Date:   Tue, 28 Jun 2022 16:38:50 +0200
From:   Miquel Raynal <miquel.raynal@...tlin.com>
To:     Martin Blumenstingl <martin.blumenstingl@...glemail.com>
Cc:     linux-mtd@...ts.infradead.org, devicetree@...r.kernel.org,
        linux-kernel@...r.kernel.org, tlanger@...linear.com,
        rtanwar@...linear.com, richard@....at, vigneshr@...com
Subject: Re: [PATCH RFC v1 0/8] intel-nand-controller: Fixes, cleanups and
 questions

Hi Martin,

martin.blumenstingl@...glemail.com wrote on Tue, 28 Jun 2022 13:27:23
+0200:

> Hello,
> 
> I am trying to replace the xway_nand driver (which is still using the
> legacy NAND API) with the intel-nand-controller driver. The Intel LGM
> IP (for which intel-nand-controller was implemented) uses a newer
> version of the EBU NAND and HSNAND IP found in Lantiq XWAY SoCs. The
> most notable change is the addition of HSNAND Intel LGM SoCs (it's not
> clear to me if/which Lantiq SoCs also have this DMA engine).
> 
> While testing my changes on a Lantiq xRX200 SoC I came across some
> issues with the intel-nand-controller driver. The problems I found are:
> 1) Mismatch between dt-bindings and driver implementation (compatible
>    string, patch #1 and patch #4) and hardware capabilities (number of
>    CS lines, patch #1).
> 2) The driver reads the CS (chip select) line from the NAND controller's
>    reg property. In the dt-bindings example this is 0xe0f00000. I don't
>    understand how this can even work on Intel SoCs. My understanding is
>    that it must be read from the NAND chip (child node).

Yes

> 3) A few smaller code cleanups to make the driver easier to understand
>    (patches #5 to #8)
> 4) I tried to understand the timing parameter calculation code but found
>    that it probably doesn't work on the Intel LGM SoCs either. The
>    dt-bindings example use clock ID 125 which is LGM_GCLK_EBU. So far
>    this is fine because EBU is the actual IP block for the NAND
>    interface. However, drivers/clk/x86/clk-lgm.c defines this clock as
>    a gate without a parent, so it's rate (as read by Linux) is always 0.
>    The intel-nand-controller driver then tries to calculate:
>      rate = clk_get_rate(ctrl->clk) / HZ_PER_MHZ
>    (rate will be 0 because clk_get_rate() returns 0) and then:
>      DIV_ROUND_UP(USEC_PER_SEC, rate)
>    (this then tries to divide by zero)
> 
> Before this series is applied it would be great to have these questions
> answered:
> - My understanding is that the chip select line (reg property of the
>   NAND controller's child node) refers to the chip select line of the
>   controller. Let's say we have a controller with two CS lines. A NAND
>   flash chip (which a single chip select line) is connected to the
>   second CS line of the controller. Is my understanding correct that
>   the NAND chip device-tree node should get reg = <1> in this case?

Yes

> - Who from Maxlinear (who took over Intel's AnyWAN division, which
>   previously worked on the drivers for the Intel LGM SoCs) can send a
>   patch to correct the LGM_GCLK_EBU clock rate in
>   drivers/clk/x86/clk-lgm.c? Or is LGM dead and the various drivers
>   should be removed instead?
> - Who from Maxlinear can provide insights into which clock is connected
>   to the EBU NAND controller on Lantiq XWAY (Danube, xRX100, xRX200,
>   xRX300) SoCs as well as newer GRX350/GRX550 SoCs so that I can make
>   the intel-nand-controller work without hardcoded timing settings on
>   the XWAY SoCs?
> 
> Due to the severity of issues 2) and 4) above I am targeting linux-next
> with this series. In my opinion there's no point in backporting these
> fixes to a driver which has been broken since it was upstreamed.

The changes look good to me, please resend with the bindings file name
fixed and we should be good.

Thanks,
Miquèl

Powered by blists - more mailing lists