On 13/02/2024 04:33, Ritesh Harjani (IBM) wrote:
diff --git a/block/blk-settings.c b/block/blk-settings.c
index 11c0361c2313..176f26374abc 100644
--- a/block/blk-settings.c
+++ b/block/blk-settings.c
@@ -108,18 +108,42 @@ void blk_queue_bounce_limit(struct request_queue *q, enum blk_bounce bounce)
}
EXPORT_SYMBOL(blk_queue_bounce_limit);
+
+/*
+ * Returns max guaranteed sectors which we can fit in a bio. For convenience of
+ * users, rounddown_pow_of_two() the return value.
+ *
+ * We always assume that we can fit in at least PAGE_SIZE in a segment, apart
+ * from first and last segments.
+ */
It took sometime to really understand what is special about the first
and the last vector. Looks like what we are discussing here is the
I/O covering a partial page, i.e. the starting offset and the end
boundary might not cover the whole page.
It still isn't very clear that why do we need to consider
queue_logical_block_size(q) and not the PAGE_SIZE for those 2 vectors
(1. given atomic writes starting offset and length has alignment
restrictions?
We are using the direct IO alignment restriction, and that is the iovecs
need to be bdev logical block size aligned - please see
bdev_iter_is_aligned().
We are also supporting a single iovec currently. As such, the middle
bvecs will always contain at least PAGE_SIZE, and the first/last must
have at least LBS data.
Note that we will want to support atomic writes in future for buffered
IO, but it would be sensible to keep this direct IO alignment
restriction there as well.
Let me know if this needs to be made clearer in the code/commit message.
Thanks,
John
