Re: [PATCH 1/3] spi/qspi: Add memory mapped read support.

[Sourav Poddar <sourav.poddar@ti.com>]

On Wednesday 09 October 2013 11:10 PM, Mark Brown wrote:
> On Wed, Oct 09, 2013 at 10:24:33PM +0530, Sourav Poddar wrote:
>
>> Here is the exact feature usecase..
>> TI qspi controller supports memory mapped read. These memory
>> mapped read configuration depends on the set_up_register, which
>> can be configured once during in setup apis based on the dt node
>> specifying whether the qspi controller supports memory mapped read
>> or not.
> What is "set_up_register"?
set_up_register is a memory mapped specific register with the
following fields:
- RCMD (Read command, which is actually the flash read command)
- NUmber of Address bytes
- Number of dummy bytes
- Read type(Single, dual and quad read).
- Write command.
>> Once, the qspi controller is configured for a memory mapped read, the qspi
>> controller does not depend on the flash command that comes from the
>> mtd layer.
>> Because, this command are already configured in QSPI set up register.
> So this does depend on the flash commands for the specific chip, which
> means that this has all the same problems as the Freescale chip had with
> requiring the user to replicate the information about the commands that
> the chip supports into the device tree.  It therefore seems like all the
> same concerns should apply, though in this case it seems like it's
> harder for the driver to infer things from looking at the operations
> being sent to it.
>
> Presumably this also only works for flash chips, or things that look
> like them...
>
Yes, true, it depends on flash command, though it is getting filled now in
my driver itself.
>> Basically, its not the commands which need to be communicated from
>> the mtd layer,its just
>> the buffer to fill, len of buffer, offset from where to fill need to
>> be communicated.
> This appears to be based on an assumption that the commands would be
> replicated into the device trees which seems like it's both more work
> for users and harder to deploy.
>
>>> I'm also concerned about the interface here, it looks like this is being
>>> made visible to SPI devices (via a dependency on patch 3/3...) but only
>>> as a flag they can set - how would devices know to enable this and why
>>> would they want to avoid it?
>> Set  spi->mode in qspi driver based on dt entry and use that in mtd layer to
>> decide whether to use memory mapped or not.
> But why would anything not want to use memory mapped mode if it can and
> why is this something that should be hard coded into the device tree?
> Especially with the current API...
>
Thats true, by default also we can keep the memory mapped read. Though, 
according
to the current implementation spi->mode can be set so that in mtd layer, 
we might
use that to selectively used t[o].memory_map.
>> The idea is whenever, we call mtd_read api from mtd layer, if memory
>> mapped is set
>> then sending the commands does not matter, what matters is the len
>> to read, buffer to fill and
>> "from" offset to read. Then, the intention was to use the memory_map
>> transfer parameter to
>> communicate to the driver that memory mapped read is used so that we
>> can just use memcopy and
>> return without going through the entire SPI based xfer function.
> I'm not convinced that this is the most useful API, it sounds like the
> hardware can "memory map" the entire flash chip so the whole SPI
> framework seems like overhead.
>
But this memory map read will work only with read opcodes.(mtd_read
path). For all other operations, normal SPI operations will be used.

As for this, I also though of bypassing the SPI frameowrk, and doing a 
memcopy
at the beginning of the mtd_read api. But, then before doing a memory mapped
read -
1. Controller need to be switched to memory mapped port using control module
      register and ti qspi switch register.
2. There is SOC specific memory mapped address space from where read 
should happen,
     this is SOC specific and should be known to mtd layer the adreess 
to read for.
So, I thought of going this way using t.memory map flag.
> It also seems seems like it's going to involve the CPU being stalled
> waiting for reads to complete instead of asking the SPI controller to
> DMA the data to RAM and allowing the CPU to get on with other things -
> replacing the explicit transmission of commands with memory to memory
> DMAs might be advantageous but replacing DMA with memcpy() would need
> numbers to show that it was a win.


______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/