dpif-netdev: Add PMD load based sleeping.

Sleep for an incremental amount of time if none of the Rx queues
assigned to a PMD have at least half a batch of packets (i.e. 16 pkts)
on an polling iteration of the PMD.

Upon detecting the threshold of >= 16 pkts on an Rxq, reset the
sleep time to zero (i.e. no sleep).

Sleep time will be increased on each iteration where the low load
conditions remain up to a total of the max sleep time which is set
by the user e.g:
ovs-vsctl set Open_vSwitch . other_config:pmd-maxsleep=500

The default pmd-maxsleep value is 0, which means that no sleeps
will occur and the default behaviour is unchanged from previously.

Also add new stats to pmd-perf-show to get visibility of operation
e.g.
...
   - sleep iterations:       153994  ( 76.8 % of iterations)
   Sleep time (us):         9159399  ( 59 us/iteration avg.)
...

Reviewed-by: Robin Jarry <rjarry@redhat.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ilya Maximets <i.maximets@ovn.org>

7 files changed, 213 insertions, 10 deletions

diff --git a/Documentation/topics/dpdk/pmd.rst b/Documentation/topics/dpdk/pmd.rst
index 9006fd40f..604ac3f6b 100644
--- a/Documentation/topics/dpdk/pmd.rst
+++ b/Documentation/topics/dpdk/pmd.rst
@@ -324,5 +324,59 @@ A user can use this option to set a minimum frequency of Rx queue to PMD
 reassignment due to PMD Auto Load Balance. For example, this could be set
 (in min) such that a reassignment is triggered at most every few hours.
 
+PMD load based sleeping (Experimental)
+--------------------------------------
+
+PMD threads constantly poll Rx queues which are assigned to them. In order to
+reduce the CPU cycles they use, they can sleep for small periods of time
+when there is no load or very-low load on all the Rx queues they poll.
+
+This can be enabled by setting the max requested sleep time (in microseconds)
+for a PMD thread::
+
+    $ ovs-vsctl set open_vswitch . other_config:pmd-maxsleep=500
+
+Non-zero values will be rounded up to the nearest 10 microseconds to avoid
+requesting very small sleep times.
+
+With a non-zero max value a PMD may request to sleep by an incrementing amount
+of time up to the maximum time. If at any point the threshold of at least half
+a batch of packets (i.e. 16) is received from an Rx queue that the PMD is
+polling is met, the requested sleep time will be reset to 0. At that point no
+sleeps will occur until the no/low load conditions return.
+
+Sleeping in a PMD thread will mean there is a period of time when the PMD
+thread will not process packets. Sleep times requested are not guaranteed
+and can differ significantly depending on system configuration. The actual
+time not processing packets will be determined by the sleep and processor
+wake-up times and should be tested with each system configuration.
+
+Sleep time statistics for 10 secs can be seen with::
+
+    $ ovs-appctl dpif-netdev/pmd-stats-clear \
+        && sleep 10 && ovs-appctl dpif-netdev/pmd-perf-show
+
+Example output, showing that during the last 10 seconds, 76.8% of iterations
+had a sleep of some length. The total amount of sleep time was 9.15 seconds and
+the average sleep time per iteration was 46 microseconds::
+
+   - sleep iterations:       153994  ( 76.8 % of iterations)
+   Sleep time (us):         9159399  ( 59 us/iteration avg.)
+
+Any potential power saving from PMD load based sleeping is dependent on the
+system configuration (e.g. enabling processor C-states) and workloads.
+
+.. note::
+
+    If there is a sudden spike of packets while the PMD thread is sleeping and
+    the processor is in a low-power state it may result in some lost packets or
+    extra latency before the PMD thread returns to processing packets at full
+    rate.
+
+.. note::
+
+    By default Linux kernel groups timer expirations and this can add an
+    overhead of up to 50 microseconds to a requested timer expiration.
+
 .. _ovs-vswitchd(8):
     http://openvswitch.org/support/dist-docs/ovs-vswitchd.8.html
diff --git a/NEWS b/NEWS
index 2f6ededfe..4f9291bf1 100644
--- a/NEWS
+++ b/NEWS
@@ -30,6 +30,9 @@ Post-v3.0.0
    - Userspace datapath:
      * Add '-secs' argument to appctl 'dpif-netdev/pmd-rxq-show' to show
        the pmd usage of an Rx queue over a configurable time period.
+     * Add new experimental PMD load based sleeping feature. PMD threads can
+       request to sleep up to a user configured 'pmd-maxsleep' value under
+       low load conditions.
 
 
 v3.0.0 - 15 Aug 2022
diff --git a/lib/dpif-netdev-perf.c b/lib/dpif-netdev-perf.c
index a2a7d8f0b..1a7bab04c 100644
--- a/lib/dpif-netdev-perf.c
+++ b/lib/dpif-netdev-perf.c
@@ -230,18 +230,26 @@ pmd_perf_format_overall_stats(struct ds *str, struct pmd_perf_stats *s,
     uint64_t tot_iter = histogram_samples(&s->pkts);
     uint64_t idle_iter = s->pkts.bin[0];
     uint64_t busy_iter = tot_iter >= idle_iter ? tot_iter - idle_iter : 0;
+    uint64_t sleep_iter = stats[PMD_SLEEP_ITER];
+    uint64_t tot_sleep_cycles = stats[PMD_CYCLES_SLEEP];
 
     ds_put_format(str,
             "  Iterations:         %12"PRIu64"  (%.2f us/it)\n"
             "  - Used TSC cycles:  %12"PRIu64"  (%5.1f %% of total cycles)\n"
             "  - idle iterations:  %12"PRIu64"  (%5.1f %% of used cycles)\n"
-            "  - busy iterations:  %12"PRIu64"  (%5.1f %% of used cycles)\n",
-            tot_iter, tot_cycles * us_per_cycle / tot_iter,
+            "  - busy iterations:  %12"PRIu64"  (%5.1f %% of used cycles)\n"
+            "  - sleep iterations: %12"PRIu64"  (%5.1f %% of iterations)\n"
+            "  Sleep time (us):    %12.0f  (%3.0f us/iteration avg.)\n",
+            tot_iter,
+            (tot_cycles + tot_sleep_cycles) * us_per_cycle / tot_iter,
             tot_cycles, 100.0 * (tot_cycles / duration) / tsc_hz,
             idle_iter,
             100.0 * stats[PMD_CYCLES_ITER_IDLE] / tot_cycles,
             busy_iter,
-            100.0 * stats[PMD_CYCLES_ITER_BUSY] / tot_cycles);
+            100.0 * stats[PMD_CYCLES_ITER_BUSY] / tot_cycles,
+            sleep_iter, tot_iter ? 100.0 * sleep_iter / tot_iter : 0,
+            tot_sleep_cycles * us_per_cycle,
+            sleep_iter ? (tot_sleep_cycles * us_per_cycle) / sleep_iter : 0);
     if (rx_packets > 0) {
         ds_put_format(str,
             "  Rx packets:         %12"PRIu64"  (%.0f Kpps, %.0f cycles/pkt)\n"
@@ -518,14 +526,15 @@ OVS_REQUIRES(s->stats_mutex)
 
 void
 pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets,
-                       int tx_packets, bool full_metrics)
+                       int tx_packets, uint64_t sleep_cycles,
+                       bool full_metrics)
 {
     uint64_t now_tsc = cycles_counter_update(s);
     struct iter_stats *cum_ms;
     uint64_t cycles, cycles_per_pkt = 0;
     char *reason = NULL;
 
-    cycles = now_tsc - s->start_tsc;
+    cycles = now_tsc - s->start_tsc - sleep_cycles;
     s->current.timestamp = s->iteration_cnt;
     s->current.cycles = cycles;
     s->current.pkts = rx_packets;
@@ -539,6 +548,11 @@ pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets,
     histogram_add_sample(&s->cycles, cycles);
     histogram_add_sample(&s->pkts, rx_packets);
 
+    if (sleep_cycles) {
+        pmd_perf_update_counter(s, PMD_SLEEP_ITER, 1);
+        pmd_perf_update_counter(s, PMD_CYCLES_SLEEP, sleep_cycles);
+    }
+
     if (!full_metrics) {
         return;
     }
diff --git a/lib/dpif-netdev-perf.h b/lib/dpif-netdev-perf.h
index 9673dddd8..84beced15 100644
--- a/lib/dpif-netdev-perf.h
+++ b/lib/dpif-netdev-perf.h
@@ -80,6 +80,8 @@ enum pmd_stat_type {
     PMD_CYCLES_ITER_IDLE,   /* Cycles spent in idle iterations. */
     PMD_CYCLES_ITER_BUSY,   /* Cycles spent in busy iterations. */
     PMD_CYCLES_UPCALL,      /* Cycles spent processing upcalls. */
+    PMD_SLEEP_ITER,         /* Iterations where a sleep has taken place. */
+    PMD_CYCLES_SLEEP,       /* Total cycles slept to save power. */
     PMD_N_STATS
 };
 
@@ -408,7 +410,8 @@ void
 pmd_perf_start_iteration(struct pmd_perf_stats *s);
 void
 pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets,
-                       int tx_packets, bool full_metrics);
+                       int tx_packets, uint64_t sleep_cycles,
+                       bool full_metrics);
 
 /* Formatting the output of commands. */
 
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 7127068fe..a47d54c6f 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -171,6 +171,11 @@ static struct odp_support dp_netdev_support = {
 /* Time in microseconds to try RCU quiescing. */
 #define PMD_RCU_QUIESCE_INTERVAL 10000LL
 
+/* Number of pkts Rx on an interface that will stop pmd thread sleeping. */
+#define PMD_SLEEP_THRESH (NETDEV_MAX_BURST / 2)
+/* Time in uS to increment a pmd thread sleep time. */
+#define PMD_SLEEP_INC_US 10
+
 struct dpcls {
     struct cmap_node node;      /* Within dp_netdev_pmd_thread.classifiers */
     odp_port_t in_port;
@@ -279,6 +284,8 @@ struct dp_netdev {
     atomic_uint32_t emc_insert_min;
     /* Enable collection of PMD performance metrics. */
     atomic_bool pmd_perf_metrics;
+    /* Max load based sleep request. */
+    atomic_uint64_t pmd_max_sleep;
     /* Enable the SMC cache from ovsdb config */
     atomic_bool smc_enable_db;
 
@@ -4821,8 +4828,10 @@ dpif_netdev_set_config(struct dpif *dpif, const struct smap *other_config)
     uint64_t rebalance_intvl;
     uint8_t cur_rebalance_load;
     uint32_t rebalance_load, rebalance_improve;
+    uint64_t  pmd_max_sleep, cur_pmd_max_sleep;
     bool log_autolb = false;
     enum sched_assignment_type pmd_rxq_assign_type;
+    static bool first_set_config = true;
 
     tx_flush_interval = smap_get_int(other_config, "tx-flush-interval",
                                      DEFAULT_TX_FLUSH_INTERVAL);
@@ -4969,6 +4978,19 @@ dpif_netdev_set_config(struct dpif *dpif, const struct smap *other_config)
     bool autolb_state = smap_get_bool(other_config, "pmd-auto-lb", false);
 
     set_pmd_auto_lb(dp, autolb_state, log_autolb);
+
+    pmd_max_sleep = smap_get_ullong(other_config, "pmd-maxsleep", 0);
+    pmd_max_sleep = ROUND_UP(pmd_max_sleep, 10);
+    pmd_max_sleep = MIN(PMD_RCU_QUIESCE_INTERVAL, pmd_max_sleep);
+    atomic_read_relaxed(&dp->pmd_max_sleep, &cur_pmd_max_sleep);
+    if (first_set_config || pmd_max_sleep != cur_pmd_max_sleep) {
+        atomic_store_relaxed(&dp->pmd_max_sleep, pmd_max_sleep);
+        VLOG_INFO("PMD max sleep request is %"PRIu64" usecs.", pmd_max_sleep);
+        VLOG_INFO("PMD load based sleeps are %s.",
+                  pmd_max_sleep ? "enabled" : "disabled" );
+    }
+
+    first_set_config  = false;
     return 0;
 }
 
@@ -6929,6 +6951,7 @@ pmd_thread_main(void *f_)
     int poll_cnt;
     int i;
     int process_packets = 0;
+    uint64_t sleep_time = 0;
 
     poll_list = NULL;
 
@@ -6989,10 +7012,13 @@ reload:
     ovs_mutex_lock(&pmd->perf_stats.stats_mutex);
     for (;;) {
         uint64_t rx_packets = 0, tx_packets = 0;
+        uint64_t time_slept = 0;
+        uint64_t max_sleep;
 
         pmd_perf_start_iteration(s);
 
         atomic_read_relaxed(&pmd->dp->smc_enable_db, &pmd->ctx.smc_enable_db);
+        atomic_read_relaxed(&pmd->dp->pmd_max_sleep, &max_sleep);
 
         for (i = 0; i < poll_cnt; i++) {
 
@@ -7011,6 +7037,9 @@ reload:
                 dp_netdev_process_rxq_port(pmd, poll_list[i].rxq,
                                            poll_list[i].port_no);
             rx_packets += process_packets;
+            if (process_packets >= PMD_SLEEP_THRESH) {
+                sleep_time = 0;
+            }
         }
 
         if (!rx_packets) {
@@ -7018,7 +7047,30 @@ reload:
              * Check if we need to send something.
              * There was no time updates on current iteration. */
             pmd_thread_ctx_time_update(pmd);
-            tx_packets = dp_netdev_pmd_flush_output_packets(pmd, false);
+            tx_packets = dp_netdev_pmd_flush_output_packets(pmd,
+                                                   max_sleep && sleep_time
+                                                   ? true : false);
+        }
+
+        if (max_sleep) {
+            /* Check if a sleep should happen on this iteration. */
+            if (sleep_time) {
+                struct cycle_timer sleep_timer;
+
+                cycle_timer_start(&pmd->perf_stats, &sleep_timer);
+                xnanosleep_no_quiesce(sleep_time * 1000);
+                time_slept = cycle_timer_stop(&pmd->perf_stats, &sleep_timer);
+                pmd_thread_ctx_time_update(pmd);
+            }
+            if (sleep_time < max_sleep) {
+                /* Increase sleep time for next iteration. */
+                sleep_time += PMD_SLEEP_INC_US;
+            } else {
+                sleep_time = max_sleep;
+            }
+        } else {
+            /* Reset sleep time as max sleep policy may have been changed. */
+            sleep_time = 0;
         }
 
         /* Do RCU synchronization at fixed interval.  This ensures that
@@ -7058,7 +7110,7 @@ reload:
             break;
         }
 
-        pmd_perf_end_iteration(s, rx_packets, tx_packets,
+        pmd_perf_end_iteration(s, rx_packets, tx_packets, time_slept,
                                pmd_perf_metrics_enabled(pmd));
     }
     ovs_mutex_unlock(&pmd->perf_stats.stats_mutex);
@@ -9909,7 +9961,7 @@ dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd,
                            struct polled_queue *poll_list, int poll_cnt)
 {
     struct dpcls *cls;
-    uint64_t tot_idle = 0, tot_proc = 0;
+    uint64_t tot_idle = 0, tot_proc = 0, tot_sleep = 0;
     unsigned int pmd_load = 0;
 
     if (pmd->ctx.now > pmd->next_cycle_store) {
@@ -9926,10 +9978,13 @@ dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd,
                        pmd->prev_stats[PMD_CYCLES_ITER_IDLE];
             tot_proc = pmd->perf_stats.counters.n[PMD_CYCLES_ITER_BUSY] -
                        pmd->prev_stats[PMD_CYCLES_ITER_BUSY];
+            tot_sleep = pmd->perf_stats.counters.n[PMD_CYCLES_SLEEP] -
+                        pmd->prev_stats[PMD_CYCLES_SLEEP];
 
             if (pmd_alb->is_enabled && !pmd->isolated) {
                 if (tot_proc) {
-                    pmd_load = ((tot_proc * 100) / (tot_idle + tot_proc));
+                    pmd_load = ((tot_proc * 100) /
+                                    (tot_idle + tot_proc + tot_sleep));
                 }
 
                 atomic_read_relaxed(&pmd_alb->rebalance_load_thresh,
@@ -9946,6 +10001,8 @@ dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd,
                         pmd->perf_stats.counters.n[PMD_CYCLES_ITER_IDLE];
         pmd->prev_stats[PMD_CYCLES_ITER_BUSY] =
                         pmd->perf_stats.counters.n[PMD_CYCLES_ITER_BUSY];
+        pmd->prev_stats[PMD_CYCLES_SLEEP] =
+                        pmd->perf_stats.counters.n[PMD_CYCLES_SLEEP];
 
         /* Get the cycles that were used to process each queue and store. */
         for (unsigned i = 0; i < poll_cnt; i++) {
diff --git a/tests/pmd.at b/tests/pmd.at
index ed90f88c4..e0f58f7a6 100644
--- a/tests/pmd.at
+++ b/tests/pmd.at
@@ -1254,3 +1254,49 @@ ovs-appctl: ovs-vswitchd: server returned an error
 
 OVS_VSWITCHD_STOP
 AT_CLEANUP
+
+dnl Check default state
+AT_SETUP([PMD - pmd sleep])
+OVS_VSWITCHD_START
+
+dnl Check default
+OVS_WAIT_UNTIL([tail ovs-vswitchd.log | grep "PMD max sleep request is 0 usecs."])
+OVS_WAIT_UNTIL([tail ovs-vswitchd.log | grep "PMD load based sleeps are disabled."])
+
+dnl Check low value max sleep
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="1"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 10 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are enabled."])
+
+dnl Check high value max sleep
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="10000"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 10000 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are enabled."])
+
+dnl Check setting max sleep to zero
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="0"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 0 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are disabled."])
+
+dnl Check above high value max sleep
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="10001"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 10000 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are enabled."])
+
+dnl Check rounding
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="490"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 490 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are enabled."])
+dnl Check rounding
+get_log_next_line_num
+AT_CHECK([ovs-vsctl set open_vswitch . other_config:pmd-maxsleep="491"])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD max sleep request is 500 usecs."])
+OVS_WAIT_UNTIL([tail -n +$LINENUM ovs-vswitchd.log | grep "PMD load based sleeps are enabled."])
+
+OVS_VSWITCHD_STOP
+AT_CLEANUP
diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml
index f9bdb2d92..8c4acfb18 100644
--- a/vswitchd/vswitch.xml
+++ b/vswitchd/vswitch.xml
@@ -788,6 +788,32 @@
          The default value is <code>25%</code>.
         </p>
       </column>
+      <column name="other_config" key="pmd-maxsleep"
+              type='{"type": "integer",
+                     "minInteger": 0, "maxInteger": 10000}'>
+        <p>
+          Specifies the maximum sleep time that will be requested in
+          microseconds per iteration for a PMD thread which has received zero
+          or a small amount of packets from the Rx queues it is polling.
+        </p>
+        <p>
+          The actual sleep time requested is based on the load
+          of the Rx queues that the PMD polls and may be less than
+          the maximum value.
+        </p>
+        <p>
+          The default value is <code>0 microseconds</code>, which means
+          that the PMD will not sleep regardless of the load from the
+          Rx queues that it polls.
+        </p>
+        <p>
+          To avoid requesting very small sleeps (e.g. less than 10 us) the
+          value will be rounded up to the nearest 10 us.
+        </p>
+        <p>
+          The maximum value is <code>10000 microseconds</code>.
+        </p>
+      </column>
       <column name="other_config" key="userspace-tso-enable"
               type='{"type": "boolean"}'>
         <p>