[PATCH] RDMA/rtrs-clt: use find_next_zero_bit() for permit allocation

Liu Zhenlong <[email protected]>
Newsgroups org.kernel.vger.linux-rdma,org.kernel.vger.linux-kernel
Message-ID <[email protected]>
__rtrs_get_permit() allocates a free permit from a bitmap under lockless
contention: it scans with find_first_zero_bit() and claims the bit with
test_and_set_bit_lock(), restarting the whole scan when it loses the
race.  Each retry rewinds to bit 0 and re-walks every already-set low
bit before reaching the free region again.

Under high queue depth - the RTRS/RNBD data path - the low part of
permits_map is densely set, so a lost race wastes a scan proportional
to the number of in-use permits on every retry.

Use find_next_zero_bit(), resuming from the last position, so a lost
race continues scanning from where it left off instead of from the
beginning.  When the scan reaches the end, wrap to the beginning to
exhaust the map, so a permit freed below the cursor is still found
and NULL is returned only when the map is actually full, matching the
original behavior.  The scan remains non-atomic, so the
test_and_set_bit_lock() retry is still required and the race
handling is unchanged.

A userspace model of the bitmap-allocation algorithm (not the kernel
find_*_bit primitives) quantifies the mechanism: with qdepth=512 and 14
threads holding ~98% of the bits set, a lost race in the baseline
re-scans the densely-set low region, traversing ~1.3k bit-positions per
allocation, while the patched version resumes and traverses ~260.  The
benefit is contention- and density-dependent: under low contention
(sparse map) the two are equivalent, and the wrap adds a small amount
of code over the single-scan baseline.

End-to-end RNBD/fio throughput was not measured (no RDMA hardware
available); the model isolates the allocation mechanism, not the full
IO path.

Compile-tested: arm64 defconfig + INFINIBAND_RTRS_CLIENT=m, rtrs-clt.o
Assisted-by: Claude:claude-opus-5
Signed-off-by: Liu Zhenlong <[email protected]>
---
 drivers/infiniband/ulp/rtrs/rtrs-clt.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)

diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
index d34d7e5f34d6..a1df90243c41 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
+++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
@@ -70,19 +70,24 @@ __rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
 {
 	size_t max_depth = clt->queue_depth;
 	struct rtrs_permit *permit;
-	int bit;
+	unsigned long bit = 0;
 
 	/*
-	 * Adapted from null_blk get_tag(). Callers from different cpus may
-	 * grab the same bit, since find_first_zero_bit is not atomic.
-	 * But then the test_and_set_bit_lock will fail for all the
-	 * callers but one, so that they will loop again.
-	 * This way an explicit spinlock is not required.
+	 * Callers from different CPUs may grab the same bit, since the bitmap
+	 * scan is not atomic. But then the test_and_set_bit_lock() will fail
+	 * for all the callers but one, so that they loop again. This way an
+	 * explicit spinlock is not required. find_next_zero_bit() resumes
+	 * from the last position so that a lost race does not rescan the
+	 * already-set low bits; if it reaches the end, wrap to the beginning
+	 * to exhaust the map and still find a permit freed below the cursor.
 	 */
 	do {
-		bit = find_first_zero_bit(clt->permits_map, max_depth);
-		if (bit >= max_depth)
-			return NULL;
+		bit = find_next_zero_bit(clt->permits_map, max_depth, bit);
+		if (bit >= max_depth) {
+			bit = find_first_zero_bit(clt->permits_map, max_depth);
+			if (bit >= max_depth)
+				return NULL;
+		}
 	} while (test_and_set_bit_lock(bit, clt->permits_map));
 
 	permit = get_permit(clt, bit);
-- 
2.55.0
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.