On 03/06/2024 13:31, Hannes Reinecke wrote:
It seems ok in principle - we would just need to ensure that it is
watertight.
We currently use chunk_sectors for quite some different things, most
notably zones boundaries, NIOIB, raid stripes etc.
So I don't have an issue adding another use-case for it.
Q2: If we don't, shouldn't we align the atomic write boundary to the
chunk_sectors setting to ensure both match up?
Yeah, right. But we can only handle what HW tells.
The atomic write boundary is only relevant to NVMe. NVMe NOIOB - which
we use to set chunk_sectors - is an IO optimization hint, AFAIK.
However the atomic write boundary is a hard limit. So if NOIOB is not
aligned with the atomic write boundary - which seems unlikely - then
the atomic write boundary takes priority.
Which is what I said; we need to check. And I would treat a NOIOB value
not aligned to the atomic write boundary as an error.
Yeah, maybe we can reject that in blk_validate_limits(), by error'ing or
disabling atomic writes there.
But the real issue here is that the atomic write boundary only matters
for requests, and not for the entire queue.
So using chunk_sectors is out of question as this would affect all
requests, and my comment was actually wrong.
I'll retract it.
I think that some of the logic could be re-used.
rq_straddles_atomic_write_boundary() is checked in merging of reqs/bios
(to see if the resultant req straddles a boundary).
So instead of saying: "will the resultant req straddle a boundary",
re-using path like blk_rq_get_max_sectors() -> blk_chunk_sectors_left(),
we check "is there space within the boundary limit to add this req/bio".
We need to take care of front and back merges, though.
Thanks,
John
