osd_do_bio() fills each bio until the disk blocks stop being contiguous or bio_add_folio() refuses the next page. Where a full bio ends then depends on how many of the I/O buffer pages merge into multi-page bvecs, not on the RAID geometry, so on RAID OSTs it usually ends in the middle of a row (full stripe). The RAID layer receives that row in two bios and has to take its partial-stripe path for it: read-modify-write, or a slower stripe-cache path instead of a full-row write.

Seen on md RAID6 8+2 OSTs with a 128 KiB chunk (1 MiB rows) under 4 MiB O_DIRECT writes: about half of the requests reached md as 2-4 bios cut mid-row, and 23% of rows were written as partial stripes. Tracing block_bio_queue on the md device shows bios ending at arbitrary offsets within a row.

The ldiskfs stripe (mount -o stripe=, or the superblock stripe width/stride set by mkfs -E stride=,stripe_width=) already gives the row size. Proposal: for writes, at each row boundary end the bio when the next row (or the rest of the I/O) will not fit in the bvecs left in it, counting the bvecs from the row's pages so bios still span several rows when pages merge. Reads and OSTs without a stripe are unchanged.

With the patch on the site above, 100% of rows were written as full stripes, and client writes improved by 26% (2 OSTs on one server) and 7% (4 OSTs on two servers).

Related:

• LU-20564 (mkfs.lustre probes md geometry): sets the stripe width this change relies on, for OSTs where mke2fs cannot derive it. That aligns block allocation to rows; this change aligns the bios that write them.
• LU-20199 (higher-order I/O buffers): reduces bio fragmentation from the bvec limit. Independent of this change, which keeps bios row-aligned at any bio size.