Hi,
在 2025/7/25 20:03, Jan Kara 写道:
On Fri 25-07-25 19:21:06, Damien Le Moal wrote:
On 7/25/25 16:05, Yu Kuai wrote:
From: Yu Kuai <yukuai3@xxxxxxxxxx>
ioc_lookup_icq() is used by bfq to lookup bfqq from IO path, the helper
have to be protected by queue_lock, which is too heavy. Hence add a new
helper that is lookless, this is safe because both request_queue and ioc
can be pinged by IO that is still issuing.
Signed-off-by: Yu Kuai <yukuai3@xxxxxxxxxx>
---
block/blk-ioc.c | 34 ++++++++++++++++++++++++++++++++++
block/blk.h | 1 +
2 files changed, 35 insertions(+)
diff --git a/block/blk-ioc.c b/block/blk-ioc.c
index ce82770c72ab..4945b48dfdb6 100644
--- a/block/blk-ioc.c
+++ b/block/blk-ioc.c
@@ -343,6 +343,40 @@ struct io_cq *ioc_lookup_icq(struct request_queue *q)
}
EXPORT_SYMBOL(ioc_lookup_icq);
+/**
+ * ioc_lookup_icq_rcu - lookup io_cq from ioc in io path
+ * @q: the associated request_queue
+ *
+ * Look up io_cq associated with @ioc - @q pair from @ioc. Must be called from
+ * io issue path, either return NULL if current issue io to @q for the first
+ * time, or return a valid icq.
+ */
+struct io_cq *ioc_lookup_icq_rcu(struct request_queue *q)
+{
+ struct io_context *ioc = current->io_context;
+ struct io_cq *icq;
+
+ WARN_ON_ONCE(percpu_ref_is_zero(&q->q_usage_counter));
I do not think this is necessary.
This is used to indicate this is from IO issue path, I can remove it.
+
+ if (!ioc)
+ return NULL;
+
+ icq = rcu_dereference(ioc->icq_hint);
+ if (icq && icq->q == q)
+ return icq;
+
+ icq = radix_tree_lookup(&ioc->icq_tree, q->id);
+ if (!icq)
+ return NULL;
+
+ if (WARN_ON_ONCE(icq->q != q))
+ return NULL;
+
+ rcu_assign_pointer(ioc->icq_hint, icq);
+ return icq;
+}
+EXPORT_SYMBOL(ioc_lookup_icq_rcu);
Patch 2 calls this function with the rcu_read_lock() held. Why not move that rcu
read lock here inside this function ? That is how ioc_lookup_icq() was doing
things, with code that is more compact than this.
And since ioc_lookup_icq() was already using RCU, it seems that the only change
you need is to remove the "lockdep_assert_held(&q->queue_lock);" from that
function to endup with the same above functionality. So why all the churn ?
Yes, I agree, just dropping the assert and updating callers should be fine.
Yes, this is much simpler.
Another question is: is it safe to call radix_tree_lookup() without any lock
held ? What if this races with a radix tree insertion ? (I may be wrong here as
I am not familiar with that code).
Yes, radix_tree_lookup() is fine to call with just rcu protection.
The insertion is protected by queue_lock, and look up is fine with rcu
protection.
Thanks,
Kuai
Honza
