+ Michal
> -----Original Message-----
> From: Anurag Kumar Vulisha
> Sent: Friday, February 26, 2016 7:18 PM
> To: 'Rob Herring'
> Cc: Arnd Bergmann; [email protected]; [email protected];
> [email protected]; [email protected]; [email protected];
> [email protected]; [email protected]; linux-
> [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah Choudary
> Kalluri
> Subject: RE: [RFC PATCH] drivers: ata: Read Rx water mark value from device-
> tree
>
> Hi Rob,
>
> > -----Original Message-----
> > From: Rob Herring [mailto:[email protected]]
> > Sent: Wednesday, February 24, 2016 1:00 AM
> > To: Anurag Kumar Vulisha
> > Cc: Arnd Bergmann; [email protected]; [email protected];
> > [email protected]; [email protected]; [email protected];
> > [email protected]; [email protected]; linux-
> > [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah
> > Choudary Kalluri
> > Subject: Re: [RFC PATCH] drivers: ata: Read Rx water mark value from
> > device- tree
> >
> > On Tue, Feb 23, 2016 at 03:29:55PM +0000, Anurag Kumar Vulisha wrote:
> > > Hi Arnd,
> > >
> > > > -----Original Message-----
> > > > From: Arnd Bergmann [mailto:[email protected]]
> > > > Sent: Tuesday, February 23, 2016 3:51 PM
> > > > To: Anurag Kumar Vulisha
> > > > Cc: [email protected]; [email protected];
> [email protected];
> > > > [email protected]; [email protected];
> > > > [email protected];
> > > > [email protected]; [email protected]; linux-
> > > > [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah
> > > > Choudary Kalluri
> > > > Subject: Re: [RFC PATCH] drivers: ata: Read Rx water mark value
> > > > from
> > > > device- tree
> > > >
> > > > On Tuesday 23 February 2016 05:58:32 Anurag Kumar Vulisha wrote:
> > > > > >
> > > > > > I don't know what is appropriate because I have no idea what
> > > > > > Rxwatermark is good for. Can you try describing why we can't
> > > > > > just set it to the correct value for everyone automatically?
> > > > > >
> > > > >
> > > > > This RX watermark level sets the minimum number of free
> > > > > locations within the RX FIFO .When the rx fifo level crosses the
> > > > > programmed watermark level ,sata controller will transmit HOLDS
> > > > > to the device asking it
> > > > to wait. This happens when dma reads the rx fifo data slower than
> > > > the device is sending the data. Note that it can take some time
> > > > for the HOLDs to get to the other end and in the time there must
> > > > be enough room in the FIFO to absorb all data that could arrive
> > > > from the
> > device.
> > > > > Currently we are using 0x40 for this value, which works fine
> > > > > with all hardware designs we are currently having. But hoping
> > > > > that this value may vary for future silicon versions, I wanted
> > > > > to make this as a configurable
> > > > value. So for this reason I thought of moving it either to
> > > > device-tree or making it as a module_param() property.
> > > > >
> > > >
> > > > Ok, so if this depends on the silicon version, your initial
> > > > approach would be better than the module_param.
> > > >
> > > > I would probably make this dependent on the compatible string
> > > > instead, and have a table in the device driver that uses a
> > > > specific value for each variant of the device, but either way should be
> fine.
> > > >
> > > > Having a separate property is most appropriate if for each
> > > > hardware revision there is exactly one ideal value, while a table
> > > > in the driver makes more sense if this takes a bit of tuning and
> > > > the driver might choose to optimize it differently based on other
> > > > constraints, such as its own interrupt handler implementation.
> > > >
> > >
> > > Since we are currently having one value in common for all the
> > > hardware and also changing the rx water mark does not require any
> > > changes other than vendor specific PTC register update , I think it
> > > would be better to use device tree property for that rx watermark
> value.
> > Doing this makes the updating of rx watermark value easy, if any
> > changes required.
> > >
> > > In future, if any silicon version rx water mark value doesn't work
> > > with the current versions, then I will do as you said by
> > > maintaining the table in device driver. But at present I feel that
> > > single rx watermark property in device tree would be enough, since
> > > it works with all
> > the hardware versions we have.
> >
> > If you currently have no reason to modify it now, then add it later
> > when you actually have a use case.
> >
>
> This property may vary from board to board . It also depends on the phy
> reference clock frequency which can be changed based on use case. So
> moving this property to device tree will make it easy for the user to
> configure this property based on the requirement.
>
> Thanks,
> Anuarg Kumar V
This email and any attachments are intended for the sole use of the named recipient(s) and contain(s) confidential information that may be proprietary, privileged or copyrighted under applicable law. If you are not the intended recipient, do not read, copy, or forward this email message or any attachments. Delete this email message and any attachments immediately.
Hi Rob and Arnd,
On 2.3.2016 06:53, Anurag Kumar Vulisha wrote:
> + Michal
>
>> -----Original Message-----
>> From: Anurag Kumar Vulisha
>> Sent: Friday, February 26, 2016 7:18 PM
>> To: 'Rob Herring'
>> Cc: Arnd Bergmann; [email protected]; [email protected];
>> [email protected]; [email protected]; [email protected];
>> [email protected]; [email protected]; linux-
>> [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah Choudary
>> Kalluri
>> Subject: RE: [RFC PATCH] drivers: ata: Read Rx water mark value from device-
>> tree
>>
>> Hi Rob,
>>
>>> -----Original Message-----
>>> From: Rob Herring [mailto:[email protected]]
>>> Sent: Wednesday, February 24, 2016 1:00 AM
>>> To: Anurag Kumar Vulisha
>>> Cc: Arnd Bergmann; [email protected]; [email protected];
>>> [email protected]; [email protected]; [email protected];
>>> [email protected]; [email protected]; linux-
>>> [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah
>>> Choudary Kalluri
>>> Subject: Re: [RFC PATCH] drivers: ata: Read Rx water mark value from
>>> device- tree
>>>
>>> On Tue, Feb 23, 2016 at 03:29:55PM +0000, Anurag Kumar Vulisha wrote:
>>>> Hi Arnd,
>>>>
>>>>> -----Original Message-----
>>>>> From: Arnd Bergmann [mailto:[email protected]]
>>>>> Sent: Tuesday, February 23, 2016 3:51 PM
>>>>> To: Anurag Kumar Vulisha
>>>>> Cc: [email protected]; [email protected];
>> [email protected];
>>>>> [email protected]; [email protected];
>>>>> [email protected];
>>>>> [email protected]; [email protected]; linux-
>>>>> [email protected]; Anirudha Sarangi; Srikanth Vemula; Punnaiah
>>>>> Choudary Kalluri
>>>>> Subject: Re: [RFC PATCH] drivers: ata: Read Rx water mark value
>>>>> from
>>>>> device- tree
>>>>>
>>>>> On Tuesday 23 February 2016 05:58:32 Anurag Kumar Vulisha wrote:
>>>>>>>
>>>>>>> I don't know what is appropriate because I have no idea what
>>>>>>> Rxwatermark is good for. Can you try describing why we can't
>>>>>>> just set it to the correct value for everyone automatically?
>>>>>>>
>>>>>>
>>>>>> This RX watermark level sets the minimum number of free
>>>>>> locations within the RX FIFO .When the rx fifo level crosses the
>>>>>> programmed watermark level ,sata controller will transmit HOLDS
>>>>>> to the device asking it
>>>>> to wait. This happens when dma reads the rx fifo data slower than
>>>>> the device is sending the data. Note that it can take some time
>>>>> for the HOLDs to get to the other end and in the time there must
>>>>> be enough room in the FIFO to absorb all data that could arrive
>>>>> from the
>>> device.
>>>>>> Currently we are using 0x40 for this value, which works fine
>>>>>> with all hardware designs we are currently having. But hoping
>>>>>> that this value may vary for future silicon versions, I wanted
>>>>>> to make this as a configurable
>>>>> value. So for this reason I thought of moving it either to
>>>>> device-tree or making it as a module_param() property.
>>>>>>
>>>>>
>>>>> Ok, so if this depends on the silicon version, your initial
>>>>> approach would be better than the module_param.
>>>>>
>>>>> I would probably make this dependent on the compatible string
>>>>> instead, and have a table in the device driver that uses a
>>>>> specific value for each variant of the device, but either way should be
>> fine.
>>>>>
>>>>> Having a separate property is most appropriate if for each
>>>>> hardware revision there is exactly one ideal value, while a table
>>>>> in the driver makes more sense if this takes a bit of tuning and
>>>>> the driver might choose to optimize it differently based on other
>>>>> constraints, such as its own interrupt handler implementation.
that 0x40 is value choose based on testing that it is not causing any
visible problem and this is used as default value in the driver
(PTC_RX_WM_VAL - ahci_ceva.c)
Values which you can setup are in range 0x0 - 0x7f (7bits). It means
hardware fifo size is probably 0x80.
And this dt/module parameter is IMHO just sw setting setup by user.
It means I am not quite sure that this is DT parameter because it is
just SW setting.
I expect this range will be valid for all silicon revisions.
If happen that any silicon revision can't setup certain level because of
HW bug we can handle it via DT parameter or specific compatible string.
But setting up watermark SW level via DT doesn't look correct to me.
Please let me know what you think.
Thanks,
Michal
On Wednesday 02 March 2016 09:05:49 Michal Simek wrote:
> On 2.3.2016 06:53, Anurag Kumar Vulisha wrote:
> >>>>>
> >>>>> I would probably make this dependent on the compatible string
> >>>>> instead, and have a table in the device driver that uses a
> >>>>> specific value for each variant of the device, but either way should be
> >> fine.
> >>>>>
> >>>>> Having a separate property is most appropriate if for each
> >>>>> hardware revision there is exactly one ideal value, while a table
> >>>>> in the driver makes more sense if this takes a bit of tuning and
> >>>>> the driver might choose to optimize it differently based on other
> >>>>> constraints, such as its own interrupt handler implementation.
>
> that 0x40 is value choose based on testing that it is not causing any
> visible problem and this is used as default value in the driver
> (PTC_RX_WM_VAL - ahci_ceva.c)
>
> Values which you can setup are in range 0x0 - 0x7f (7bits). It means
> hardware fifo size is probably 0x80.
>
> And this dt/module parameter is IMHO just sw setting setup by user.
> It means I am not quite sure that this is DT parameter because it is
> just SW setting.
> I expect this range will be valid for all silicon revisions.
> If happen that any silicon revision can't setup certain level because of
> HW bug we can handle it via DT parameter or specific compatible string.
> But setting up watermark SW level via DT doesn't look correct to me.
>
> Please let me know what you think.
Ok, thanks for the background. I think we should just leave it to be
set by the driver then. Please make sure that each SoC specific .dtsi
file has a unique "compatible" string for the device though, so that
the driver can later override it based on the specific variant if
that ends up being necessary for performance or bug-avoidance.
Arnd
On 2.3.2016 10:11, Arnd Bergmann wrote:
> On Wednesday 02 March 2016 09:05:49 Michal Simek wrote:
>> On 2.3.2016 06:53, Anurag Kumar Vulisha wrote:
>>>>>>>
>>>>>>> I would probably make this dependent on the compatible string
>>>>>>> instead, and have a table in the device driver that uses a
>>>>>>> specific value for each variant of the device, but either way should be
>>>> fine.
>>>>>>>
>>>>>>> Having a separate property is most appropriate if for each
>>>>>>> hardware revision there is exactly one ideal value, while a table
>>>>>>> in the driver makes more sense if this takes a bit of tuning and
>>>>>>> the driver might choose to optimize it differently based on other
>>>>>>> constraints, such as its own interrupt handler implementation.
>>
>> that 0x40 is value choose based on testing that it is not causing any
>> visible problem and this is used as default value in the driver
>> (PTC_RX_WM_VAL - ahci_ceva.c)
>>
>> Values which you can setup are in range 0x0 - 0x7f (7bits). It means
>> hardware fifo size is probably 0x80.
>>
>> And this dt/module parameter is IMHO just sw setting setup by user.
>> It means I am not quite sure that this is DT parameter because it is
>> just SW setting.
>> I expect this range will be valid for all silicon revisions.
>> If happen that any silicon revision can't setup certain level because of
>> HW bug we can handle it via DT parameter or specific compatible string.
>> But setting up watermark SW level via DT doesn't look correct to me.
>>
>> Please let me know what you think.
>
> Ok, thanks for the background. I think we should just leave it to be
> set by the driver then. Please make sure that each SoC specific .dtsi
> file has a unique "compatible" string for the device though, so that
> the driver can later override it based on the specific variant if
> that ends up being necessary for performance or bug-avoidance.
No problem with default value in driver. Something has to be setup.
Reset value based on reg spec I was checking is 0x20. Based on our
testing we saw some issues that's why 0x40 was setup as default value.
There is a need to be able to configure this value for example for
testing different values that's why I think module parameter should be
the right way to go.
If this should be DT parameters there should be different ceva IP which
allows different fifo size and different watermark level to be setup by
user.
What do you think? Does it sound reasonable.
Thanks,
Michal
On Wednesday 02 March 2016 10:27:51 Michal Simek wrote:
>
> No problem with default value in driver. Something has to be setup.
> Reset value based on reg spec I was checking is 0x20. Based on our
> testing we saw some issues that's why 0x40 was setup as default value.
> There is a need to be able to configure this value for example for
> testing different values that's why I think module parameter should be
> the right way to go.
I don't object to the module parameter, but I don't understand how important
that kind of testing is to normal users. Who would set it, aside from
the person writing that driver to come up with the correct default?
> If this should be DT parameters there should be different ceva IP which
> allows different fifo size and different watermark level to be setup by
> user.
>
> What do you think? Does it sound reasonable.
Having a property for the actual hardware fifo size once you get
different implementations seems like the correct approach, but it's
moot as long as all implementations are hardwired to 128 entries.
Arnd
On 2.3.2016 12:42, Arnd Bergmann wrote:
> On Wednesday 02 March 2016 10:27:51 Michal Simek wrote:
>>
>> No problem with default value in driver. Something has to be setup.
>> Reset value based on reg spec I was checking is 0x20. Based on our
>> testing we saw some issues that's why 0x40 was setup as default value.
>> There is a need to be able to configure this value for example for
>> testing different values that's why I think module parameter should be
>> the right way to go.
>
> I don't object to the module parameter, but I don't understand how important
> that kind of testing is to normal users. Who would set it, aside from
> the person writing that driver to come up with the correct default?
>
>> If this should be DT parameters there should be different ceva IP which
>> allows different fifo size and different watermark level to be setup by
>> user.
>>
>> What do you think? Does it sound reasonable.
>
> Having a property for the actual hardware fifo size once you get
> different implementations seems like the correct approach, but it's
> moot as long as all implementations are hardwired to 128 entries.
yep right now and we don't know what can happen in future. I just wanted
to point to example where this property can be specified or wired to
particular compatible string.
Anurag: Please make it as module parameter instead of DT parameter.
Thanks,
Michal