[PATCH linux-next v5 1/5] mtd: spi-nor: notify (Q)SPI controller about protocol change

[Cyrille Pitchen]

Hi Jonas,

Le 26/08/2015 16:02, Jonas Gorski a écrit :
> On Wed, Aug 26, 2015 at 2:30 PM, Cyrille Pitchen
> <cyrille.pitchen at atmel.com> wrote:
>> Once the Quad SPI mode has been enabled on a Micron flash memory, this
>> device expects ALL the following commands to use the SPI 4-4-4 protocol.
>> The (Q)SPI controller needs to be notified about the protocol change so it
>> can adapt and keep on dialoging with the Micron memory.
> 
> Doesn't that mean you need to disable quad mode on removal? Else the
> following will break/fail:
> 
> insmod atmel-quadspi.ko ~> spi-nor attaches -> sees micron -> enables quad mode
> rmmod atmel-quadspi.ko ~> spi-nor detaches
> insmod atmel-quadspi.ko ~> spi-nor attaches -> fails to read the id
> because flash is still in quad mode.
>

Indeed you're right such an issue does exist. So as you said one solution
could be create a new function to "clean" what have been done by spi_nor_scan()
then call it from remove() function of each driver calling spi_nor_scan() from
its probe().
However we could also enhance the probing of the memory. It is true that Micron
spi nor memories only accept SPI 4-4-4 commands once switched in quad mode but
actually they also provide a new command for this purpose:
"Multiple I/O Read ID" (0xAF).
Hence we could first try to probe using the regular Read ID (0x9F) command then
change the protocol, for instance to SPI 4-4-4, and try the 0xAF command.
I don't think all combinations for command/protocol need to be tested: for
Micron memories, their datasheets claim the 0x9F command is only supported in
"extended spi mode" so for the SPI 1-1-1 protocol whereas the 0xAF command
only works in dual or quad modes.
On the other hand Spansion memories still reply to the regular 0x9F command in
SPI 1-1-1 protocol even after their quad mode had been enabled.
For other manufacturers, well... I don't know!

Some of the advantages of the probe solution as compared to the remove one are:
- we don't need to patch all drivers using spi_nor_scan() to call a new
  function from their remove().
- it doesn't rely on the assumption that the spi-nor memory starts in
  SPI 1-1-1 protocol. As a matter of fact the remove() won't be called for
  built-in modules or in many (all ?) cases of reset. Moreover some bootloaders
  may have already enabled the quad mode before starting the Linux kernel. This
  is what the sama5d2 romcode does when it is configured to boot from a QSPI
  memory.

Anyway you're right and the issue need to be addressed but maybe in another
dedicated patch ?
 
>> Signed-off-by: Cyrille Pitchen <cyrille.pitchen at atmel.com>
>> Acked-by: Marek Vasut <marex at denx.de>
>> Acked-by: Bean Huo <beanhuo at micron.com>
>> ---
>>  drivers/mtd/spi-nor/spi-nor.c | 21 +++++++++++++++++++++
>>  include/linux/mtd/spi-nor.h   | 13 +++++++++++++
>>  2 files changed, 34 insertions(+)
>>
>> diff --git a/drivers/mtd/spi-nor/spi-nor.c b/drivers/mtd/spi-nor/spi-nor.c
>> index c27d427fead4..c8810313a752 100644
>> --- a/drivers/mtd/spi-nor/spi-nor.c
>> +++ b/drivers/mtd/spi-nor/spi-nor.c
>> @@ -165,6 +165,22 @@ static inline int write_disable(struct spi_nor *nor)
>>         return nor->write_reg(nor, SPINOR_OP_WRDI, NULL, 0, 0);
>>  }
>>
>> +/*
>> + * Let the spi-nor framework notify lower layers, especially the driver of the
>> + * (Q)SPI controller, about the new protocol to be used. Indeed, once the
>> + * spi-nor framework has sent manufacturer specific commands to a memory to
>> + * enable its Quad SPI mode, it should immediately after tell the QSPI
>> + * controller to use the very same Quad SPI protocol as expected by the memory.
>> + */
>> +static inline int spi_nor_set_protocol(struct spi_nor *nor,
>> +                                      enum spi_protocol proto)
>> +{
>> +       if (nor->set_protocol)
>> +               return nor->set_protocol(nor, proto);
>> +
>> +       return 0;
> 
> Shouldn't the default assumption be that it won't support it? Also it
> might make sense to first check if it's supported before enabling it
> in the chip, so that we don't enable something just to then find out
> we can't back out of it.
> 
> I also wonder if we need an extra flag for that as at least SPI has
> RX_{DUAL,QUAD} and TX_{DUAL,QUAD} separated, so in theory there could
> be a controller that supports quad read, but not quad write, so we
> shouldn't be using the quad mode in that case. m25p80 currently sets
> SPI_NOR_{DUAL,QUAD} only based on SPI_RX_{DUAL,QUAD}, so that would
> then fail.
> 
> At least the n25q32 seems to support the "boring" 1_1_2 and 1_1_4
> commands, so these should work in case the spi controller does not
> support quad tx, but quad rx.
> 
> So maybe an additional flag for the "full" dual/quad modes would be in order.
> 

My first thought was to return -ENOSYS when the set_protocol() callback is not
implemented but logically none of the already existing drivers implements it.
So I made this new callback optional. This way, micron_quad_enable() works the
exact same way as before for all the existing drivers so no regression or side-
effect should be introduced by this patch.

Besides, the purpose of this callback is to notify the SPI controller that the
protocol change has been done at the memory side but not to decide whether such
a change is possible. Indeed, the capabilities of both the controller and the
memory are checked before calling set_quad_mode() so the decision has already
been taken when micron_quad_enable() is eventually called. Once the Quad Mode
bit set in the Enhanced Volatile Configuration Register of Micron memory, it's
too late: the memory now expects commands in SPI 4-4-4 protocol whether or not
the controller supports this protocol.

So more accurate checks should be done before calling set_quad_mode(). Maybe
the range of values for the mode parameter of spi_nor_scan() is too small.
SPI_NOR_QUAD doesn't allow to make the difference between SPI protocols 1-1-4,
1-4-4 or 4-4-4. Knowing the exact set of protocols supported by both the
controller and the memory could help making the right decision and choosing the
best suited Fast Read command.
The RX_{DUAL_QUAD} and TX_{DUAL,QUAD} flags from the spi->mode in the m25p80
driver provide more information than the limited "mode" argument of
spi_nor_scan().

> Last but not least, the protocol probably should be stored somewhere
> in the nor struct, so that users don't have to store it themselves (I
> assume they will need to check it for each command again to know if
> the cmd/address should be send in serial or quad mode).
> 
> 
I agree with you, this could be an improvement for some spi controller drivers :)

> Jonas
> 

thanks for your review!

Best regards,

Cyrille