Re: i.MX28 nand driver broken in Linux 4.18

[Miquel Raynal <miquel.raynal@xxxxxxxxxxx>]

Hi Wolfgang,

Wolfgang Grandegger <wg@xxxxxxxxxxxxxx> wrote on Mon, 1 Apr 2019
22:08:45 +0200:

> Hello Miquel;
> 
> Am 01.04.19 um 11:23 schrieb Miquel Raynal:
> > Hi Wolfgang,
> > 
> > Wolfgang Grandegger <wg@xxxxxxxxxxxxxx> wrote on Sat, 23 Mar 2019
> > 20:55:19 +0100:
> >   
> >> Hello Miquel,
> >>
> >> Am 06.03.19 um 14:59 schrieb Miquel Raynal:  
> >>> Hi Wolfgang,
> >>>
> >>> Wolfgang Grandegger <wg@xxxxxxxxxxxxxx> wrote on Tue, 5 Mar 2019
> >>> 15:52:52 +0100:
> >>>     
> >>>> Hello,
> >>>>
> >>>> I will bisect the problem next week when I have access to the
> >>>> hardware... more soon...
> >>>>    
> >>>
> >>> Great, thanks.    
> >>
> >> Here is the result of git bisection:
> >>
> >> wolf@bernex:~/git/linux$ git bisect good
> >> 76e1a0086a0c3276b384f77905345e0fcc886fdd is the first bad commit
> >> commit 76e1a0086a0c3276b384f77905345e0fcc886fdd
> >> Author: Miquel Raynal <miquel.raynal@xxxxxxxxxxx>
> >> Date:   Fri Mar 2 15:38:39 2018 +0100
> >>
> >>     mtd: rawnand: gpmi: support ->setup_data_interface()
> >>     
> >>     Until now the GPMI driver had its own timings logic while the core
> >>     already handles that and request the NAND controller drivers to support
> >>     the ->setup_data_interface() hook. Implement that hook by reusing the
> >>     already existing function. No real glue is necessary between core timing
> >>     delays and GPMI registers because the driver already translates the
> >>     ONFI timing modes into register values.
> >>     
> >>     Make use of the core's tREA, tRLOH and tRHOH values that allow computing
> >>     more precise timings for mode [0-3] and get significantly better values
> >>     (+20% with an i.MX6 Sabre Auto board). Otherwise use the existing logic.
> >>     
> >>     Signed-off-by: Miquel Raynal <miquel.raynal@xxxxxxxxxxx>
> >>     Tested-by: Han Xu <han.xu@xxxxxxx>
> >>     Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxx>  
> > 
> > Thank you for the bisection, there is definitely something wrong with
> > this commit but it worked for me and for Han so it's quite difficult to
> > find out what is failing if I cannot reproduce. Could you please dump
> > the timing registers in both cases (working/not working) and observer if
> > there are odd values ? (0, too short or too big values, etc).  
> 
> here are some first figures:
> 
> 76e1a0086a0c3276b384f77905345e0fcc886fdd^:
> [    1.911760] clock_period_in_ns : 41
> [    1.922818] address_setup_in_cycles : 1
> [    1.915343] data_setup_in_cycles : 3
> [    1.919254] data_hold_in_cycles : 2
> [    1.926709] HW_GPMI_TIMING0 : 0x10203
> [    1.930641] HW_GPMI_TIMING1 : 0x5000000
> 
> v4.18:
> [    2.090621] period_ps : 45454
> [    2.076601] addr_setup_cycles : 1
> [    2.080002] data_setup_cycles : 1
> [    2.083598] data_hold_cycles : 1
> [    2.093849] HW_GPMI_TIMING0 : 0x10101
> [    2.096890] HW_GPMI_TIMING1 : 0x90000000
> 
> Hope that's what you are looking for. Unfortunately, the code of both
> versions is very different (complete rewrite). I will have a closer look
> tomorrow.

Just looking at these values it does not look like there is a big
difference...

Can you please force the timing registers to working values (taken from
v4.18) to be sure the problem comes from the derivations made in
->setup_data_interface() and not something else?

Also, in setup_data_interface(), you may refuse to support the highest
modes (4 and 5) and observe if it solves something? I would test:
forcing to maximum mode 4, 3 and 0. Please continue to dump the above
values/registers to compare.

Thanks for your time,
Miquèl

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/