U-Boot SPI DM framework issues

U-Boot SPI DM framework issues

[Andre Przywara]

Hi,

while trying to debug some nasty SPI flash issue on sunxi, I came
across some oddities in the SPI DM framework, and wanted to check some
things, since I am not sure whether I miss things here:

- When I do a simple "sf probe" call, I see *two* calls to the .claim_bus
  callback: the first one from:
	drivers/mtd/spi/sf_probe.c:spi_flash_probe_slave(), which is
  expected, but then also, shortly afterwards, a second call from:
	drivers/spi/spi-mem.c:spi_mem_exec_op().
  After the probe operation finished, I see the corresponding
  two spi_release_bus() calls. Seeing *two* claim calls in that nested
  way looks rather odd, and the name *claim* sounds like it should be only
  one slave device/driver being able to, well, claim the bus.
  From digging through the code it looks like we call spi_mem_exec_op()
  only from *inside* drivers that already claimed the bus, so shall we
  drop the call inside spi_mem_exec_op()?
  Allowing those nested calls has some consequences for the
  implementation of .claim_bus in the controller drivers, as it does not
  sound indicated to enabled/disable hardware in there. Enabling clock
  gates *again* might not be harmful per se, but if the first .release_bus
  call already disables clocks and resets, this might cause problems.

- In many controller drivers I see the .of_to_plat implementation checking
  for a "spi-max-frequency" property in the *controller's* DT node, mostly
  with a fallback value, and using that frequency as an upper bus
  frequency limit for the whole controller operation. But looking at the
  official DT bindings, this property is supposed to be a *slave device*
  property, and indeed checking a few dozen .dts file I see it only being
  used in *child* nodes of SPI controllers, never in controller nodes.
  There is some proper usage in
  drivers/spi/spi-uclass.c:spi_slave_of_to_plat(), but to me it looks like
  the bus node code is wrong and should be removed, from the other
  invocation in spi-uclass.c and all the controller drivers.
  For sunxi (and probably others) this has the consequence of limiting the
  whole bus to some "default" frequency limit, which is rather low:
	#define SUN4I_SPI_DEFAULT_RATE          1000000
        plat->max_hz = fdtdec_get_int(gd->fdt_blob, node,
                                      "spi-max-frequency",
                                      SUN4I_SPI_DEFAULT_RATE);
  So we never go beyond this 1 MHz, even though the SPI flash chips
  explicitly advertise 40MHz or more in their child nodes.

Please let me know if I am missing something here! If not, I am happy to
send some patches aiming at fixing those things.

Cheers,
Andre

Re: U-Boot SPI DM framework issues

[Simon Glass]

Hi Andre,

On Mon, 11 Jul 2022 at 09:52, Andre Przywara <andre.przywara@arm.com> wrote:
>
> Hi,
>
> while trying to debug some nasty SPI flash issue on sunxi, I came
> across some oddities in the SPI DM framework, and wanted to check some
> things, since I am not sure whether I miss things here:
>
> - When I do a simple "sf probe" call, I see *two* calls to the .claim_bus
>   callback: the first one from:
>         drivers/mtd/spi/sf_probe.c:spi_flash_probe_slave(), which is
>   expected, but then also, shortly afterwards, a second call from:
>         drivers/spi/spi-mem.c:spi_mem_exec_op().
>   After the probe operation finished, I see the corresponding
>   two spi_release_bus() calls. Seeing *two* claim calls in that nested
>   way looks rather odd, and the name *claim* sounds like it should be only
>   one slave device/driver being able to, well, claim the bus.
>   From digging through the code it looks like we call spi_mem_exec_op()
>   only from *inside* drivers that already claimed the bus, so shall we
>   drop the call inside spi_mem_exec_op()?
>   Allowing those nested calls has some consequences for the
>   implementation of .claim_bus in the controller drivers, as it does not
>   sound indicated to enabled/disable hardware in there. Enabling clock
>   gates *again* might not be harmful per se, but if the first .release_bus
>   call already disables clocks and resets, this might cause problems.
>
> - In many controller drivers I see the .of_to_plat implementation checking
>   for a "spi-max-frequency" property in the *controller's* DT node, mostly
>   with a fallback value, and using that frequency as an upper bus
>   frequency limit for the whole controller operation. But looking at the
>   official DT bindings, this property is supposed to be a *slave device*
>   property, and indeed checking a few dozen .dts file I see it only being
>   used in *child* nodes of SPI controllers, never in controller nodes.
>   There is some proper usage in
>   drivers/spi/spi-uclass.c:spi_slave_of_to_plat(), but to me it looks like
>   the bus node code is wrong and should be removed, from the other
>   invocation in spi-uclass.c and all the controller drivers.
>   For sunxi (and probably others) this has the consequence of limiting the
>   whole bus to some "default" frequency limit, which is rather low:
>         #define SUN4I_SPI_DEFAULT_RATE          1000000
>         plat->max_hz = fdtdec_get_int(gd->fdt_blob, node,
>                                       "spi-max-frequency",
>                                       SUN4I_SPI_DEFAULT_RATE);
>   So we never go beyond this 1 MHz, even though the SPI flash chips
>   explicitly advertise 40MHz or more in their child nodes.
>
> Please let me know if I am missing something here! If not, I am happy to
> send some patches aiming at fixing those things.
>

This all sounds right to me and patches are welcome.

Re the first issue, I suspect something went haywire with the spimem conversion.

For the second, the binding predates Linux (I think?) so it would be
good to tidy this up.

Regards,
Simon