Fix deadlock over hsk->protect_count in homa_rpc_reap

johnousterhout · johnousterhout · commit ad2acbd77925 · 2025-09-10T16:27:54.000-07:00
Must return from homa_rpc_reap if hsk-&gt;protect_count &gt; 0, even
if reap_all is specified. Otherwise self-deadlock can occur, where
homa_timer protects a socket, then gets interrupted by
homa_softirq, which calls homa_rpc_reap with reap_all.
diff --git a/homa_rpc.c b/homa_rpc.c
@@ -502,8 +502,6 @@ int homa_rpc_reap(struct homa_sock *hsk, bool reap_all)
 				   hsk->port, atomic_read(&hsk->protect_count),
 				   hsk->dead_skbs);
 			homa_sock_unlock(hsk);
-			if (reap_all)
-				continue;
 			return 0;
 		}
 
diff --git a/test/unit_homa_rpc.c b/test/unit_homa_rpc.c
@@ -14,34 +14,6 @@
 #define n(x) htons(x)
 #define N(x) htonl(x)
 
-static struct homa_sock *hook_hsk;
-void unprotect_hsk_hook(char *id)
-{
-	if (strcmp(id, "unlock") != 0)
-		return;
-	if (hook_hsk) {
-		homa_unprotect_rpcs(hook_hsk);
-		hook_hsk = NULL;
-	}
-}
-
-#if 0
-static struct homa_rpc *hook_rpc;
-static int hook_count;
-static void unlink_rpc_hook(char *id)
-{
-	if (strcmp(id, "spin_lock")!= 0)
-		return;
-	if (hook_count == 0)
-		return;
-	hook_count--;
-	if (hook_count == 0) {
-		list_del_init(&hook_rpc->ready_links);
-		homa_rpc_put(hook_rpc);
-	}
-}
-#endif
-
 FIXTURE(homa_rpc) {
 	struct in6_addr client_ip[1];
 	int client_port;
@@ -606,22 +578,6 @@ TEST_F(homa_rpc, homa_rpc_reap__protected)
 	homa_unprotect_rpcs(&self->hsk);
 	EXPECT_STREQ("", unit_log_get());
 }
-TEST_F(homa_rpc, homa_rpc_reap__protected_and_reap_all)
-{
-	struct homa_rpc *crpc1 = unit_client_rpc(&self->hsk,
-			UNIT_RCVD_ONE_PKT, self->client_ip, self->server_ip,
-			self->server_port, self->client_id, 5000, 2000);
-
-	ASSERT_NE(NULL, crpc1);
-	homa_rpc_end(crpc1);
-	unit_log_clear();
-	homa_protect_rpcs(&self->hsk);
-	hook_hsk = &self->hsk;
-	unit_hook_register(unprotect_hsk_hook);
-	EXPECT_EQ(0, homa_rpc_reap(&self->hsk, true));
-	EXPECT_STREQ("reaped 1234", unit_log_get());
-	EXPECT_EQ(0, self->hsk.dead_skbs);
-}
 TEST_F(homa_rpc, homa_rpc_reap__skip_rpc_because_locked)
 {
 	struct homa_rpc *crpc1 = unit_client_rpc(&self->hsk,

Original file line number	Diff line number	Diff line change
`@@ -502,8 +502,6 @@ int homa_rpc_reap(struct homa_sock *hsk, bool reap_all)`
`502`	`502`	`hsk->port, atomic_read(&hsk->protect_count),`
`503`	`503`	`hsk->dead_skbs);`
`504`	`504`	`homa_sock_unlock(hsk);`
`505`		`- if (reap_all)`
`506`		`- continue;`
`507`	`505`	`return 0;`
`508`	`506`	`}`
`509`	`507`