[PATCHv3 0/2] block+nvme: reducing virtual boundary mask reliance

[Keith Busch <kbusch@meta.com>]

From: Keith Busch <kbusch@kernel.org>

Previous version is here:

  https://lore.kernel.org/linux-nvme/20250805195608.2379107-1-kbusch@meta.com/

This patch set depends on this unmerged series for flexible direct-io
here:

  https://lore.kernel.org/linux-block/20250819164922.640964-1-kbusch@meta.com/

The purpose of this is to allow optimization decisions to happen per IO.
The virtual boundary that NVMe reports provides specific guarantees
about the data alignment, but that might not be large enough for some
CPU architectures to take advantage of even iif an applications uses
aligned data buffers that could use it.

At the same time, the virtual boundary prevents the driver from directly
using memory in ways the hardware may be capable of accessing. This
creates unnecessary needs on applications to double buffer their data
into the more restrictive virtually contiguous format.

This patch series provides an efficient way to track page boundary gaps
per-IO so that the optimizations can be decided per-IO. This provides
flexibility to use all hardware to their abilities beyond what the
virtual boundary mask can provide.

Note, abuse of this capability may result in worse performance compared
to the bounce buffer solutions. Sending a bunch of tiny vectors for one
IO incurs significant protocol overhead, so while this patch set allows
you to do that, I recommend that you don't. We can't enforce a minimum
size though because vectors may straddle pages with only a few words in
the first and/or last pages, which we do need to support.

Changes from v2:

  - We only need to know about the lowest set bit in any bio vector page
    gap. Use that to avoid increasing the bio size

  - Fixed back merges; the previous was potentially missing one of the
    bio's gaps

  - Use pointers instead of relying on the inline to generate good code.

  - Trivial name changes

  - Comments explaing the new bio field, and the nvme usage for deciding
    on SGL vs PRP DMA.

Keith Busch (2):
  block: accumulate segment page gaps per bio
  nvme: remove virtual boundary for sgl capable devices

 block/bio.c               |  1 +
 block/blk-merge.c         | 39 ++++++++++++++++++++++++++++++++++++---
 block/blk-mq-dma.c        |  3 +--
 block/blk-mq.c            | 10 ++++++++++
 drivers/nvme/host/core.c  | 21 ++++++++++++++++-----
 drivers/nvme/host/pci.c   | 16 +++++++++++++---
 include/linux/blk-mq.h    |  2 ++
 include/linux/blk_types.h |  8 ++++++++
 8 files changed, 87 insertions(+), 13 deletions(-)

-- 
2.47.3