| Message ID | 1515489982-24380-3-git-send-email-jan.scheurich@ericsson.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | Refactor PMD stats and cycle counting | expand |
PMD_CYCLES_UPCALL PMD_CYCLES_POLL_IDLE PMD_CYCLES_POLL_BUSY are not used in current series and should be removed. PMD_CYCLES_OVERHEAD implemented in patch #1 and removed in patch #2. Should be removed from patch #1 to simplify it. Other comments inline. Best regards, Ilya Maximets. On 09.01.2018 12:26, Jan Scheurich wrote: > Simplify the historically grown TSC cycle counting in PMD threads. > Cycles are currently counted for the following purposes: > > 1. Measure PMD ustilization > > PMD utilization is defined as ratio of cycles spent in busy iterations > (at least one packet received or sent) over the total number of cycles. > > This is already done in pmd_perf_start_iteration() and > pmd_perf_end_iteration() based on a TSC timestamp saved in current > iteration at start_iteration() and the actual TSC at end_iteration(). > No dependency on intermediate cycle accounting. > > 2. Measure the processing load per RX queue > > This comprises cycles spend on polling and processing packets received > from the rx queue. > > The previous scheme using cycles_count_start(), cycles_count_intermediate() > and cycles-count_end() originally introduced to simplify cycle counting > and saving calls to rte_get_tsc_cycles() was rather obscuring things. > > Replace by a nestable cycle_timer with with start and stop functions to > embrace a code segment to be timed. The timed code may contain arbitrary > nested cycle_timers. The duration of nested timers is excluded from the > outer timer. > > The caller must ensure that each call to cycle_timer_start() is > followed by a call to cycle_timer_end(). Failure to do so will lead to > assertion failure or a memory leak. > > The new cycle_timer is used to measure the processing cycles per rx queue. > This is not yet strictly necessary but will be made use of in subsequent > commits. > > All cycle count functions and data are relocated to module > dpif-netdev-perf. > > Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com> > > --- > lib/dpif-netdev-perf.h | 94 +++++++++++++++++++++++++++++++++++--- > lib/dpif-netdev.c | 121 ++++++++++++++----------------------------------- > 2 files changed, 123 insertions(+), 92 deletions(-) > > diff --git a/lib/dpif-netdev-perf.h b/lib/dpif-netdev-perf.h > index b31bf49..a6dfcfd 100644 > --- a/lib/dpif-netdev-perf.h > +++ b/lib/dpif-netdev-perf.h > @@ -90,9 +90,92 @@ struct pmd_perf_stats { > period. */ > uint64_t last_tsc; /* Start of the current PMD iteration in TSC > cycles. */ > + uint64_t last_cycles; /* Latest TSC time stamp taken in PMD. */ > + struct cycle_timer *cur_timer; > + /* If non-NULL, outermost cycle timer currently > + running in PMD. */ Same comment as in patch #1. This one looks the most ugly, but could be written in one line above the element. > struct pmd_counters counters; > }; > > +/* Support for accurate timing of PMD execution on TSC clock cycle level. > + * These functions are intended to be invoked in the context of pmd threads. */ > + > +/* Read the TSC cycle register and cache it. Any function not requiring clock > + * cycle accuracy should read the cached value using cycles_counter_get(). */ > + > +static inline uint64_t > +cycles_counter_update(struct pmd_perf_stats *s) > +{ > +#ifdef DPDK_NETDEV > + return s->last_cycles = rte_get_tsc_cycles(); > +#else > + return s->last_cycles = 0; > +#endif > +} > + > +static inline uint64_t > +cycles_counter_get(struct pmd_perf_stats *s) > +{ > + return s->last_cycles; > +} This function not used. > + > +/* A nestable timer for measuring execution time in TSC cycles. > + * > + * Usage: > + * struct cycle_timer timer; > + * > + * cycle_timer_start(pmd, &timer); > + * <Timed execution> > + * uint64_t cycles = cycle_timer_stop(pmd, &timer); > + * > + * The caller must guarantee that a call to cycle_timer_start() is always > + * paired with a call to cycle_stimer_stop(). > + * > + * Is is possible to have nested cycles timers within the timed code. The > + * execution time measured by the nested timers is excluded from the time > + * measured by the embracing timer. > + */ > + > +struct cycle_timer { > + uint64_t start; > + uint64_t suspended; > + struct cycle_timer *interrupted; > +}; > + > +static inline void > +cycle_timer_start(struct pmd_perf_stats *s, > + struct cycle_timer *timer) > +{ > + struct cycle_timer *cur_timer = s->cur_timer; > + uint64_t now = cycles_counter_update(s); > + > + if (cur_timer) { > + cur_timer->suspended = now; > + } > + timer->interrupted = cur_timer; > + timer->start = now; > + timer->suspended = 0; > + s->cur_timer = timer; > +} > + > +static inline uint64_t > +cycle_timer_stop(struct pmd_perf_stats *s, > + struct cycle_timer *timer) > +{ > + /* Assert that this is the current cycle timer. */ > + ovs_assert(s->cur_timer == timer); > + uint64_t now = cycles_counter_update(s); > + struct cycle_timer *intr_timer = timer->interrupted; > + > + if (intr_timer) { > + /* Adjust the start offset by the suspended cycles. */ > + intr_timer->start += now - intr_timer->suspended; > + } > + /* Restore suspended timer, if any. */ > + s->cur_timer = intr_timer; > + return now - timer->start; > +} > + > void pmd_perf_stats_init(struct pmd_perf_stats *s); > void pmd_perf_stats_clear(struct pmd_perf_stats *s); > void pmd_perf_read_counters(struct pmd_perf_stats *s, > @@ -109,23 +192,22 @@ pmd_perf_update_counter(struct pmd_perf_stats *s, > } > > static inline void > -pmd_perf_start_iteration(struct pmd_perf_stats *s, uint64_t now_tsc) > +pmd_perf_start_iteration(struct pmd_perf_stats *s) > { > if (OVS_UNLIKELY(s->start_ms == 0)) { > /* Stats not yet fully initialized. */ > return; > } > - s->last_tsc = now_tsc; > + s->last_tsc = s->last_cycles; > } > > static inline void > -pmd_perf_end_iteration(struct pmd_perf_stats *s, uint64_t now_tsc, > - int packets) > +pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets) > { > - uint64_t cycles = now_tsc - s->last_tsc; > + uint64_t cycles = cycles_counter_update(s) - s->last_tsc; > > /* No need for atomic updates as only invoked within PMD thread. */ > - if (packets > 0) { > + if (rx_packets > 0) { I guess, this renaming should be done in patch #1. > s->counters.n[PMD_CYCLES_ITER_BUSY] += cycles; > } else { > s->counters.n[PMD_CYCLES_ITER_IDLE] += cycles; > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c > index ad050f2..709d7ee 100644 > --- a/lib/dpif-netdev.c > +++ b/lib/dpif-netdev.c > @@ -509,8 +509,6 @@ struct tx_port { > struct dp_netdev_pmd_thread_ctx { > /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ > long long now; > - /* Used to count cycles. See 'cycles_count_end()' */ > - unsigned long long last_cycles; > }; > > /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate > @@ -873,8 +871,8 @@ pmd_info_show_stats(struct ds *reply, > ds_put_format(reply, > "\tavg processing cycles per packet: " > "%.02f (%"PRIu64"/%"PRIu64")\n", > - stats[PMD_CYCLES_POLL_BUSY] / (double) total_packets, > - stats[PMD_CYCLES_POLL_BUSY], total_packets); > + stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, > + stats[PMD_CYCLES_ITER_BUSY], total_packets); This is unclear. Why semantics changed? Is it typo or this should be done in patch #1? > } > > static int > @@ -3110,64 +3108,20 @@ dp_netdev_actions_free(struct dp_netdev_actions *actions) > free(actions); > } > > -static inline unsigned long long > -cycles_counter(void) > -{ > -#ifdef DPDK_NETDEV > - return rte_get_tsc_cycles(); > -#else > - return 0; > -#endif > -} > - > -/* Fake mutex to make sure that the calls to cycles_count_* are balanced */ > -extern struct ovs_mutex cycles_counter_fake_mutex; > - > -/* Start counting cycles. Must be followed by 'cycles_count_end()' */ > -static inline void > -cycles_count_start(struct dp_netdev_pmd_thread *pmd) > - OVS_ACQUIRES(&cycles_counter_fake_mutex) > - OVS_NO_THREAD_SAFETY_ANALYSIS > -{ > - pmd->ctx.last_cycles = cycles_counter(); > -} > - > -/* Stop counting cycles and add them to the counter 'type' */ > -static inline void > -cycles_count_end(struct dp_netdev_pmd_thread *pmd, > - enum pmd_stat_type type) > - OVS_RELEASES(&cycles_counter_fake_mutex) > - OVS_NO_THREAD_SAFETY_ANALYSIS > -{ > - unsigned long long interval = cycles_counter() - pmd->ctx.last_cycles; > - > - pmd_perf_update_counter(&pmd->perf_stats, type, interval); > -} > - > -/* Calculate the intermediate cycle result and add to the counter 'type' */ > -static inline void > -cycles_count_intermediate(struct dp_netdev_pmd_thread *pmd, > - struct dp_netdev_rxq *rxq, > - enum pmd_stat_type type) > - OVS_NO_THREAD_SAFETY_ANALYSIS > +static void > +dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, > + enum rxq_cycles_counter_type type, > + unsigned long long cycles) > { > - unsigned long long new_cycles = cycles_counter(); > - unsigned long long interval = new_cycles - pmd->ctx.last_cycles; > - pmd->ctx.last_cycles = new_cycles; > - > - pmd_perf_update_counter(&pmd->perf_stats, type, interval); > - if (rxq && (type == PMD_CYCLES_POLL_BUSY)) { > - /* Add to the amount of current processing cycles. */ > - non_atomic_ullong_add(&rxq->cycles[RXQ_CYCLES_PROC_CURR], interval); > - } > + atomic_store_relaxed(&rx->cycles[type], cycles); > } > > static void > -dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, > +dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, > enum rxq_cycles_counter_type type, > unsigned long long cycles) > { > - atomic_store_relaxed(&rx->cycles[type], cycles); > + non_atomic_ullong_add(&rx->cycles[type], cycles); > } > > static uint64_t > @@ -3235,27 +3189,40 @@ dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd) > > static int > dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, > - struct netdev_rxq *rx, > + struct dp_netdev_rxq *rxq, > odp_port_t port_no) > { > struct dp_packet_batch batch; > + struct cycle_timer timer; > int error; > int batch_cnt = 0; > > + /* Measure duration for polling and processing rx burst. */ > + cycle_timer_start(&pmd->perf_stats, &timer); > dp_packet_batch_init(&batch); > - error = netdev_rxq_recv(rx, &batch); > + error = netdev_rxq_recv(rxq->rx, &batch); > if (!error) { > + /* At least one packet received. */ > *recirc_depth_get() = 0; > pmd_thread_ctx_time_update(pmd); > > batch_cnt = batch.count; > dp_netdev_input(pmd, &batch, port_no); > dp_netdev_pmd_flush_output_packets(pmd); > - } else if (error != EAGAIN && error != EOPNOTSUPP) { > - static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); > > - VLOG_ERR_RL(&rl, "error receiving data from %s: %s", > - netdev_rxq_get_name(rx), ovs_strerror(error)); > + /* Assign processing cycles to rx queue. */ > + uint64_t cycles = cycle_timer_stop(&pmd->perf_stats, &timer); > + dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); > + > + } else { > + /* Discard cycles. */ > + cycle_timer_stop(&pmd->perf_stats, &timer); > + if (error != EAGAIN && error != EOPNOTSUPP) { > + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); > + > + VLOG_ERR_RL(&rl, "error receiving data from %s: %s", > + netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); > + } > } > > return batch_cnt; > @@ -3881,30 +3848,22 @@ dpif_netdev_run(struct dpif *dpif) > struct dp_netdev *dp = get_dp_netdev(dpif); > struct dp_netdev_pmd_thread *non_pmd; > uint64_t new_tnl_seq; > - int process_packets = 0; > > ovs_mutex_lock(&dp->port_mutex); > non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); > if (non_pmd) { > ovs_mutex_lock(&dp->non_pmd_mutex); > - cycles_count_start(non_pmd); > HMAP_FOR_EACH (port, node, &dp->ports) { > if (!netdev_is_pmd(port->netdev)) { > int i; > > for (i = 0; i < port->n_rxq; i++) { > - process_packets = > - dp_netdev_process_rxq_port(non_pmd, > - port->rxqs[i].rx, > - port->port_no); > - cycles_count_intermediate(non_pmd, NULL, > - process_packets > - ? PMD_CYCLES_POLL_BUSY > - : PMD_CYCLES_POLL_IDLE); > + dp_netdev_process_rxq_port(non_pmd, > + &port->rxqs[i], > + port->port_no); > } > } > } > - cycles_count_end(non_pmd, PMD_CYCLES_POLL_IDLE); > pmd_thread_ctx_time_update(non_pmd); > dpif_netdev_xps_revalidate_pmd(non_pmd, false); > ovs_mutex_unlock(&dp->non_pmd_mutex); > @@ -4083,18 +4042,13 @@ reload: > lc = UINT_MAX; > } > > - cycles_count_start(pmd); > for (;;) { > uint64_t iter_packets = 0; > - pmd_perf_start_iteration(s, pmd->ctx.last_cycles); > + pmd_perf_start_iteration(s); > for (i = 0; i < poll_cnt; i++) { > process_packets = > - dp_netdev_process_rxq_port(pmd, poll_list[i].rxq->rx, > + dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, > poll_list[i].port_no); > - cycles_count_intermediate(pmd, poll_list[i].rxq, > - process_packets > - ? PMD_CYCLES_POLL_BUSY > - : PMD_CYCLES_POLL_IDLE); > iter_packets += process_packets; > } > > @@ -4116,13 +4070,10 @@ reload: > if (reload) { > break; > } > - cycles_count_intermediate(pmd, NULL, PMD_CYCLES_OVERHEAD); > } > - pmd_perf_end_iteration(s, pmd->ctx.last_cycles, iter_packets); > + pmd_perf_end_iteration(s, iter_packets); > } > > - cycles_count_end(pmd, PMD_CYCLES_OVERHEAD); > - > poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); > exiting = latch_is_set(&pmd->exit_latch); > /* Signal here to make sure the pmd finishes > @@ -5067,9 +5018,7 @@ handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, > ovs_mutex_unlock(&pmd->flow_mutex); > emc_probabilistic_insert(pmd, key, netdev_flow); > } > - /* Only error ENOSPC can reach here. We process the packet but do not > - * install a datapath flow. Treat as successful. */ > - return 0; > + return error; > } > > static inline void >
Hi Ilya, Please find answers inline. Regards, Jan > -----Original Message----- > From: Ilya Maximets [mailto:i.maximets@samsung.com] > Sent: Wednesday, 10 January, 2018 12:53 > To: Jan Scheurich <jan.scheurich@ericsson.com>; dev@openvswitch.org > Cc: ktraynor@redhat.com; ian.stokes@intel.com > Subject: Re: [PATCH v6 2/2] dpif-netdev: Refactor cycle counting > > PMD_CYCLES_UPCALL > PMD_CYCLES_POLL_IDLE > PMD_CYCLES_POLL_BUSY > are not used in current series and should be removed. OK > > PMD_CYCLES_OVERHEAD implemented in patch #1 and removed in patch #2. > Should be removed from patch #1 to simplify it. It is needed for correctness in patch #1. > > Other comments inline. > > Best regards, Ilya Maximets. > > On 09.01.2018 12:26, Jan Scheurich wrote: > > Simplify the historically grown TSC cycle counting in PMD threads. > > Cycles are currently counted for the following purposes: > > > > 1. Measure PMD ustilization > > > > PMD utilization is defined as ratio of cycles spent in busy iterations > > (at least one packet received or sent) over the total number of cycles. > > > > This is already done in pmd_perf_start_iteration() and > > pmd_perf_end_iteration() based on a TSC timestamp saved in current > > iteration at start_iteration() and the actual TSC at end_iteration(). > > No dependency on intermediate cycle accounting. > > > > 2. Measure the processing load per RX queue > > > > This comprises cycles spend on polling and processing packets received > > from the rx queue. > > > > The previous scheme using cycles_count_start(), cycles_count_intermediate() > > and cycles-count_end() originally introduced to simplify cycle counting > > and saving calls to rte_get_tsc_cycles() was rather obscuring things. > > > > Replace by a nestable cycle_timer with with start and stop functions to > > embrace a code segment to be timed. The timed code may contain arbitrary > > nested cycle_timers. The duration of nested timers is excluded from the > > outer timer. > > > > The caller must ensure that each call to cycle_timer_start() is > > followed by a call to cycle_timer_end(). Failure to do so will lead to > > assertion failure or a memory leak. > > > > The new cycle_timer is used to measure the processing cycles per rx queue. > > This is not yet strictly necessary but will be made use of in subsequent > > commits. > > > > All cycle count functions and data are relocated to module > > dpif-netdev-perf. > > > > Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com> > > > > --- > > lib/dpif-netdev-perf.h | 94 +++++++++++++++++++++++++++++++++++--- > > lib/dpif-netdev.c | 121 ++++++++++++++----------------------------------- > > 2 files changed, 123 insertions(+), 92 deletions(-) > > > > diff --git a/lib/dpif-netdev-perf.h b/lib/dpif-netdev-perf.h > > index b31bf49..a6dfcfd 100644 > > --- a/lib/dpif-netdev-perf.h > > +++ b/lib/dpif-netdev-perf.h > > @@ -90,9 +90,92 @@ struct pmd_perf_stats { > > period. */ > > uint64_t last_tsc; /* Start of the current PMD iteration in TSC > > cycles. */ > > + uint64_t last_cycles; /* Latest TSC time stamp taken in PMD. */ > > + struct cycle_timer *cur_timer; > > + /* If non-NULL, outermost cycle timer currently > > + running in PMD. */ > > Same comment as in patch #1. > This one looks the most ugly, but could be written in one line above the element. Yes > > > struct pmd_counters counters; > > }; > > > > +/* Support for accurate timing of PMD execution on TSC clock cycle level. > > + * These functions are intended to be invoked in the context of pmd threads. */ > > + > > +/* Read the TSC cycle register and cache it. Any function not requiring clock > > + * cycle accuracy should read the cached value using cycles_counter_get(). */ > > + > > +static inline uint64_t > > +cycles_counter_update(struct pmd_perf_stats *s) > > +{ > > +#ifdef DPDK_NETDEV > > + return s->last_cycles = rte_get_tsc_cycles(); > > +#else > > + return s->last_cycles = 0; > > +#endif > > +} > > + > > +static inline uint64_t > > +cycles_counter_get(struct pmd_perf_stats *s) > > +{ > > + return s->last_cycles; > > +} > > This function not used. True, but needed for a complete cycle count API for dpif_netdev, now that the last_cycles has been moved to struct pmd_perf_stats, which is sort of "opaque" outside of dpif-netdev-perf (see comment above). > > > + > > +/* A nestable timer for measuring execution time in TSC cycles. > > + * > > + * Usage: > > + * struct cycle_timer timer; > > + * > > + * cycle_timer_start(pmd, &timer); > > + * <Timed execution> > > + * uint64_t cycles = cycle_timer_stop(pmd, &timer); > > + * > > + * The caller must guarantee that a call to cycle_timer_start() is always > > + * paired with a call to cycle_stimer_stop(). > > + * > > + * Is is possible to have nested cycles timers within the timed code. The > > + * execution time measured by the nested timers is excluded from the time > > + * measured by the embracing timer. > > + */ > > + > > +struct cycle_timer { > > + uint64_t start; > > + uint64_t suspended; > > + struct cycle_timer *interrupted; > > +}; > > + > > +static inline void > > +cycle_timer_start(struct pmd_perf_stats *s, > > + struct cycle_timer *timer) > > +{ > > + struct cycle_timer *cur_timer = s->cur_timer; > > + uint64_t now = cycles_counter_update(s); > > + > > + if (cur_timer) { > > + cur_timer->suspended = now; > > + } > > + timer->interrupted = cur_timer; > > + timer->start = now; > > + timer->suspended = 0; > > + s->cur_timer = timer; > > +} > > + > > +static inline uint64_t > > +cycle_timer_stop(struct pmd_perf_stats *s, > > + struct cycle_timer *timer) > > +{ > > + /* Assert that this is the current cycle timer. */ > > + ovs_assert(s->cur_timer == timer); > > + uint64_t now = cycles_counter_update(s); > > + struct cycle_timer *intr_timer = timer->interrupted; > > + > > + if (intr_timer) { > > + /* Adjust the start offset by the suspended cycles. */ > > + intr_timer->start += now - intr_timer->suspended; > > + } > > + /* Restore suspended timer, if any. */ > > + s->cur_timer = intr_timer; > > + return now - timer->start; > > +} > > + > > void pmd_perf_stats_init(struct pmd_perf_stats *s); > > void pmd_perf_stats_clear(struct pmd_perf_stats *s); > > void pmd_perf_read_counters(struct pmd_perf_stats *s, > > @@ -109,23 +192,22 @@ pmd_perf_update_counter(struct pmd_perf_stats *s, > > } > > > > static inline void > > -pmd_perf_start_iteration(struct pmd_perf_stats *s, uint64_t now_tsc) > > +pmd_perf_start_iteration(struct pmd_perf_stats *s) > > { > > if (OVS_UNLIKELY(s->start_ms == 0)) { > > /* Stats not yet fully initialized. */ > > return; > > } > > - s->last_tsc = now_tsc; > > + s->last_tsc = s->last_cycles; > > } > > > > static inline void > > -pmd_perf_end_iteration(struct pmd_perf_stats *s, uint64_t now_tsc, > > - int packets) > > +pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets) > > { > > - uint64_t cycles = now_tsc - s->last_tsc; > > + uint64_t cycles = cycles_counter_update(s) - s->last_tsc; > > > > /* No need for atomic updates as only invoked within PMD thread. */ > > - if (packets > 0) { > > + if (rx_packets > 0) { > > I guess, this renaming should be done in patch #1. Does that really matter? > > > s->counters.n[PMD_CYCLES_ITER_BUSY] += cycles; > > } else { > > s->counters.n[PMD_CYCLES_ITER_IDLE] += cycles; > > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c > > index ad050f2..709d7ee 100644 > > --- a/lib/dpif-netdev.c > > +++ b/lib/dpif-netdev.c > > @@ -509,8 +509,6 @@ struct tx_port { > > struct dp_netdev_pmd_thread_ctx { > > /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ > > long long now; > > - /* Used to count cycles. See 'cycles_count_end()' */ > > - unsigned long long last_cycles; > > }; > > > > /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate > > @@ -873,8 +871,8 @@ pmd_info_show_stats(struct ds *reply, > > ds_put_format(reply, > > "\tavg processing cycles per packet: " > > "%.02f (%"PRIu64"/%"PRIu64")\n", > > - stats[PMD_CYCLES_POLL_BUSY] / (double) total_packets, > > - stats[PMD_CYCLES_POLL_BUSY], total_packets); > > + stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, > > + stats[PMD_CYCLES_ITER_BUSY], total_packets); > > This is unclear. Why semantics changed? Is it typo or this should be done in patch #1? Will move this correction to patch #1. > > > } > > > > static int > > @@ -3110,64 +3108,20 @@ dp_netdev_actions_free(struct dp_netdev_actions *actions) > > free(actions); > > } > > > > -static inline unsigned long long > > -cycles_counter(void) > > -{ > > -#ifdef DPDK_NETDEV > > - return rte_get_tsc_cycles(); > > -#else > > - return 0; > > -#endif > > -} > > - > > -/* Fake mutex to make sure that the calls to cycles_count_* are balanced */ > > -extern struct ovs_mutex cycles_counter_fake_mutex; > > - > > -/* Start counting cycles. Must be followed by 'cycles_count_end()' */ > > -static inline void > > -cycles_count_start(struct dp_netdev_pmd_thread *pmd) > > - OVS_ACQUIRES(&cycles_counter_fake_mutex) > > - OVS_NO_THREAD_SAFETY_ANALYSIS > > -{ > > - pmd->ctx.last_cycles = cycles_counter(); > > -} > > - > > -/* Stop counting cycles and add them to the counter 'type' */ > > -static inline void > > -cycles_count_end(struct dp_netdev_pmd_thread *pmd, > > - enum pmd_stat_type type) > > - OVS_RELEASES(&cycles_counter_fake_mutex) > > - OVS_NO_THREAD_SAFETY_ANALYSIS > > -{ > > - unsigned long long interval = cycles_counter() - pmd->ctx.last_cycles; > > - > > - pmd_perf_update_counter(&pmd->perf_stats, type, interval); > > -} > > - > > -/* Calculate the intermediate cycle result and add to the counter 'type' */ > > -static inline void > > -cycles_count_intermediate(struct dp_netdev_pmd_thread *pmd, > > - struct dp_netdev_rxq *rxq, > > - enum pmd_stat_type type) > > - OVS_NO_THREAD_SAFETY_ANALYSIS > > +static void > > +dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, > > + enum rxq_cycles_counter_type type, > > + unsigned long long cycles) > > { > > - unsigned long long new_cycles = cycles_counter(); > > - unsigned long long interval = new_cycles - pmd->ctx.last_cycles; > > - pmd->ctx.last_cycles = new_cycles; > > - > > - pmd_perf_update_counter(&pmd->perf_stats, type, interval); > > - if (rxq && (type == PMD_CYCLES_POLL_BUSY)) { > > - /* Add to the amount of current processing cycles. */ > > - non_atomic_ullong_add(&rxq->cycles[RXQ_CYCLES_PROC_CURR], interval); > > - } > > + atomic_store_relaxed(&rx->cycles[type], cycles); > > } > > > > static void > > -dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, > > +dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, > > enum rxq_cycles_counter_type type, > > unsigned long long cycles) > > { > > - atomic_store_relaxed(&rx->cycles[type], cycles); > > + non_atomic_ullong_add(&rx->cycles[type], cycles); > > } > > > > static uint64_t > > @@ -3235,27 +3189,40 @@ dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd) > > > > static int > > dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, > > - struct netdev_rxq *rx, > > + struct dp_netdev_rxq *rxq, > > odp_port_t port_no) > > { > > struct dp_packet_batch batch; > > + struct cycle_timer timer; > > int error; > > int batch_cnt = 0; > > > > + /* Measure duration for polling and processing rx burst. */ > > + cycle_timer_start(&pmd->perf_stats, &timer); > > dp_packet_batch_init(&batch); > > - error = netdev_rxq_recv(rx, &batch); > > + error = netdev_rxq_recv(rxq->rx, &batch); > > if (!error) { > > + /* At least one packet received. */ > > *recirc_depth_get() = 0; > > pmd_thread_ctx_time_update(pmd); > > > > batch_cnt = batch.count; > > dp_netdev_input(pmd, &batch, port_no); > > dp_netdev_pmd_flush_output_packets(pmd); > > - } else if (error != EAGAIN && error != EOPNOTSUPP) { > > - static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); > > > > - VLOG_ERR_RL(&rl, "error receiving data from %s: %s", > > - netdev_rxq_get_name(rx), ovs_strerror(error)); > > + /* Assign processing cycles to rx queue. */ > > + uint64_t cycles = cycle_timer_stop(&pmd->perf_stats, &timer); > > + dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); > > + > > + } else { > > + /* Discard cycles. */ > > + cycle_timer_stop(&pmd->perf_stats, &timer); > > + if (error != EAGAIN && error != EOPNOTSUPP) { > > + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); > > + > > + VLOG_ERR_RL(&rl, "error receiving data from %s: %s", > > + netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); > > + } > > } > > > > return batch_cnt; > > @@ -3881,30 +3848,22 @@ dpif_netdev_run(struct dpif *dpif) > > struct dp_netdev *dp = get_dp_netdev(dpif); > > struct dp_netdev_pmd_thread *non_pmd; > > uint64_t new_tnl_seq; > > - int process_packets = 0; > > > > ovs_mutex_lock(&dp->port_mutex); > > non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); > > if (non_pmd) { > > ovs_mutex_lock(&dp->non_pmd_mutex); > > - cycles_count_start(non_pmd); > > HMAP_FOR_EACH (port, node, &dp->ports) { > > if (!netdev_is_pmd(port->netdev)) { > > int i; > > > > for (i = 0; i < port->n_rxq; i++) { > > - process_packets = > > - dp_netdev_process_rxq_port(non_pmd, > > - port->rxqs[i].rx, > > - port->port_no); > > - cycles_count_intermediate(non_pmd, NULL, > > - process_packets > > - ? PMD_CYCLES_POLL_BUSY > > - : PMD_CYCLES_POLL_IDLE); > > + dp_netdev_process_rxq_port(non_pmd, > > + &port->rxqs[i], > > + port->port_no); > > } > > } > > } > > - cycles_count_end(non_pmd, PMD_CYCLES_POLL_IDLE); > > pmd_thread_ctx_time_update(non_pmd); > > dpif_netdev_xps_revalidate_pmd(non_pmd, false); > > ovs_mutex_unlock(&dp->non_pmd_mutex); > > @@ -4083,18 +4042,13 @@ reload: > > lc = UINT_MAX; > > } > > > > - cycles_count_start(pmd); > > for (;;) { > > uint64_t iter_packets = 0; > > - pmd_perf_start_iteration(s, pmd->ctx.last_cycles); > > + pmd_perf_start_iteration(s); > > for (i = 0; i < poll_cnt; i++) { > > process_packets = > > - dp_netdev_process_rxq_port(pmd, poll_list[i].rxq->rx, > > + dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, > > poll_list[i].port_no); > > - cycles_count_intermediate(pmd, poll_list[i].rxq, > > - process_packets > > - ? PMD_CYCLES_POLL_BUSY > > - : PMD_CYCLES_POLL_IDLE); > > iter_packets += process_packets; > > } > > > > @@ -4116,13 +4070,10 @@ reload: > > if (reload) { > > break; > > } > > - cycles_count_intermediate(pmd, NULL, PMD_CYCLES_OVERHEAD); > > } > > - pmd_perf_end_iteration(s, pmd->ctx.last_cycles, iter_packets); > > + pmd_perf_end_iteration(s, iter_packets); > > } > > > > - cycles_count_end(pmd, PMD_CYCLES_OVERHEAD); > > - > > poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); > > exiting = latch_is_set(&pmd->exit_latch); > > /* Signal here to make sure the pmd finishes > > @@ -5067,9 +5018,7 @@ handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, > > ovs_mutex_unlock(&pmd->flow_mutex); > > emc_probabilistic_insert(pmd, key, netdev_flow); > > } > > - /* Only error ENOSPC can reach here. We process the packet but do not > > - * install a datapath flow. Treat as successful. */ > > - return 0; > > + return error; > > } > > > > static inline void > >
On 11.01.2018 02:34, Jan Scheurich wrote: > Hi Ilya, > > Please find answers inline. > > Regards, Jan > >> -----Original Message----- >> From: Ilya Maximets [mailto:i.maximets@samsung.com] >> Sent: Wednesday, 10 January, 2018 12:53 >> To: Jan Scheurich <jan.scheurich@ericsson.com>; dev@openvswitch.org >> Cc: ktraynor@redhat.com; ian.stokes@intel.com >> Subject: Re: [PATCH v6 2/2] dpif-netdev: Refactor cycle counting >> >> PMD_CYCLES_UPCALL >> PMD_CYCLES_POLL_IDLE >> PMD_CYCLES_POLL_BUSY >> are not used in current series and should be removed. > > OK > >> >> PMD_CYCLES_OVERHEAD implemented in patch #1 and removed in patch #2. >> Should be removed from patch #1 to simplify it. > > It is needed for correctness in patch #1. > >> >> Other comments inline. >> >> Best regards, Ilya Maximets. >> >> On 09.01.2018 12:26, Jan Scheurich wrote: >>> Simplify the historically grown TSC cycle counting in PMD threads. >>> Cycles are currently counted for the following purposes: >>> >>> 1. Measure PMD ustilization >>> >>> PMD utilization is defined as ratio of cycles spent in busy iterations >>> (at least one packet received or sent) over the total number of cycles. >>> >>> This is already done in pmd_perf_start_iteration() and >>> pmd_perf_end_iteration() based on a TSC timestamp saved in current >>> iteration at start_iteration() and the actual TSC at end_iteration(). >>> No dependency on intermediate cycle accounting. >>> >>> 2. Measure the processing load per RX queue >>> >>> This comprises cycles spend on polling and processing packets received >>> from the rx queue. >>> >>> The previous scheme using cycles_count_start(), cycles_count_intermediate() >>> and cycles-count_end() originally introduced to simplify cycle counting >>> and saving calls to rte_get_tsc_cycles() was rather obscuring things. >>> >>> Replace by a nestable cycle_timer with with start and stop functions to >>> embrace a code segment to be timed. The timed code may contain arbitrary >>> nested cycle_timers. The duration of nested timers is excluded from the >>> outer timer. >>> >>> The caller must ensure that each call to cycle_timer_start() is >>> followed by a call to cycle_timer_end(). Failure to do so will lead to >>> assertion failure or a memory leak. >>> >>> The new cycle_timer is used to measure the processing cycles per rx queue. >>> This is not yet strictly necessary but will be made use of in subsequent >>> commits. >>> >>> All cycle count functions and data are relocated to module >>> dpif-netdev-perf. >>> >>> Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com> >>> >>> --- >>> lib/dpif-netdev-perf.h | 94 +++++++++++++++++++++++++++++++++++--- >>> lib/dpif-netdev.c | 121 ++++++++++++++----------------------------------- >>> 2 files changed, 123 insertions(+), 92 deletions(-) >>> >>> diff --git a/lib/dpif-netdev-perf.h b/lib/dpif-netdev-perf.h >>> index b31bf49..a6dfcfd 100644 >>> --- a/lib/dpif-netdev-perf.h >>> +++ b/lib/dpif-netdev-perf.h >>> @@ -90,9 +90,92 @@ struct pmd_perf_stats { >>> period. */ >>> uint64_t last_tsc; /* Start of the current PMD iteration in TSC >>> cycles. */ >>> + uint64_t last_cycles; /* Latest TSC time stamp taken in PMD. */ >>> + struct cycle_timer *cur_timer; >>> + /* If non-NULL, outermost cycle timer currently >>> + running in PMD. */ >> >> Same comment as in patch #1. >> This one looks the most ugly, but could be written in one line above the element. > > Yes > >> >>> struct pmd_counters counters; >>> }; >>> >>> +/* Support for accurate timing of PMD execution on TSC clock cycle level. >>> + * These functions are intended to be invoked in the context of pmd threads. */ >>> + >>> +/* Read the TSC cycle register and cache it. Any function not requiring clock >>> + * cycle accuracy should read the cached value using cycles_counter_get(). */ >>> + >>> +static inline uint64_t >>> +cycles_counter_update(struct pmd_perf_stats *s) >>> +{ >>> +#ifdef DPDK_NETDEV >>> + return s->last_cycles = rte_get_tsc_cycles(); >>> +#else >>> + return s->last_cycles = 0; >>> +#endif >>> +} >>> + >>> +static inline uint64_t >>> +cycles_counter_get(struct pmd_perf_stats *s) >>> +{ >>> + return s->last_cycles; >>> +} >> >> This function not used. > > True, but needed for a complete cycle count API for dpif_netdev, now that the last_cycles has been moved to struct pmd_perf_stats, which is sort of "opaque" outside of dpif-netdev-perf (see comment above). OK. > >> >>> + >>> +/* A nestable timer for measuring execution time in TSC cycles. >>> + * >>> + * Usage: >>> + * struct cycle_timer timer; >>> + * >>> + * cycle_timer_start(pmd, &timer); >>> + * <Timed execution> >>> + * uint64_t cycles = cycle_timer_stop(pmd, &timer); >>> + * >>> + * The caller must guarantee that a call to cycle_timer_start() is always >>> + * paired with a call to cycle_stimer_stop(). >>> + * >>> + * Is is possible to have nested cycles timers within the timed code. The >>> + * execution time measured by the nested timers is excluded from the time >>> + * measured by the embracing timer. >>> + */ >>> + >>> +struct cycle_timer { >>> + uint64_t start; >>> + uint64_t suspended; >>> + struct cycle_timer *interrupted; >>> +}; >>> + >>> +static inline void >>> +cycle_timer_start(struct pmd_perf_stats *s, >>> + struct cycle_timer *timer) >>> +{ >>> + struct cycle_timer *cur_timer = s->cur_timer; >>> + uint64_t now = cycles_counter_update(s); >>> + >>> + if (cur_timer) { >>> + cur_timer->suspended = now; >>> + } >>> + timer->interrupted = cur_timer; >>> + timer->start = now; >>> + timer->suspended = 0; >>> + s->cur_timer = timer; >>> +} >>> + >>> +static inline uint64_t >>> +cycle_timer_stop(struct pmd_perf_stats *s, >>> + struct cycle_timer *timer) >>> +{ >>> + /* Assert that this is the current cycle timer. */ >>> + ovs_assert(s->cur_timer == timer); >>> + uint64_t now = cycles_counter_update(s); >>> + struct cycle_timer *intr_timer = timer->interrupted; >>> + >>> + if (intr_timer) { >>> + /* Adjust the start offset by the suspended cycles. */ >>> + intr_timer->start += now - intr_timer->suspended; >>> + } >>> + /* Restore suspended timer, if any. */ >>> + s->cur_timer = intr_timer; >>> + return now - timer->start; >>> +} >>> + >>> void pmd_perf_stats_init(struct pmd_perf_stats *s); >>> void pmd_perf_stats_clear(struct pmd_perf_stats *s); >>> void pmd_perf_read_counters(struct pmd_perf_stats *s, >>> @@ -109,23 +192,22 @@ pmd_perf_update_counter(struct pmd_perf_stats *s, >>> } >>> >>> static inline void >>> -pmd_perf_start_iteration(struct pmd_perf_stats *s, uint64_t now_tsc) >>> +pmd_perf_start_iteration(struct pmd_perf_stats *s) >>> { >>> if (OVS_UNLIKELY(s->start_ms == 0)) { >>> /* Stats not yet fully initialized. */ >>> return; >>> } >>> - s->last_tsc = now_tsc; >>> + s->last_tsc = s->last_cycles; >>> } >>> >>> static inline void >>> -pmd_perf_end_iteration(struct pmd_perf_stats *s, uint64_t now_tsc, >>> - int packets) >>> +pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets) >>> { >>> - uint64_t cycles = now_tsc - s->last_tsc; >>> + uint64_t cycles = cycles_counter_update(s) - s->last_tsc; >>> >>> /* No need for atomic updates as only invoked within PMD thread. */ >>> - if (packets > 0) { >>> + if (rx_packets > 0) { >> >> I guess, this renaming should be done in patch #1. > > Does that really matter? It saves few lines of the patch. I see no reason to introduce this local variable and rename it in the next patch in so insignificant way. > >> >>> s->counters.n[PMD_CYCLES_ITER_BUSY] += cycles; >>> } else { >>> s->counters.n[PMD_CYCLES_ITER_IDLE] += cycles; >>> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c >>> index ad050f2..709d7ee 100644 >>> --- a/lib/dpif-netdev.c >>> +++ b/lib/dpif-netdev.c >>> @@ -509,8 +509,6 @@ struct tx_port { >>> struct dp_netdev_pmd_thread_ctx { >>> /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ >>> long long now; >>> - /* Used to count cycles. See 'cycles_count_end()' */ >>> - unsigned long long last_cycles; >>> }; >>> >>> /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate >>> @@ -873,8 +871,8 @@ pmd_info_show_stats(struct ds *reply, >>> ds_put_format(reply, >>> "\tavg processing cycles per packet: " >>> "%.02f (%"PRIu64"/%"PRIu64")\n", >>> - stats[PMD_CYCLES_POLL_BUSY] / (double) total_packets, >>> - stats[PMD_CYCLES_POLL_BUSY], total_packets); >>> + stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, >>> + stats[PMD_CYCLES_ITER_BUSY], total_packets); >> >> This is unclear. Why semantics changed? Is it typo or this should be done in patch #1? > > Will move this correction to patch #1. > >> >>> } >>> >>> static int >>> @@ -3110,64 +3108,20 @@ dp_netdev_actions_free(struct dp_netdev_actions *actions) >>> free(actions); >>> } >>> > >>> -static inline unsigned long long >>> -cycles_counter(void) >>> -{ >>> -#ifdef DPDK_NETDEV >>> - return rte_get_tsc_cycles(); >>> -#else >>> - return 0; >>> -#endif >>> -} >>> - >>> -/* Fake mutex to make sure that the calls to cycles_count_* are balanced */ >>> -extern struct ovs_mutex cycles_counter_fake_mutex; >>> - >>> -/* Start counting cycles. Must be followed by 'cycles_count_end()' */ >>> -static inline void >>> -cycles_count_start(struct dp_netdev_pmd_thread *pmd) >>> - OVS_ACQUIRES(&cycles_counter_fake_mutex) >>> - OVS_NO_THREAD_SAFETY_ANALYSIS >>> -{ >>> - pmd->ctx.last_cycles = cycles_counter(); >>> -} >>> - >>> -/* Stop counting cycles and add them to the counter 'type' */ >>> -static inline void >>> -cycles_count_end(struct dp_netdev_pmd_thread *pmd, >>> - enum pmd_stat_type type) >>> - OVS_RELEASES(&cycles_counter_fake_mutex) >>> - OVS_NO_THREAD_SAFETY_ANALYSIS >>> -{ >>> - unsigned long long interval = cycles_counter() - pmd->ctx.last_cycles; >>> - >>> - pmd_perf_update_counter(&pmd->perf_stats, type, interval); >>> -} >>> - >>> -/* Calculate the intermediate cycle result and add to the counter 'type' */ >>> -static inline void >>> -cycles_count_intermediate(struct dp_netdev_pmd_thread *pmd, >>> - struct dp_netdev_rxq *rxq, >>> - enum pmd_stat_type type) >>> - OVS_NO_THREAD_SAFETY_ANALYSIS >>> +static void >>> +dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, >>> + enum rxq_cycles_counter_type type, >>> + unsigned long long cycles) >>> { >>> - unsigned long long new_cycles = cycles_counter(); >>> - unsigned long long interval = new_cycles - pmd->ctx.last_cycles; >>> - pmd->ctx.last_cycles = new_cycles; >>> - >>> - pmd_perf_update_counter(&pmd->perf_stats, type, interval); >>> - if (rxq && (type == PMD_CYCLES_POLL_BUSY)) { >>> - /* Add to the amount of current processing cycles. */ >>> - non_atomic_ullong_add(&rxq->cycles[RXQ_CYCLES_PROC_CURR], interval); >>> - } >>> + atomic_store_relaxed(&rx->cycles[type], cycles); >>> } >>> >>> static void >>> -dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, >>> +dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, >>> enum rxq_cycles_counter_type type, >>> unsigned long long cycles) >>> { >>> - atomic_store_relaxed(&rx->cycles[type], cycles); >>> + non_atomic_ullong_add(&rx->cycles[type], cycles); >>> } >>> >>> static uint64_t >>> @@ -3235,27 +3189,40 @@ dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd) >>> >>> static int >>> dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, >>> - struct netdev_rxq *rx, >>> + struct dp_netdev_rxq *rxq, >>> odp_port_t port_no) >>> { >>> struct dp_packet_batch batch; >>> + struct cycle_timer timer; >>> int error; >>> int batch_cnt = 0; >>> >>> + /* Measure duration for polling and processing rx burst. */ >>> + cycle_timer_start(&pmd->perf_stats, &timer); >>> dp_packet_batch_init(&batch); >>> - error = netdev_rxq_recv(rx, &batch); >>> + error = netdev_rxq_recv(rxq->rx, &batch); >>> if (!error) { >>> + /* At least one packet received. */ >>> *recirc_depth_get() = 0; >>> pmd_thread_ctx_time_update(pmd); >>> >>> batch_cnt = batch.count; >>> dp_netdev_input(pmd, &batch, port_no); >>> dp_netdev_pmd_flush_output_packets(pmd); >>> - } else if (error != EAGAIN && error != EOPNOTSUPP) { >>> - static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); >>> >>> - VLOG_ERR_RL(&rl, "error receiving data from %s: %s", >>> - netdev_rxq_get_name(rx), ovs_strerror(error)); >>> + /* Assign processing cycles to rx queue. */ >>> + uint64_t cycles = cycle_timer_stop(&pmd->perf_stats, &timer); >>> + dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); >>> + >>> + } else { >>> + /* Discard cycles. */ >>> + cycle_timer_stop(&pmd->perf_stats, &timer); >>> + if (error != EAGAIN && error != EOPNOTSUPP) { >>> + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); >>> + >>> + VLOG_ERR_RL(&rl, "error receiving data from %s: %s", >>> + netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); >>> + } >>> } >>> >>> return batch_cnt; >>> @@ -3881,30 +3848,22 @@ dpif_netdev_run(struct dpif *dpif) >>> struct dp_netdev *dp = get_dp_netdev(dpif); >>> struct dp_netdev_pmd_thread *non_pmd; >>> uint64_t new_tnl_seq; >>> - int process_packets = 0; >>> >>> ovs_mutex_lock(&dp->port_mutex); >>> non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); >>> if (non_pmd) { >>> ovs_mutex_lock(&dp->non_pmd_mutex); >>> - cycles_count_start(non_pmd); >>> HMAP_FOR_EACH (port, node, &dp->ports) { >>> if (!netdev_is_pmd(port->netdev)) { >>> int i; >>> >>> for (i = 0; i < port->n_rxq; i++) { >>> - process_packets = >>> - dp_netdev_process_rxq_port(non_pmd, >>> - port->rxqs[i].rx, >>> - port->port_no); >>> - cycles_count_intermediate(non_pmd, NULL, >>> - process_packets >>> - ? PMD_CYCLES_POLL_BUSY >>> - : PMD_CYCLES_POLL_IDLE); >>> + dp_netdev_process_rxq_port(non_pmd, >>> + &port->rxqs[i], >>> + port->port_no); >>> } >>> } >>> } >>> - cycles_count_end(non_pmd, PMD_CYCLES_POLL_IDLE); >>> pmd_thread_ctx_time_update(non_pmd); >>> dpif_netdev_xps_revalidate_pmd(non_pmd, false); >>> ovs_mutex_unlock(&dp->non_pmd_mutex); >>> @@ -4083,18 +4042,13 @@ reload: >>> lc = UINT_MAX; >>> } >>> >>> - cycles_count_start(pmd); >>> for (;;) { >>> uint64_t iter_packets = 0; >>> - pmd_perf_start_iteration(s, pmd->ctx.last_cycles); >>> + pmd_perf_start_iteration(s); >>> for (i = 0; i < poll_cnt; i++) { >>> process_packets = >>> - dp_netdev_process_rxq_port(pmd, poll_list[i].rxq->rx, >>> + dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, >>> poll_list[i].port_no); >>> - cycles_count_intermediate(pmd, poll_list[i].rxq, >>> - process_packets >>> - ? PMD_CYCLES_POLL_BUSY >>> - : PMD_CYCLES_POLL_IDLE); >>> iter_packets += process_packets; >>> } >>> >>> @@ -4116,13 +4070,10 @@ reload: >>> if (reload) { >>> break; >>> } >>> - cycles_count_intermediate(pmd, NULL, PMD_CYCLES_OVERHEAD); >>> } >>> - pmd_perf_end_iteration(s, pmd->ctx.last_cycles, iter_packets); >>> + pmd_perf_end_iteration(s, iter_packets); >>> } >>> >>> - cycles_count_end(pmd, PMD_CYCLES_OVERHEAD); >>> - >>> poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); >>> exiting = latch_is_set(&pmd->exit_latch); >>> /* Signal here to make sure the pmd finishes >>> @@ -5067,9 +5018,7 @@ handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, >>> ovs_mutex_unlock(&pmd->flow_mutex); >>> emc_probabilistic_insert(pmd, key, netdev_flow); >>> } >>> - /* Only error ENOSPC can reach here. We process the packet but do not >>> - * install a datapath flow. Treat as successful. */ >>> - return 0; >>> + return error; >>> } >>> >>> static inline void >>>
diff --git a/lib/dpif-netdev-perf.h b/lib/dpif-netdev-perf.h index b31bf49..a6dfcfd 100644 --- a/lib/dpif-netdev-perf.h +++ b/lib/dpif-netdev-perf.h @@ -90,9 +90,92 @@ struct pmd_perf_stats { period. */ uint64_t last_tsc; /* Start of the current PMD iteration in TSC cycles. */ + uint64_t last_cycles; /* Latest TSC time stamp taken in PMD. */ + struct cycle_timer *cur_timer; + /* If non-NULL, outermost cycle timer currently + running in PMD. */ struct pmd_counters counters; }; +/* Support for accurate timing of PMD execution on TSC clock cycle level. + * These functions are intended to be invoked in the context of pmd threads. */ + +/* Read the TSC cycle register and cache it. Any function not requiring clock + * cycle accuracy should read the cached value using cycles_counter_get(). */ + +static inline uint64_t +cycles_counter_update(struct pmd_perf_stats *s) +{ +#ifdef DPDK_NETDEV + return s->last_cycles = rte_get_tsc_cycles(); +#else + return s->last_cycles = 0; +#endif +} + +static inline uint64_t +cycles_counter_get(struct pmd_perf_stats *s) +{ + return s->last_cycles; +} + +/* A nestable timer for measuring execution time in TSC cycles. + * + * Usage: + * struct cycle_timer timer; + * + * cycle_timer_start(pmd, &timer); + * <Timed execution> + * uint64_t cycles = cycle_timer_stop(pmd, &timer); + * + * The caller must guarantee that a call to cycle_timer_start() is always + * paired with a call to cycle_stimer_stop(). + * + * Is is possible to have nested cycles timers within the timed code. The + * execution time measured by the nested timers is excluded from the time + * measured by the embracing timer. + */ + +struct cycle_timer { + uint64_t start; + uint64_t suspended; + struct cycle_timer *interrupted; +}; + +static inline void +cycle_timer_start(struct pmd_perf_stats *s, + struct cycle_timer *timer) +{ + struct cycle_timer *cur_timer = s->cur_timer; + uint64_t now = cycles_counter_update(s); + + if (cur_timer) { + cur_timer->suspended = now; + } + timer->interrupted = cur_timer; + timer->start = now; + timer->suspended = 0; + s->cur_timer = timer; +} + +static inline uint64_t +cycle_timer_stop(struct pmd_perf_stats *s, + struct cycle_timer *timer) +{ + /* Assert that this is the current cycle timer. */ + ovs_assert(s->cur_timer == timer); + uint64_t now = cycles_counter_update(s); + struct cycle_timer *intr_timer = timer->interrupted; + + if (intr_timer) { + /* Adjust the start offset by the suspended cycles. */ + intr_timer->start += now - intr_timer->suspended; + } + /* Restore suspended timer, if any. */ + s->cur_timer = intr_timer; + return now - timer->start; +} + void pmd_perf_stats_init(struct pmd_perf_stats *s); void pmd_perf_stats_clear(struct pmd_perf_stats *s); void pmd_perf_read_counters(struct pmd_perf_stats *s, @@ -109,23 +192,22 @@ pmd_perf_update_counter(struct pmd_perf_stats *s, } static inline void -pmd_perf_start_iteration(struct pmd_perf_stats *s, uint64_t now_tsc) +pmd_perf_start_iteration(struct pmd_perf_stats *s) { if (OVS_UNLIKELY(s->start_ms == 0)) { /* Stats not yet fully initialized. */ return; } - s->last_tsc = now_tsc; + s->last_tsc = s->last_cycles; } static inline void -pmd_perf_end_iteration(struct pmd_perf_stats *s, uint64_t now_tsc, - int packets) +pmd_perf_end_iteration(struct pmd_perf_stats *s, int rx_packets) { - uint64_t cycles = now_tsc - s->last_tsc; + uint64_t cycles = cycles_counter_update(s) - s->last_tsc; /* No need for atomic updates as only invoked within PMD thread. */ - if (packets > 0) { + if (rx_packets > 0) { s->counters.n[PMD_CYCLES_ITER_BUSY] += cycles; } else { s->counters.n[PMD_CYCLES_ITER_IDLE] += cycles; diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c index ad050f2..709d7ee 100644 --- a/lib/dpif-netdev.c +++ b/lib/dpif-netdev.c @@ -509,8 +509,6 @@ struct tx_port { struct dp_netdev_pmd_thread_ctx { /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ long long now; - /* Used to count cycles. See 'cycles_count_end()' */ - unsigned long long last_cycles; }; /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate @@ -873,8 +871,8 @@ pmd_info_show_stats(struct ds *reply, ds_put_format(reply, "\tavg processing cycles per packet: " "%.02f (%"PRIu64"/%"PRIu64")\n", - stats[PMD_CYCLES_POLL_BUSY] / (double) total_packets, - stats[PMD_CYCLES_POLL_BUSY], total_packets); + stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, + stats[PMD_CYCLES_ITER_BUSY], total_packets); } static int @@ -3110,64 +3108,20 @@ dp_netdev_actions_free(struct dp_netdev_actions *actions) free(actions); } -static inline unsigned long long -cycles_counter(void) -{ -#ifdef DPDK_NETDEV - return rte_get_tsc_cycles(); -#else - return 0; -#endif -} - -/* Fake mutex to make sure that the calls to cycles_count_* are balanced */ -extern struct ovs_mutex cycles_counter_fake_mutex; - -/* Start counting cycles. Must be followed by 'cycles_count_end()' */ -static inline void -cycles_count_start(struct dp_netdev_pmd_thread *pmd) - OVS_ACQUIRES(&cycles_counter_fake_mutex) - OVS_NO_THREAD_SAFETY_ANALYSIS -{ - pmd->ctx.last_cycles = cycles_counter(); -} - -/* Stop counting cycles and add them to the counter 'type' */ -static inline void -cycles_count_end(struct dp_netdev_pmd_thread *pmd, - enum pmd_stat_type type) - OVS_RELEASES(&cycles_counter_fake_mutex) - OVS_NO_THREAD_SAFETY_ANALYSIS -{ - unsigned long long interval = cycles_counter() - pmd->ctx.last_cycles; - - pmd_perf_update_counter(&pmd->perf_stats, type, interval); -} - -/* Calculate the intermediate cycle result and add to the counter 'type' */ -static inline void -cycles_count_intermediate(struct dp_netdev_pmd_thread *pmd, - struct dp_netdev_rxq *rxq, - enum pmd_stat_type type) - OVS_NO_THREAD_SAFETY_ANALYSIS +static void +dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, + enum rxq_cycles_counter_type type, + unsigned long long cycles) { - unsigned long long new_cycles = cycles_counter(); - unsigned long long interval = new_cycles - pmd->ctx.last_cycles; - pmd->ctx.last_cycles = new_cycles; - - pmd_perf_update_counter(&pmd->perf_stats, type, interval); - if (rxq && (type == PMD_CYCLES_POLL_BUSY)) { - /* Add to the amount of current processing cycles. */ - non_atomic_ullong_add(&rxq->cycles[RXQ_CYCLES_PROC_CURR], interval); - } + atomic_store_relaxed(&rx->cycles[type], cycles); } static void -dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, +dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, enum rxq_cycles_counter_type type, unsigned long long cycles) { - atomic_store_relaxed(&rx->cycles[type], cycles); + non_atomic_ullong_add(&rx->cycles[type], cycles); } static uint64_t @@ -3235,27 +3189,40 @@ dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd) static int dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, - struct netdev_rxq *rx, + struct dp_netdev_rxq *rxq, odp_port_t port_no) { struct dp_packet_batch batch; + struct cycle_timer timer; int error; int batch_cnt = 0; + /* Measure duration for polling and processing rx burst. */ + cycle_timer_start(&pmd->perf_stats, &timer); dp_packet_batch_init(&batch); - error = netdev_rxq_recv(rx, &batch); + error = netdev_rxq_recv(rxq->rx, &batch); if (!error) { + /* At least one packet received. */ *recirc_depth_get() = 0; pmd_thread_ctx_time_update(pmd); batch_cnt = batch.count; dp_netdev_input(pmd, &batch, port_no); dp_netdev_pmd_flush_output_packets(pmd); - } else if (error != EAGAIN && error != EOPNOTSUPP) { - static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); - VLOG_ERR_RL(&rl, "error receiving data from %s: %s", - netdev_rxq_get_name(rx), ovs_strerror(error)); + /* Assign processing cycles to rx queue. */ + uint64_t cycles = cycle_timer_stop(&pmd->perf_stats, &timer); + dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); + + } else { + /* Discard cycles. */ + cycle_timer_stop(&pmd->perf_stats, &timer); + if (error != EAGAIN && error != EOPNOTSUPP) { + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); + + VLOG_ERR_RL(&rl, "error receiving data from %s: %s", + netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); + } } return batch_cnt; @@ -3881,30 +3848,22 @@ dpif_netdev_run(struct dpif *dpif) struct dp_netdev *dp = get_dp_netdev(dpif); struct dp_netdev_pmd_thread *non_pmd; uint64_t new_tnl_seq; - int process_packets = 0; ovs_mutex_lock(&dp->port_mutex); non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); if (non_pmd) { ovs_mutex_lock(&dp->non_pmd_mutex); - cycles_count_start(non_pmd); HMAP_FOR_EACH (port, node, &dp->ports) { if (!netdev_is_pmd(port->netdev)) { int i; for (i = 0; i < port->n_rxq; i++) { - process_packets = - dp_netdev_process_rxq_port(non_pmd, - port->rxqs[i].rx, - port->port_no); - cycles_count_intermediate(non_pmd, NULL, - process_packets - ? PMD_CYCLES_POLL_BUSY - : PMD_CYCLES_POLL_IDLE); + dp_netdev_process_rxq_port(non_pmd, + &port->rxqs[i], + port->port_no); } } } - cycles_count_end(non_pmd, PMD_CYCLES_POLL_IDLE); pmd_thread_ctx_time_update(non_pmd); dpif_netdev_xps_revalidate_pmd(non_pmd, false); ovs_mutex_unlock(&dp->non_pmd_mutex); @@ -4083,18 +4042,13 @@ reload: lc = UINT_MAX; } - cycles_count_start(pmd); for (;;) { uint64_t iter_packets = 0; - pmd_perf_start_iteration(s, pmd->ctx.last_cycles); + pmd_perf_start_iteration(s); for (i = 0; i < poll_cnt; i++) { process_packets = - dp_netdev_process_rxq_port(pmd, poll_list[i].rxq->rx, + dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, poll_list[i].port_no); - cycles_count_intermediate(pmd, poll_list[i].rxq, - process_packets - ? PMD_CYCLES_POLL_BUSY - : PMD_CYCLES_POLL_IDLE); iter_packets += process_packets; } @@ -4116,13 +4070,10 @@ reload: if (reload) { break; } - cycles_count_intermediate(pmd, NULL, PMD_CYCLES_OVERHEAD); } - pmd_perf_end_iteration(s, pmd->ctx.last_cycles, iter_packets); + pmd_perf_end_iteration(s, iter_packets); } - cycles_count_end(pmd, PMD_CYCLES_OVERHEAD); - poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); exiting = latch_is_set(&pmd->exit_latch); /* Signal here to make sure the pmd finishes @@ -5067,9 +5018,7 @@ handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, ovs_mutex_unlock(&pmd->flow_mutex); emc_probabilistic_insert(pmd, key, netdev_flow); } - /* Only error ENOSPC can reach here. We process the packet but do not - * install a datapath flow. Treat as successful. */ - return 0; + return error; } static inline void
Simplify the historically grown TSC cycle counting in PMD threads. Cycles are currently counted for the following purposes: 1. Measure PMD ustilization PMD utilization is defined as ratio of cycles spent in busy iterations (at least one packet received or sent) over the total number of cycles. This is already done in pmd_perf_start_iteration() and pmd_perf_end_iteration() based on a TSC timestamp saved in current iteration at start_iteration() and the actual TSC at end_iteration(). No dependency on intermediate cycle accounting. 2. Measure the processing load per RX queue This comprises cycles spend on polling and processing packets received from the rx queue. The previous scheme using cycles_count_start(), cycles_count_intermediate() and cycles-count_end() originally introduced to simplify cycle counting and saving calls to rte_get_tsc_cycles() was rather obscuring things. Replace by a nestable cycle_timer with with start and stop functions to embrace a code segment to be timed. The timed code may contain arbitrary nested cycle_timers. The duration of nested timers is excluded from the outer timer. The caller must ensure that each call to cycle_timer_start() is followed by a call to cycle_timer_end(). Failure to do so will lead to assertion failure or a memory leak. The new cycle_timer is used to measure the processing cycles per rx queue. This is not yet strictly necessary but will be made use of in subsequent commits. All cycle count functions and data are relocated to module dpif-netdev-perf. Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com> --- lib/dpif-netdev-perf.h | 94 +++++++++++++++++++++++++++++++++++--- lib/dpif-netdev.c | 121 ++++++++++++++----------------------------------- 2 files changed, 123 insertions(+), 92 deletions(-)