Re: [PATCH v3 10/10] mtd: rawnand: brcmnand: allow for on-die ecc

[Miquel Raynal <miquel.raynal@bootlin.com>]

Hi William,

> >>>>>>>> This is a double check to turn on/off our hardware ECC.  
> >>>>>>>
> >>>>>>> I expect the engine to be always disabled. Enable it only when you
> >>>>>>> need (may require an additional patch before this one).  
> >>>>>>
> >>>>>> We are already turning on the ECC enable at this point,
> >>>>>> this is just adding the option to turn it off if the NAND chip
> >>>>>> itself will be doing the ECC instead of our controller.  
> >>>>>
> >>>>> Sorry if I have not been clear.
> >>>>>
> >>>>> This sequence:
> >>>>> - init
> >>>>> - enable hw ECC engine
> >>>>> Is broken.  
> >>>>>   >>>> ECC engine is not enabled for all the cases. Here we only intended to enable it for the nand chip that is set to use NAND_ECC_ENGINE_TYPE_ON_HOST. The logic here should better change to:  
> >>>> if (chip->ecc.engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST)
> >>>>       brcmnand_set_ecc_enabled(host, 1);
> >>>> else
> >>>>       brcmnand_set_ecc_enabled(host, 0);  
> >>>>   >>>>> It *cannot* work as any operation going through exec_op now may  
> >>>>> perform page reads which should be unmodified by the ECC engine. You > driver *must* follow the following sequence:
> >>>>> - init and disable (or keep disabled) the hw ECC engine
> >>>>> - when you perform a page operation with correction you need to
> >>>>>      - enable the engine
> >>>>>      - perform the operation
> >>>>>      - disable the engine
> >>>>> Maybe I am missing something here but are you saying the exec_op can have different ecc type for page read/write at run time on the same nand chip? I don't see the op instr structure has the ecc type field and thought it is only bind to the nand chip and won't change at run time. So looks to me the init time setting to the engine based on ecc.engine_type should be sufficient.  
> >>>>
> >>>> What you described here can work for the hw.ecc read path (ecc.read_page = brcmnand_read_page) which always assumes ecc is enabled. Although it is probably not too bad with these two extra operation, it would be better if we don't have to add anything as our current code does. For the brcmnand_read_page_raw,  we currently disable the engine and then re-enable it(but we need to fix it to only enable it with hw ecc engine type).  So it is just opposite of you logic but works the same with no impact on the most performance critical path.  
> >>>
> >>> This is not "my" logic, this is the "core's" logic. I am saying: your
> >>> approach is broken because that is not how the API is supposed to work,
> >>> but it mostly works in the standard case.  
> >>
> >> In the interest of minimizing register writes, would it be acceptable to
> >> enable/disable ECC at the beginning of a standard
> >> path transfer but not, after the transfer, turn off the ECC? This should not
> >> affect other standard path operations nor affect the exec_op path as those
> >> are low level transfers which our ECC engine would not touch and the NAND
> >> device driver should be responsible for turning on/off its own ECC.  
> > 
> > Do you have legitimate concerns about this register write taking way
> > more time than I could expect? Because compared to the transfer of a
> > NAND page + tR/tPROG it should not be noticeable. I don't see how you
> > could even measure such impact actually, unless the register write does
> > way more than usual. I'm fine with the above idea if you show me it has
> > an interest.
> >   
> Dave did the mtd_speed test and we can see we get consistently ~35KB/s slower with the extra enable and disable ecc engine calls in ecc read page path.
> 
> With the change:
> [   28.148355] mtd_speedtest:   page read speed is 9857 KiB/s
> [   31.754258] mtd_speedtest: 2 page read speed is 9865 KiB/s
> Without the change
> [   56.444735] mtd_speedtest:   page read speed is 9892 KiB/s
> [   60.042262] mtd_speedtest: 2 page read speed is 9897 KiB/s

I believe if you repeat this 10 times you'll get totally different
results. I don't think this test on a non RT machine is precise enough
so that a unique 35kiB difference can be interpreted as being
significant.

> Although it is only less than 1% drop, it is still something. I understand the procedure you laid out above is the preferred way but with our driver fully control the chip ecc read/write page, ecc read_raw/write_raw page function and exec_op path, I don't see where it may not work.

I just told you, the exec_op path runs with ECC enabled. I don't know
how this controller works. Now if you don't care and are 100% sure this
is working and future proof, just keep it like this.

Cheers,
Miquèl