lkml.org 
[lkml]   [2020]   [Jun]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/2] mtd: rawnand: brcmnand: Ecc error handling on EDU transfers
Hi Kamal,

Kamal Dasu <kdasu.kdev@gmail.com> wrote on Mon, 15 Jun 2020 11:11:00
-0400:

> On Mon, Jun 15, 2020 at 3:19 AM Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> >
> > Hi Kamal,
> >
> > Kamal Dasu <kdasu.kdev@gmail.com> wrote on Fri, 12 Jun 2020 12:34:22
> > -0400:
> >
> > > On Fri, Jun 12, 2020 at 3:07 AM Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> > > >
> > > > Hi Kamal,
> > > >
> > > > Kamal Dasu <kdasu.kdev@gmail.com> wrote on Thu, 11 Jun 2020 12:04:29
> > > > -0400:
> > > >
> > > > > On Thu, Jun 11, 2020 at 3:27 AM Miquel Raynal <miquel.raynal@bootlin.com> wrote:
> > > > > >
> > > > > > Hi Kamal,
> > > > > >
> > > > > > Kamal Dasu <kdasu.kdev@gmail.com> wrote on Thu, 11 Jun 2020 01:44:54
> > > > > > -0400:
> > > > > >
> > > > > > > Implemented ECC correctable and uncorrectable error handling for EDU
> > > > > >
> > > > > > Implement?
> > > > > >
> > > > > > > reads. If ECC correctable bitflips are encountered on EDU transfer,
> > > > > >
> > > > > > extra space ^
> > > > > >
> > > > > > > read page again using pio, This is needed due to a controller lmitation
> > > > > >
> > > > > > s/pio/PIO/
> > > > > >
> > > > > > > where read and corrected data is not transferred to the DMA buffer on ECC
> > > > > > > errors. This holds true for ECC correctable errors beyond set threshold.
> > > > > >
> > > > > > error.
> > > > > >
> > > > > > Not sure what the last sentence means?
> > > > > >
> > > > >
> > > > > NAND controller allows for setting a correctable ECC threshold number
> > > > > of bits beyond which it will actually report the error to the driver.
> > > > > e.g. for BCH-4 the threshold is 3, so 3-bit and 4-bit errors will
> > > > > generate correctable ECC interrupt however 1-bit and 2-bit errors will
> > > > > be corrected silently.
> > > > > From the above example EDU hardware will not transfer corrected data
> > > > > to the DMA buffer for 3-bit and 4-bit errors that get reported. So
> > > > > once we detect
> > > > > the error duing EDU we read the page again using pio.
> > > >
> > > > Ok I see what you mean, can't you fake the threshold instead? The NAND
> > > > controller in Linux is not supposed to handle this threshold, the NAND
> > > > core is in charge. So what the controller driver should do is just:
> > > > increase the number of bitflips + return the maximum number or bitflip
> > > > or increase the failure counter. Is this already the case?
> > > >
> > > /* threshold = ceil(BCH-level * 0.75) */
> > > brcmnand_wr_corr_thresh(host, DIV_ROUND_UP(chip->ecc.strength * 3, 4));
> > > This how the threshold is set, all it means is that for high BCH
> > > levels don't interrupt on low number (less than threshold) of
> > > bit_flips. Yes the controller driver only increments correctable ECC
> > > count. But due the EDU design an EDU operation is disrupted when the
> > > controller interrupts on correctable ECC errors during subpage ECC
> > > calculations. Hence the driver needs to read the page again with PIO
> > > to transfer corrected data.
> >
> > IIUC, you are doing the job twice: you should just return a number of
> > bitflips or an error to the NAND core. So that's why I'm telling that
> > you should get rid of this threshold. It would avoid the need for the
> > PIO transfer too.
>
> I think you are reading some statements in isolation that probably are
> causing some confusion.
> EDU design has a flaw in case of reported ECC error interrupt in that
> corrected data is not transferred to the DMA buffer. The PIO is needed
> to read corrected data into the NAND data buffer and only for the
> reported errors. So there is no need to change the threshold
> calculation logic, if we get rid of the threshold then we will have to
> do the PIO read on any correctable bit error if it occurs during EDU
> reads.
>
> >
> > You also say that the controller "only increments correctable ECC
> > count", what do you mean exactly?
>
> Maybe that statement was a bit misleading. To be clear when an ECC
> error is reported the controller gives the bit_flips count as well as
> updates the ECC error address Register and ecc error status registers.
> This logic works as expected in the hardware.
>
> >The controller does not report errors
> > when the number of bitflips happens to be above the BCH threshold? This
> > would be the only case where what is currently done would be actually
> > needed though.
>
> It's the other way. The controller only reports bit errors beyond
> >=threshold value, will not report otherwise and silently correct the
> data. There is no problem in cases where erros are corrected
> silently. Now ECC (un)correctable on EDU reads are detected by simply
> reading back the ECC Error address register. And in case of reported
> uncorrectable ECC errors are treated as usual. And for reported
> correctable ECC errors we need to read the page again using PIO so
> that the corrected data is properly transferred. All this applies to
> EDU transfer only.
>

Thank you very much for the explanation, I understand better how this
controller works now.


Cheers,
Miquèl

\
 
 \ /
  Last update: 2020-06-15 19:38    [W:0.150 / U:0.292 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site