It looks like if we add these to the end, we don't break binary compat...
But does camcontrol even know about mmc_data yet?

Well, if you don't zero those fields when passing structure to the kernel, these fields end up being filled with the random garbage, and this changes the behavior of the SDHC driver because now it thinks that the upper layer requested the custom block size.
And this happens regardless of the position of these fields in the struct.
camcontrol does know about mmc_data, see the code in this change:-)

Well, given there is a memset(,0,) above it you don't need the extra assignments, do you?

On a side-note; that malloc can fail...

The softc mmcdata is allocated with M_ZERO; if it is not used anywhere else than here block_count and block_size are never touched so wouldn't need to be set to 0.
In general, if needed, I'd prefer an memset(,0,) on the structure as that way it'll not repeat the problem with other struct fields in the future?

These all are independent on the change kind-of and only needed for using the change later; really nothing to do with block_count and block_size. Can you commit them on their own please?

But we bzero the CCB, which is larger than mmc_data so it won't be garbage...
And you could also add a flag MMC_DATA_BLOCK_SIZE that needs to be set for these values to be used.

that memset() is for mmc_data variable, not for mmc_d. But I will better do another memset(,0,) for mmc_d as well.

Added a check for that.

That mmcdata ends up being used in each sddastart(), so we _do_ need to initialize the fields. But you are right, I'd rather add memset(,0,).

CCB contains struct mmc_command and mmc_command contains a _pointer_ to mmc_data, not the actual mmc_data. So it will be garbaged unless initialized properly.

Done

In theory, we should retry at some other level... In practice, this will never happen and it's likely super hard to do that retry right anyway with any kind of meaningful testing... Other periph drivers do this so it's fine.

so you have this flag, but don't use it.
And you've not moved the new elements to the end.

Currently there is no driver that can use this flag. mmc_da is not one of them.

Moving elements -- ack.