Not necessarily: an offset within the same page is acceptable.
Aren't you just guessing there? Those kunmap_atomics in at91_mci.c
may look wrong to you, but they're not incorrect (so long as sg->offset
falls within the page, as it must do here to make sense of the page).
Especially not on ARM, where kunmap_atomic actually has no interest
in the argument passed. And the oops was in the flush_dcache_page.
If you actually reproduced Nicolas' problem on ARM, and verified
that your patch then fixes it, please let us know: that will be
I believe I posted the correct fix last night (or at least a safe fix
for now: Christoph Lameter may prefer to undo it and change ARM's
dma_mapping at leisure later), checking PageSlab in page_mapping;
and Linus has already put that one into his git tree.
But it would be good to hear from Nicolas whether that indeed fixes
his oops: I couldn't actually try my patch on ARM either.