Re: [PATCH v4 06/11] mtd: intel-dg: align 64bit read and write

[David Laight <david.laight.linux@gmail.com>]

On Wed,  1 Jan 2025 17:39:20 +0200
Alexander Usyskin <alexander.usyskin@intel.com> wrote:

> GSC NVM controller HW errors on quad access overlapping 1K border.
> Align 64bit read and write to avoid readq/writeq over 1K border.
> 
> Acked-by: Miquel Raynal <miquel.raynal@bootlin.com>
> Signed-off-by: Alexander Usyskin <alexander.usyskin@intel.com>
> ---
>  drivers/mtd/devices/mtd-intel-dg.c | 35 ++++++++++++++++++++++++++++++
>  1 file changed, 35 insertions(+)
> 
> diff --git a/drivers/mtd/devices/mtd-intel-dg.c b/drivers/mtd/devices/mtd-intel-dg.c
> index 76ef7198fff8..230bf444b7fe 100644
> --- a/drivers/mtd/devices/mtd-intel-dg.c
> +++ b/drivers/mtd/devices/mtd-intel-dg.c
> @@ -238,6 +238,24 @@ static ssize_t idg_write(struct intel_dg_nvm *nvm, u8 region,
>  		len_s -= to_shift;
>  	}
>  
> +	if (!IS_ALIGNED(to, sizeof(u64)) &&
> +	    ((to ^ (to + len_s)) & GENMASK(31, 10))) {

I'm sure that should be (to + len_s - 1).
Using GENMASK(31, 10) completely fails to indicate what is being tested.

> +		/*
> +		 * Workaround reads/writes across 1k-aligned addresses
> +		 * (start u32 before 1k, end u32 after)
> +		 * as this fails on hardware.
> +		 */
> +		u32 data;
> +
> +		memcpy(&data, &buf[0], sizeof(u32));

	get_unaligned_u32()

> +		idg_nvm_write32(nvm, to, data);
> +		if (idg_nvm_error(nvm))
> +			return -EIO;
> +		buf += sizeof(u32);
> +		to += sizeof(u32);
> +		len_s -= sizeof(u32);
> +	}

It isn't at all obvious why copying 4 bytes helps.
Indeed, if 'to' is 1023 and 'len_s' is 2 it goes terribly wrong.

	David

> +
>  	len8 = ALIGN_DOWN(len_s, sizeof(u64));
>  	for (i = 0; i < len8; i += sizeof(u64)) {
>  		u64 data;
> @@ -295,6 +313,23 @@ static ssize_t idg_read(struct intel_dg_nvm *nvm, u8 region,
>  		from += from_shift;
>  	}
>  
> +	if (!IS_ALIGNED(from, sizeof(u64)) &&
> +	    ((from ^ (from + len_s)) & GENMASK(31, 10))) {
> +		/*
> +		 * Workaround reads/writes across 1k-aligned addresses
> +		 * (start u32 before 1k, end u32 after)
> +		 * as this fails on hardware.
> +		 */
> +		u32 data = idg_nvm_read32(nvm, from);
> +
> +		if (idg_nvm_error(nvm))
> +			return -EIO;
> +		memcpy(&buf[0], &data, sizeof(data));
> +		len_s -= sizeof(u32);
> +		buf += sizeof(u32);
> +		from += sizeof(u32);
> +	}
> +
>  	len8 = ALIGN_DOWN(len_s, sizeof(u64));
>  	for (i = 0; i < len8; i += sizeof(u64)) {
>  		u64 data = idg_nvm_read64(nvm, from + i);