在 2026/8/18 15:06, Yu Kuai 写道:
> blkg lookup currently uses a per-blkcg radix tree keyed by request queue
> ID, plus a lookup hint for the common case. This spreads the queue-local
> blkcg association index across every blkcg and requires radix-tree
> preloading before creating a blkg while holding q->queue_lock.
>
> Replace the radix tree and lookup hint with a request_queue-owned
> rhashtable keyed by the blkcg CSS ID. Cache the ID in each blkg; the blkg
> holds a CSS reference until after it leaves the hash, so the ID cannot be
> reused while it is hash-visible. The integer key also reduces hashing and
> comparison work relative to a pointer-sized key on 64-bit systems.
>
> Keep entries until blkg_release() and provide blkg_lookup_any() for callers
> which need to find dying entries. blkg_lookup() filters offline entries so
> existing lookup semantics remain unchanged.
>
> Keep q->blkg_list for ordered policy and scheduler walks. All current
> walkers are cgroupfs or sysfs slow paths, so they can move to rhashtable
> iteration once the q->queue_lock to q->blkcg_mutex conversion lands.
> Initialize and destroy the hash with request_queue, and remove the
> radix-tree preload paths which are no longer needed.
>
> blkg_release() removes the hash entry only when the blkg was successfully
> inserted into q->blkg_list; the list_empty case covers allocation or
> creation failure before insertion.
Reviewed-by: Tao Cui <[email protected]>
