service: fix race in service on app lcore function
authorHarry van Haaren <harry.van.haaren@intel.com>
Wed, 1 Nov 2017 18:48:01 +0000 (18:48 +0000)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 7 Nov 2017 00:42:04 +0000 (01:42 +0100)
This commit fixes a possible race condition if an application
uses the service-cores infrastructure and the function to run
a service on an application lcore at the same time.

The fix is to change the num_mapped_cores variable to be an
atomic variable. This causes concurrent accesses by multiple
threads to a service using rte_service_run_iter_on_app_lcore()
to detect if another core is currently mapped to the service,
and refuses to run if it is not multi-thread safe.

The run iteration on app lcore function has two arguments, the
service id to run, and if atomics should be used to serialize access
to multi-thread unsafe services. This allows applications to choose
if they wish to use use the service-cores feature, or if they
take responsibility themselves for serializing invoking a service.
See doxygen documentation for more details.

Two unit tests were added to verify the behaviour of the
function to run a service on an application core, testing both
a multi-thread safe service, and a multi-thread unsafe service.

The doxygen API documentation for the function has been updated
to reflect the current and correct behaviour.

Fixes: e9139a32f6e8 ("service: add function to run on app lcore")

Signed-off-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
examples/eventdev_pipeline_sw_pmd/main.c
lib/librte_eal/common/include/rte_service.h
lib/librte_eal/common/rte_service.c
test/test/test_eventdev_sw.c
test/test/test_service_cores.c

index 948be67..5f431d8 100644 (file)
@@ -245,7 +245,7 @@ schedule_devices(unsigned int lcore_id)
 
        if (fdata->sched_core[lcore_id] && (fdata->sched_single ||
            rte_atomic32_cmpset(&(fdata->sched_lock), 0, 1))) {
-               rte_service_run_iter_on_app_lcore(fdata->evdev_service_id);
+               rte_service_run_iter_on_app_lcore(fdata->evdev_service_id, 1);
                if (cdata.dump_dev_signal) {
                        rte_event_dev_dump(0, stdout);
                        cdata.dump_dev_signal = 0;
index d9de5ad..9272440 100644 (file)
@@ -232,11 +232,28 @@ int32_t rte_service_set_runstate_mapped_check(uint32_t id, int32_t enable);
  * @warning
  * @b EXPERIMENTAL: this API may change without prior notice
  *
- * This function runs a service callback from a non-service lcore context.
- * The *id* of the service to be run is passed in, and the service-callback
- * is executed on the calling lcore immediately if possible. If the service is
- * not multi-thread capable and another thread is currently executing it, this
- * function returns without running the callback.
+ * This function runs a service callback from a non-service lcore.
+ *
+ * This function is designed to enable gradual porting to service cores, and
+ * to enable unit tests to verify a service behaves as expected.
+ *
+ * When called, this function ensures that the service identified by *id* is
+ * safe to run on this lcore. Multi-thread safe services are invoked even if
+ * other cores are simultaneously running them as they are multi-thread safe.
+ *
+ * Multi-thread unsafe services are handled depending on the variable
+ * *serialize_multithread_unsafe*:
+ * - When set, the function will check if a service is already being invoked
+ *   on another lcore, refusing to run it and returning -EBUSY.
+ * - When zero, the application takes responsibility to ensure that the service
+ *   indicated by *id* is not going to be invoked by another lcore. This setting
+ *   avoids atomic operations, so is likely to be more performant.
+ *
+ * @param id The ID of the service to run
+ * @param serialize_multithread_unsafe This parameter indicates to the service
+ *           cores library if it is required to use atomics to serialize access
+ *           to mult-thread unsafe services. As there is an overhead in using
+ *           atomics, applications can choose to enable or disable this feature
  *
  * Note that any thread calling this function MUST be a DPDK EAL thread, as
  * the *rte_lcore_id* function is used to access internal data structures.
@@ -244,10 +261,11 @@ int32_t rte_service_set_runstate_mapped_check(uint32_t id, int32_t enable);
  * @retval 0 Service was run on the calling thread successfully
  * @retval -EBUSY Another lcore is executing the service, and it is not a
  *         multi-thread safe service, so the service was not run on this lcore
- * @retval -ENOEXEC Service is not in a runnable state
+ * @retval -ENOEXEC Service is not in a run-able state
  * @retval -EINVAL Invalid service id
  */
-int32_t rte_service_run_iter_on_app_lcore(uint32_t id);
+int32_t rte_service_run_iter_on_app_lcore(uint32_t id,
+               uint32_t serialize_multithread_unsafe);
 
 /**
  * @warning
index f17bf4b..09b758c 100644 (file)
@@ -77,7 +77,7 @@ struct rte_service_spec_impl {
        uint8_t internal_flags;
 
        /* per service statistics */
-       uint32_t num_mapped_cores;
+       rte_atomic32_t num_mapped_cores;
        uint64_t calls;
        uint64_t cycles_spent;
 } __rte_cache_aligned;
@@ -325,7 +325,7 @@ rte_service_runstate_get(uint32_t id)
        rte_smp_rmb();
 
        int check_disabled = !(s->internal_flags & SERVICE_F_START_CHECK);
-       int lcore_mapped = (s->num_mapped_cores > 0);
+       int lcore_mapped = (rte_atomic32_read(&s->num_mapped_cores) > 0);
 
        return (s->app_runstate == RUNSTATE_RUNNING) &&
                (s->comp_runstate == RUNSTATE_RUNNING) &&
@@ -365,7 +365,7 @@ service_run(uint32_t i, struct core_state *cs, uint64_t service_mask)
         * mapped, atomic ops are not required.
         */
        const int use_atomics = (service_mt_safe(s) == 0) &&
-                               (s->num_mapped_cores > 1);
+                               (rte_atomic32_read(&s->num_mapped_cores) > 1);
        if (use_atomics) {
                if (!rte_atomic32_cmpset((uint32_t *)&s->execute_lock, 0, 1))
                        return -EBUSY;
@@ -378,11 +378,36 @@ service_run(uint32_t i, struct core_state *cs, uint64_t service_mask)
        return 0;
 }
 
-int32_t rte_service_run_iter_on_app_lcore(uint32_t id)
+int32_t rte_service_run_iter_on_app_lcore(uint32_t id,
+               uint32_t serialize_mt_unsafe)
 {
        /* run service on calling core, using all-ones as the service mask */
+       if (!service_valid(id))
+               return -EINVAL;
+
        struct core_state *cs = &lcore_states[rte_lcore_id()];
-       return service_run(id, cs, UINT64_MAX);
+       struct rte_service_spec_impl *s = &rte_services[id];
+
+       /* Atomically add this core to the mapped cores first, then examine if
+        * we can run the service. This avoids a race condition between
+        * checking the value, and atomically adding to the mapped count.
+        */
+       if (serialize_mt_unsafe)
+               rte_atomic32_inc(&s->num_mapped_cores);
+
+       if (service_mt_safe(s) == 0 &&
+                       rte_atomic32_read(&s->num_mapped_cores) > 1) {
+               if (serialize_mt_unsafe)
+                       rte_atomic32_dec(&s->num_mapped_cores);
+               return -EBUSY;
+       }
+
+       int ret = service_run(id, cs, UINT64_MAX);
+
+       if (serialize_mt_unsafe)
+               rte_atomic32_dec(&s->num_mapped_cores);
+
+       return ret;
 }
 
 static int32_t
@@ -522,10 +547,10 @@ service_update(struct rte_service_spec *service, uint32_t lcore,
        if (set) {
                if (*set) {
                        lcore_states[lcore].service_mask |= sid_mask;
-                       rte_services[sid].num_mapped_cores++;
+                       rte_atomic32_inc(&rte_services[sid].num_mapped_cores);
                } else {
                        lcore_states[lcore].service_mask &= ~(sid_mask);
-                       rte_services[sid].num_mapped_cores--;
+                       rte_atomic32_dec(&rte_services[sid].num_mapped_cores);
                }
        }
 
@@ -568,7 +593,7 @@ int32_t rte_service_lcore_reset_all(void)
                lcore_states[i].runstate = RUNSTATE_STOPPED;
        }
        for (i = 0; i < RTE_SERVICE_NUM_MAX; i++)
-               rte_services[i].num_mapped_cores = 0;
+               rte_atomic32_set(&rte_services[i].num_mapped_cores, 0);
 
        rte_smp_wmb();
 
@@ -664,7 +689,8 @@ rte_service_lcore_stop(uint32_t lcore)
        for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
                int32_t enabled = service_mask & (UINT64_C(1) << i);
                int32_t service_running = rte_service_runstate_get(i);
-               int32_t only_core = rte_services[i].num_mapped_cores == 1;
+               int32_t only_core = (1 ==
+                       rte_atomic32_read(&rte_services[i].num_mapped_cores));
 
                /* if the core is mapped, and the service is running, and this
                 * is the only core that is mapped, the service would cease to
index b86b137..01aa4d9 100644 (file)
@@ -417,7 +417,7 @@ run_prio_packet_test(struct test *t)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        struct test_event_dev_stats stats;
        err = test_event_dev_stats_get(evdev, &stats);
@@ -509,7 +509,7 @@ test_single_directed_packet(struct test *t)
        }
 
        /* Run schedule() as dir packets may need to be re-ordered */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        struct test_event_dev_stats stats;
        err = test_event_dev_stats_get(evdev, &stats);
@@ -576,7 +576,7 @@ test_directed_forward_credits(struct test *t)
                        printf("%d: error failed to enqueue\n", __LINE__);
                        return -1;
                }
-               rte_service_run_iter_on_app_lcore(t->service_id);
+               rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
                uint32_t deq_pkts;
                deq_pkts = rte_event_dequeue_burst(evdev, 0, &ev, 1, 0);
@@ -738,7 +738,7 @@ burst_packets(struct test *t)
                        return -1;
                }
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* Check stats for all NUM_PKTS arrived to sched core */
        struct test_event_dev_stats stats;
@@ -827,7 +827,7 @@ abuse_inflights(struct test *t)
        }
 
        /* schedule */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        struct test_event_dev_stats stats;
 
@@ -965,7 +965,7 @@ xstats_tests(struct test *t)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* Device names / values */
        int num_stats = rte_event_dev_xstats_names_get(evdev,
@@ -1292,7 +1292,7 @@ port_reconfig_credits(struct test *t)
                        }
                }
 
-               rte_service_run_iter_on_app_lcore(t->service_id);
+               rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
                struct rte_event ev[NPKTS];
                int deq = rte_event_dequeue_burst(evdev, t->port[0], ev,
@@ -1518,7 +1518,7 @@ xstats_id_reset_tests(struct test *t)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        static const char * const dev_names[] = {
                "dev_rx", "dev_tx", "dev_drop", "dev_sched_calls",
@@ -1909,7 +1909,7 @@ qid_priorities(struct test *t)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* dequeue packets, verify priority was upheld */
        struct rte_event ev[32];
@@ -1990,7 +1990,7 @@ load_balancing(struct test *t)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        struct test_event_dev_stats stats;
        err = test_event_dev_stats_get(evdev, &stats);
@@ -2090,7 +2090,7 @@ load_balancing_history(struct test *t)
        }
 
        /* call the scheduler */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* Dequeue the flow 0 packet from port 1, so that we can then drop */
        struct rte_event ev;
@@ -2107,7 +2107,7 @@ load_balancing_history(struct test *t)
        rte_event_enqueue_burst(evdev, t->port[1], &release_ev, 1);
 
        /* call the scheduler */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /*
         * Set up the next set of flows, first a new flow to fill up
@@ -2140,7 +2140,7 @@ load_balancing_history(struct test *t)
        }
 
        /* schedule */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (err) {
@@ -2184,7 +2184,7 @@ load_balancing_history(struct test *t)
                while (rte_event_dequeue_burst(evdev, i, &ev, 1, 0))
                        rte_event_enqueue_burst(evdev, i, &release_ev, 1);
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        cleanup(t);
        return 0;
@@ -2250,7 +2250,7 @@ invalid_qid(struct test *t)
        }
 
        /* call the scheduler */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (err) {
@@ -2335,7 +2335,7 @@ single_packet(struct test *t)
                return -1;
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (err) {
@@ -2378,7 +2378,7 @@ single_packet(struct test *t)
                printf("%d: Failed to enqueue\n", __LINE__);
                return -1;
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (stats.port_inflight[wrk_enq] != 0) {
@@ -2466,7 +2466,7 @@ inflight_counts(struct test *t)
        }
 
        /* schedule */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (err) {
@@ -2522,7 +2522,7 @@ inflight_counts(struct test *t)
         * As the scheduler core decrements inflights, it needs to run to
         * process packets to act on the drop messages
         */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (stats.port_inflight[p1] != 0) {
@@ -2557,7 +2557,7 @@ inflight_counts(struct test *t)
         * As the scheduler core decrements inflights, it needs to run to
         * process packets to act on the drop messages
         */
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        err = test_event_dev_stats_get(evdev, &stats);
        if (stats.port_inflight[p2] != 0) {
@@ -2651,7 +2651,7 @@ parallel_basic(struct test *t, int check_order)
                }
        }
 
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* use extra slot to make logic in loops easier */
        struct rte_event deq_ev[w3_port + 1];
@@ -2678,7 +2678,7 @@ parallel_basic(struct test *t, int check_order)
                        return -1;
                }
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* dequeue from the tx ports, we should get 3 packets */
        deq_pkts = rte_event_dequeue_burst(evdev, t->port[tx_port], deq_ev,
@@ -2756,7 +2756,7 @@ holb(struct test *t) /* test to check we avoid basic head-of-line blocking */
                printf("%d: Error doing first enqueue\n", __LINE__);
                goto err;
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        if (rte_event_dev_xstats_by_name_get(evdev, "port_0_cq_ring_used", NULL)
                        != 1)
@@ -2781,7 +2781,7 @@ holb(struct test *t) /* test to check we avoid basic head-of-line blocking */
                        printf("%d: Error with enqueue\n", __LINE__);
                        goto err;
                }
-               rte_service_run_iter_on_app_lcore(t->service_id);
+               rte_service_run_iter_on_app_lcore(t->service_id, 1);
        } while (rte_event_dev_xstats_by_name_get(evdev,
                                rx_port_free_stat, NULL) != 0);
 
@@ -2791,7 +2791,7 @@ holb(struct test *t) /* test to check we avoid basic head-of-line blocking */
                printf("%d: Error with enqueue\n", __LINE__);
                goto err;
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        /* check that the other port still has an empty CQ */
        if (rte_event_dev_xstats_by_name_get(evdev, other_port_used_stat, NULL)
@@ -2814,7 +2814,7 @@ holb(struct test *t) /* test to check we avoid basic head-of-line blocking */
                printf("%d: Error with enqueue\n", __LINE__);
                goto err;
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
        if (rte_event_dev_xstats_by_name_get(evdev, other_port_used_stat, NULL)
                        != 1) {
@@ -3004,7 +3004,7 @@ worker_loopback(struct test *t)
        while (rte_eal_get_lcore_state(p_lcore) != FINISHED ||
                        rte_eal_get_lcore_state(w_lcore) != FINISHED) {
 
-               rte_service_run_iter_on_app_lcore(t->service_id);
+               rte_service_run_iter_on_app_lcore(t->service_id, 1);
 
                uint64_t new_cycles = rte_get_timer_cycles();
 
@@ -3031,7 +3031,7 @@ worker_loopback(struct test *t)
                        cycles = new_cycles;
                }
        }
-       rte_service_run_iter_on_app_lcore(t->service_id);
+       rte_service_run_iter_on_app_lcore(t->service_id, 1);
        /* ensure all completions are flushed */
 
        rte_eal_mp_wait_lcore();
index ee8313e..311c704 100644 (file)
@@ -548,6 +548,113 @@ service_mt_unsafe_poll(void)
        return TEST_SUCCESS;
 }
 
+static int32_t
+delay_as_a_mt_safe_service(void *args)
+{
+       RTE_SET_USED(args);
+       uint32_t *params = args;
+
+       /* retrieve done flag and atomic lock to inc/dec */
+       uint32_t *done = &params[0];
+       rte_atomic32_t *lock = (rte_atomic32_t *)&params[1];
+
+       while (!*done) {
+               rte_atomic32_inc(lock);
+               rte_delay_us(500);
+               if (rte_atomic32_read(lock) > 1)
+                       /* pass: second core has simultaneously incremented */
+                       *done = 1;
+               rte_atomic32_dec(lock);
+       }
+
+       return 0;
+}
+
+static int32_t
+delay_as_a_service(void *args)
+{
+       uint32_t *done = (uint32_t *)args;
+       while (!*done)
+               rte_delay_ms(5);
+       return 0;
+}
+
+static int
+service_run_on_app_core_func(void *arg)
+{
+       uint32_t *delay_service_id = (uint32_t *)arg;
+       return rte_service_run_iter_on_app_lcore(*delay_service_id, 1);
+}
+
+static int
+service_app_lcore_poll_impl(const int mt_safe)
+{
+       uint32_t params[2] = {0};
+
+       struct rte_service_spec service;
+       memset(&service, 0, sizeof(struct rte_service_spec));
+       snprintf(service.name, sizeof(service.name), MT_SAFE_SERVICE_NAME);
+       if (mt_safe) {
+               service.callback = delay_as_a_mt_safe_service;
+               service.callback_userdata = params;
+               service.capabilities |= RTE_SERVICE_CAP_MT_SAFE;
+       } else {
+               service.callback = delay_as_a_service;
+               service.callback_userdata = &params;
+       }
+
+       uint32_t id;
+       TEST_ASSERT_EQUAL(0, rte_service_component_register(&service, &id),
+                       "Register of app lcore delay service failed");
+
+       rte_service_component_runstate_set(id, 1);
+       rte_service_runstate_set(id, 1);
+
+       uint32_t app_core2 = rte_get_next_lcore(slcore_id, 1, 1);
+       int app_core2_ret = rte_eal_remote_launch(service_run_on_app_core_func,
+                                                 &id, app_core2);
+
+       rte_delay_ms(100);
+
+       int app_core1_ret = service_run_on_app_core_func(&id);
+
+       /* flag done, then wait for the spawned 2nd core to return */
+       params[0] = 1;
+       rte_eal_mp_wait_lcore();
+
+       /* core two gets launched first - and should hold the service lock */
+       TEST_ASSERT_EQUAL(0, app_core2_ret,
+                       "App core2 : run service didn't return zero");
+
+       if (mt_safe) {
+               /* mt safe should have both cores return 0 for success */
+               TEST_ASSERT_EQUAL(0, app_core1_ret,
+                               "MT Safe: App core1 didn't return 0");
+       } else {
+               /* core one attempts to run later - should be blocked */
+               TEST_ASSERT_EQUAL(-EBUSY, app_core1_ret,
+                               "MT Unsafe: App core1 didn't return -EBUSY");
+       }
+
+       unregister_all();
+
+       return TEST_SUCCESS;
+}
+
+static int
+service_app_lcore_mt_safe(void)
+{
+       const int mt_safe = 1;
+       return service_app_lcore_poll_impl(mt_safe);
+}
+
+static int
+service_app_lcore_mt_unsafe(void)
+{
+       const int mt_safe = 0;
+       return service_app_lcore_poll_impl(mt_safe);
+}
+
 /* start and stop a service core - ensuring it goes back to sleep */
 static int
 service_lcore_start_stop(void)
@@ -613,6 +720,8 @@ static struct unit_test_suite service_tests  = {
                TEST_CASE_ST(dummy_register, NULL, service_lcore_en_dis_able),
                TEST_CASE_ST(dummy_register, NULL, service_mt_unsafe_poll),
                TEST_CASE_ST(dummy_register, NULL, service_mt_safe_poll),
+               TEST_CASE_ST(dummy_register, NULL, service_app_lcore_mt_safe),
+               TEST_CASE_ST(dummy_register, NULL, service_app_lcore_mt_unsafe),
                TEST_CASES_END() /**< NULL terminate unit test array */
        }
 };