Re: [PATCH 0/3] mtd: Make sure UBIFS does not do multi-pass page programming on flashes that don't support it

From: Pratyush Yadav <p.yadav@ti.com>
To: Vignesh Raghavendra <vigneshr@ti.com>
Cc: Tudor Ambarus <tudor.ambarus@microchip.com>,
	Richard Weinberger <richard.weinberger@gmail.com>,
	Richard Weinberger <richard@nod.at>,
	LKML <linux-kernel@vger.kernel.org>,
	linux-mtd@lists.infradead.org,
	Miquel Raynal <miquel.raynal@bootlin.com>
Subject: Re: [PATCH 0/3] mtd: Make sure UBIFS does not do multi-pass page programming on flashes that don't support it
Date: Tue, 3 Nov 2020 18:15:29 +0530	[thread overview]
Message-ID: <20201103124527.x6mp6slck44aotzn@ti.com> (raw)
In-Reply-To: <fa578bda-132a-320a-264c-d973bae194dd@ti.com>

On 03/11/20 05:05PM, Vignesh Raghavendra wrote:
> 
> 
> On 11/1/20 3:14 AM, Richard Weinberger wrote:
> > On Tue, Oct 27, 2020 at 12:24 PM Pratyush Yadav <p.yadav@ti.com> wrote:
> >>> [0] https://lore.kernel.org/linux-mtd/20201005153138.6437-1-p.yadav@ti.com/
> >>
> >> Ping. Any comments on the series?
> > 
> > From the UBIFS point of view I'd like to avoid as many device specific
> > settings as possible.
> > We check already for NOR flash, checking for NOR *and* SPI_NOR_NO_MULTI_PASS_PP
> > feels a bit clumsy.
> > 
> > Tudor, what do you think about SPI_NOR_NO_MULTI_PASS_PP?
> > This kind of NOR seems to be a little NAND'ish. Maybe we can hide this detail
> > in the mtd framework?
> > 
> 
> Agree with Richard. I don't see need for SPI_NOR_NO_MULTI_PASS_PP. From
> MTD point of view setting mtd->writesize to be equal to pagesize should
> be enough. Its upto clients of MTD devices to ensure there is no multi
> pass programming within a "writesize" block.

That is what I initially thought too but then I realized that multi-pass 
programming is completely different from page-size programming. Instead 
of writing 4 bytes twice, you can zero out the entire page in one single 
operation. You would be compliant with the write size requirement but 
you still do multi-pass programming because you did not erase the page 
before this operation.

It is also not completely correct to say the Cypress S28 flash has a 
write size of 256. You _can_ write one byte if you want. You just can't 
write to that page again without erasing it first. For example, if a 
file system only wants to write 128 bytes on a page, it can do so 
without having to write the whole page. It just needs to make sure it 
doesn't write to it again without erasing first.

nor_erase_prepare() was written to handle quirks of some specific 
devices. Not every device starts filling zeroes from the end of a page. 
So we have device-specific code in UBIFS already. You will obviously 
need device-specific settings to have control over that code.

One might argue that we should move nor_erase_prepare() out of UBIFS. 
But requiring a flash to start erasing from the start of the page is a 
UBIFS-specific requirement. Other users of a flash might not care about 
it at all.

And so we have ourselves a bit of a conundrum. Adding 
SPI_NOR_NO_MULTI_PASS_PP is IMHO the least disruptive answer. If the 
file system wants to do multi-pass page programming on NOR flashes, how 
else do we tell it not to do it for this specific flash?

> If this is not clear in the current documentation of struct mtd, then
> that can be updated.

-- 
Regards,
Pratyush Yadav
Texas Instruments India

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