's happening under the hood.
>
>> That is by design and expected. This is the behavior one would see if the drive
>> is set to use standby timers (to go to standby on its own after some idle time)
>> or if runtime suspend is used. I do not see any issue with this behavior. Is
>> this causing any issue on your end ? Do you have concerns about this approach ?
>>
>> Having the resume process wait for the drive to fully spin-up is easy to do. But
>> as I mentioned, many users are really not happy about how slow resuming become
>> with that. If I do that, you will get back the previous behavior you mention,
>> but I will be again getting emails about "resume is broken".
>>
>> I made a decision: no waiting for spinup. That causes a delay for the user on
>> first access, but that makes resume overall far faster. I do not want to go back
>> on that, unless you can confirm that there is a real regression/error/data
>> corruption happening.
>
> I agree that is fine.
Still, R-CAR is not working for you with the patch adding the link. Looking into
it now. Trying to verify things with qemu & regular AHCI on PCs too as I do not
have an R-CAR board. Hard to debug :)
Preparing some patches for you to get more debug info.
>
> Gr{oetje,eeting}s,
>
> Geert
>
--
Damien Le Moal
Western Digital Research
