eal: fix memory ordering around lcore task accesses

Ensure that the memory operations before the call to
rte_eal_remote_launch are visible to the worker thread.
Use the function pointer to execute in worker thread
as the guard variable.

Ensure that the memory operations in worker thread, that happen
before it returns the status of the assigned function, are
visible to the main thread. Use the variable containing the
lcore's state as the guard variable.

Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Ola Liljedahl <ola.liljedahl@arm.com>
Reviewed-by: Feifei Wang <feifei.wang2@arm.com>

diff --git a/lib/eal/common/eal_common_launch.c b/lib/eal/common/eal_common_launch.c
index 78fd9402677151413df5b69c6709214d70d8a1d9..e95dadffb343e289323f60a38fe95549a4786c03 100644 (file)
--- a/lib/eal/common/eal_common_launch.c
+++ b/lib/eal/common/eal_common_launch.c
@@ -23,14 +23,10 @@
 int
 rte_eal_wait_lcore(unsigned worker_id)
 {
-       if (lcore_config[worker_id].state == WAIT)
-               return 0;
-
-       while (lcore_config[worker_id].state != WAIT)
+       while (__atomic_load_n(&lcore_config[worker_id].state,
+                                       __ATOMIC_ACQUIRE) != WAIT)
                rte_pause();
 
-       rte_rmb();
-
        return lcore_config[worker_id].ret;
 }
 

diff --git a/lib/eal/freebsd/eal_thread.c b/lib/eal/freebsd/eal_thread.c
index ce314c1dd4ae7447c65d56fbdba18c30b1c174c2..3b18030d735bf5df340d6ddc01494b3ffd01a8be 100644 (file)
--- a/lib/eal/freebsd/eal_thread.c
+++ b/lib/eal/freebsd/eal_thread.c
@@ -39,11 +39,19 @@ rte_eal_remote_launch(int (*f)(void *), void *arg, unsigned worker_id)
        int w2m = lcore_config[worker_id].pipe_worker2main[0];
        int rc = -EBUSY;
 
-       if (lcore_config[worker_id].state != WAIT)
+       /* Check if the worker is in 'WAIT' state. Use acquire order
+        * since 'state' variable is used as the guard variable.
+        */
+       if (__atomic_load_n(&lcore_config[worker_id].state,
+                                       __ATOMIC_ACQUIRE) != WAIT)
                goto finish;
 
-       lcore_config[worker_id].f = f;
        lcore_config[worker_id].arg = arg;
+       /* Ensure that all the memory operations are completed
+        * before the worker thread starts running the function.
+        * Use worker thread function as the guard variable.
+        */
+       __atomic_store_n(&lcore_config[worker_id].f, f, __ATOMIC_RELEASE);
 
        /* send message */
        n = 0;
@@ -100,6 +108,7 @@ eal_thread_loop(__rte_unused void *arg)
 
        /* read on our pipe to get commands */
        while (1) {
+               lcore_function_t *f;
                void *fct_arg;
 
                /* wait command */
@@ -110,7 +119,11 @@ eal_thread_loop(__rte_unused void *arg)
                if (n <= 0)
                        rte_panic("cannot read on configuration pipe\n");
 
-               lcore_config[lcore_id].state = RUNNING;
+               /* Set the state to 'RUNNING'. Use release order
+                * since 'state' variable is used as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, RUNNING,
+                                       __ATOMIC_RELEASE);
 
                /* send ack */
                n = 0;
@@ -119,17 +132,29 @@ eal_thread_loop(__rte_unused void *arg)
                if (n < 0)
                        rte_panic("cannot write on configuration pipe\n");
 
-               if (lcore_config[lcore_id].f == NULL)
-                       rte_panic("NULL function pointer\n");
+               /* Load 'f' with acquire order to ensure that
+                * the memory operations from the main thread
+                * are accessed only after update to 'f' is visible.
+                * Wait till the update to 'f' is visible to the worker.
+                */
+               while ((f = __atomic_load_n(&lcore_config[lcore_id].f,
+                       __ATOMIC_ACQUIRE)) == NULL)
+                       rte_pause();
 
                /* call the function and store the return value */
                fct_arg = lcore_config[lcore_id].arg;
-               ret = lcore_config[lcore_id].f(fct_arg);
+               ret = f(fct_arg);
                lcore_config[lcore_id].ret = ret;
                lcore_config[lcore_id].f = NULL;
                lcore_config[lcore_id].arg = NULL;
-               rte_wmb();
-               lcore_config[lcore_id].state = WAIT;
+
+               /* Store the state with release order to ensure that
+                * the memory operations from the worker thread
+                * are completed before the state is updated.
+                * Use 'state' as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, WAIT,
+                                       __ATOMIC_RELEASE);
        }
 
        /* never reached */

diff --git a/lib/eal/linux/eal_thread.c b/lib/eal/linux/eal_thread.c
index e2c9e5a3f7c1e049afe4f0dad19165fd42e1dbf9..c7f0f9b2f79b12000146c7f26d0a48853e866a2d 100644 (file)
--- a/lib/eal/linux/eal_thread.c
+++ b/lib/eal/linux/eal_thread.c
@@ -39,11 +39,19 @@ rte_eal_remote_launch(int (*f)(void *), void *arg, unsigned int worker_id)
        int w2m = lcore_config[worker_id].pipe_worker2main[0];
        int rc = -EBUSY;
 
-       if (lcore_config[worker_id].state != WAIT)
+       /* Check if the worker is in 'WAIT' state. Use acquire order
+        * since 'state' variable is used as the guard variable.
+        */
+       if (__atomic_load_n(&lcore_config[worker_id].state,
+                                       __ATOMIC_ACQUIRE) != WAIT)
                goto finish;
 
-       lcore_config[worker_id].f = f;
        lcore_config[worker_id].arg = arg;
+       /* Ensure that all the memory operations are completed
+        * before the worker thread starts running the function.
+        * Use worker thread function pointer as the guard variable.
+        */
+       __atomic_store_n(&lcore_config[worker_id].f, f, __ATOMIC_RELEASE);
 
        /* send message */
        n = 0;
@@ -100,6 +108,7 @@ eal_thread_loop(__rte_unused void *arg)
 
        /* read on our pipe to get commands */
        while (1) {
+               lcore_function_t *f;
                void *fct_arg;
 
                /* wait command */
@@ -110,7 +119,11 @@ eal_thread_loop(__rte_unused void *arg)
                if (n <= 0)
                        rte_panic("cannot read on configuration pipe\n");
 
-               lcore_config[lcore_id].state = RUNNING;
+               /* Set the state to 'RUNNING'. Use release order
+                * since 'state' variable is used as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, RUNNING,
+                                       __ATOMIC_RELEASE);
 
                /* send ack */
                n = 0;
@@ -119,18 +132,29 @@ eal_thread_loop(__rte_unused void *arg)
                if (n < 0)
                        rte_panic("cannot write on configuration pipe\n");
 
-               if (lcore_config[lcore_id].f == NULL)
-                       rte_panic("NULL function pointer\n");
+               /* Load 'f' with acquire order to ensure that
+                * the memory operations from the main thread
+                * are accessed only after update to 'f' is visible.
+                * Wait till the update to 'f' is visible to the worker.
+                */
+               while ((f = __atomic_load_n(&lcore_config[lcore_id].f,
+                       __ATOMIC_ACQUIRE)) == NULL)
+                       rte_pause();
 
                /* call the function and store the return value */
                fct_arg = lcore_config[lcore_id].arg;
-               ret = lcore_config[lcore_id].f(fct_arg);
+               ret = f(fct_arg);
                lcore_config[lcore_id].ret = ret;
                lcore_config[lcore_id].f = NULL;
                lcore_config[lcore_id].arg = NULL;
-               rte_wmb();
 
-               lcore_config[lcore_id].state = WAIT;
+               /* Store the state with release order to ensure that
+                * the memory operations from the worker thread
+                * are completed before the state is updated.
+                * Use 'state' as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, WAIT,
+                                       __ATOMIC_RELEASE);
        }
 
        /* never reached */

diff --git a/lib/eal/windows/eal_thread.c b/lib/eal/windows/eal_thread.c
index 0028123c954f652d7c80c9e00974cd4c91c22774..54fa93fa62157282035b662549927fdfbf19300b 100644 (file)
--- a/lib/eal/windows/eal_thread.c
+++ b/lib/eal/windows/eal_thread.c
@@ -29,11 +29,19 @@ rte_eal_remote_launch(lcore_function_t *f, void *arg, unsigned int worker_id)
        int m2w = lcore_config[worker_id].pipe_main2worker[1];
        int w2m = lcore_config[worker_id].pipe_worker2main[0];
 
-       if (lcore_config[worker_id].state != WAIT)
+       /* Check if the worker is in 'WAIT' state. Use acquire order
+        * since 'state' variable is used as the guard variable.
+        */
+       if (__atomic_load_n(&lcore_config[worker_id].state,
+                                       __ATOMIC_ACQUIRE) != WAIT)
                return -EBUSY;
 
-       lcore_config[worker_id].f = f;
        lcore_config[worker_id].arg = arg;
+       /* Ensure that all the memory operations are completed
+        * before the worker thread starts running the function.
+        * Use worker thread function as the guard variable.
+        */
+       __atomic_store_n(&lcore_config[worker_id].f, f, __ATOMIC_RELEASE);
 
        /* send message */
        n = 0;
@@ -84,6 +92,7 @@ eal_thread_loop(void *arg __rte_unused)
 
        /* read on our pipe to get commands */
        while (1) {
+               lcore_function_t *f;
                void *fct_arg;
 
                /* wait command */
@@ -94,7 +103,11 @@ eal_thread_loop(void *arg __rte_unused)
                if (n <= 0)
                        rte_panic("cannot read on configuration pipe\n");
 
-               lcore_config[lcore_id].state = RUNNING;
+               /* Set the state to 'RUNNING'. Use release order
+                * since 'state' variable is used as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, RUNNING,
+                                       __ATOMIC_RELEASE);
 
                /* send ack */
                n = 0;
@@ -103,18 +116,29 @@ eal_thread_loop(void *arg __rte_unused)
                if (n < 0)
                        rte_panic("cannot write on configuration pipe\n");
 
-               if (lcore_config[lcore_id].f == NULL)
-                       rte_panic("NULL function pointer\n");
+               /* Load 'f' with acquire order to ensure that
+                * the memory operations from the main thread
+                * are accessed only after update to 'f' is visible.
+                * Wait till the update to 'f' is visible to the worker.
+                */
+               while ((f = __atomic_load_n(&lcore_config[lcore_id].f,
+                       __ATOMIC_ACQUIRE)) == NULL)
+                       rte_pause();
 
                /* call the function and store the return value */
                fct_arg = lcore_config[lcore_id].arg;
-               ret = lcore_config[lcore_id].f(fct_arg);
+               ret = f(fct_arg);
                lcore_config[lcore_id].ret = ret;
                lcore_config[lcore_id].f = NULL;
                lcore_config[lcore_id].arg = NULL;
-               rte_wmb();
 
-               lcore_config[lcore_id].state = WAIT;
+               /* Store the state with release order to ensure that
+                * the memory operations from the worker thread
+                * are completed before the state is updated.
+                * Use 'state' as the guard variable.
+                */
+               __atomic_store_n(&lcore_config[lcore_id].state, WAIT,
+                                       __ATOMIC_RELEASE);
        }
 }