[PATCH][RFC] mtd: spinand: fix detection of GD5FxGQ4xA flash
boris.brezillon at collabora.com
Sun Nov 3 05:27:41 PST 2019
On Sun, 3 Nov 2019 20:03:21 +0800
Chuanhong Guo <gch981213 at gmail.com> wrote:
> On Tue, Oct 29, 2019 at 12:41 AM Miquel Raynal
> <miquel.raynal at bootlin.com> wrote:
> > Hello,
> > Chuanhong Guo <gch981213 at gmail.com> wrote on Wed, 16 Oct 2019 09:38:24
> > +0800:
> > > GD5FxGQ4xA didn't follow the SPI spec to keep MISO low while slave is
> > > reading, and instead MISO is kept high. As a result, the first byte
> > > of id becomes 0xFF.
> > > Since the first byte isn't supposed to be checked at all, this patch
> > > just removed that check.
> > >
> > > While at it, redo the comment above to better explain what's happening.
> > >
> > > Fixes: cfd93d7c908e ("mtd: spinand: Add support for GigaDevice GD5F1GQ4UFxxG")
> > > Signed-off-by: Chuanhong Guo <gch981213 at gmail.com>
> > > CC: Jeff Kletsky <git-commits at allycomm.com>
> > > ---
> > > RFC:
> > > I doubt whether this patch is a proper fix for the underlying problem:
> > > The actual problem is that we have two different implementation of read id
> > > command: One replies immediately after master sending 0x9f and the other
> > > need to send 0x9f and an offset byte (found in winbond and early GD flashes.)
> Correction: Only early GigaDevice nand chips uses this implementation.
> Winbond chips uses a dummy byte instead of an address byte so there's
> no problem for Winbond chips.
> > > Current code only works if SPI master is properly implemented (i.e. keep MOSI
> > > low while reading.)
> > I am not entirely against the fix, but this is a SPI host controller
> > issue, right? Can you try to fix the controller driver instead?
> I think this is a spi nand framework issue. GigaDevice uses an unusual
> READ ID implementation, and as a result, both host controller and chip
> are reading during the first byte after 0x9f command: chip is reading
> the address/offset byte and host is expecting the first ID byte.
> Here lies two problems:
> 1. According to the sequence diagram in their datasheet, MISO pin is
> in High-Z state during the 0x9f command and the offset byte, and host
> could read anything during this time instead of a fixed 0x0 or 0xff
> byte, so the check of first byte should be removed. This is what this
> patch is doing.
> 2. If there's a buggy SPI host controller that didn't keep MOSI low
> during reading operation, the chip will get 0xff as ID offset, causing
> the read vendor/device ID to be swapped. I never met such a controller
> so far, but if there is one, it will be a silicon bug that can't be
> fixed by software. To fix this one, we'll have to make a second
> read-id implementation in spi nand framework.
I realize how fragile this ID-based detection is when manufacturers
decide to not follow the standard READID semantic (one 0x9f command byte
followed by 1 or more input cycles encoding the ID). Let's imagine you
have a valid manuf ID byte in ID, and the device ID (ID) matches
the Winbond or Gigadevice manufacturer ID, and ID (extended Device ID
byte?) matches a valid Winbond/Gigadevice device ID. If you skip the
check on ID you might erroneously detect a Winbond or Gigadevice
NAND, while it's actually something else.
Note that I don't really have a solution to make this detection more
> The second problem only exist in theory, so my preference is to apply
> this patch and fix only the first problem for now.
I think we should fix that problem now. Maybe by doing a 3 steps
1/ READID + ID
2/ READID + DUMMY + ID
3/ READID + ADDR + ID
At each step we would check if the returned ID matches a valid NAND,
and if it does, stop there.
More information about the linux-mtd