Message-ID: <20230614082205.74a04de1@xps-13>
Date:   Wed, 14 Jun 2023 08:22:05 +0200
From:   Miquel Raynal <miquel.raynal@...tlin.com>
To:     William Zhang <william.zhang@...adcom.com>
Cc:     Broadcom Kernel List <bcm-kernel-feedback-list@...adcom.com>,
        Linux MTD List <linux-mtd@...ts.infradead.org>,
        f.fainelli@...il.com, rafal@...ecki.pl, kursad.oney@...adcom.com,
        joel.peshkin@...adcom.com, computersforpeace@...il.com,
        anand.gore@...adcom.com, dregan@...l.com, kamal.dasu@...adcom.com,
        tomer.yacoby@...adcom.com, dan.beygelman@...adcom.com,
        linux-kernel@...r.kernel.org,
        Vignesh Raghavendra <vigneshr@...com>,
        Richard Weinberger <richard@....at>,
        Kamal Dasu <kdasu.kdev@...il.com>,
        linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH 10/12] mtd: rawnand: brcmnand: Add BCMBCA read data bus
 interface

Hi William,

william.zhang@...adcom.com wrote on Tue, 13 Jun 2023 17:00:19 -0700:

> Hi Miquel,
> 
> On 06/12/2023 11:42 PM, Miquel Raynal wrote:
> > Hi William,
> > 
> > william.zhang@...adcom.com wrote on Mon, 12 Jun 2023 12:18:58 -0700:
> >   
> >> On 06/12/2023 10:53 AM, Miquel Raynal wrote:  
> >>> Hello again,  
> >>>    >>>>>>>>>> Perhaps we could have a single function that is statically assigned at  
> >>>>>>>>>> probe time instead of a first helper with two conditions which calls in
> >>>>>>>>>> one case another hook... This can be simplified I guess.  
> >>>>>>>>>>       >> Well this will need to be done at the SoC specific implementation level (bcm<xxx>_nand.c) and each SoC will need to have either general data bus read func with is_param option or data_bus_read_page, data_bus_read_param.  
> >>>>>>>>
> >>>>>>>> You told me in case we would use exec_op we could avoid the param
> >>>>>>>> cache. If that's true then the whole support can be simplified.  
> >>>>>>>>      >> Correct we may possibly unified the parameter data read but exec_op is long shot and we are not fully ready for that yet. It also depends on if the low level data register has endianess difference for the parameter data between difference SoCs.  
> >>>>>>>
> >>>>>>> So I would like to push the current implementation and we can explore the exec_op option late which will be a much big and complete different implementation.  
> >>>>>>
> >>>>>> I am sorry but this series is totally backwards, you're trying to guess
> >>>>>> what comes next with the 'is_param' thing, it's exactly what we are
> >>>>>> fighting against since 2017. There are plenty of ->exec_op()
> >>>>>> conversions out there, I don't believe this one will be harder. You
> >>>>>> need to convert the driver to this new API and get rid of this whole
> >>>>>> endianness non-sense to simplify a lot the driver.  
> >>>>>>     >>> I am not guessing anything but just factor out the existing common nand cache read logic into the single default function(or one for page read and another for parameter read as I mentioned in another thread) and allow SoC to overrides the implementation when needed.  
> >>>>
> >>>> No, you are trying to guess what type of read the core is performing,
> >>>> either a regular data page read or a parameter page read.  
> >>>>   >>>>> I agree ->exec_op can possibly get rid of the parameter page read function and is the way to go. But it won't help on the page read for endianess.  
> >>>>
> >>>> You told me there is no endianess issue with the data pages, so why it
> >>>> won't help on the page read?  
> >>>>   >>>>> It's not that I am against exec_op but I want to take one step a time  
> >>>>> and I'd like to get these fixes  
> >>>>
> >>>> I don't see any fix here? Let me know if I am missing something but
> >>>> right now I see a new version of the controller being supported with
> >>>> its own constraints. If you are fixing existing code for already
> >>>> supported platform, then make it clear and we can discuss this. But if
> >>>> you just want to support the bcmbca flavor, then there is no risk
> >>>> mitigation involved here, and a conversion is the right step :)  
> >>>>   >>>  
> >>> I forgot to mention: the exec_op conversion is almost ready, Boris
> >>> worked on it but he lacked the hardware so maybe you'll just need to
> >>> revive the few patches which target your platform and do a little bit of
> >>> debugging?
> >>>
> >>> https://github.com/bbrezillon/linux/commits/nand/exec-op-conversion?after=8a3cf6fd25d5e15c6667f9e95c1fc86e4cb735e6+34&branch=nand%2Fexec-op-conversion&qualified_name=refs%2Fheads%2Fnand%2Fexec-op-conversion  
> >>>    >> Yes this is the patch what our exec_op work is based on. Thanks Boris! The issue with patch is that performance is very slow for anything that rely on nand_read_page_op as the patch implementing it using the low level cmd and data register to transfer the data byte by byte.  
> > 
> > You don't need to use exec_op for your read_page/write_page hooks,
> > quite the opposite actually. exec_op is not meant for high throughput.
> > exec_op is meant to be simple. You can have fast I/Os with a different
> > mechanism in your read/write_page hooks.
> >   
> Right it does not impact our fast path: controller based ecc read/write. But things like on-chip ecc nand driver that uses exec_op API get impacted badly. We need to add nand op parser, several matching rules and other logics to use fast path page read/write instead of the low level data register read/write.
> 
> >>   I actually sent out email regarding this to Boris and he cc'ed you in
> >>   sept last year. We have to use the nand parser to match the page read
> >>   from exec_op so we can actually match and use the brcmnand_page_read
> >>   fast path. But there are many situations that we need to match so the
> >>   project to migrate exce_op are still work in progress just on our
> >>   bcmbca chip as of now.  Just forward that email again to you and I
> >>   appreciate it if you have any inputs there. So IMHO it is just too
> >>   risky and too big of scope to have the exec_op added to this patch
> >>   series and definitively better to do it afterwards with a dedicated
> >>   patch.  
> > 
> > As long as you add small and orthogonal changes to cmd_ctrl/cmd_func
> > I don't mind, but what you want now is to force me to pull dirty
> > changes "first", the type of change we are refusing since 2018, making
> > me expect you'll perform the conversion after. It would have been
> > terribly less dirty and you would have all your code already upstreamed
> > if you had performed the exec_op conversion since September.
> >   
> I didn't work on open source 5 years ago. I am sorry that I missed the background of the rejected changes since then but I do not agree that this change is dirty change just because I factor out the code with is_param argument(and I offered an alternative to remove is_param with two data read functions).

This _is_ dirty because you cannot know with the cmd_ctrl/cmdfunc
API whether we read a parameter page or a page of data. So your are
_guessing_. There are plenty ways of reading one of the others, the
heuristics on the controller side will _always_ be wrong. That is why
exec_op() was introduced.

> I see your point with exec_op and agree that is the way to go.  We had an initial look of the Borris exec_op patch last Sept and noticed the performance issue but we haven't got the chance to actively work on improving the performance and prepare for up-streaming until recently. What if we bring in the original exec_op patch in this series so we don't need to add the parameter data read function(if we verify it works on difference SoCs without endianess)?  Or better to have exec_op as separate patch first and then this series?

This one is my favorite:
1/ Add exec_op support
2/ Remove legacy hooks
3/ Add support for the bcmbca SoC

Then you can improve the performance for on-die ECC situations, but to
be honest this improvement looks little a very little addition. You can
take example from the existing hooks, how they match specific
operations in the parser and then hook them to specific helpers.
Nothing terribly complex, there are dozens of conversions available
now.

Good luck :)
Miquèl

Powered by blists - more mailing lists