From 2efbe2828bc507896248598ffb1026653cbfed2c Mon Sep 17 00:00:00 2001
From: riastradh <campbell+github@mumble.net>
Date: Mon, 23 Jan 2023 15:53:58 +0000
Subject: [PATCH] Use store-release, load-consume operations. (#128)

Co-authored-by: Taylor R Campbell <campbell+npf@mumble.net>
---
 src/kern/npf_conf.c        | 11 +++++------
 src/kern/npf_conn.c        |  9 ++++-----
 src/kern/npf_ifaddr.c      |  4 +++-
 src/kern/stand/npf_stand.h |  2 ++
 4 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/src/kern/npf_conf.c b/src/kern/npf_conf.c
index db3ea79..6158c65 100644
--- a/src/kern/npf_conf.c
+++ b/src/kern/npf_conf.c
@@ -131,8 +131,7 @@ npf_config_load(npf_t *npf, npf_config_t *nc, npf_conndb_t *conns, bool flush)
 	/*
 	 * Set the new config and release the lock.
 	 */
-	membar_sync();
-	atomic_store_relaxed(&npf->config, nc);
+	atomic_store_release(&npf->config, nc);
 	if (onc == NULL) {
 		/* Initial load, done. */
 		npf_ifmap_flush(npf);
@@ -225,7 +224,7 @@ npf_config_read_exit(npf_t *npf, int s)
 npf_ruleset_t *
 npf_config_ruleset(npf_t *npf)
 {
-	npf_config_t *config = atomic_load_relaxed(&npf->config);
+	npf_config_t *config = atomic_load_consume(&npf->config);
 	KASSERT(npf_config_locked_p(npf) || npf_ebr_incrit_p(npf->ebr));
 	return config->ruleset;
 }
@@ -233,7 +232,7 @@ npf_config_ruleset(npf_t *npf)
 npf_ruleset_t *
 npf_config_natset(npf_t *npf)
 {
-	npf_config_t *config = atomic_load_relaxed(&npf->config);
+	npf_config_t *config = atomic_load_consume(&npf->config);
 	KASSERT(npf_config_locked_p(npf) || npf_ebr_incrit_p(npf->ebr));
 	return config->nat_ruleset;
 }
@@ -241,7 +240,7 @@ npf_config_natset(npf_t *npf)
 npf_tableset_t *
 npf_config_tableset(npf_t *npf)
 {
-	npf_config_t *config = atomic_load_relaxed(&npf->config);
+	npf_config_t *config = atomic_load_consume(&npf->config);
 	KASSERT(npf_config_locked_p(npf) || npf_ebr_incrit_p(npf->ebr));
 	return config->tableset;
 }
@@ -249,7 +248,7 @@ npf_config_tableset(npf_t *npf)
 bool
 npf_default_pass(npf_t *npf)
 {
-	npf_config_t *config = atomic_load_relaxed(&npf->config);
+	npf_config_t *config = atomic_load_consume(&npf->config);
 	KASSERT(npf_config_locked_p(npf) || npf_ebr_incrit_p(npf->ebr));
 	return config->default_pass;
 }
diff --git a/src/kern/npf_conn.c b/src/kern/npf_conn.c
index 791a66c..585f6fd 100644
--- a/src/kern/npf_conn.c
+++ b/src/kern/npf_conn.c
@@ -199,8 +199,7 @@ npf_conn_load(npf_t *npf, npf_conndb_t *ndb, bool track)
 		KASSERT(atomic_load_relaxed(&npf->conn_tracking)
 		    == CONN_TRACKING_OFF);
 		odb = atomic_load_relaxed(&npf->conn_db);
-		membar_sync();
-		atomic_store_relaxed(&npf->conn_db, ndb);
+		atomic_store_release(&npf->conn_db, ndb);
 	}
 	if (track) {
 		/* After this point lookups start flying in. */
@@ -470,7 +469,7 @@ npf_conn_establish(npf_cache_t *npc, const unsigned di, bool global)
 	 * the connection later.
 	 */
 	mutex_enter(&con->c_lock);
-	conn_db = atomic_load_relaxed(&npf->conn_db);
+	conn_db = atomic_load_consume(&npf->conn_db);
 	if (!npf_conndb_insert(conn_db, fw, con, NPF_FLOW_FORW)) {
 		error = EISCONN;
 		goto err;
@@ -575,7 +574,7 @@ npf_conn_setnat(const npf_cache_t *npc, npf_conn_t *con,
 	}
 
 	/* Remove the "backwards" key. */
-	conn_db = atomic_load_relaxed(&npf->conn_db);
+	conn_db = atomic_load_consume(&npf->conn_db);
 	bk = npf_conn_getbackkey(con, con->c_alen);
 	ret = npf_conndb_remove(conn_db, bk);
 	KASSERT(ret == con);
@@ -740,7 +739,7 @@ npf_conn_remove(npf_conndb_t *cd, npf_conn_t *con)
 void
 npf_conn_worker(npf_t *npf)
 {
-	npf_conndb_t *conn_db = atomic_load_relaxed(&npf->conn_db);
+	npf_conndb_t *conn_db = atomic_load_consume(&npf->conn_db);
 	npf_conndb_gc(npf, conn_db, false, true);
 }
 
diff --git a/src/kern/npf_ifaddr.c b/src/kern/npf_ifaddr.c
index 57720fd..cd94306 100644
--- a/src/kern/npf_ifaddr.c
+++ b/src/kern/npf_ifaddr.c
@@ -65,9 +65,11 @@ lookup_ifnet_table(npf_t *npf, ifnet_t *ifp)
 static void
 replace_ifnet_table(npf_t *npf, npf_table_t *newt)
 {
-	npf_tableset_t *ts = npf->config->tableset;
+	npf_tableset_t *ts = atomic_load_relaxed(&npf->config)->tableset;
 	npf_table_t *oldt;
 
+	KASSERT(npf_config_locked_p(npf));
+
 	KERNEL_UNLOCK_ONE(NULL);
 
 	/*
diff --git a/src/kern/stand/npf_stand.h b/src/kern/stand/npf_stand.h
index 9538530..c43900b 100644
--- a/src/kern/stand/npf_stand.h
+++ b/src/kern/stand/npf_stand.h
@@ -154,6 +154,8 @@ npfkern_atomic_swap_ptr(volatile void *ptr, void *newval)
 
 #define	atomic_load_acquire(x)		\
     atomic_load_explicit((x), memory_order_acquire)
+#define	atomic_load_consume(x)		\
+    atomic_load_explicit((x), memory_order_consume)
 #define	atomic_store_release(x, y)	\
     atomic_store_explicit((x), (y), memory_order_release)