Re: [PATCH v3 4/6] mtd: rawnand: add NVIDIA Tegra NAND Flash controller driver

[Stefan Agner]

On 09.06.2018 08:41, Boris Brezillon wrote:
> On Sat, 09 Jun 2018 08:23:51 +0200
> Stefan Agner <stefan@xxxxxxxx> wrote:
> 
>> On 09.06.2018 07:52, Boris Brezillon wrote:
>> > On Fri, 08 Jun 2018 23:51:01 +0200
>> > Stefan Agner <stefan@xxxxxxxx> wrote:
>> >
>> >
>> >> >
>> >> > void tegra_nand_controller_reset(struct tegra_nand_controller *ctrl)
>> >> > {
>> >> > 	int err;
>> >> >
>> >> > 	disable_irq(ctrl->irq);
>> >> >
>> >> > 	err = reset_control_reset(ctrl->rst);
>> >> > 	if (err) {
>> >> > 		dev_err(ctrl->dev, "Failed to reset HW: %d\n", err);
>> >> > 		msleep(HW_TIMEOUT);
>> >> > 	}
>> >> >
>> >> > 	writel_relaxed(NAND_CMD_STATUS, ctrl->regs + HWSTATUS_CMD);
>> >> > 	writel_relaxed(HWSTATUS_MASK, ctrl->regs + HWSTATUS_MASK);
>> >> > 	writel_relaxed(INT_MASK, ctrl->regs + ISR);
>> >>
>> >> If we do a controller reset, there is much more state than that which
>> >> needs to be restored. A lot of it is not readily available currently
>> >> (timing, ECC settings...)
>> >
>> > This is actually a good test to detect what is not properly initialized
>> > by the driver. Timings should be configured correctly through
>> > ->setup_data_interface(). ECC engine should be disabled by default and
>> > only enabled when ->{read,write}_page() is called.
>> >
>>
>> Is setup_data_interface guaranteed to be called after a failed
>> ->exec_op()/{read,write}_page()?
> 
> No. Maybe I misunderstood when tegra_nand_controller_reset() was
> supposed to be called. That's something I would call only once, early
> in the probe function, so that the controller is placed in a well-known
> state before we start using it. Definitely not something you should
> call after each error.
> 

Dmitry suggests to make use of it in the error handling path in case the
command/DMA  timed out.

Which makes sense in general I guess, just to make sure that the state
is properly set. It just isn't entirely trivial to do, since state is
setup during probe, chip detect and timing setup...

>>
>> >>
>> >> That seems a lot of work for a code path I do not intend to ever use :-)
>> >>
>> >
>> > Not so sure it's a lot of work. If ECC and timing settings are the
>> > only thing you need to initialize then it should work just fine.
>> > Try with a controller reset and you'll know if you miss something ;-).
>>
>> Currently the setting gets written directly to the registers. Only the
>> enable flag is set in the HW ECC {read,write}_page() functions. So I
>> will have to store the complete register in the chip structure and write
>> them on every {read,write}_page()?
> 
> Well, your solution works as long as you only have one chip connected
> to the controller. What we usually set the ECC config in
> ->select_chip() (or at least make sure the current setting matches the
> one we expect) and then enable the engine in read/write_page() (as you
> seem to already do).

I did not plan to make the driver multi chip capable, as I am not aware
of any real hardware using it.

But to properly reset state, we would have to have all the chip settings
stored somewhere...

--
Stefan