Re: [PATCH v4 8/9] blk-merge: Ensure bios aren't split in middle of a crypto data unit
From: Eric Biggers <ebiggers@kernel.org>
Date: 2021-07-23 18:11:49
Also in:
lkml
On Tue, Jul 06, 2021 at 10:29:42PM -0700, Satya Tangirala wrote:
quoted hunk ↗ jump to hunk
From: Satya Tangirala <redacted> Update get_max_io_size() to return a value aligned to bio_required_sector_alignment(). With this change, and the changes to blk_ksm_register() that restrict the supported data unit sizes based on the queue's limits, blk_bio_segment_split() won't split bios in the middle of a data unit anymore Signed-off-by: Satya Tangirala <redacted> --- block/blk-merge.c | 49 ++++++++++++++++++++++++++++++----------------- 1 file changed, 31 insertions(+), 18 deletions(-)diff --git a/block/blk-merge.c b/block/blk-merge.c index a11b3b53717e..68c96dec5680 100644 --- a/block/blk-merge.c +++ b/block/blk-merge.c@@ -135,27 +135,39 @@ static struct bio *blk_bio_write_same_split(struct request_queue *q, /* * Return the maximum number of sectors from the start of a bio that may be - * submitted as a single request to a block device. If enough sectors remain, - * align the end to the physical block size. Otherwise align the end to the - * logical block size. This approach minimizes the number of non-aligned - * requests that are submitted to a block device if the start of a bio is not - * aligned to a physical block boundary. + * submitted as a single request to a block device. Tries to align the end to + * the physical block size, while also aligning the returned number of sectors + * to bio_required_sector_alignment(). This approach minimizes the number of + * non-aligned requests that are submitted to a block device if the start of a + * bio is not aligned to a physical block boundary. + * + * More clearly, there are two conditions we're interested in satisfying. + * + * Condition 1) We absolutely must have @return divisible by the + * bio_required_sector_alignment(bio). + * + * Condition 2) *If possible*, while still satisfying Condition 1, we would like + * to have start_offset + @return divisible by physical block size in sectors + * (pbs). */
This comment is basically saying the same thing in two different ways. Could we just say it once clearly? Maybe: /* * Return the maximum number of sectors from the start of a bio that may be * submitted as a single request to a block device. The returned value will be * a multiple of bio_required_sector_alignment(). If possible, the returned * value will also make the request end on a physical block boundary; this * minimizes the number of non-aligned requests that are submitted to a block * device when a bio doesn't start on a physical block boundary. */
static inline unsigned get_max_io_size(struct request_queue *q,
struct bio *bio)
{
- unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0);
- unsigned max_sectors = sectors;
- unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
- unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
- unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
-
- max_sectors += start_offset;
- max_sectors &= ~(pbs - 1);
- if (max_sectors > start_offset)
- return max_sectors - start_offset;
-
- return sectors & ~(lbs - 1);
+ unsigned int start_offset = bio->bi_iter.bi_sector;
+ unsigned int sectors = blk_max_size_offset(q, start_offset, 0);
+ unsigned int pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
+ unsigned int req_sector_align = bio_required_sector_alignment(bio);
+ unsigned int pbs_aligned_sector = round_down(start_offset + sectors, pbs);
+
+ /*
+ * If we can return a pbs aligned endpoint while satisfying Condition 1,
+ * then do so.
+ */
+ if (pbs_aligned_sector > start_offset &&
+ IS_ALIGNED(pbs_aligned_sector - start_offset, req_sector_align))
+ return pbs_aligned_sector - start_offset;
+
+ return round_down(sectors, req_sector_align);
}'start_offset' and 'pbs_aligned_sector' need to be 'sector_t', not 'unsigned int'. Also 'start_offset' probably should be called 'start_sector'. And 'req_sector_align' could be shortened to just 'align'. - Eric