Sorry, not familiar with this project.
The scratch buffer method I originally developed for the F4 SPL addressed the unaligned memory issue when using DMA, which needs 32-bit alignment.
It would decompose read/writes into single sector accesses via an aligned buffer, either on the stack or statically allocated.
If you do f_read/f_write with large requests that are sector multiples, sector aligned, and with 32-bit aligned source/dest memory the scratch buffer won't be needed.
Where things get ugly is if you do a lot of small or spanning writes which require FatFs to shuffle and deblock, using portions of your original buffer at a byte level.
SDIO is going to be woefully inefficient with non-sector aligned file IO, you should buffer 8KB or up to 32KB and manage/wrap your output functionality to optimize writing in multiple sectors, on sector or cluster boundaries within the file.
If you're using DMA, and randomly sized f_read/f_write, you're going to want to "double buffer" rather than "transfer in place"
You could add an assert() to the routines in DISKIO layer to confirm this as a potential failure mechanism.
ST's implementation for POLLED transfers is also VERY WEAK, rather than optimize the normative case, it cripples all transfers to handle the potentially unaligned ones.
