Re: [PATCH v2 3/3] ata: ahci: Skip 200 ms debounce delay for AMD 300 Series Chipset SATA Controller

[Damien Le Moal <damien.lemoal@opensource.wdc.com>]

On 9/14/22 00:28, Limonciello, Mario wrote:
> [Public]
> 
> 
> 
>> -----Original Message-----
>> From: Paul Menzel <pmenzel@molgen.mpg.de>
>> Sent: Tuesday, September 13, 2022 10:23
>> To: Damien Le Moal <damien.lemoal@opensource.wdc.com>
>> Cc: Limonciello, Mario <Mario.Limonciello@amd.com>; Hans de Goede
>> <hdegoede@redhat.com>; linux-ide@vger.kernel.org; LKML <linux-
>> kernel@vger.kernel.org>
>> Subject: Re: [PATCH v2 3/3] ata: ahci: Skip 200 ms debounce delay for AMD 300
>> Series Chipset SATA Controller
>>
>> Dear Damien,
>>
>>
>> Am 01.09.22 um 00:13 schrieb Damien Le Moal:
>>> On 8/30/22 18:05, Paul Menzel wrote:
>>
>> […]
>>
>>>> Am 01.06.22 um 10:58 schrieb Damien Le Moal:
>>>>> On 6/1/22 01:18, Paul Menzel wrote:
>>>>>>>>> With that in mind, I am not planning to apply your previous patches
>>>>>>>>> for 5.18, as they would conflict and would only end up being churn
>>>>>>>>> since the delay removal by default will undo your changes.
>>>>>>>> Obviously, I do not agree, as this would give the a little bit more
>>>>>>>> testing already, if changing the default is a good idea. Also, if the
>>>>>>>> conflict will be hard to resolve, I happily do it (the patches could
>>>>>>>> even be reverted on top – git commits are cheap and easy to handle).
>>>>>>>
>>>>>>> The conflict is not hard to resolve. The point is that my patches changing
>>>>>>> the default to no debounce delay completely remove the changes of your
>>>>>>> patch to do the same for one or some adapters. So adding your patches
>> now
>>>>>>> and then my patches on top does not make much sense at all.
>>>>>>>
>>>>>>> If too many problems show up and I end up reverting/removing the
>> patches,
>>>>>>> then I will be happy to take your patches for the adapter you tested. Note
>>>>>>> that *all* the machines I have tested so far are OK without a debounce
>>>>>>> delay too. So we could add them too... And endup with a long list of
>>>>>>> adapters that use the default ahci driver without debounce delay. The
>> goal
>>>>>>> of changing the default to no delay is to avoid that. So far, the adapters
>>>>>>> I have identified that need the delay have their own declaration, so we
>>>>>>> only need to add a flag there. Simpler change that listing up adapters
>>>>>>> that are OK without the delay.
>>>>>>>
>>>>>>>> Anyway, I wrote my piece, but you are the maintainer, so it’s your call
>>>>>>>> and I stop bothering you.
>>>>>>
>>>>>> I just wanted to inquire about the status of your changes? I do not find
>>>>>> them in your `for-5.19` branch. As they should be tested in linux-next
>>>>>> before the merge window opens, if these are not ready yet, could you
>>>>>> please apply my (tested) patches?
>>>>>
>>>>> I could, but 5.19 now has an updated libata.force kernel parameter that
>>>>> allows one to disable the debounce delay for a particular port or for all
>>>>> ports of an adapter. See libata.force=x.y:nodbdelay for a port y of
>>>>> adapter x or libata.force=x:nodbdelay for all ports of adapter x.
>>>>
>>>> This is commit 3af9ca4d341d (ata: libata-core: Improve link flags forced
>>>> settings) [1]. Thank you, this is really useful, but easily overlooked. ;-)
>>>>
>>>>> I still plan to revisit the arbitrary link debounce timers but I prefer to
>>>>> have the power management cleanup applied first. The reason is that link
>>>>> debounce depends on PHY readiness, which itself depends heavily on power
>>>>> mode transitions. My plan is to get this done during this cycle for
>>>>> release with 5.20 and then fix on top the arbitrary delays for 5.21.
>>>>
>>>> Nice. Can you share the current status?
>>>
>>> No progress. I need to put together a series with all the patches that
>>> were sent already. Unless Mario can resend something ?
>>
>> No reply from Mario.
> 
> I think what happened here is there was related patches from another party
> that got tangled up with this.

Niklas and I are investigating this again now because Niklas discovered 
that one AMD AHCI adapter leads to drive resets when the drive goes to 
low power mode. The adapter is:

  SATA controller [0106]: Advanced Micro Devices, Inc. [AMD] FCH SATA 
Controller [AHCI mode] [1022:7901] (rev 51)

If we switch to performance mode (no LPM), the reset disapears. But if 
LPM is enabled, any command sent after the disk goes to low poer mode 
(device initiated), there is a link reset...


-- 
Damien Le Moal
Western Digital Research