[Devel] [PATCH RHEL9 COMMIT] fuse: pcs: new rpc affinity mode - RSS

[Konstantin Khorenko]

The commit is pushed to "branch-rh9-5.14.0-284.25.1.vz9.30.x-ovz" and will appear at https://src.openvz.org/scm/ovz/vzkernel.git
after rh9-5.14.0-284.25.1.vz9.30.8
------>
commit a00af8d0e5837dabe0eabab9fce5fe382da811e8
Author: Alexey Kuznetsov <kuznet at virtuozzo.com>
Date:   Fri Oct 6 18:44:08 2023 +0800

    fuse: pcs: new rpc affinity mode - RSS
    
    The mode aligns socket io jobs to RSS, receive/transmit jobs
    are scheduled at cpus which is mapped by RSS from rpc socket.
    Precondition is multiqueue device with enabled RSS and XPS.
    
    If RSS and XPS are enabled, sockets are entirely localized
    to one cpu, they are not accessed from other cpus, which
    minimizes lock contention and keep perfect cache locality
    for socket data. Nevertheless, we have to pay, data, prepared
    by user at some cpu is forced to be accessed from rss cpu,
    which can be a problem, especially with NUMA. This has to
    be measured.
    
    When it does not work:
    1. Single queue devices. It is not recommended to use such dumb
       hardware with vstorage.
    2. RDMA. Yet not. It requires some experimenting to figure out
       how rdma completion jobs are scheduled.
    3. IPsec without binding CSes to separate IP addresses.
    
    https://pmc.acronis.work/browse/VSTOR-54040
    
    Signed-off-by: Alexey Kuznetsov <kuznet at acronis.com>
    
    Feature: vStorage
---
 fs/fuse/kio/pcs/pcs_rpc.c     | 8 ++++++--
 fs/fuse/kio/pcs/pcs_rpc.h     | 3 +++
 fs/fuse/kio/pcs/pcs_sock_io.c | 4 +++-
 3 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/fs/fuse/kio/pcs/pcs_rpc.c b/fs/fuse/kio/pcs/pcs_rpc.c
index 3aca05fd6c20..3b763cea9c01 100644
--- a/fs/fuse/kio/pcs/pcs_rpc.c
+++ b/fs/fuse/kio/pcs/pcs_rpc.c
@@ -36,7 +36,7 @@
 #include "fuse_ktrace.h"
 
 
-static unsigned int rpc_affinity_mode = RPC_AFFINITY_RETENT;
+unsigned int rpc_affinity_mode = RPC_AFFINITY_RETENT;
 module_param(rpc_affinity_mode, uint, 0644);
 MODULE_PARM_DESC(rpc_affinity_mode, "RPC affinity mode");
 
@@ -731,6 +731,10 @@ static void pcs_rpc_affinity(struct pcs_rpc *ep, bool was_idle)
 				ep->cpu = WORK_CPU_UNBOUND;
 			}
 			break;
+		case RPC_AFFINITY_RSS:
+			if (!(ep->flags & PCS_RPC_F_LOCAL) && ep->addr.type != PCS_ADDRTYPE_RDMA)
+				break;
+			fallthrough;
 		case RPC_AFFINITY_RETENT:
 			/* Naive socket-to-cpu binding approach */
 			if (time_is_before_jiffies(ep->cpu_stamp) && was_idle) {
@@ -745,7 +749,7 @@ static void pcs_rpc_affinity(struct pcs_rpc *ep, bool was_idle)
 			}
 			break;
 		default:
-			pr_err("Unknown affninity mode: %u\n", rpc_affinity_mode);
+			pr_err("Unknown affinity mode: %u\n", rpc_affinity_mode);
 	}
 }
 
diff --git a/fs/fuse/kio/pcs/pcs_rpc.h b/fs/fuse/kio/pcs/pcs_rpc.h
index 2ff4494a7cd4..ef4ab26b9d44 100644
--- a/fs/fuse/kio/pcs/pcs_rpc.h
+++ b/fs/fuse/kio/pcs/pcs_rpc.h
@@ -39,8 +39,11 @@ enum {
 	RPC_AFFINITY_NONE   = 0,
 	RPC_AFFINITY_RETENT = 1,
 	RPC_AFFINITY_SPREAD = 2,
+	RPC_AFFINITY_RSS    = 3,
 };
 
+extern unsigned int rpc_affinity_mode;
+
 struct pcs_rpc_params
 {
 	unsigned int	alloc_hdr_size;
diff --git a/fs/fuse/kio/pcs/pcs_sock_io.c b/fs/fuse/kio/pcs/pcs_sock_io.c
index fc76e326bef5..46de1023c876 100644
--- a/fs/fuse/kio/pcs/pcs_sock_io.c
+++ b/fs/fuse/kio/pcs/pcs_sock_io.c
@@ -548,7 +548,9 @@ static void pcs_sk_kick_queue(struct sock *sk)
 	sio = rcu_dereference_sk_user_data(sk);
 	if (sio) {
 		struct pcs_rpc *ep = sio->netio.parent;
-		TRACE(PEER_FMT" queue\n", PEER_ARGS(ep));
+		TRACE(PEER_FMT" queue cpu=%d\n", PEER_ARGS(ep), smp_processor_id());
+		if (rpc_affinity_mode == RPC_AFFINITY_RSS && !(ep->flags & PCS_RPC_F_LOCAL))
+			ep->cpu = smp_processor_id();
 		pcs_rpc_kick_queue(ep);
 	}
 	rcu_read_unlock();