Endpoint lookup in the receive path acts as a demultiplexer routing
incoming messages to the appropriate endpoint. This is a read-heavy
operation (frequent message receives) with infrequent writes
(endpoint creation/destruction). Since idr_find() is safe under RCU
read-side protection, RCU can be used to optimize this path.

Convert endpoint lookup to use RCU:
- Read path: Use rcu_read_lock/unlock for lockless lookup
- Destroy path: Add synchronize_rcu() after endpoint removal

This reduces lock contention in the hot receive path.

RCU safety note:
When idr_alloc() returns, the endpoint becomes immediately visible to
idr_find(), but ept->addr might not yet be set. This creates a theoretical
window where RX could find an endpoint with uninitialized addr.

This is safe because:
1) When endpoints are created via rpmsg core callbacks, initialization
   completes before announce_create() is sent. Remote processors only
   send messages after receiving the announcement.
2) For manually created endpoints, drivers control timing and typically
   do not announce until ready.

Thus, messages only arrive after ept->addr is initialized, making this
RCU optimization safe.

No functional change except reduced contention.

Signed-off-by: Zhongqiu Han <[email protected]>
---
 drivers/rpmsg/virtio_rpmsg_bus.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/drivers/rpmsg/virtio_rpmsg_bus.c b/drivers/rpmsg/virtio_rpmsg_bus.c
index 79d983055b4d..4cbb8a8aaec5 100644
--- a/drivers/rpmsg/virtio_rpmsg_bus.c
+++ b/drivers/rpmsg/virtio_rpmsg_bus.c
@@ -17,6 +17,7 @@
 #include <linux/kernel.h>
 #include <linux/module.h>
 #include <linux/mutex.h>
+#include <linux/rcupdate.h>
 #include <linux/rpmsg.h>
 #include <linux/rpmsg/byteorder.h>
 #include <linux/rpmsg/ns.h>
@@ -297,6 +298,12 @@ __rpmsg_destroy_ept(struct virtproc_info *vrp, struct 
rpmsg_endpoint *ept)
        idr_remove(&vrp->endpoints, ept->addr);
        mutex_unlock(&vrp->endpoints_lock);
 
+       /*
+        * Wait for any ongoing RCU read-side critical sections to complete.
+        * This ensures no one is accessing the endpoint after removal.
+        */
+       synchronize_rcu();
+
        /* make sure in-flight inbound messages won't invoke cb anymore */
        mutex_lock(&ept->cb_lock);
        ept->cb = NULL;
@@ -680,7 +687,7 @@ static int rpmsg_recv_single(struct virtproc_info *vrp, 
struct device *dev,
        }
 
        /* use the dst addr to fetch the callback of the appropriate user */
-       mutex_lock(&vrp->endpoints_lock);
+       rcu_read_lock();
 
        ept = idr_find(&vrp->endpoints, __rpmsg32_to_cpu(little_endian, 
msg->dst));
 
@@ -688,7 +695,7 @@ static int rpmsg_recv_single(struct virtproc_info *vrp, 
struct device *dev,
        if (ept)
                kref_get(&ept->refcount);
 
-       mutex_unlock(&vrp->endpoints_lock);
+       rcu_read_unlock();
 
        if (ept) {
                /* make sure ept->cb doesn't go away while we use it */
-- 
2.43.0


Reply via email to