4 * Copyright(c) 2016-2017 Intel Corporation. All rights reserved.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
10 * * Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * * Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in
14 * the documentation and/or other materials provided with the
16 * * Neither the name of Intel Corporation nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
23 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
24 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
25 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
26 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
27 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
28 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
29 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
30 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34 #include <rte_hash_crc.h>
35 #include <rte_event_ring.h>
39 #define SW_IQS_MASK (SW_IQS_MAX-1)
41 /* Retrieve the highest priority IQ or -1 if no pkts available. Doing the
42 * CLZ twice is faster than caching the value due to data dependencies
44 #define PKT_MASK_TO_IQ(pkts) \
45 (__builtin_ctz(pkts | (1 << SW_IQS_MAX)))
48 #error Misconfigured PRIO_TO_IQ caused by SW_IQS_MAX value change
50 #define PRIO_TO_IQ(prio) (prio >> 6)
52 #define MAX_PER_IQ_DEQUEUE 48
53 #define FLOWID_MASK (SW_QID_NUM_FIDS-1)
54 /* use cheap bit mixing, we only need to lose a few bits */
55 #define SW_HASH_FLOWID(f) (((f) ^ (f >> 10)) & FLOWID_MASK)
57 static inline uint32_t
58 sw_schedule_atomic_to_cq(struct sw_evdev *sw, struct sw_qid * const qid,
59 uint32_t iq_num, unsigned int count)
61 struct rte_event qes[MAX_PER_IQ_DEQUEUE]; /* count <= MAX */
62 struct rte_event blocked_qes[MAX_PER_IQ_DEQUEUE];
63 uint32_t nb_blocked = 0;
66 if (count > MAX_PER_IQ_DEQUEUE)
67 count = MAX_PER_IQ_DEQUEUE;
69 /* This is the QID ID. The QID ID is static, hence it can be
70 * used to identify the stage of processing in history lists etc
72 uint32_t qid_id = qid->id;
74 iq_ring_dequeue_burst(qid->iq[iq_num], qes, count);
75 for (i = 0; i < count; i++) {
76 const struct rte_event *qe = &qes[i];
77 const uint16_t flow_id = SW_HASH_FLOWID(qes[i].flow_id);
78 struct sw_fid_t *fid = &qid->fids[flow_id];
82 uint32_t cq_idx = qid->cq_next_tx++;
83 if (qid->cq_next_tx == qid->cq_num_mapped_cqs)
85 cq = qid->cq_map[cq_idx];
88 int cq_free_cnt = sw->cq_ring_space[cq];
89 for (cq_idx = 0; cq_idx < qid->cq_num_mapped_cqs;
91 int test_cq = qid->cq_map[cq_idx];
92 int test_cq_free = sw->cq_ring_space[test_cq];
93 if (test_cq_free > cq_free_cnt) {
95 cq_free_cnt = test_cq_free;
99 fid->cq = cq; /* this pins early */
102 if (sw->cq_ring_space[cq] == 0 ||
103 sw->ports[cq].inflights == SW_PORT_HIST_LIST) {
104 blocked_qes[nb_blocked++] = *qe;
108 struct sw_port *p = &sw->ports[cq];
110 /* at this point we can queue up the packet on the cq_buf */
112 p->cq_buf[p->cq_buf_count++] = *qe;
114 sw->cq_ring_space[cq]--;
116 int head = (p->hist_head++ & (SW_PORT_HIST_LIST-1));
117 p->hist_list[head].fid = flow_id;
118 p->hist_list[head].qid = qid_id;
121 qid->stats.tx_pkts++;
124 /* if we just filled in the last slot, flush the buffer */
125 if (sw->cq_ring_space[cq] == 0) {
126 struct rte_event_ring *worker = p->cq_worker_ring;
127 rte_event_ring_enqueue_burst(worker, p->cq_buf,
129 &sw->cq_ring_space[cq]);
133 iq_ring_put_back(qid->iq[iq_num], blocked_qes, nb_blocked);
135 return count - nb_blocked;
138 static inline uint32_t
139 sw_schedule_parallel_to_cq(struct sw_evdev *sw, struct sw_qid * const qid,
140 uint32_t iq_num, unsigned int count, int keep_order)
143 uint32_t cq_idx = qid->cq_next_tx;
145 /* This is the QID ID. The QID ID is static, hence it can be
146 * used to identify the stage of processing in history lists etc
148 uint32_t qid_id = qid->id;
150 if (count > MAX_PER_IQ_DEQUEUE)
151 count = MAX_PER_IQ_DEQUEUE;
154 /* only schedule as many as we have reorder buffer entries */
155 count = RTE_MIN(count,
156 rte_ring_count(qid->reorder_buffer_freelist));
158 for (i = 0; i < count; i++) {
159 const struct rte_event *qe = iq_ring_peek(qid->iq[iq_num]);
160 uint32_t cq_check_count = 0;
164 * for parallel, just send to next available CQ in round-robin
165 * fashion. So scan for an available CQ. If all CQs are full
166 * just return and move on to next QID
169 if (++cq_check_count > qid->cq_num_mapped_cqs)
171 cq = qid->cq_map[cq_idx];
172 if (++cq_idx == qid->cq_num_mapped_cqs)
174 } while (rte_event_ring_free_count(
175 sw->ports[cq].cq_worker_ring) == 0 ||
176 sw->ports[cq].inflights == SW_PORT_HIST_LIST);
178 struct sw_port *p = &sw->ports[cq];
179 if (sw->cq_ring_space[cq] == 0 ||
180 p->inflights == SW_PORT_HIST_LIST)
183 sw->cq_ring_space[cq]--;
185 qid->stats.tx_pkts++;
187 const int head = (p->hist_head & (SW_PORT_HIST_LIST-1));
188 p->hist_list[head].fid = SW_HASH_FLOWID(qe->flow_id);
189 p->hist_list[head].qid = qid_id;
192 rte_ring_sc_dequeue(qid->reorder_buffer_freelist,
193 (void *)&p->hist_list[head].rob_entry);
195 sw->ports[cq].cq_buf[sw->ports[cq].cq_buf_count++] = *qe;
196 iq_ring_pop(qid->iq[iq_num]);
198 rte_compiler_barrier();
204 qid->cq_next_tx = cq_idx;
209 sw_schedule_dir_to_cq(struct sw_evdev *sw, struct sw_qid * const qid,
210 uint32_t iq_num, unsigned int count __rte_unused)
212 uint32_t cq_id = qid->cq_map[0];
213 struct sw_port *port = &sw->ports[cq_id];
215 /* get max burst enq size for cq_ring */
216 uint32_t count_free = sw->cq_ring_space[cq_id];
220 /* burst dequeue from the QID IQ ring */
221 struct iq_ring *ring = qid->iq[iq_num];
222 uint32_t ret = iq_ring_dequeue_burst(ring,
223 &port->cq_buf[port->cq_buf_count], count_free);
224 port->cq_buf_count += ret;
226 /* Update QID, Port and Total TX stats */
227 qid->stats.tx_pkts += ret;
228 port->stats.tx_pkts += ret;
230 /* Subtract credits from cached value */
231 sw->cq_ring_space[cq_id] -= ret;
237 sw_schedule_qid_to_cq(struct sw_evdev *sw)
242 sw->sched_cq_qid_called++;
244 for (qid_idx = 0; qid_idx < sw->qid_count; qid_idx++) {
245 struct sw_qid *qid = sw->qids_prioritized[qid_idx];
247 int type = qid->type;
248 int iq_num = PKT_MASK_TO_IQ(qid->iq_pkt_mask);
250 /* zero mapped CQs indicates directed */
251 if (iq_num >= SW_IQS_MAX)
254 uint32_t pkts_done = 0;
255 uint32_t count = iq_ring_count(qid->iq[iq_num]);
258 if (type == SW_SCHED_TYPE_DIRECT)
259 pkts_done += sw_schedule_dir_to_cq(sw, qid,
261 else if (type == RTE_SCHED_TYPE_ATOMIC)
262 pkts_done += sw_schedule_atomic_to_cq(sw, qid,
265 pkts_done += sw_schedule_parallel_to_cq(sw, qid,
267 type == RTE_SCHED_TYPE_ORDERED);
270 /* Check if the IQ that was polled is now empty, and unset it
271 * in the IQ mask if its empty.
273 int all_done = (pkts_done == count);
275 qid->iq_pkt_mask &= ~(all_done << (iq_num));
282 /* This function will perform re-ordering of packets, and injecting into
283 * the appropriate QID IQ. As LB and DIR QIDs are in the same array, but *NOT*
284 * contiguous in that array, this function accepts a "range" of QIDs to scan.
287 sw_schedule_reorder(struct sw_evdev *sw, int qid_start, int qid_end)
289 /* Perform egress reordering */
290 struct rte_event *qe;
291 uint32_t pkts_iter = 0;
293 for (; qid_start < qid_end; qid_start++) {
294 struct sw_qid *qid = &sw->qids[qid_start];
295 int i, num_entries_in_use;
297 if (qid->type != RTE_SCHED_TYPE_ORDERED)
300 num_entries_in_use = rte_ring_free_count(
301 qid->reorder_buffer_freelist);
303 for (i = 0; i < num_entries_in_use; i++) {
304 struct reorder_buffer_entry *entry;
307 entry = &qid->reorder_buffer[qid->reorder_buffer_index];
312 for (j = 0; j < entry->num_fragments; j++) {
316 int idx = entry->fragment_index + j;
317 qe = &entry->fragments[idx];
319 dest_qid = qe->queue_id;
320 dest_iq = PRIO_TO_IQ(qe->priority);
322 if (dest_qid >= sw->qid_count) {
323 sw->stats.rx_dropped++;
327 struct sw_qid *dest_qid_ptr =
329 const struct iq_ring *dest_iq_ptr =
330 dest_qid_ptr->iq[dest_iq];
331 if (iq_ring_free_count(dest_iq_ptr) == 0)
336 struct sw_qid *q = &sw->qids[dest_qid];
337 struct iq_ring *r = q->iq[dest_iq];
339 /* we checked for space above, so enqueue must
342 iq_ring_enqueue(r, qe);
343 q->iq_pkt_mask |= (1 << (dest_iq));
344 q->iq_pkt_count[dest_iq]++;
348 entry->ready = (j != entry->num_fragments);
349 entry->num_fragments -= j;
350 entry->fragment_index += j;
353 entry->fragment_index = 0;
356 qid->reorder_buffer_freelist,
359 qid->reorder_buffer_index++;
360 qid->reorder_buffer_index %= qid->window_size;
367 static __rte_always_inline void
368 sw_refill_pp_buf(struct sw_evdev *sw, struct sw_port *port)
371 struct rte_event_ring *worker = port->rx_worker_ring;
372 port->pp_buf_start = 0;
373 port->pp_buf_count = rte_event_ring_dequeue_burst(worker, port->pp_buf,
374 RTE_DIM(port->pp_buf), NULL);
377 static __rte_always_inline uint32_t
378 __pull_port_lb(struct sw_evdev *sw, uint32_t port_id, int allow_reorder)
380 static struct reorder_buffer_entry dummy_rob;
381 uint32_t pkts_iter = 0;
382 struct sw_port *port = &sw->ports[port_id];
384 /* If shadow ring has 0 pkts, pull from worker ring */
385 if (port->pp_buf_count == 0)
386 sw_refill_pp_buf(sw, port);
388 while (port->pp_buf_count) {
389 const struct rte_event *qe = &port->pp_buf[port->pp_buf_start];
390 struct sw_hist_list_entry *hist_entry = NULL;
391 uint8_t flags = qe->op;
392 const uint16_t eop = !(flags & QE_FLAG_NOT_EOP);
393 int needs_reorder = 0;
394 /* if no-reordering, having PARTIAL == NEW */
395 if (!allow_reorder && !eop)
396 flags = QE_FLAG_VALID;
399 * if we don't have space for this packet in an IQ,
400 * then move on to next queue. Technically, for a
401 * packet that needs reordering, we don't need to check
402 * here, but it simplifies things not to special-case
404 uint32_t iq_num = PRIO_TO_IQ(qe->priority);
405 struct sw_qid *qid = &sw->qids[qe->queue_id];
407 if ((flags & QE_FLAG_VALID) &&
408 iq_ring_free_count(qid->iq[iq_num]) == 0)
411 /* now process based on flags. Note that for directed
412 * queues, the enqueue_flush masks off all but the
413 * valid flag. This makes FWD and PARTIAL enqueues just
414 * NEW type, and makes DROPS no-op calls.
416 if ((flags & QE_FLAG_COMPLETE) && port->inflights > 0) {
417 const uint32_t hist_tail = port->hist_tail &
418 (SW_PORT_HIST_LIST - 1);
420 hist_entry = &port->hist_list[hist_tail];
421 const uint32_t hist_qid = hist_entry->qid;
422 const uint32_t hist_fid = hist_entry->fid;
424 struct sw_fid_t *fid =
425 &sw->qids[hist_qid].fids[hist_fid];
427 if (fid->pcount == 0)
431 /* set reorder ready if an ordered QID */
433 (uintptr_t)hist_entry->rob_entry;
434 const uintptr_t valid = (rob_ptr != 0);
435 needs_reorder = valid;
437 ((valid - 1) & (uintptr_t)&dummy_rob);
438 struct reorder_buffer_entry *tmp_rob_ptr =
439 (struct reorder_buffer_entry *)rob_ptr;
440 tmp_rob_ptr->ready = eop * needs_reorder;
443 port->inflights -= eop;
444 port->hist_tail += eop;
446 if (flags & QE_FLAG_VALID) {
447 port->stats.rx_pkts++;
449 if (allow_reorder && needs_reorder) {
450 struct reorder_buffer_entry *rob_entry =
451 hist_entry->rob_entry;
453 hist_entry->rob_entry = NULL;
454 /* Although fragmentation not currently
455 * supported by eventdev API, we support it
456 * here. Open: How do we alert the user that
457 * they've exceeded max frags?
459 int num_frag = rob_entry->num_fragments;
460 if (num_frag == SW_FRAGMENTS_MAX)
461 sw->stats.rx_dropped++;
463 int idx = rob_entry->num_fragments++;
464 rob_entry->fragments[idx] = *qe;
469 /* Use the iq_num from above to push the QE
470 * into the qid at the right priority
473 qid->iq_pkt_mask |= (1 << (iq_num));
474 iq_ring_enqueue(qid->iq[iq_num], qe);
475 qid->iq_pkt_count[iq_num]++;
476 qid->stats.rx_pkts++;
481 port->pp_buf_start++;
482 port->pp_buf_count--;
483 } /* while (avail_qes) */
489 sw_schedule_pull_port_lb(struct sw_evdev *sw, uint32_t port_id)
491 return __pull_port_lb(sw, port_id, 1);
495 sw_schedule_pull_port_no_reorder(struct sw_evdev *sw, uint32_t port_id)
497 return __pull_port_lb(sw, port_id, 0);
501 sw_schedule_pull_port_dir(struct sw_evdev *sw, uint32_t port_id)
503 uint32_t pkts_iter = 0;
504 struct sw_port *port = &sw->ports[port_id];
506 /* If shadow ring has 0 pkts, pull from worker ring */
507 if (port->pp_buf_count == 0)
508 sw_refill_pp_buf(sw, port);
510 while (port->pp_buf_count) {
511 const struct rte_event *qe = &port->pp_buf[port->pp_buf_start];
512 uint8_t flags = qe->op;
514 if ((flags & QE_FLAG_VALID) == 0)
517 uint32_t iq_num = PRIO_TO_IQ(qe->priority);
518 struct sw_qid *qid = &sw->qids[qe->queue_id];
519 struct iq_ring *iq_ring = qid->iq[iq_num];
521 if (iq_ring_free_count(iq_ring) == 0)
522 break; /* move to next port */
524 port->stats.rx_pkts++;
526 /* Use the iq_num from above to push the QE
527 * into the qid at the right priority
529 qid->iq_pkt_mask |= (1 << (iq_num));
530 iq_ring_enqueue(iq_ring, qe);
531 qid->iq_pkt_count[iq_num]++;
532 qid->stats.rx_pkts++;
536 port->pp_buf_start++;
537 port->pp_buf_count--;
538 } /* while port->pp_buf_count */
544 sw_event_schedule(struct rte_eventdev *dev)
546 struct sw_evdev *sw = sw_pmd_priv(dev);
547 uint32_t in_pkts, out_pkts;
548 uint32_t out_pkts_total = 0, in_pkts_total = 0;
549 int32_t sched_quanta = sw->sched_quanta;
557 uint32_t in_pkts_this_iteration = 0;
559 /* Pull from rx_ring for ports */
562 for (i = 0; i < sw->port_count; i++)
563 if (sw->ports[i].is_directed)
564 in_pkts += sw_schedule_pull_port_dir(sw, i);
565 else if (sw->ports[i].num_ordered_qids > 0)
566 in_pkts += sw_schedule_pull_port_lb(sw, i);
568 in_pkts += sw_schedule_pull_port_no_reorder(sw, i);
570 /* QID scan for re-ordered */
571 in_pkts += sw_schedule_reorder(sw, 0,
573 in_pkts_this_iteration += in_pkts;
574 } while (in_pkts > 4 &&
575 (int)in_pkts_this_iteration < sched_quanta);
578 out_pkts += sw_schedule_qid_to_cq(sw);
579 out_pkts_total += out_pkts;
580 in_pkts_total += in_pkts_this_iteration;
582 if (in_pkts == 0 && out_pkts == 0)
584 } while ((int)out_pkts_total < sched_quanta);
586 /* push all the internal buffered QEs in port->cq_ring to the
587 * worker cores: aka, do the ring transfers batched.
589 for (i = 0; i < sw->port_count; i++) {
590 struct rte_event_ring *worker = sw->ports[i].cq_worker_ring;
591 rte_event_ring_enqueue_burst(worker, sw->ports[i].cq_buf,
592 sw->ports[i].cq_buf_count,
593 &sw->cq_ring_space[i]);
594 sw->ports[i].cq_buf_count = 0;
597 sw->stats.tx_pkts += out_pkts_total;
598 sw->stats.rx_pkts += in_pkts_total;
600 sw->sched_no_iq_enqueues += (in_pkts_total == 0);
601 sw->sched_no_cq_enqueues += (out_pkts_total == 0);