1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2016-2017 Intel Corporation 3 */ 4 5 #ifndef _SW_EVDEV_H_ 6 #define _SW_EVDEV_H_ 7 8 #include "sw_evdev_log.h" 9 #include <rte_eventdev.h> 10 #include <rte_eventdev_pmd_vdev.h> 11 #include <rte_atomic.h> 12 13 #define SW_DEFAULT_CREDIT_QUANTA 32 14 #define SW_DEFAULT_SCHED_QUANTA 128 15 #define SW_QID_NUM_FIDS 16384 16 #define SW_IQS_MAX 4 17 #define SW_Q_PRIORITY_MAX 255 18 #define SW_PORTS_MAX 64 19 #define MAX_SW_CONS_Q_DEPTH 128 20 #define SW_INFLIGHT_EVENTS_TOTAL 4096 21 /* allow for lots of over-provisioning */ 22 #define MAX_SW_PROD_Q_DEPTH 4096 23 #define SW_FRAGMENTS_MAX 16 24 25 /* Should be power-of-two minus one, to leave room for the next pointer */ 26 #define SW_EVS_PER_Q_CHUNK 255 27 #define SW_Q_CHUNK_SIZE ((SW_EVS_PER_Q_CHUNK + 1) * sizeof(struct rte_event)) 28 29 /* report dequeue burst sizes in buckets */ 30 #define SW_DEQ_STAT_BUCKET_SHIFT 2 31 /* how many packets pulled from port by sched */ 32 #define SCHED_DEQUEUE_BURST_SIZE 32 33 34 #define SW_PORT_HIST_LIST (MAX_SW_PROD_Q_DEPTH) /* size of our history list */ 35 #define NUM_SAMPLES 64 /* how many data points use for average stats */ 36 37 #define EVENTDEV_NAME_SW_PMD event_sw 38 #define SW_PMD_NAME RTE_STR(event_sw) 39 #define SW_PMD_NAME_MAX 64 40 41 #define SW_SCHED_TYPE_DIRECT (RTE_SCHED_TYPE_PARALLEL + 1) 42 43 #define SW_NUM_POLL_BUCKETS (MAX_SW_CONS_Q_DEPTH >> SW_DEQ_STAT_BUCKET_SHIFT) 44 45 enum { 46 QE_FLAG_VALID_SHIFT = 0, 47 QE_FLAG_COMPLETE_SHIFT, 48 QE_FLAG_NOT_EOP_SHIFT, 49 _QE_FLAG_COUNT 50 }; 51 52 #define QE_FLAG_VALID (1 << QE_FLAG_VALID_SHIFT) /* for NEW FWD, FRAG */ 53 #define QE_FLAG_COMPLETE (1 << QE_FLAG_COMPLETE_SHIFT) /* set for FWD, DROP */ 54 #define QE_FLAG_NOT_EOP (1 << QE_FLAG_NOT_EOP_SHIFT) /* set for FRAG only */ 55 56 static const uint8_t sw_qe_flag_map[] = { 57 QE_FLAG_VALID /* NEW Event */, 58 QE_FLAG_VALID | QE_FLAG_COMPLETE /* FWD Event */, 59 QE_FLAG_COMPLETE /* RELEASE Event */, 60 61 /* Values which can be used for future support for partial 62 * events, i.e. where one event comes back to the scheduler 63 * as multiple which need to be tracked together 64 */ 65 QE_FLAG_VALID | QE_FLAG_COMPLETE | QE_FLAG_NOT_EOP, 66 }; 67 68 /* Records basic event stats at a given point. Used in port and qid structs */ 69 struct sw_point_stats { 70 uint64_t rx_pkts; 71 uint64_t rx_dropped; 72 uint64_t tx_pkts; 73 }; 74 75 /* structure used to track what port a flow (FID) is pinned to */ 76 struct sw_fid_t { 77 /* which CQ this FID is currently pinned to */ 78 int32_t cq; 79 /* number of packets gone to the CQ with this FID */ 80 uint32_t pcount; 81 }; 82 83 struct reorder_buffer_entry { 84 uint16_t num_fragments; /**< Number of packet fragments */ 85 uint16_t fragment_index; /**< Points to the oldest valid frag */ 86 uint8_t ready; /**< Entry is ready to be reordered */ 87 struct rte_event fragments[SW_FRAGMENTS_MAX]; 88 }; 89 90 struct sw_iq { 91 struct sw_queue_chunk *head; 92 struct sw_queue_chunk *tail; 93 uint16_t head_idx; 94 uint16_t tail_idx; 95 uint16_t count; 96 }; 97 98 struct sw_qid { 99 /* set when the QID has been initialized */ 100 uint8_t initialized; 101 /* The type of this QID */ 102 int8_t type; 103 /* Integer ID representing the queue. This is used in history lists, 104 * to identify the stage of processing. 105 */ 106 uint32_t id; 107 struct sw_point_stats stats; 108 109 /* Internal priority rings for packets */ 110 struct sw_iq iq[SW_IQS_MAX]; 111 uint32_t iq_pkt_mask; /* A mask to indicate packets in an IQ */ 112 uint64_t iq_pkt_count[SW_IQS_MAX]; 113 114 /* Information on what CQs are polling this IQ */ 115 uint32_t cq_num_mapped_cqs; 116 uint32_t cq_next_tx; /* cq to write next (non-atomic) packet */ 117 uint32_t cq_map[SW_PORTS_MAX]; 118 uint64_t to_port[SW_PORTS_MAX]; 119 120 /* Track flow ids for atomic load balancing */ 121 struct sw_fid_t fids[SW_QID_NUM_FIDS]; 122 123 /* Track packet order for reordering when needed */ 124 struct reorder_buffer_entry *reorder_buffer; /*< pkts await reorder */ 125 struct rte_ring *reorder_buffer_freelist; /* available reorder slots */ 126 uint32_t reorder_buffer_index; /* oldest valid reorder buffer entry */ 127 uint32_t window_size; /* Used to wrap reorder_buffer_index */ 128 129 uint8_t priority; 130 }; 131 132 struct sw_hist_list_entry { 133 int32_t qid; 134 int32_t fid; 135 struct reorder_buffer_entry *rob_entry; 136 }; 137 138 struct sw_evdev; 139 140 struct sw_port { 141 /* new enqueue / dequeue API doesn't have an instance pointer, only the 142 * pointer to the port being enqueue/dequeued from 143 */ 144 struct sw_evdev *sw; 145 146 /* set when the port is initialized */ 147 uint8_t initialized; 148 /* A numeric ID for the port */ 149 uint8_t id; 150 151 /* An atomic counter for when the port has been unlinked, and the 152 * scheduler has not yet acked this unlink - hence there may still be 153 * events in the buffers going to the port. When the unlinks in 154 * progress is read by the scheduler, no more events will be pushed to 155 * the port - hence the scheduler core can just assign zero. 156 */ 157 uint8_t unlinks_in_progress; 158 159 int16_t is_directed; /** Takes from a single directed QID */ 160 /** 161 * For loadbalanced we can optimise pulling packets from 162 * producers if there is no reordering involved 163 */ 164 int16_t num_ordered_qids; 165 166 /** Ring and buffer for pulling events from workers for scheduling */ 167 struct rte_event_ring *rx_worker_ring __rte_cache_aligned; 168 /** Ring and buffer for pushing packets to workers after scheduling */ 169 struct rte_event_ring *cq_worker_ring; 170 171 /* hole */ 172 173 /* num releases yet to be completed on this port */ 174 uint16_t outstanding_releases __rte_cache_aligned; 175 uint16_t inflight_max; /* app requested max inflights for this port */ 176 uint16_t inflight_credits; /* num credits this port has right now */ 177 uint8_t implicit_release; /* release events before dequeueing */ 178 179 uint16_t last_dequeue_burst_sz; /* how big the burst was */ 180 uint64_t last_dequeue_ticks; /* used to track burst processing time */ 181 uint64_t avg_pkt_ticks; /* tracks average over NUM_SAMPLES burst */ 182 uint64_t total_polls; /* how many polls were counted in stats */ 183 uint64_t zero_polls; /* tracks polls returning nothing */ 184 uint32_t poll_buckets[SW_NUM_POLL_BUCKETS]; 185 /* bucket values in 4s for shorter reporting */ 186 187 /* History list structs, containing info on pkts egressed to worker */ 188 uint16_t hist_head __rte_cache_aligned; 189 uint16_t hist_tail; 190 uint16_t inflights; 191 struct sw_hist_list_entry hist_list[SW_PORT_HIST_LIST]; 192 193 /* track packets in and out of this port */ 194 struct sw_point_stats stats; 195 196 197 uint32_t pp_buf_start; 198 uint32_t pp_buf_count; 199 uint16_t cq_buf_count; 200 struct rte_event pp_buf[SCHED_DEQUEUE_BURST_SIZE]; 201 struct rte_event cq_buf[MAX_SW_CONS_Q_DEPTH]; 202 203 uint8_t num_qids_mapped; 204 }; 205 206 struct sw_evdev { 207 struct rte_eventdev_data *data; 208 209 uint32_t port_count; 210 uint32_t qid_count; 211 uint32_t xstats_count; 212 struct sw_xstats_entry *xstats; 213 uint32_t xstats_count_mode_dev; 214 uint32_t xstats_count_mode_port; 215 uint32_t xstats_count_mode_queue; 216 217 /* Contains all ports - load balanced and directed */ 218 struct sw_port ports[SW_PORTS_MAX] __rte_cache_aligned; 219 220 rte_atomic32_t inflights __rte_cache_aligned; 221 222 /* 223 * max events in this instance. Cached here for performance. 224 * (also available in data->conf.nb_events_limit) 225 */ 226 uint32_t nb_events_limit; 227 228 /* Internal queues - one per logical queue */ 229 struct sw_qid qids[RTE_EVENT_MAX_QUEUES_PER_DEV] __rte_cache_aligned; 230 struct sw_queue_chunk *chunk_list_head; 231 struct sw_queue_chunk *chunks; 232 233 /* Cache how many packets are in each cq */ 234 uint16_t cq_ring_space[SW_PORTS_MAX] __rte_cache_aligned; 235 236 /* Array of pointers to load-balanced QIDs sorted by priority level */ 237 struct sw_qid *qids_prioritized[RTE_EVENT_MAX_QUEUES_PER_DEV]; 238 239 /* Stats */ 240 struct sw_point_stats stats __rte_cache_aligned; 241 uint64_t sched_called; 242 int32_t sched_quanta; 243 uint64_t sched_no_iq_enqueues; 244 uint64_t sched_no_cq_enqueues; 245 uint64_t sched_cq_qid_called; 246 247 uint8_t started; 248 uint32_t credit_update_quanta; 249 250 /* store num stats and offset of the stats for each port */ 251 uint16_t xstats_count_per_port[SW_PORTS_MAX]; 252 uint16_t xstats_offset_for_port[SW_PORTS_MAX]; 253 /* store num stats and offset of the stats for each queue */ 254 uint16_t xstats_count_per_qid[RTE_EVENT_MAX_QUEUES_PER_DEV]; 255 uint16_t xstats_offset_for_qid[RTE_EVENT_MAX_QUEUES_PER_DEV]; 256 257 uint32_t service_id; 258 char service_name[SW_PMD_NAME_MAX]; 259 }; 260 261 static inline struct sw_evdev * 262 sw_pmd_priv(const struct rte_eventdev *eventdev) 263 { 264 return eventdev->data->dev_private; 265 } 266 267 static inline const struct sw_evdev * 268 sw_pmd_priv_const(const struct rte_eventdev *eventdev) 269 { 270 return eventdev->data->dev_private; 271 } 272 273 uint16_t sw_event_enqueue(void *port, const struct rte_event *ev); 274 uint16_t sw_event_enqueue_burst(void *port, const struct rte_event ev[], 275 uint16_t num); 276 277 uint16_t sw_event_dequeue(void *port, struct rte_event *ev, uint64_t wait); 278 uint16_t sw_event_dequeue_burst(void *port, struct rte_event *ev, uint16_t num, 279 uint64_t wait); 280 void sw_event_schedule(struct rte_eventdev *dev); 281 int sw_xstats_init(struct sw_evdev *dev); 282 int sw_xstats_uninit(struct sw_evdev *dev); 283 int sw_xstats_get_names(const struct rte_eventdev *dev, 284 enum rte_event_dev_xstats_mode mode, uint8_t queue_port_id, 285 struct rte_event_dev_xstats_name *xstats_names, 286 unsigned int *ids, unsigned int size); 287 int sw_xstats_get(const struct rte_eventdev *dev, 288 enum rte_event_dev_xstats_mode mode, uint8_t queue_port_id, 289 const unsigned int ids[], uint64_t values[], unsigned int n); 290 uint64_t sw_xstats_get_by_name(const struct rte_eventdev *dev, 291 const char *name, unsigned int *id); 292 int sw_xstats_reset(struct rte_eventdev *dev, 293 enum rte_event_dev_xstats_mode mode, 294 int16_t queue_port_id, 295 const uint32_t ids[], 296 uint32_t nb_ids); 297 298 int test_sw_eventdev(void); 299 300 #endif /* _SW_EVDEV_H_ */ 301