Re: [PATCH] drivers: brcmaxi: provide amba axi functionality in separate module

["Arend van Spriel" <arend@broadcom.com>]

On Thu, 21 Apr 2011 16:12:49 +0200, Arnd Bergmann <arnd@arndb.de> wrote:

> On Wednesday 20 April 2011, Arend van Spriel wrote:
>> The open-source community is looking for a library which will detect
>> cores in a chip using axi backplane. Another proposal has been
>> sent by Rafał Miłecki, which registers detected cores in the linux
>> device tree to be claimed by device drivers. This implies cores will
>> always provide a system function to the kernel which is indepent from
>> other cores and have very loose or no coupling. If this is not true,
>> exceptions need to be added in the device registration process. This
>> means knowledge of specific devices from specific vendors is sitting
>> in a bus driver. Whether the exceptions are rarely or likely is a
>> pending question.
>
> Hi Arend,
>
> I have two very general comments about this:
>
>> To feed the discussion this implementation takes a different approach.
>> A calling entity (being a pci device driver, or SoC initialization
>> sequence) registers a table with core identities and a callback  
>> function.
>> It then starts the scan and for each detected core with a callback
>> function it does the call providing the core information. Apart from
>> that it provides some basic operations on the core.
>>
>> It has been tested using the brcmsmac driver (in  
>> drivers/staging/brcm80211).
>
> The API split between PCI and non-PCI devices appears to be
> unhelpful. Can't you abstract the interface so that a user
> would apply the exact same interfaces in both cases, and handle
> the differences internally?

Ok. that can be arranged ;-)

>> +/* Core Codes */
>> +#define	NODEV_CORE_ID		0x700	/* Invalid coreid */
>> +#define	CC_CORE_ID		0x800	/* chipcommon core */
>> +#define	ILINE20_CORE_ID		0x801	/* iline20 core */
>> +#define	SRAM_CORE_ID		0x802	/* sram core */
>> +#define	SDRAM_CORE_ID		0x803	/* sdram core */
>> +#define	PCI_CORE_ID		0x804	/* pci core */
>> +#define	MIPS_CORE_ID		0x805	/* mips core */
>> +#define	ENET_CORE_ID		0x806	/* enet mac core */
>> +#define	CODEC_CORE_ID		0x807	/* v90 codec core */
>> +#define	USB_CORE_ID		0x808	/* usb 1.1 host/device core */
>> +#define	ADSL_CORE_ID		0x809	/* ADSL core */
>> +#define	ILINE100_CORE_ID	0x80a	/* iline100 core */
>> +#define	IPSEC_CORE_ID		0x80b	/* ipsec core */
>> +#define	UTOPIA_CORE_ID		0x80c	/* utopia core */
>> +#define	PCMCIA_CORE_ID		0x80d	/* pcmcia core */
>> +#define	SOCRAM_CORE_ID		0x80e	/* internal memory core */
>> +#define	MEMC_CORE_ID		0x80f	/* memc sdram core */
>> +#define	OFDM_CORE_ID		0x810	/* OFDM phy core */
>> ...
>
> This list to me is a strong hint that the cores behind the AXI bridge
> should normally be actual devices in Linux, i.e. the approach that
> Rafał suggested.  The vast majority of these is something that in Linux
> would be operated by a device driver. The exceptions that I can see
> are CPU cores and bus bridges, both of which we typically also represent
> as devices in the flattened device tree, even though they typically
> don't have a Linux driver attached to them.

Fine. Your providing the kind of feedback I was looking for. The  
OFDM_CORE_ID is also an exception.

So could a device driver claim multiple cores/devices to assure other  
drivers are not accessing those? I would prefer that over having  
exceptions coded in the axi bus driver, like the chipcommon core  
(CC_CORE_ID). I assume it can be done by the device table. Is that correct?

Rafał,

Would it be possible to make chipcommon driver optional (not doing the  
initialization)?

Gr. AvS
-- 
"The world is indeed comic, but the joke is on mankind." — H.P. Lovecraft