On 4/7/25 6:59 AM, Christoph Hellwig wrote:
On Mon, Apr 07, 2025 at 02:46:32PM +0100, Keith Busch wrote:
O_DIRECT only requires each user iovec be a multiple of the logical
block size with the address aligned to the dma_alignment. If the
dma_alignment is smaller than the logical block size, then this could
create bvec segments that are smaller. For nvme where we have 4-byte
dma alignment, you could have the first segment be the last 4 bytes of a
page, then the remaing 508 bytes from a different page in the next
segment.
Oh, right - with a smaller dma alignment this can actually happen.
Some time ago I added src/discontiguous-io.cpp to the blktests project
to trigger this scenario. This test program submits an SG_IO request
with multiple 4 byte segments. Maybe this test program should be
modified such that it uses O_DIRECT instead of SG_IO.
Bart.
