[PATCH v3 0/3] blk-cgroup: store blkcg in bio before blkcg_mutex conversion

From: Yu Kuai <[email protected]>

This is v3 of the preparatory series for the blkcg_mutex conversion
proposed in the related series [1]. That conversion moves queue-local
blkg topology synchronization from q->queue_lock to q->blkcg_mutex,
which is awkward while bios directly store queue-local blkg references.

Storing a blkg also ties a bio to one request_queue. A remapped bio must
drop that reference and recreate the association for the new queue even
when no blkcg policy needs it. Dying blkgs add another complication:
they disappear from the per-blkcg lookup tree before all pinned bio
references drain.

Patch 1 makes request_queue the authoritative lookup owner. A
queue-owned rhashtable keyed by the blkcg CSS ID keeps dying blkgs
discoverable until their references drain, while q->blkg_list remains
available for ordered walks.

Patch 2 stores and pins the queue-independent blkcg CSS in each bio.
Queue-local blkgs are created lazily by policy users, then pinned until
the bio changes devices or releases its cgroup state. Completion and
accounting paths which do not create a blkg remain lookup-only.

Patch 3 moves async bio punt state from blkg to blkcg, so punting alone
does not instantiate a queue-local blkg.

Changes since v2:

  - Rebase onto the latest for-7.4/block branch.
  - In patch 2, clarify why dm_bio_record() excludes BIO_BLKG_REF and
    add Nilay Shroff's Reviewed-by.
  - In patch 3, replace scoped spinlock guards with explicit
    spin_lock()/spin_unlock(), drop the redundant teardown check, and
    rely on zero-initialization for async_bios before publication, as
    suggested by Nilay Shroff and Christoph Hellwig.

Changes since v1:

  - Rebase onto the latest block-7.3 branch.
  - In patch 2, require every blkg rhashtable lookup to run inside an
    explicit RCU read-side critical section. Annotate blkg_lookup_any(),
    blkg_lookup(), and recursive wrappers with __must_hold_shared(RCU)
    so Clang thread-safety analysis verifies the contract.
  - In patch 2, exclude BIO_BLKG_REF from the flags saved by
    dm_bio_record(), since restoring an ownership bit cannot recreate
    the associated blkg reference.
  - In patch 3, use scoped spinlock guards for async_bios and annotate
    the list with __guarded_by().
  - Add Nilay Shroff's Reviewed-by to patch 1.

Changes since RFC v3:

  - Drop the RFC prefix.
  - Add Christoph Hellwig's Reviewed-by to patch 1.
  - Add Tao Cui's Reviewed-by to patch 2.
  - In patch 2, relax blkg_lookup_any()/blkg_lookup() to also allow
    q->queue_lock as an alternative to the RCU read lock.

Changes since RFC v2:

  - Drop the old patch 1 and send it separately as a bugfix for 7.3 and
    -stable, as suggested by Christoph Hellwig.
  - Rebase onto the latest for-7.3/block branch.
  - In the new patch 1, note that the remaining q->blkg_list walkers are
    cgroupfs/sysfs slow paths and can move to rhashtable iteration after
    the queue_lock-to-blkcg_mutex conversion.
  - In the new patch 1, explain the list_empty case in blkg_release(), and
    add an RCU lockdep assertion plus documentation that blkg_lookup_any()
    does not acquire a reference.

Changes since RFC v1:

  - Add patch 1 to wait for every old blkg to leave q->blkg_list before a
    shared request_queue is rebound, instead of treating root_blkg == NULL
    as completion of asynchronous blkg teardown.
  - Add patch 2 to replace the per-blkcg radix tree and lookup hint with a
    request_queue rhashtable keyed by blkcg->css.id, as suggested by
    Christoph Hellwig. Keep dying pinned blkgs in the hash until
    blkg_release() so bio-owned references remain discoverable.
  - Fold the v1 helper-only patch into patch 3, as suggested by Jan Kara
    and Christoph Hellwig, and make bio_blkcg() naturally return NULL for
    an unassociated bio.
  - Rework patch 3 so blkg_lookup_create() acquires the bio-owned
    reference, falls back to a live parent when creation or tryget fails,
    and updates bi_blkcg when the returned blkg belongs to an ancestor.
  - Make bio_blkg_lookup() lookup-only: it returns NULL unless BIO_BLKG_REF
    is already set. Use bio_blkg() in the BFQ and IOCOST merge paths which
    may need to create a blkg, while keeping blk_cgroup_bio_start() and
    completion paths lookup-only.
  - Move CSS online-reference handling into
    bio_associate_blkcg_from_css(), including fallback to the root blkcg,
    so bio_associate_blkcg() does not take a redundant reference.
  - Keep the v1 async bio punt conversion as patch 4 and document that
    async_bio_lock protects async_bios.

Previous versions:
  - v2: https://lore.kernel.org/r/[email protected]
  - v1: https://lore.kernel.org/r/[email protected]
  - RFC v3:
    https://lore.kernel.org/r/[email protected]
  - RFC v2:
    https://lore.kernel.org/r/[email protected]
  - RFC v1:
    https://lore.kernel.org/r/[email protected]

Related series:
  [1] RFC v2 blk-cgroup: protect blkgs with blkcg_mutex
      https://lore.kernel.org/r/[email protected]

Yu Kuai (3):
  blk-cgroup: use a request_queue rhashtable for blkg lookup
  blk-cgroup: store blkcg in bio instead of blkg
  blk-cgroup: move async bio punt state to blkcg

 Documentation/admin-guide/cgroup-v2.rst |   2 +-
 block/bfq-cgroup.c                      |  18 +-
 block/bfq-iosched.c                     |  19 +-
 block/bio.c                             |  22 +-
 block/blk-cgroup-fc-appid.c             |  10 +-
 block/blk-cgroup-rwstat.h               |   2 +-
 block/blk-cgroup.c                      | 346 ++++++++++++++----------
 block/blk-cgroup.h                      |  93 +++++--
 block/blk-core.c                        |   9 +-
 block/blk-crypto-fallback.c             |   2 +-
 block/blk-iocost.c                      |  12 +-
 block/blk-iolatency.c                   |  11 +-
 block/blk-ioprio.c                      |   2 +-
 block/blk-throttle.c                    |   3 +-
 block/blk-throttle.h                    |   2 +-
 drivers/md/bcache/request.c             |   2 +-
 drivers/md/dm-bio-record.h              |   5 +-
 drivers/md/dm.c                         |   2 +-
 drivers/md/md.c                         |   2 +-
 fs/gfs2/lops.c                          |   3 +-
 include/linux/bio.h                     |  26 +-
 include/linux/blk_types.h               |   9 +-
 include/linux/blkdev.h                  |   2 +
 include/linux/writeback.h               |   2 +-
 mm/page_io.c                            |  12 +-
 25 files changed, 381 insertions(+), 237 deletions(-)


base-commit: af78ca007be5f63dc34b8180c8063c1113697192
-- 
2.51.0