Re: [PATCH] mmc: core: add auto bkops support

[Ulf Hansson]

On 22 June 2016 at 16:20, Alex Lemberg <Alex.Lemberg@xxxxxxxxxxx> wrote:
>
>
> On 6/22/16, 1:21 PM, "Ulf Hansson" <ulf.hansson@xxxxxxxxxx> wrote:
>
>>On 13 June 2016 at 14:25, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>>> On 13/06/16 11:58, Shawn Lin wrote:
>>>> å 2016/6/13 16:17, Adrian Hunter åé:
>>>>> On 13/06/16 10:48, Shawn Lin wrote:
>>>>>> On 2016/6/13 14:29, Adrian Hunter wrote:
>>>>>>> On 06/06/16 06:07, Shawn Lin wrote:
>>>>>>>> JEDEC eMMC v5.1 introduce an autonomously initiated method
>>>>>>>> for background operations.
>>>>>>>>
>>>>>>>> Host that wants to enable the device to perform background
>>>>>>>> operations during device idle time, should signal the device
>>>>>>>> by setting AUTO_EN in BKOPS_EN field EXT_CSD[163] to 1b. When
>>>>>>>> this bit is set, the device may start or stop background operations
>>>>>>>> whenever it sees fit, without any notification to the host.
>>>>>>>>
>>>>>>>> When AUTO_EN bit is set, the host should keep the device power
>>>>>>>> active. The host may set or clear this bit at any time based on
>>>>>>>> its power constraints or other considerations.
>>>>>>>>
>>>>>>>> Currently the manual bkops is only be used under the async req
>>>>>>>> circumstances and it's a bit complicated to be controlled as the
>>>>>>>> perfect method is that we should do some idle monitor just as rpm
>>>>>>>> and send HPI each time if receiving rd/wr req. But it will impact
>>>>>>>> performance significantly, especially for random iops since the
>>>>>>>> weight of executing HPI against r/w small piece of LBAs is
>>>>>>>> nonnegligible.
>>>>>>>>
>>>>>>>> So we now prefer to select the auto one unconditionally if supported
>>>>>>>> which makes it as simple as possible. It should really good enough
>>>>>>>> for devices to manage its internal policy for bkops rather than the
>>>>>>>> host, which makes us believe that we could achieve the best
>>>>>>>> performance for all the devices implementing auto bkops and the only
>>>>>>>> thing we should do is to disable it when cutting off the power.
>>>>>>>
>>>>>>> Do you know if there is really a requirement to do that?
>>>>>>
>>>>>> Even without bkops enable, no matter for manual or auto one, FTL should
>>>>>> always do bkops like GC internally when needed to guarantee the
>>>>>> performance and balance the wear leveling. What I thought to do is to
>>>>>> make it more explicitly.
>>>>>>
>>>>>> Because then, what
>>>>>>> is the point of power off notification?
>>>>>>
>>>>>> When power off notification is sent, bkops will be stopped
>>>>>> in _mmc_suspend. So I don't undertand your point here?
>>>>>
>>>>> I am trying to understand why we need to do anything for auto bkops.
>>>>> Since AUTO_EN is persistent, we can leave the decision whether to turn it on
>>>>> to whomever provisions the device. Then we just leave it alone.
>>>>>
>>>>
>>>> Hrm..
>>>>
>>>> one possible way is to control it by mmc-utils on
>>>> user space?  So we should add a cmd for mmc-utils
>>>> there?
>>>
>>> That would be consistent with manual bkops.
>>>
>>
> >From my first impression I agree, as that is the policy we have been
>>sticking to when writing to persistent EXT_CSD registers.
>>Although, in this case, I am actually wondering on what is the best approach.
>>
>>Is there really ever a case when we don't want auto BKOPS to be default enabled?
>>I think BKOPS is a fundamental feature of an FTL and I can't see a
>>reason to why we need to involve mmc-utils/userspace to enable it. Am
>>I wrong?
>
> The even worst case is â involve mmc-utils/userspace to DISABLE it.
> I think this register need to be set by vendor and no need to be changed on runtime.

If it is set by the Vendor, that's of course the best.

Are you saying that we shouldn't enable it during the card init
sequence from the kernel, in case it is disabled?

Kind regards
Uffe