Revert "mtd: spi-nor: core: replace dummy buswidth from addr to data"

This reverts commit 98d1fb94ce75f39febd456d6d3cbbe58b6678795. The commit

uses data nbits instead of addr nbits for dummy phase. This causes a regression for all boards where spi-tx-bus-width is smaller than spi-rx-bus-width. It is a common pattern for boards to have spi-tx-bus-width == 1 and spi-rx-bus-width > 1. The regression causes all reads with a dummy phase to become unavailable for such boards, leading to a usually slower 0-dummy-cycle read being selected.

Most controllers' supports_op hooks call spi_mem_default_supports_op(). In spi_mem_default_supports_op(), spi_mem_check_buswidth() is called to check if the buswidths for the op can actually be supported by the board's wiring. This wiring information comes from (among other things) the spi-{tx,rx}-bus-width DT properties. Based on these properties, SPI_TX_* or SPI_RX_* flags are set by of_spi_parse_dt(). spi_mem_check_buswidth() then uses these flags to make the decision whether an op can be supported by the board's wiring (in a way, indirectly checking against spi-{rx,tx}-bus-width).

Now the tricky bit here is that spi_mem_check_buswidth() does:

	if (op->dummy.nbytes &&
	    spi_check_buswidth_req(mem, op->dummy.buswidth, true))
		return false;

The true argument to spi_check_buswidth_req() means the op is treated as a TX op. For a board that has say 1-bit TX and 4-bit RX, a 4-bit dummy TX is considered as unsupported, and the op gets rejected.

The commit being reverted uses the data buswidth for dummy buswidth. So for reads, the RX buswidth gets used for the dummy phase, uncovering this issue. In reality, a dummy phase is neither RX nor TX. As the name suggests, these are just dummy cycles that send or receive no data, and thus don't really need to have any buswidth at all.

Ideally, dummy phases should not be checked against the board's wiring capabilities at all, and should only be sanity-checked for having a sane buswidth value. Since we are now at rc7 and such a change might introduce many unexpected bugs, revert the commit for now. It can be sent out later along with the spi_mem_check_buswidth() fix.

Reported-by Alexander Stein <[email protected]>

Closes https://lore.kernel.org/linux-mtd/3342163.44csPzL39Z@steina-w/ open here

Quoted 1 line · Pratyush Yadav

This reverts commit 98d1fb94ce75f39febd456d6d3cbbe58b6678795. The commit

"The commit" on a new paragraph please

Quoted 40 lines

uses data nbits instead of addr nbits for dummy phase. This causes a regression for all boards where spi-tx-bus-width is smaller than spi-rx-bus-width. It is a common pattern for boards to have spi-tx-bus-width == 1 and spi-rx-bus-width > 1. The regression causes all reads with a dummy phase to become unavailable for such boards, leading to a usually slower 0-dummy-cycle read being selected.

Most controllers' supports_op hooks call spi_mem_default_supports_op(). In spi_mem_default_supports_op(), spi_mem_check_buswidth() is called to check if the buswidths for the op can actually be supported by the board's wiring. This wiring information comes from (among other things) the spi-{tx,rx}-bus-width DT properties. Based on these properties, SPI_TX_* or SPI_RX_* flags are set by of_spi_parse_dt(). spi_mem_check_buswidth() then uses these flags to make the decision whether an op can be supported by the board's wiring (in a way, indirectly checking against spi-{rx,tx}-bus-width).

Now the tricky bit here is that spi_mem_check_buswidth() does:

	if (op->dummy.nbytes &&
	    spi_check_buswidth_req(mem, op->dummy.buswidth, true))
		return false;

The true argument to spi_check_buswidth_req() means the op is treated as a TX op. For a board that has say 1-bit TX and 4-bit RX, a 4-bit dummy TX is considered as unsupported, and the op gets rejected.

The commit being reverted uses the data buswidth for dummy buswidth. So for reads, the RX buswidth gets used for the dummy phase, uncovering this issue. In reality, a dummy phase is neither RX nor TX. As the name suggests, these are just dummy cycles that send or receive no data, and thus don't really need to have any buswidth at all.

Ideally, dummy phases should not be checked against the board's wiring capabilities at all, and should only be sanity-checked for having a sane buswidth value. Since we are now at rc7 and such a change might introduce many unexpected bugs, revert the commit for now. It can be sent out later along with the spi_mem_check_buswidth() fix.

Quoted 2 lines

Reported-by Alexander Stein <[email protected]>

Closes https://lore.kernel.org/linux-mtd/3342163.44csPzL39Z@steina-w/ open here

fixes tag please

with that:

Reviewed-by Tudor Ambarus <[email protected]>

Quoted 1 line

fixes tag please

Hmm, I thought a fixes for a revert might be pointless since we have the commit hash in the message anyway. I don't have a strong opinion, so will add the below when applying:

Quoted 4 lines · Tudor Ambarus

Quoted 1 line · Pratyush Yadav

This reverts commit 98d1fb94ce75f39febd456d6d3cbbe58b6678795. The commit

"The commit" on a new paragraph please

ACK, will do when applying.

This reverts commit 98d1fb94ce75f39febd456d6d3cbbe58b6678795. The commit uses data nbits instead of addr nbits for dummy phase. This causes a regression for all boards where spi-tx-bus-width is smaller than spi-rx-bus-width. It is a common pattern for boards to have spi-tx-bus-width == 1 and spi-rx-bus-width > 1. The regression causes all reads with a dummy phase to become unavailable for such boards, leading to a usually slower 0-dummy-cycle read being selected.

Most controllers' supports_op hooks call spi_mem_default_supports_op(). In spi_mem_default_supports_op(), spi_mem_check_buswidth() is called to check if the buswidths for the op can actually be supported by the board's wiring. This wiring information comes from (among other things) the spi-{tx,rx}-bus-width DT properties. Based on these properties, SPI_TX_* or SPI_RX_* flags are set by of_spi_parse_dt(). spi_mem_check_buswidth() then uses these flags to make the decision whether an op can be supported by the board's wiring (in a way, indirectly checking against spi-{rx,tx}-bus-width).

Now the tricky bit here is that spi_mem_check_buswidth() does:

	if (op->dummy.nbytes &&
	    spi_check_buswidth_req(mem, op->dummy.buswidth, true))
		return false;

The true argument to spi_check_buswidth_req() means the op is treated as a TX op. For a board that has say 1-bit TX and 4-bit RX, a 4-bit dummy TX is considered as unsupported, and the op gets rejected.

The commit being reverted uses the data buswidth for dummy buswidth. So for reads, the RX buswidth gets used for the dummy phase, uncovering this issue. In reality, a dummy phase is neither RX nor TX. As the name suggests, these are just dummy cycles that send or receive no data, and thus don't really need to have any buswidth at all.

Ideally, dummy phases should not be checked against the board's wiring capabilities at all, and should only be sanity-checked for having a sane buswidth value. Since we are now at rc7 and such a change might introduce many unexpected bugs, revert the commit for now. It can be sent out later along with the spi_mem_check_buswidth() fix.

Reported-by Alexander Stein <[email protected]>

Closes https://lore.kernel.org/linux-mtd/3342163.44csPzL39Z@steina-w/ open here

Signed-off-by Pratyush Yadav <[email protected]>

---

I have only compile-tested this patch. Alexander, if you can send your Tested-by it would be great!

Miquel, since the bug was introduced in v6.13-rc1, I want to have this patch go in v6.13. So you would have to send Linus a pull request before he likely releases it this Sunday. How do you want to take the patch? Do you want to apply it directly to mtd/fixes or should I send out a pull request from a spi-nor/fixes branch?

 drivers/mtd/spi-nor/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

drivers/mtd/spi-nor/core.c +1 -1

index 66949d9f0cc5a..b6f374ded390a 100644
@@ -89,7 +89,7 @@ void spi_nor_spimem_setup_op(const struct spi_nor *nor,
 		op->addr.buswidth = spi_nor_get_protocol_addr_nbits(proto);
 
 	if (op->dummy.nbytes)
-		op->dummy.buswidth = spi_nor_get_protocol_data_nbits(proto);
+		op->dummy.buswidth = spi_nor_get_protocol_addr_nbits(proto);
 
 	if (op->data.nbytes)
 		op->data.buswidth = spi_nor_get_protocol_data_nbits(proto);