Re: [PATCH 1/3] spi: spi-mem: add automatic poll status functions

[Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx>]

On Mon, 3 May 2021 14:59:37 +0530
Pratyush Yadav <p.yadav@xxxxxx> wrote:

> On 03/05/21 11:11AM, Boris Brezillon wrote:
> > On Mon, 3 May 2021 14:17:44 +0530
> > Pratyush Yadav <p.yadav@xxxxxx> wrote:
> >   
> > > On 30/04/21 06:51PM, Boris Brezillon wrote:  
> > > > On Mon, 26 Apr 2021 16:39:32 +0200
> > > > <patrice.chotard@xxxxxxxxxxx> wrote:
> > > >     
> > > > > From: Christophe Kerello <christophe.kerello@xxxxxxxxxxx>
> > > > > 
> > > > > With STM32 QSPI, it is possible to poll the status register of the device.
> > > > > This could be done to offload the CPU during an operation (erase or
> > > > > program a SPI NAND for example).
> > > > > 
> > > > > spi_mem_poll_status API has been added to handle this feature.
> > > > > 
> > > > > Signed-off-by: Christophe Kerello <christophe.kerello@xxxxxxxxxxx>
> > > > > Signed-off-by: Patrice Chotard <patrice.chotard@xxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/spi/spi-mem.c       | 34 ++++++++++++++++++++++++++++++++++
> > > > >  include/linux/spi/spi-mem.h |  8 ++++++++
> > > > >  2 files changed, 42 insertions(+)
> > > > > 
> > > > > diff --git a/drivers/spi/spi-mem.c b/drivers/spi/spi-mem.c
> > > > > index 1513553e4080..43dce4b0efa4 100644
> > > > > --- a/drivers/spi/spi-mem.c
> > > > > +++ b/drivers/spi/spi-mem.c
> > > > > @@ -743,6 +743,40 @@ static inline struct spi_mem_driver *to_spi_mem_drv(struct device_driver *drv)
> > > > >  	return container_of(drv, struct spi_mem_driver, spidrv.driver);
> > > > >  }
> > > > >  
> > > > > +/**
> > > > > + * spi_mem_poll_status() - Poll memory device status
> > > > > + * @mem: SPI memory device
> > > > > + * @op: the memory operation to execute
> > > > > + * @mask: status bitmask to ckeck
> > > > > + * @match: status expected value
> > > > > + * @timeout: timeout
> > > > > + *
> > > > > + * This function send a polling status request to the controller driver
> > > > > + *
> > > > > + * Return: 0 in case of success, -ETIMEDOUT in case of error,
> > > > > + *         -EOPNOTSUPP if not supported.
> > > > > + */
> > > > > +int spi_mem_poll_status(struct spi_mem *mem,
> > > > > +			const struct spi_mem_op *op,
> > > > > +			u8 mask, u8 match, u16 timeout)
> > > > > +{
> > > > > +	struct spi_controller *ctlr = mem->spi->controller;
> > > > > +	int ret = -EOPNOTSUPP;
> > > > > +
> > > > > +	if (ctlr->mem_ops && ctlr->mem_ops->poll_status) {
> > > > > +		ret = spi_mem_access_start(mem);    
> > > > 
> > > > You should probably check that op is a single byte read before
> > > > accepting the command.    
> > > 
> > > Please do not discriminate against 8D-8D-8D flashes ;-).  
> > 
> > Then mask and match should probably be u16 :P. And the check as it is
> > seems a bit lax to me. Drivers will of course be able to reject the op
> > when there's more than one byte (or 16bit word in case of 8D) to read,
> > but it feels like the core could automate that a bit.  
> 
> The two 8D flashes that are currently supported in SPI NOR both have a 
> 1-byte status register. But to read it, the read op should be 2-byte 
> long to avoid partial cycles at the end. The second byte is simply 
> discarded.
> 
> 2-byte wide registers might show up in the future, but for now at least 
> we don't have to worry about them.

Well, I guess it doesn't hurt to take it into account now. I mean,
what's happening on the bus in that case is a 2byte transfer, with the
second byte being ignored, which you can describe with a 16bit mask
of 0xMM00 (assuming big endian transfers here, as done for other ops).

> 
> >   
> > >   
> > > >     
> > > > > +		if (ret)
> > > > > +			return ret;
> > > > > +
> > > > > +		ret = ctlr->mem_ops->poll_status(mem, op, mask, match, timeout);    
> > > > 
> > > > You also need some sort of ->poll_status_is_supported() to validate
> > > > that the controller supports the status polling for this specific op (I    
> > > 
> > > I don't think a separate function is needed for checking if the poll 
> > > status op is supported. Return value of -EOPNOTSUPP should be able to 
> > > signal that. This can also be used to check if Octal DDR capable 
> > > controllers are able to poll using 2-byte reads.  
> > 
> > Yeah, I had something more complex in mind to avoid doing this 'try
> > native mode and fall back on sw-based more if not supported' dance
> > every time a status poll is requested (something similar to what we do
> > for dirmaps, with a status poll desc), but I guess that's a bit
> > premature (and probably uneeded).  
> 
> I think Mark also suggested something similar. Make the CPU/non-CPU case 
> transparent to the caller. I agree with with this direction. Makes the 
> caller simpler.

It's kind of orthogonal to what I was suggesting, but yes, that's
definitely a good idea. We certainly don't want the spi-nor layer to
open code the same logic if the spi-mem layer can do it for us.

> 
> I also mentioned in a reply to this patch that supports_op() should be 
> called before the op is executed. That should take care of "base" 
> support for the op. The poll-specific checks can go in the poll_status() 
> function itself. If either of those say the op is not supported, it 
> should fall back to CPU based polling. That's the design that makes the 
> most sense to me.

What I had in mind was more:

1/ create a poll desc with spi_mem_create_poll_status_desc(). The
   "operation supported" check is done here. The controller can store
   all its HW-specific state in there. If the operation is not natively
   supported, a SW-based poll descriptor (similar to the SW-based
   dirmap) is created
2/ poll the status with spi_mem_poll_status(). This function is passed
   a poll descriptor which helps select the path that should be taken
   without having to check every time whether the hardware supports a
   specific status polling op. I can also imagine some preparation
   being done during the desc creation if that makes sense (preparing
   reg values to be written when a status poll request is issued for
   instance)

Anyway, as I said, this sort of optimization might be a bit premature.