Markus,
I'm on a DM814x project, and am using NAND geometry just like yours I think. I have not grabbed the latest Arago patches that support sub-page accees, for fear that the bugs aren't flushed out yet. (Those patches went in after the last EZSDK and PSP releases, so I'm not sure how many people are using it.) So we have sub-page access completely disabled. After more that a year on the project, we started to get reports flaky behavior, at least one of which is a NAND issue: UBIFS fails to ubiattach because the ubi_io_read fails with -74, because omap_correct_data() fails, because elm_decode_bch_error() fails, because ECC_CORRECTABLE bit in the ELM_LOCATION_STATUS register doesn't go to 1 - indicating that the 814x's error location process failed. However, the block where this failure occurs is not being detected as a bad block.
Hunting around, I noticed this commit at kernel.org that specifically affects Bad Block Marker checking on devices with 2kB blocks:
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=cdbec0508699e3346052bf098c5c330a711a86a9
It's not a smoking gun, but I'm wondering if a similar change is needed in TI's Arago kernel (used by 814x and 816x) ...
Dan -