Re: [PATCH] pci: rename *host* directory to *controller*

From: Kishon Vijay Abraham I
Date: Fri Dec 30 2016 - 01:25:10 EST


Hi,

On Thursday 29 December 2016 06:49 PM, Joao Pinto wrote:
>
> Hi,
>
> Às 12:20 PM de 12/29/2016, Kishon Vijay Abraham I escreveu:
>> Hi,
>>
>> On Thursday 29 December 2016 05:38 PM, Joao Pinto wrote:
>>> Às 11:58 AM de 12/29/2016, Kishon Vijay Abraham I escreveu:
>>>> Hi,
>>>>
>>>> On Thursday 29 December 2016 05:23 PM, Joao Pinto wrote:
>>>>> Às 11:48 AM de 12/29/2016, Kishon Vijay Abraham I escreveu:
>>>>>> Hi,
>>>>>>
>>>>>> On Thursday 29 December 2016 04:08 PM, Joao Pinto wrote:
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Às 5:46 AM de 12/29/2016, Kishon Vijay Abraham I escreveu:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> On Wednesday 28 December 2016 10:50 PM, Joao Pinto wrote:
>>>>>>>>> Às 5:17 PM de 12/28/2016, Joao Pinto escreveu:
>>>>>>>>>> Às 4:41 PM de 12/28/2016, Bjorn Helgaas escreveu:
>>>>>>>>>>> On Wed, Dec 28, 2016 at 01:57:13PM +0000, Joao Pinto wrote:
>>>>>>>>>>>> Às 9:22 AM de 12/28/2016, Christoph Hellwig escreveu:
>>>>>>>>>>>>> On Wed, Dec 28, 2016 at 01:39:37PM +0530, Kishon Vijay Abraham I wrote:
>>>>>>>>>>>>>> As discussed during our LPC discussions, I'm posting the rename patch
>>>>>>>>>>>>>> here. I'll post the rest of EP series before the next merge window.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> There might be hiccups because of this renaming but feel this is
>>>>>>>>>>>>>> necessary for long-term maintenance.
>>>>>>>>>>>>>
>>>>>>>>>>>>> if we do this rename it would be great to get it to Linus NOW after
>>>>>>>>>>>>> -rc1 as that minimizes the impact on the 4.11 merge window.
>>>>>>>>>>>>
>>>>>>>>>>>> Rename it to controller is a bit vague I thing since we have the PCI Endpoint IP
>>>>>>>>>>>> also. Wouldn't be better to name it rc_controller?
>>>>>>>>>>>
>>>>>>>>>>> I think Kishon's whole goal is to add PCI Endpoint IP, so he wants a
>>>>>>>>>>> neutral name that can cover both RC and Endpoint.
>>>>>>>>
>>>>>>>> right.
>>>>>>>>>>>
>>>>>>>>>>> I'm not a huge fan of "controller" because it feels a little bit long
>>>>>>>>>>> and while I suppose it technically does include the concept of the PCI
>>>>>>>>>>> interface of an endpoint, it still suggests more of the host side to
>>>>>>>>>>> me.
>>>>>>>>>>>
>>>>>>>>>>> Doesn't USB have a similar situation? I see there's a
>>>>>>>>>>> drivers/usb/host/ (probably where we copied from in the first place).
>>>>>>>>>>> Is a USB gadget the USB analog of what you're doing? How do they
>>>>>>>>>>> share code between the master/slave sides?
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> The usb/host contains the implemnentations by the spec of the several
>>>>>>>>>> *hci (USB Host) and then you can have for example the USB 3.0 Designware
>>>>>>>>>> Host specific ops in dwc3 and for USB 2.0 in dwc2/.
>>>>>>>>
>>>>>>>> right, each IP have a separate directory in USB. I thought of doing something
>>>>>>>> similar for PCI but decided against it since that would involve identifying all
>>>>>>>> the PCI IPs used and eventually result in more directories.
>>>>>>>>>> For device purposes it uses the core/ and then some of the device functions
>>>>>>>>>> are extended from the gadget/ package in which you can use mass_storage and
>>>>>>>>>> other types of functions.
>>>>>>>>
>>>>>>>> That would be similar for PCI endpoint. All endpoint specific core
>>>>>>>> functionality will be added in drivers/pci/endpoint (see RFC [1]).
>>>>>>>>>>
>>>>>>>>>> In our case in PCI we have the core functions inside /drivers/pci and the host
>>>>>>>>>> mangled inside host. I suggest:
>>>>>>>>>>
>>>>>>>>>> drivers/pci
>>>>>>>>>> drivers/pci/core/
>>>>>>>>>> drivers/pci/core/hotplug
>>>>>>>>>> drivers/pci/core/pcie
>>>>>>>>>> drivers/pci/core/<all other files inside pci/ today>
>>>>>>>>>> drivers/pci/host
>>>>>>>>>> drivers/pci/dwc -> here would be pcie-designware and the specific vendor drivers
>>>>>>>>>
>>>>>>>>> Correction:
>>>>>>>>> drivers/pci/host/dwc -> here would be pcie-designware and the specific vendor
>>>>>>>>> drivers
>>>>>>>>>
>>>>>>>>>> drivers/pci/<vendorN> -> here would be the drivers for vendorN controller
>>>>>>>>>
>>>>>>>>> Correction:
>>>>>>>>> drivers/pci/host/<vendorN> -> here would be the drivers for vendorN controller
>>>>>>>>>
>>>>>>>>>> drivers/pci/endpoint -> common code
>>>>>>>>>> drivers/pci/endpoint/dwc -> implementation of Synopsys specific endpoint ops
>>>>>>>>>> drivers/pci/<vendorN> -> implementation of other vendors specific endpoint ops
>>>>>>>>
>>>>>>>> There are some parts of the dwc driver that is common to both root complex and
>>>>>>>> endpoint. Where should that be? I'm sure no one wants to duplicate the common
>>>>>>>> piece in both root complex and endpoint.
>>>>>>>
>>>>>>> You are right, the config space is almost the same and some ops also common.
>>>>>>> I would suggest:
>>>>>>>
>>>>>>> drivers/pci
>>>>>>> drivers/pci/core/
>>>>>>> drivers/pci/core/hotplug
>>>>>>> drivers/pci/core/pcie
>>>>>>> drivers/pci/core/<all other files inside pci/ today>
>>>>>>> drivers/pci/dwc
>>>>>>> drivers/pci/dwc/common.c -> common ops and registers between RC and endpoint
>>>>>>> drivers/pci/dwc/host/
>>>>>>> drivers/pci/dwc/endpoint/
>>>>>>>
>>>>>>> What do you think?
>>>>>>
>>>>>> I don't think we should have sub-directories within dwc (USB too doesn't have
>>>>>> sub-directories). Where should the platform specific driver be kept? For
>>>>>> example pci-dra7xx.c (which use dwc) has both rc and ep specific parts but the
>>>>>> changes are so minimal that splitting the file won't make much sense.
>>>>>>
>>>>>> And such a change would also mean we create a separate directory for every
>>>>>> other driver present right now in pci/host.
>>>>>
>>>>> I understand you idea. We can simplify it this way:
>>>>>
>>>>> drivers/pci
>>>>> drivers/pci/core/
>>>>> drivers/pci/core/hotplug
>>>>> drivers/pci/core/pcie
>>>>> drivers/pci/core/<all other files inside pci/ today>
>>>>> drivers/pci/dwc -> Common files (RC and EP), specific vendor drivers for EP
>>>>> and EP
>>>>>
>>>>> BTW dwc states for DesigWare Controller.
>>>>>
>>>>> What do you think?
>>>>
>>>> I'd like to avoid using different directory structures for different IPs. Lets
>>>> try to make it uniform.
>>>
>>> I understand, but mixing them all up is not a good aproach in my opinion, since
>>> a SoCs using Synopsys IP will only use the common files for that IP. Today in
>>> the host/ folder you have a bunch of drivers that is not clear which are using
>>> Synopsys IP or not. Of course I mention Synopsys, because it is where I work,
>>> but the opinion would be the same for other IP vendor.
>>
>> right, but it has been that way always. My point is why should we disturb it
>> for the sake of adding endpoint support.
>>>
>>> I understand that you want to do a common Endpoint framework to be used by any
>>> IP vendor based, and maybe this partition makes it a bit harder, but in my
>>> personal opinion each IP vendor should have its own folder for clarity and
>>> organization of the code.
>>>
>>> You framework should be outside those IP vendor folders and be available for
>>> their drivers to use it, and so it should be completely vendor agnostic.
>>
>> right, that's how it was designed (please see drivers/pci/endpoint/ directories
>> in RFC [1] which has the endpoint framework).
>>
>> [1] -> https://urldefense.proofpoint.com/v2/url?u=https-3A__lkml.org_lkml_2016_9_14_27&d=DgID-g&c=DPL6_X_6JkXFx7AXWqB0tg&r=s2fO0hii0OGNOv9qQy_HRXy-xAJUD1NNoEcc3io_kx0&m=Usq-eV6vJQ4rVg_Jj-JzgYyH26VfwLaqXDrE_54JCmg&s=wSindfuj1wjD3QeOSzLJNPmgMyqcZNtNX0X-D5yCpqw&e=
>
> I have checked your patch and seems very useful for certain that I am going to
> use to add support a DWC reference Endpoint driver.
> Synopsys Endpoint has a DMA engine available that can be used or not by the IP
> client. To configure and use it the IP as a set of registers that like other IP
> must be configured and managed. Other IP vendors won't have this feature for
> sure. So I am saying is, although it gets simpler to have a host/ and a
> endpoint/ folder with everything inside, I still feel that we should isolate IP
> vendor stuff in isolated ecosystems creating host/dwc and endpoint/dwc where
> Synopsys clients can put their drivers and use the common functions. Current
> host/ is just a place where every Root Complex driver is deployed. I know it
> works of course, but if you are improving structure why not organize it better?

I'm not inclined to have host/dwc or endpoint/dwc as that would mean the driver
code is spread across multiple directories.
I would be rather okay to have separate directory for dwc (and have both RC/EP
controller drivers there).
But let's get the opinion of others. Bjorn? Jingoo?

Thanks
Kishon