[U-Boot] [PATCH 03/16] spi: Add non DM version of SPI_MEM

[Miquel Raynal]

Hi Vignesh,

Vignesh R <vigneshr at ti.com> wrote on Thu, 13 Dec 2018 15:08:32 +0530:

> On 13/12/18 2:55 AM, Boris Brezillon wrote:
> > On Wed, 12 Dec 2018 22:07:44 +0100
> > Jagan Teki <jagan at amarulasolutions.com> wrote:
> >   
> >> On Wed 12 Dec, 2018, 10:02 PM Boris Brezillon <boris.brezillon at bootlin.com
> >> wrote:
> >>  
> >>> On Thu, 13 Dec 2018 02:15:16 +0530
> >>> Jagan Teki <jagan at amarulasolutions.com> wrote:
> >>>    
> >>>> On Thu, Dec 13, 2018 at 2:10 AM Boris Brezillon
> >>>> <boris.brezillon at bootlin.com> wrote:    
> >>>>>
> >>>>> Hi Jagan,
> >>>>>
> >>>>> On Thu, 13 Dec 2018 01:55:08 +0530
> >>>>> Jagan Teki <jagan at amarulasolutions.com> wrote:
> >>>>>    
> >>>>>> On Wed, Dec 12, 2018 at 11:08 PM Vignesh R <vigneshr at ti.com>    
> >>> wrote:    
> >>>>>>>
> >>>>>>> Add non DM version of SPI_MEM to support easy migration to new SPI    
> >>> NOR    
> >>>>>>> framework. This can be removed once DM_SPI conversion is    
> >>> complete.    
> >>>>>>
> >>>>>> Our intention to use new driver to follow dm, why we need to support
> >>>>>> non-dm? any usecases?    
> >>>>>
> >>>>> Looks like we're having the same discussion over and over. Vignesh is
> >>>>> dropping spi_flash.c which AFAICT was not depending on DM_SPI, so, if
> >>>>> we want to keep everyone happy while getting rid of some legacy code,
> >>>>> that's the only solution. DM conversion is a nice goal, but it's kind
> >>>>> of orthogonal to what Vignesh is working on. If DM_SPI conversion
> >>>>> happens before the spi-nor stuff is merged  (which I doubt) then this
> >>>>> patch can simply be dropped.    
> >>>>
> >>>> spi_flash.c is a core code not a specific driver it belongs. spi-mem
> >>>> is new feature driver how come new driver will support legacy non-dm
> >>>> do we have legacy use for that(ie what I'm asking about usecase)    
> >>>
> >>> I recommend that you read the spi-mem code carefully. spi-mem is not
> >>> driver specific, it's a thin layer on top of spi and driver *can* (but
> >>> are not forced to) provide optimized methods to execute spi-mem
> >>> operations. When that's not the case, the implementation falls back to
> >>> regular spi transfers. AFAIK, both DM and non-DM drivers support
> >>> regular spi transfers, right? So why should we depend on DM_SPI? And
> >>> more importantly, if we do that, that means we can't get rid of
> >>> spi_flash.c since some users might still have non-DM SPI drivers, which
> >>> in turn means we keep more legacy code for no good reasons.
> >>>    
> >>
> >> I understand spi-mem is core file, but new code too.  
> > 
> > Sorry, I don't get it.
> >   
> >>
> >>  
> >>> You want non-DM SPI controller drivers to go away, then remove them,
> >>> instead of blocking other changes using this excuse.
> >>>    
> >>
> >> Please understand uboot development flow, legacy driver can be removed if
> >> possible once migration expire and NEW drivers or code must be dm driven.  
> > 
> > Sorry, but I think you're the one misunderstanding what we are trying
> > to do here. Vignesh changes have simply no impact on the DM SPI
> > conversion you're aiming at. All Vignesh does is provide a dummy
> > wrapper for non-DM drivers, which would probably have been implemented
> > by Miquel if you had not been so insistent on your precious DM_SPI
> > conversion. That was not really a problem for spi-nand, as we were
> > adding support for a new feature. This is not the case here. SPI NORs
> > are already partially supported by the u-boot spi flash layer, and we
> > need to keep things in a working state for those that were using it and
> > didn't have their SPI controller drivers converted to the DM. This
> > leaves us 2 options:
> > 
> > 1/ keep the sf_flash code as is and add a new spi-nor code base 
> > 2/ replace spi_flash code by the spi-nor layer imported from Linux
> > 
> > Vignesh chose option #2 which has the benefit of avoiding code
> > duplication. Given the discussion we're having right now, I'm wondering
> > if it wouldn't be easier to go for option #1 in order to avoid those
> > endless discussions...
> >   
> 
> Boris, thanks for chiming in! This is exactly what I had in mind.
> 
> To add, I did start with #1 by simply adding support for 4 byte
> addressing. But, then released I need spi-mem to communicate this
> protocol info to SPI drivers, then found Quad Enable detection logic to
> be incomplete and so on.. Finally released I would end up with code
> exactly similar to Linux SPI-NOR(with addition of SFDP logic). Therefore
> switched to #2 ;)
> 
> 

I also share (strongly) the same point of view: #2 is the best way
IMHO. 


Thanks,
Miquèl