[U-Boot] [PATCH 00/16] SF: Migrate to Linux SPI NOR framework

[Vignesh R <vigneshr@ti.com>]

On 18-Dec-18 6:02 PM, Jagan Teki wrote:
> On Sat, Dec 15, 2018 at 9:13 PM Vignesh R <vigneshr@ti.com> wrote:
>>
>>>>> 2) For BAR support, lets place it as it is and support via spi-nor
>>>>
>>>> Problem is, it not desirable to use BAR as default because its not
>>>> stateless and does not work with all flash parts. OTOH, it seems like 4
>>>> byte addressing (stateless dedicated opcode or with enter/exit 4 byte
>>>> mode) seems to be standard.
>>>> Also, Linux doesn't support BAR and haven't seen any request for BAR
>>>> support. Why support additional feature and burden of maintaining when
>>>> it may not be needed.
>>>>
>>>> But if you insist, I just have to add BAR support back.
>>>
>>> It's not about insistence, ie how we support since from long and
>>> u-boot bootloader project does in general. bootloader can be some
>>> certain boot difference functionalities unlike Linux, it's better not
>>> to compare u-boot with Linux in all cases.
>>>
>>> Example we have SPI_TX_BYTE which usually not supported in Linux.
>>> Since it's ich controller specific and it require for bootloader to do
>>> byte tx on that specific controller, so we supported. Same for the
>>> case with the BAR, this specific controller(or supported boards) has
>>> been in U-Boot since from long and they do upstream well. So we need
>>> to support BAR in any case or we can find any alternative to work the
>>> same functionalities. (we tried before but ended-up adding BAR)
>>>
>> How about this instead?
>>
>> Lets use 4 byte addressing opcodes as default for >16MB flashes.
>> But if CONFIG_SPI_FLASH_BAR is enabled then, spi-nor layer will use BAR
>> registers for >16MB access instead of 4 byte addressing.
> 
> Yes, this can be normal approach. but what I'm trying to initiate here
> is if we have a generic kind of sanity check transfer for > 16MB
> flashes, then we can notify the controllers enable BAR to support >
> 16MiB.
> 

I guess you are saying to attempt 4 byte addressing transfers and if
that fails then try using BAR? AFAIK, there is _no way_ to detect
whether or not controller/flash supports 4 byte addressing and then do a
fallback.

> This way we can reduce the ifdefs on code, and if we handle BAR in
> better way to minimize the code eventually we even drop the CONFIG
> option.
> 
>> I will remove SPI_FLASH_BAR from defconfigs from all boards expect those
>> controllers that really cannot handle 4 byte addressing. From a quick
>> glance it looks following controllers support only 3 byte addresses:
>> stm32_qspi.c
>> ich.c
>> fsl_qspi.c
>> renesas_rpc_spi.c
>> mtk_qspi.c
> 
> I'm not sure whether all these support 4-byte addressing or not? 

Just to be clear, all SPI drivers _except_ those in above list can
support 4 byte addressing mode w/o any additional code change.

> since  we don't have 4-byte on spi-flash they do use BAR for accessing >
> 16MiB, zynq_qspi is the key taker for this initially.
>

I am confused here, let me put down more details:

Access >16MB of flash memory is possible by 3 ways:
1. using dedicated 4 byte addressing opcodes
2. Send "enter 4 byte" command and then use normal commands but with
   4 bytes of address
3 Provide BA24 BA25 of address via BAR register

AFAIK, all flashes support 1 or 2. But only some flashes support 3.
So, the idea is to use option 1 or 2 as default and use 3 only if
SPI_FLASH_BAR is enabled.

Ideally, there is _nothing_ in SPI controllers that would stop us from
using 4 byte addressing because a SPI controller would have no idea of
slave connected (ie. it may or may not be a flash). SPI drivers should
_blindly_ send/receive dout/din buffers passed as part of spi_xfer() and
not interpret these buffers as cmd/addr/data separately
So, the decision on which of the above 3 options to use should be in the
SPI NOR layer and has nothing to do with SPI controller drivers.

But since spi_flash.c was always used 3 byte addressing, _some_ of the
SPI controller drivers are _hard-coded_ to expect 3 byte addresses (eg.:
above 4 drivers that I mentioned) See [1] on how these drivers convert
dout buffer is back to flash offset. This is because there is no way to
communicate cmd/addr/data separately using spi_xfer(), but _intelligent_
SPI HW do need this information. This is not a HW limitation but more of
a framework limitation.

So, thus we can classify SPI controllers into roughly two types:
1. Controllers that need to know flash specific information like opcode,
address length, dummy bytes etc (controllers that supports memory mapped
access to flash(or such other acceleration interface).
Such controllers should move to spi-mem (but can optionally implement
spi_xfer to support non flash devices)

2. Traditional SPI controllers that are usually FIFO/shift register
based and provide direct access to SPI bus will continue to use
traditional SPI framework.

Moving to 4 byte addressing mode will not affect second type of SPI
controllers, but will break first type of controllers that do their own
address conversion logic like[1].
So my suggestion is to enable CONFIG_SPI_FLASH_BAR only for controllers
to do[1] until moved to spi-mem.

From my analysis:
>> stm32_qspi.c
>> ich.c
>> fsl_qspi.c
>> renesas_rpc_spi.c
>> mtk_qspi.c
>> ti_qspi.c
>> cadence_quadspi.c

are candidates to move to spi-mem layer. Rest seems to be traditional
SPI controllers.

AFAIU, zynq_qspi seems to be FIFO based SPI controller. Why would
zynq_qspi driverneed to know whether SPI NOR layer uses BAR or not? Or
did I miss something?

>>
>> So, except for boards with above controllers, I will remove
>> SPI_FLASH_BAR for all other boards. If there is a regression, then such
>> boards can go back to enabling CONFIG_SPI_FLASH_BAR.
> 
> I think we can go the BAR code in flash as you mentioned with CONFIG
> option, and don't remove any CONFIG items from board configs then we
> can add sanity check patch on top of spi-nor changes and give warning
> to boards which never need BAR(ask them to drop CONFIG_BAR). This way
> transition look promising to move BAR to 4-byte addressing.
> 

Are you suggesting board users should individually remove
CONFIG_SPI_FLASH_BAR from defconfigs based on their judgement?
How do you propose to do sanity check?
Why not use 4 byte addressing as default and enable CONFIG_SPI_FLASH_BAR
only for controllers that need them?

>>
>> AFAIU, above controller HWs(except ich perhaps) can support 4 byte
>> addressing but would need to move to spi-mem.
>t
> Even if we move to spi-mem, but BAR need to handle it flash side. isn't it?
> 

Yes, presence/configuration of BAR registers is abstracted by SPI NOR
and spi-mem drivers need not have any knowledge.

[1]
https://elixir.bootlin.com/u-boot/latest/source/drivers/spi/renesas_rpc_spi.c#L264
https://elixir.bootlin.com/u-boot/v2019.01-rc2/source/drivers/spi/mtk_qspi.c#L253

Regards
Vignesh