1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Ericsson AB
8 #include <eventdev_pmd.h>
10 #include <rte_event_ring.h>
11 #include <rte_eventdev.h>
13 #define DSW_PMD_NAME RTE_STR(event_dsw)
15 #define DSW_MAX_PORTS (64)
16 #define DSW_MAX_PORT_DEQUEUE_DEPTH (128)
17 #define DSW_MAX_PORT_ENQUEUE_DEPTH (128)
18 #define DSW_MAX_PORT_OUT_BUFFER (32)
20 #define DSW_MAX_QUEUES (16)
22 #define DSW_MAX_EVENTS (16384)
24 /* Multiple 24-bit flow ids will map to the same DSW-level flow. The
25 * number of DSW flows should be high enough make it unlikely that
26 * flow ids of several large flows hash to the same DSW-level flow.
27 * Such collisions will limit parallelism and thus the number of cores
28 * that may be utilized. However, configuring a large number of DSW
29 * flows might potentially, depending on traffic and actual
30 * application flow id value range, result in each such DSW-level flow
31 * being very small. The effect of migrating such flows will be small,
32 * in terms amount of processing load redistributed. This will in turn
33 * reduce the load balancing speed, since flow migration rate has an
34 * upper limit. Code changes are required to allow > 32k DSW-level
37 #define DSW_MAX_FLOWS_BITS (13)
38 #define DSW_MAX_FLOWS (1<<(DSW_MAX_FLOWS_BITS))
39 #define DSW_MAX_FLOWS_MASK (DSW_MAX_FLOWS-1)
41 /* Eventdev RTE_SCHED_TYPE_PARALLEL doesn't have a concept of flows,
42 * but the 'dsw' scheduler (more or less) randomly assign flow id to
43 * events on parallel queues, to be able to reuse some of the
44 * migration mechanism and scheduling logic from
45 * RTE_SCHED_TYPE_ATOMIC. By moving one of the parallel "flows" from a
46 * particular port, the likely-hood of events being scheduled to this
47 * port is reduced, and thus a kind of statistical load balancing is
50 #define DSW_PARALLEL_FLOWS (1024)
52 /* 'Background tasks' are polling the control rings for *
53 * migration-related messages, or flush the output buffer (so
54 * buffered events doesn't linger too long). Shouldn't be too low,
55 * since the system won't benefit from the 'batching' effects from
56 * the output buffer, and shouldn't be too high, since it will make
57 * buffered events linger too long in case the port goes idle.
59 #define DSW_MAX_PORT_OPS_PER_BG_TASK (128)
61 /* Avoid making small 'loans' from the central in-flight event credit
62 * pool, to improve efficiency.
64 #define DSW_MIN_CREDIT_LOAN (64)
65 #define DSW_PORT_MAX_CREDITS (2*DSW_MIN_CREDIT_LOAN)
66 #define DSW_PORT_MIN_CREDITS (DSW_MIN_CREDIT_LOAN)
68 /* The rings are dimensioned so that all in-flight events can reside
69 * on any one of the port rings, to avoid the trouble of having to
70 * care about the case where there's no room on the destination port's
73 #define DSW_IN_RING_SIZE (DSW_MAX_EVENTS)
75 #define DSW_MAX_LOAD (INT16_MAX)
76 #define DSW_LOAD_FROM_PERCENT(x) ((int16_t)(((x)*DSW_MAX_LOAD)/100))
77 #define DSW_LOAD_TO_PERCENT(x) ((100*x)/DSW_MAX_LOAD)
79 /* The thought behind keeping the load update interval shorter than
80 * the migration interval is that the load from newly migrated flows
81 * should 'show up' on the load measurement before new migrations are
82 * considered. This is to avoid having too many flows, from too many
83 * source ports, to be migrated too quickly to a lightly loaded port -
84 * in particular since this might cause the system to oscillate.
86 #define DSW_LOAD_UPDATE_INTERVAL (DSW_MIGRATION_INTERVAL/4)
87 #define DSW_OLD_LOAD_WEIGHT (1)
89 /* The minimum time (in us) between two flow migrations. What puts an
90 * upper limit on the actual migration rate is primarily the pace in
91 * which the ports send and receive control messages, which in turn is
92 * largely a function of how much cycles are spent the processing of
95 #define DSW_MIGRATION_INTERVAL (1000)
96 #define DSW_MIN_SOURCE_LOAD_FOR_MIGRATION (DSW_LOAD_FROM_PERCENT(70))
97 #define DSW_MAX_TARGET_LOAD_FOR_MIGRATION (DSW_LOAD_FROM_PERCENT(95))
98 #define DSW_REBALANCE_THRESHOLD (DSW_LOAD_FROM_PERCENT(3))
100 #define DSW_MAX_EVENTS_RECORDED (128)
102 #define DSW_MAX_FLOWS_PER_MIGRATION (8)
104 /* Only one outstanding migration per port is allowed */
105 #define DSW_MAX_PAUSED_FLOWS (DSW_MAX_PORTS*DSW_MAX_FLOWS_PER_MIGRATION)
107 /* Enough room for pause request/confirm and unpaus request/confirm for
108 * all possible senders.
110 #define DSW_CTL_IN_RING_SIZE ((DSW_MAX_PORTS-1)*4)
112 /* With DSW_SORT_DEQUEUED enabled, the scheduler will, at the point of
113 * dequeue(), arrange events so that events with the same flow id on
114 * the same queue forms a back-to-back "burst", and also so that such
115 * bursts of different flow ids, but on the same queue, also come
116 * consecutively. All this in an attempt to improve data and
117 * instruction cache usage for the application, at the cost of a
118 * scheduler overhead increase.
121 /* #define DSW_SORT_DEQUEUED */
123 struct dsw_queue_flow {
128 enum dsw_migration_state {
129 DSW_MIGRATION_STATE_IDLE,
130 DSW_MIGRATION_STATE_PAUSING,
131 DSW_MIGRATION_STATE_FORWARDING,
132 DSW_MIGRATION_STATE_UNPAUSING
138 /* Keeping a pointer here to avoid container_of() calls, which
139 * are expensive since they are very frequent and will result
140 * in an integer multiplication (since the port id is an index
141 * into the dsw_evdev port array).
143 struct dsw_evdev *dsw;
145 uint16_t dequeue_depth;
146 uint16_t enqueue_depth;
148 int32_t inflight_credits;
150 int32_t new_event_threshold;
152 uint16_t pending_releases;
154 uint16_t next_parallel_flow_id;
156 uint16_t ops_since_bg_task;
158 /* most recent 'background' processing */
161 /* For port load measurement. */
162 uint64_t next_load_update;
163 uint64_t load_update_interval;
164 uint64_t measurement_start;
166 uint64_t busy_cycles;
167 uint64_t total_busy_cycles;
169 /* For the ctl interface and flow migration mechanism. */
170 uint64_t next_emigration;
171 uint64_t migration_interval;
172 enum dsw_migration_state migration_state;
174 uint64_t emigration_start;
175 uint64_t emigrations;
176 uint64_t emigration_latency;
178 uint8_t emigration_target_port_ids[DSW_MAX_FLOWS_PER_MIGRATION];
179 struct dsw_queue_flow
180 emigration_target_qfs[DSW_MAX_FLOWS_PER_MIGRATION];
181 uint8_t emigration_targets_len;
184 uint64_t immigrations;
186 uint16_t paused_flows_len;
187 struct dsw_queue_flow paused_flows[DSW_MAX_PAUSED_FLOWS];
189 /* In a very contrived worst case all inflight events can be
190 * laying around paused here.
192 uint16_t paused_events_len;
193 struct rte_event paused_events[DSW_MAX_EVENTS];
195 uint16_t seen_events_len;
196 uint16_t seen_events_idx;
197 struct dsw_queue_flow seen_events[DSW_MAX_EVENTS_RECORDED];
199 uint64_t enqueue_calls;
200 uint64_t new_enqueued;
201 uint64_t forward_enqueued;
202 uint64_t release_enqueued;
203 uint64_t queue_enqueued[DSW_MAX_QUEUES];
205 uint64_t dequeue_calls;
207 uint64_t queue_dequeued[DSW_MAX_QUEUES];
209 uint16_t out_buffer_len[DSW_MAX_PORTS];
210 struct rte_event out_buffer[DSW_MAX_PORTS][DSW_MAX_PORT_OUT_BUFFER];
212 uint16_t in_buffer_len;
213 uint16_t in_buffer_start;
214 /* This buffer may contain events that were read up from the
215 * in_ring during the flow migration process.
217 struct rte_event in_buffer[DSW_MAX_EVENTS];
219 struct rte_event_ring *in_ring __rte_cache_aligned;
221 struct rte_ring *ctl_in_ring __rte_cache_aligned;
223 /* Estimate of current port load. */
224 int16_t load __rte_cache_aligned;
225 /* Estimate of flows currently migrating to this port. */
226 int32_t immigration_load __rte_cache_aligned;
227 } __rte_cache_aligned;
230 uint8_t schedule_type;
231 uint8_t serving_ports[DSW_MAX_PORTS];
232 uint16_t num_serving_ports;
234 uint8_t flow_to_port_map[DSW_MAX_FLOWS] __rte_cache_aligned;
238 struct rte_eventdev_data *data;
240 struct dsw_port ports[DSW_MAX_PORTS];
242 struct dsw_queue queues[DSW_MAX_QUEUES];
244 int32_t max_inflight;
246 int32_t credits_on_loan __rte_cache_aligned;
249 #define DSW_CTL_PAUS_REQ (0)
250 #define DSW_CTL_UNPAUS_REQ (1)
251 #define DSW_CTL_CFM (2)
255 uint8_t originating_port_id;
257 struct dsw_queue_flow qfs[DSW_MAX_FLOWS_PER_MIGRATION];
260 uint16_t dsw_event_enqueue(void *port, const struct rte_event *event);
261 uint16_t dsw_event_enqueue_burst(void *port,
262 const struct rte_event events[],
263 uint16_t events_len);
264 uint16_t dsw_event_enqueue_new_burst(void *port,
265 const struct rte_event events[],
266 uint16_t events_len);
267 uint16_t dsw_event_enqueue_forward_burst(void *port,
268 const struct rte_event events[],
269 uint16_t events_len);
271 uint16_t dsw_event_dequeue(void *port, struct rte_event *ev, uint64_t wait);
272 uint16_t dsw_event_dequeue_burst(void *port, struct rte_event *events,
273 uint16_t num, uint64_t wait);
274 void dsw_event_maintain(void *port, int op);
276 int dsw_xstats_get_names(const struct rte_eventdev *dev,
277 enum rte_event_dev_xstats_mode mode,
278 uint8_t queue_port_id,
279 struct rte_event_dev_xstats_name *xstats_names,
280 unsigned int *ids, unsigned int size);
281 int dsw_xstats_get(const struct rte_eventdev *dev,
282 enum rte_event_dev_xstats_mode mode, uint8_t queue_port_id,
283 const unsigned int ids[], uint64_t values[], unsigned int n);
284 uint64_t dsw_xstats_get_by_name(const struct rte_eventdev *dev,
285 const char *name, unsigned int *id);
287 static inline struct dsw_evdev *
288 dsw_pmd_priv(const struct rte_eventdev *eventdev)
290 return eventdev->data->dev_private;
293 #define DSW_LOG_DP(level, fmt, args...) \
294 RTE_LOG_DP(level, EVENTDEV, "[%s] %s() line %u: " fmt, \
296 __func__, __LINE__, ## args)
298 #define DSW_LOG_DP_PORT(level, port_id, fmt, args...) \
299 DSW_LOG_DP(level, "<Port %d> " fmt, port_id, ## args)