xref: /dpdk/drivers/net/sfc/sfc_ev.c (revision 89f0711f9ddfb5822da9d34f384b92f72a61c4dc)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  *
3  * Copyright (c) 2016-2018 Solarflare Communications Inc.
4  * All rights reserved.
5  *
6  * This software was jointly developed between OKTET Labs (under contract
7  * for Solarflare) and Solarflare Communications, Inc.
8  */
9 
10 #include <rte_debug.h>
11 #include <rte_cycles.h>
12 #include <rte_alarm.h>
13 #include <rte_branch_prediction.h>
14 
15 #include "efx.h"
16 
17 #include "sfc.h"
18 #include "sfc_debug.h"
19 #include "sfc_log.h"
20 #include "sfc_ev.h"
21 #include "sfc_rx.h"
22 #include "sfc_tx.h"
23 #include "sfc_kvargs.h"
24 
25 
26 /* Initial delay when waiting for event queue init complete event */
27 #define SFC_EVQ_INIT_BACKOFF_START_US	(1)
28 /* Maximum delay between event queue polling attempts */
29 #define SFC_EVQ_INIT_BACKOFF_MAX_US	(10 * 1000)
30 /* Event queue init approx timeout */
31 #define SFC_EVQ_INIT_TIMEOUT_US		(2 * US_PER_S)
32 
33 /* Management event queue polling period in microseconds */
34 #define SFC_MGMT_EV_QPOLL_PERIOD_US	(US_PER_S)
35 
36 static const char *
37 sfc_evq_type2str(enum sfc_evq_type type)
38 {
39 	switch (type) {
40 	case SFC_EVQ_TYPE_MGMT:
41 		return "mgmt-evq";
42 	case SFC_EVQ_TYPE_RX:
43 		return "rx-evq";
44 	case SFC_EVQ_TYPE_TX:
45 		return "tx-evq";
46 	default:
47 		SFC_ASSERT(B_FALSE);
48 		return NULL;
49 	}
50 }
51 
52 static boolean_t
53 sfc_ev_initialized(void *arg)
54 {
55 	struct sfc_evq *evq = arg;
56 
57 	/* Init done events may be duplicated on SFN7xxx (SFC bug 31631) */
58 	SFC_ASSERT(evq->init_state == SFC_EVQ_STARTING ||
59 		   evq->init_state == SFC_EVQ_STARTED);
60 
61 	evq->init_state = SFC_EVQ_STARTED;
62 
63 	return B_FALSE;
64 }
65 
66 static boolean_t
67 sfc_ev_nop_rx(void *arg, uint32_t label, uint32_t id,
68 	      uint32_t size, uint16_t flags)
69 {
70 	struct sfc_evq *evq = arg;
71 
72 	sfc_err(evq->sa,
73 		"EVQ %u unexpected Rx event label=%u id=%#x size=%u flags=%#x",
74 		evq->evq_index, label, id, size, flags);
75 	return B_TRUE;
76 }
77 
78 static boolean_t
79 sfc_ev_efx_rx(void *arg, __rte_unused uint32_t label, uint32_t id,
80 	      uint32_t size, uint16_t flags)
81 {
82 	struct sfc_evq *evq = arg;
83 	struct sfc_efx_rxq *rxq;
84 	unsigned int stop;
85 	unsigned int pending_id;
86 	unsigned int delta;
87 	unsigned int i;
88 	struct sfc_efx_rx_sw_desc *rxd;
89 
90 	if (unlikely(evq->exception))
91 		goto done;
92 
93 	rxq = sfc_efx_rxq_by_dp_rxq(evq->dp_rxq);
94 
95 	SFC_ASSERT(rxq != NULL);
96 	SFC_ASSERT(rxq->evq == evq);
97 	SFC_ASSERT(rxq->flags & SFC_EFX_RXQ_FLAG_STARTED);
98 
99 	stop = (id + 1) & rxq->ptr_mask;
100 	pending_id = rxq->pending & rxq->ptr_mask;
101 	delta = (stop >= pending_id) ? (stop - pending_id) :
102 		(rxq->ptr_mask + 1 - pending_id + stop);
103 
104 	if (delta == 0) {
105 		/*
106 		 * Rx event with no new descriptors done and zero length
107 		 * is used to abort scattered packet when there is no room
108 		 * for the tail.
109 		 */
110 		if (unlikely(size != 0)) {
111 			evq->exception = B_TRUE;
112 			sfc_err(evq->sa,
113 				"EVQ %u RxQ %u invalid RX abort "
114 				"(id=%#x size=%u flags=%#x); needs restart",
115 				evq->evq_index, rxq->dp.dpq.queue_id,
116 				id, size, flags);
117 			goto done;
118 		}
119 
120 		/* Add discard flag to the first fragment */
121 		rxq->sw_desc[pending_id].flags |= EFX_DISCARD;
122 		/* Remove continue flag from the last fragment */
123 		rxq->sw_desc[id].flags &= ~EFX_PKT_CONT;
124 	} else if (unlikely(delta > rxq->batch_max)) {
125 		evq->exception = B_TRUE;
126 
127 		sfc_err(evq->sa,
128 			"EVQ %u RxQ %u completion out of order "
129 			"(id=%#x delta=%u flags=%#x); needs restart",
130 			evq->evq_index, rxq->dp.dpq.queue_id,
131 			id, delta, flags);
132 
133 		goto done;
134 	}
135 
136 	for (i = pending_id; i != stop; i = (i + 1) & rxq->ptr_mask) {
137 		rxd = &rxq->sw_desc[i];
138 
139 		rxd->flags = flags;
140 
141 		SFC_ASSERT(size < (1 << 16));
142 		rxd->size = (uint16_t)size;
143 	}
144 
145 	rxq->pending += delta;
146 
147 done:
148 	return B_FALSE;
149 }
150 
151 static boolean_t
152 sfc_ev_dp_rx(void *arg, __rte_unused uint32_t label, uint32_t id,
153 	     __rte_unused uint32_t size, __rte_unused uint16_t flags)
154 {
155 	struct sfc_evq *evq = arg;
156 	struct sfc_dp_rxq *dp_rxq;
157 
158 	dp_rxq = evq->dp_rxq;
159 	SFC_ASSERT(dp_rxq != NULL);
160 
161 	SFC_ASSERT(evq->sa->dp_rx->qrx_ev != NULL);
162 	return evq->sa->dp_rx->qrx_ev(dp_rxq, id);
163 }
164 
165 static boolean_t
166 sfc_ev_nop_tx(void *arg, uint32_t label, uint32_t id)
167 {
168 	struct sfc_evq *evq = arg;
169 
170 	sfc_err(evq->sa, "EVQ %u unexpected Tx event label=%u id=%#x",
171 		evq->evq_index, label, id);
172 	return B_TRUE;
173 }
174 
175 static boolean_t
176 sfc_ev_tx(void *arg, __rte_unused uint32_t label, uint32_t id)
177 {
178 	struct sfc_evq *evq = arg;
179 	struct sfc_dp_txq *dp_txq;
180 	struct sfc_efx_txq *txq;
181 	unsigned int stop;
182 	unsigned int delta;
183 
184 	dp_txq = evq->dp_txq;
185 	SFC_ASSERT(dp_txq != NULL);
186 
187 	txq = sfc_efx_txq_by_dp_txq(dp_txq);
188 	SFC_ASSERT(txq->evq == evq);
189 
190 	if (unlikely((txq->flags & SFC_EFX_TXQ_FLAG_STARTED) == 0))
191 		goto done;
192 
193 	stop = (id + 1) & txq->ptr_mask;
194 	id = txq->pending & txq->ptr_mask;
195 
196 	delta = (stop >= id) ? (stop - id) : (txq->ptr_mask + 1 - id + stop);
197 
198 	txq->pending += delta;
199 
200 done:
201 	return B_FALSE;
202 }
203 
204 static boolean_t
205 sfc_ev_dp_tx(void *arg, __rte_unused uint32_t label, uint32_t id)
206 {
207 	struct sfc_evq *evq = arg;
208 	struct sfc_dp_txq *dp_txq;
209 
210 	dp_txq = evq->dp_txq;
211 	SFC_ASSERT(dp_txq != NULL);
212 
213 	SFC_ASSERT(evq->sa->dp_tx->qtx_ev != NULL);
214 	return evq->sa->dp_tx->qtx_ev(dp_txq, id);
215 }
216 
217 static boolean_t
218 sfc_ev_exception(void *arg, uint32_t code, __rte_unused uint32_t data)
219 {
220 	struct sfc_evq *evq = arg;
221 
222 	if (code == EFX_EXCEPTION_UNKNOWN_SENSOREVT)
223 		return B_FALSE;
224 
225 	evq->exception = B_TRUE;
226 	sfc_warn(evq->sa,
227 		 "hardware exception %s (code=%u, data=%#x) on EVQ %u;"
228 		 " needs recovery",
229 		 (code == EFX_EXCEPTION_RX_RECOVERY) ? "RX_RECOVERY" :
230 		 (code == EFX_EXCEPTION_RX_DSC_ERROR) ? "RX_DSC_ERROR" :
231 		 (code == EFX_EXCEPTION_TX_DSC_ERROR) ? "TX_DSC_ERROR" :
232 		 (code == EFX_EXCEPTION_FWALERT_SRAM) ? "FWALERT_SRAM" :
233 		 (code == EFX_EXCEPTION_UNKNOWN_FWALERT) ? "UNKNOWN_FWALERT" :
234 		 (code == EFX_EXCEPTION_RX_ERROR) ? "RX_ERROR" :
235 		 (code == EFX_EXCEPTION_TX_ERROR) ? "TX_ERROR" :
236 		 (code == EFX_EXCEPTION_EV_ERROR) ? "EV_ERROR" :
237 		 "UNKNOWN",
238 		 code, data, evq->evq_index);
239 
240 	return B_TRUE;
241 }
242 
243 static boolean_t
244 sfc_ev_nop_rxq_flush_done(void *arg, uint32_t rxq_hw_index)
245 {
246 	struct sfc_evq *evq = arg;
247 
248 	sfc_err(evq->sa, "EVQ %u unexpected RxQ %u flush done",
249 		evq->evq_index, rxq_hw_index);
250 	return B_TRUE;
251 }
252 
253 static boolean_t
254 sfc_ev_rxq_flush_done(void *arg, __rte_unused uint32_t rxq_hw_index)
255 {
256 	struct sfc_evq *evq = arg;
257 	struct sfc_dp_rxq *dp_rxq;
258 	struct sfc_rxq *rxq;
259 
260 	dp_rxq = evq->dp_rxq;
261 	SFC_ASSERT(dp_rxq != NULL);
262 
263 	rxq = sfc_rxq_by_dp_rxq(dp_rxq);
264 	SFC_ASSERT(rxq != NULL);
265 	SFC_ASSERT(rxq->hw_index == rxq_hw_index);
266 	SFC_ASSERT(rxq->evq == evq);
267 	sfc_rx_qflush_done(rxq);
268 
269 	return B_FALSE;
270 }
271 
272 static boolean_t
273 sfc_ev_nop_rxq_flush_failed(void *arg, uint32_t rxq_hw_index)
274 {
275 	struct sfc_evq *evq = arg;
276 
277 	sfc_err(evq->sa, "EVQ %u unexpected RxQ %u flush failed",
278 		evq->evq_index, rxq_hw_index);
279 	return B_TRUE;
280 }
281 
282 static boolean_t
283 sfc_ev_rxq_flush_failed(void *arg, __rte_unused uint32_t rxq_hw_index)
284 {
285 	struct sfc_evq *evq = arg;
286 	struct sfc_dp_rxq *dp_rxq;
287 	struct sfc_rxq *rxq;
288 
289 	dp_rxq = evq->dp_rxq;
290 	SFC_ASSERT(dp_rxq != NULL);
291 
292 	rxq = sfc_rxq_by_dp_rxq(dp_rxq);
293 	SFC_ASSERT(rxq != NULL);
294 	SFC_ASSERT(rxq->hw_index == rxq_hw_index);
295 	SFC_ASSERT(rxq->evq == evq);
296 	sfc_rx_qflush_failed(rxq);
297 
298 	return B_FALSE;
299 }
300 
301 static boolean_t
302 sfc_ev_nop_txq_flush_done(void *arg, uint32_t txq_hw_index)
303 {
304 	struct sfc_evq *evq = arg;
305 
306 	sfc_err(evq->sa, "EVQ %u unexpected TxQ %u flush done",
307 		evq->evq_index, txq_hw_index);
308 	return B_TRUE;
309 }
310 
311 static boolean_t
312 sfc_ev_txq_flush_done(void *arg, __rte_unused uint32_t txq_hw_index)
313 {
314 	struct sfc_evq *evq = arg;
315 	struct sfc_dp_txq *dp_txq;
316 	struct sfc_txq *txq;
317 
318 	dp_txq = evq->dp_txq;
319 	SFC_ASSERT(dp_txq != NULL);
320 
321 	txq = sfc_txq_by_dp_txq(dp_txq);
322 	SFC_ASSERT(txq != NULL);
323 	SFC_ASSERT(txq->hw_index == txq_hw_index);
324 	SFC_ASSERT(txq->evq == evq);
325 	sfc_tx_qflush_done(txq);
326 
327 	return B_FALSE;
328 }
329 
330 static boolean_t
331 sfc_ev_software(void *arg, uint16_t magic)
332 {
333 	struct sfc_evq *evq = arg;
334 
335 	sfc_err(evq->sa, "EVQ %u unexpected software event magic=%#.4x",
336 		evq->evq_index, magic);
337 	return B_TRUE;
338 }
339 
340 static boolean_t
341 sfc_ev_sram(void *arg, uint32_t code)
342 {
343 	struct sfc_evq *evq = arg;
344 
345 	sfc_err(evq->sa, "EVQ %u unexpected SRAM event code=%u",
346 		evq->evq_index, code);
347 	return B_TRUE;
348 }
349 
350 static boolean_t
351 sfc_ev_wake_up(void *arg, uint32_t index)
352 {
353 	struct sfc_evq *evq = arg;
354 
355 	sfc_err(evq->sa, "EVQ %u unexpected wake up event index=%u",
356 		evq->evq_index, index);
357 	return B_TRUE;
358 }
359 
360 static boolean_t
361 sfc_ev_timer(void *arg, uint32_t index)
362 {
363 	struct sfc_evq *evq = arg;
364 
365 	sfc_err(evq->sa, "EVQ %u unexpected timer event index=%u",
366 		evq->evq_index, index);
367 	return B_TRUE;
368 }
369 
370 static boolean_t
371 sfc_ev_nop_link_change(void *arg, __rte_unused efx_link_mode_t link_mode)
372 {
373 	struct sfc_evq *evq = arg;
374 
375 	sfc_err(evq->sa, "EVQ %u unexpected link change event",
376 		evq->evq_index);
377 	return B_TRUE;
378 }
379 
380 static boolean_t
381 sfc_ev_link_change(void *arg, efx_link_mode_t link_mode)
382 {
383 	struct sfc_evq *evq = arg;
384 	struct sfc_adapter *sa = evq->sa;
385 	struct rte_eth_link *dev_link = &sa->eth_dev->data->dev_link;
386 	struct rte_eth_link new_link;
387 	uint64_t new_link_u64;
388 	uint64_t old_link_u64;
389 
390 	EFX_STATIC_ASSERT(sizeof(*dev_link) == sizeof(rte_atomic64_t));
391 
392 	sfc_port_link_mode_to_info(link_mode, &new_link);
393 
394 	new_link_u64 = *(uint64_t *)&new_link;
395 	do {
396 		old_link_u64 = rte_atomic64_read((rte_atomic64_t *)dev_link);
397 		if (old_link_u64 == new_link_u64)
398 			break;
399 
400 		if (rte_atomic64_cmpset((volatile uint64_t *)dev_link,
401 					old_link_u64, new_link_u64)) {
402 			evq->sa->port.lsc_seq++;
403 			break;
404 		}
405 	} while (B_TRUE);
406 
407 	return B_FALSE;
408 }
409 
410 static const efx_ev_callbacks_t sfc_ev_callbacks = {
411 	.eec_initialized	= sfc_ev_initialized,
412 	.eec_rx			= sfc_ev_nop_rx,
413 	.eec_tx			= sfc_ev_nop_tx,
414 	.eec_exception		= sfc_ev_exception,
415 	.eec_rxq_flush_done	= sfc_ev_nop_rxq_flush_done,
416 	.eec_rxq_flush_failed	= sfc_ev_nop_rxq_flush_failed,
417 	.eec_txq_flush_done	= sfc_ev_nop_txq_flush_done,
418 	.eec_software		= sfc_ev_software,
419 	.eec_sram		= sfc_ev_sram,
420 	.eec_wake_up		= sfc_ev_wake_up,
421 	.eec_timer		= sfc_ev_timer,
422 	.eec_link_change	= sfc_ev_link_change,
423 };
424 
425 static const efx_ev_callbacks_t sfc_ev_callbacks_efx_rx = {
426 	.eec_initialized	= sfc_ev_initialized,
427 	.eec_rx			= sfc_ev_efx_rx,
428 	.eec_tx			= sfc_ev_nop_tx,
429 	.eec_exception		= sfc_ev_exception,
430 	.eec_rxq_flush_done	= sfc_ev_rxq_flush_done,
431 	.eec_rxq_flush_failed	= sfc_ev_rxq_flush_failed,
432 	.eec_txq_flush_done	= sfc_ev_nop_txq_flush_done,
433 	.eec_software		= sfc_ev_software,
434 	.eec_sram		= sfc_ev_sram,
435 	.eec_wake_up		= sfc_ev_wake_up,
436 	.eec_timer		= sfc_ev_timer,
437 	.eec_link_change	= sfc_ev_nop_link_change,
438 };
439 
440 static const efx_ev_callbacks_t sfc_ev_callbacks_dp_rx = {
441 	.eec_initialized	= sfc_ev_initialized,
442 	.eec_rx			= sfc_ev_dp_rx,
443 	.eec_tx			= sfc_ev_nop_tx,
444 	.eec_exception		= sfc_ev_exception,
445 	.eec_rxq_flush_done	= sfc_ev_rxq_flush_done,
446 	.eec_rxq_flush_failed	= sfc_ev_rxq_flush_failed,
447 	.eec_txq_flush_done	= sfc_ev_nop_txq_flush_done,
448 	.eec_software		= sfc_ev_software,
449 	.eec_sram		= sfc_ev_sram,
450 	.eec_wake_up		= sfc_ev_wake_up,
451 	.eec_timer		= sfc_ev_timer,
452 	.eec_link_change	= sfc_ev_nop_link_change,
453 };
454 
455 static const efx_ev_callbacks_t sfc_ev_callbacks_efx_tx = {
456 	.eec_initialized	= sfc_ev_initialized,
457 	.eec_rx			= sfc_ev_nop_rx,
458 	.eec_tx			= sfc_ev_tx,
459 	.eec_exception		= sfc_ev_exception,
460 	.eec_rxq_flush_done	= sfc_ev_nop_rxq_flush_done,
461 	.eec_rxq_flush_failed	= sfc_ev_nop_rxq_flush_failed,
462 	.eec_txq_flush_done	= sfc_ev_txq_flush_done,
463 	.eec_software		= sfc_ev_software,
464 	.eec_sram		= sfc_ev_sram,
465 	.eec_wake_up		= sfc_ev_wake_up,
466 	.eec_timer		= sfc_ev_timer,
467 	.eec_link_change	= sfc_ev_nop_link_change,
468 };
469 
470 static const efx_ev_callbacks_t sfc_ev_callbacks_dp_tx = {
471 	.eec_initialized	= sfc_ev_initialized,
472 	.eec_rx			= sfc_ev_nop_rx,
473 	.eec_tx			= sfc_ev_dp_tx,
474 	.eec_exception		= sfc_ev_exception,
475 	.eec_rxq_flush_done	= sfc_ev_nop_rxq_flush_done,
476 	.eec_rxq_flush_failed	= sfc_ev_nop_rxq_flush_failed,
477 	.eec_txq_flush_done	= sfc_ev_txq_flush_done,
478 	.eec_software		= sfc_ev_software,
479 	.eec_sram		= sfc_ev_sram,
480 	.eec_wake_up		= sfc_ev_wake_up,
481 	.eec_timer		= sfc_ev_timer,
482 	.eec_link_change	= sfc_ev_nop_link_change,
483 };
484 
485 
486 void
487 sfc_ev_qpoll(struct sfc_evq *evq)
488 {
489 	SFC_ASSERT(evq->init_state == SFC_EVQ_STARTED ||
490 		   evq->init_state == SFC_EVQ_STARTING);
491 
492 	/* Synchronize the DMA memory for reading not required */
493 
494 	efx_ev_qpoll(evq->common, &evq->read_ptr, evq->callbacks, evq);
495 
496 	if (unlikely(evq->exception) && sfc_adapter_trylock(evq->sa)) {
497 		struct sfc_adapter *sa = evq->sa;
498 		int rc;
499 
500 		if (evq->dp_rxq != NULL) {
501 			unsigned int rxq_sw_index;
502 
503 			rxq_sw_index = evq->dp_rxq->dpq.queue_id;
504 
505 			sfc_warn(sa,
506 				 "restart RxQ %u because of exception on its EvQ %u",
507 				 rxq_sw_index, evq->evq_index);
508 
509 			sfc_rx_qstop(sa, rxq_sw_index);
510 			rc = sfc_rx_qstart(sa, rxq_sw_index);
511 			if (rc != 0)
512 				sfc_err(sa, "cannot restart RxQ %u",
513 					rxq_sw_index);
514 		}
515 
516 		if (evq->dp_txq != NULL) {
517 			unsigned int txq_sw_index;
518 
519 			txq_sw_index = evq->dp_txq->dpq.queue_id;
520 
521 			sfc_warn(sa,
522 				 "restart TxQ %u because of exception on its EvQ %u",
523 				 txq_sw_index, evq->evq_index);
524 
525 			sfc_tx_qstop(sa, txq_sw_index);
526 			rc = sfc_tx_qstart(sa, txq_sw_index);
527 			if (rc != 0)
528 				sfc_err(sa, "cannot restart TxQ %u",
529 					txq_sw_index);
530 		}
531 
532 		if (evq->exception)
533 			sfc_panic(sa, "unrecoverable exception on EvQ %u",
534 				  evq->evq_index);
535 
536 		sfc_adapter_unlock(sa);
537 	}
538 
539 	/* Poll-mode driver does not re-prime the event queue for interrupts */
540 }
541 
542 void
543 sfc_ev_mgmt_qpoll(struct sfc_adapter *sa)
544 {
545 	if (rte_spinlock_trylock(&sa->mgmt_evq_lock)) {
546 		if (sa->mgmt_evq_running)
547 			sfc_ev_qpoll(sa->mgmt_evq);
548 
549 		rte_spinlock_unlock(&sa->mgmt_evq_lock);
550 	}
551 }
552 
553 int
554 sfc_ev_qprime(struct sfc_evq *evq)
555 {
556 	SFC_ASSERT(evq->init_state == SFC_EVQ_STARTED);
557 	return efx_ev_qprime(evq->common, evq->read_ptr);
558 }
559 
560 /* Event queue HW index allocation scheme is described in sfc_ev.h. */
561 int
562 sfc_ev_qstart(struct sfc_evq *evq, unsigned int hw_index)
563 {
564 	struct sfc_adapter *sa = evq->sa;
565 	efsys_mem_t *esmp;
566 	uint32_t evq_flags = sa->evq_flags;
567 	unsigned int total_delay_us;
568 	unsigned int delay_us;
569 	int rc;
570 
571 	sfc_log_init(sa, "hw_index=%u", hw_index);
572 
573 	esmp = &evq->mem;
574 
575 	evq->evq_index = hw_index;
576 
577 	/* Clear all events */
578 	(void)memset((void *)esmp->esm_base, 0xff, EFX_EVQ_SIZE(evq->entries));
579 
580 	if (sa->intr.lsc_intr && hw_index == sa->mgmt_evq_index)
581 		evq_flags |= EFX_EVQ_FLAGS_NOTIFY_INTERRUPT;
582 	else
583 		evq_flags |= EFX_EVQ_FLAGS_NOTIFY_DISABLED;
584 
585 	/* Create the common code event queue */
586 	rc = efx_ev_qcreate(sa->nic, hw_index, esmp, evq->entries,
587 			    0 /* unused on EF10 */, 0, evq_flags,
588 			    &evq->common);
589 	if (rc != 0)
590 		goto fail_ev_qcreate;
591 
592 	SFC_ASSERT(evq->dp_rxq == NULL || evq->dp_txq == NULL);
593 	if (evq->dp_rxq != 0) {
594 		if (strcmp(sa->dp_rx->dp.name, SFC_KVARG_DATAPATH_EFX) == 0)
595 			evq->callbacks = &sfc_ev_callbacks_efx_rx;
596 		else
597 			evq->callbacks = &sfc_ev_callbacks_dp_rx;
598 	} else if (evq->dp_txq != 0) {
599 		if (strcmp(sa->dp_tx->dp.name, SFC_KVARG_DATAPATH_EFX) == 0)
600 			evq->callbacks = &sfc_ev_callbacks_efx_tx;
601 		else
602 			evq->callbacks = &sfc_ev_callbacks_dp_tx;
603 	} else {
604 		evq->callbacks = &sfc_ev_callbacks;
605 	}
606 
607 	evq->init_state = SFC_EVQ_STARTING;
608 
609 	/* Wait for the initialization event */
610 	total_delay_us = 0;
611 	delay_us = SFC_EVQ_INIT_BACKOFF_START_US;
612 	do {
613 		(void)sfc_ev_qpoll(evq);
614 
615 		/* Check to see if the initialization complete indication
616 		 * posted by the hardware.
617 		 */
618 		if (evq->init_state == SFC_EVQ_STARTED)
619 			goto done;
620 
621 		/* Give event queue some time to init */
622 		rte_delay_us(delay_us);
623 
624 		total_delay_us += delay_us;
625 
626 		/* Exponential backoff */
627 		delay_us *= 2;
628 		if (delay_us > SFC_EVQ_INIT_BACKOFF_MAX_US)
629 			delay_us = SFC_EVQ_INIT_BACKOFF_MAX_US;
630 
631 	} while (total_delay_us < SFC_EVQ_INIT_TIMEOUT_US);
632 
633 	rc = ETIMEDOUT;
634 	goto fail_timedout;
635 
636 done:
637 	return 0;
638 
639 fail_timedout:
640 	evq->init_state = SFC_EVQ_INITIALIZED;
641 	efx_ev_qdestroy(evq->common);
642 
643 fail_ev_qcreate:
644 	sfc_log_init(sa, "failed %d", rc);
645 	return rc;
646 }
647 
648 void
649 sfc_ev_qstop(struct sfc_evq *evq)
650 {
651 	if (evq == NULL)
652 		return;
653 
654 	sfc_log_init(evq->sa, "hw_index=%u", evq->evq_index);
655 
656 	if (evq->init_state != SFC_EVQ_STARTED)
657 		return;
658 
659 	evq->init_state = SFC_EVQ_INITIALIZED;
660 	evq->callbacks = NULL;
661 	evq->read_ptr = 0;
662 	evq->exception = B_FALSE;
663 
664 	efx_ev_qdestroy(evq->common);
665 
666 	evq->evq_index = 0;
667 }
668 
669 static void
670 sfc_ev_mgmt_periodic_qpoll(void *arg)
671 {
672 	struct sfc_adapter *sa = arg;
673 	int rc;
674 
675 	sfc_ev_mgmt_qpoll(sa);
676 
677 	rc = rte_eal_alarm_set(SFC_MGMT_EV_QPOLL_PERIOD_US,
678 			       sfc_ev_mgmt_periodic_qpoll, sa);
679 	if (rc == -ENOTSUP) {
680 		sfc_warn(sa, "alarms are not supported");
681 		sfc_warn(sa, "management EVQ must be polled indirectly using no-wait link status update");
682 	} else if (rc != 0) {
683 		sfc_err(sa,
684 			"cannot rearm management EVQ polling alarm (rc=%d)",
685 			rc);
686 	}
687 }
688 
689 static void
690 sfc_ev_mgmt_periodic_qpoll_start(struct sfc_adapter *sa)
691 {
692 	sfc_ev_mgmt_periodic_qpoll(sa);
693 }
694 
695 static void
696 sfc_ev_mgmt_periodic_qpoll_stop(struct sfc_adapter *sa)
697 {
698 	rte_eal_alarm_cancel(sfc_ev_mgmt_periodic_qpoll, sa);
699 }
700 
701 int
702 sfc_ev_start(struct sfc_adapter *sa)
703 {
704 	int rc;
705 
706 	sfc_log_init(sa, "entry");
707 
708 	rc = efx_ev_init(sa->nic);
709 	if (rc != 0)
710 		goto fail_ev_init;
711 
712 	/* Start management EVQ used for global events */
713 
714 	/*
715 	 * Management event queue start polls the queue, but it cannot
716 	 * interfere with other polling contexts since mgmt_evq_running
717 	 * is false yet.
718 	 */
719 	rc = sfc_ev_qstart(sa->mgmt_evq, sa->mgmt_evq_index);
720 	if (rc != 0)
721 		goto fail_mgmt_evq_start;
722 
723 	rte_spinlock_lock(&sa->mgmt_evq_lock);
724 	sa->mgmt_evq_running = true;
725 	rte_spinlock_unlock(&sa->mgmt_evq_lock);
726 
727 	if (sa->intr.lsc_intr) {
728 		rc = sfc_ev_qprime(sa->mgmt_evq);
729 		if (rc != 0)
730 			goto fail_mgmt_evq_prime;
731 	}
732 
733 	/*
734 	 * Start management EVQ polling. If interrupts are disabled
735 	 * (not used), it is required to process link status change
736 	 * and other device level events to avoid unrecoverable
737 	 * error because the event queue overflow.
738 	 */
739 	sfc_ev_mgmt_periodic_qpoll_start(sa);
740 
741 	/*
742 	 * Rx/Tx event queues are started/stopped when corresponding
743 	 * Rx/Tx queue is started/stopped.
744 	 */
745 
746 	return 0;
747 
748 fail_mgmt_evq_prime:
749 	sfc_ev_qstop(sa->mgmt_evq);
750 
751 fail_mgmt_evq_start:
752 	efx_ev_fini(sa->nic);
753 
754 fail_ev_init:
755 	sfc_log_init(sa, "failed %d", rc);
756 	return rc;
757 }
758 
759 void
760 sfc_ev_stop(struct sfc_adapter *sa)
761 {
762 	sfc_log_init(sa, "entry");
763 
764 	sfc_ev_mgmt_periodic_qpoll_stop(sa);
765 
766 	rte_spinlock_lock(&sa->mgmt_evq_lock);
767 	sa->mgmt_evq_running = false;
768 	rte_spinlock_unlock(&sa->mgmt_evq_lock);
769 
770 	sfc_ev_qstop(sa->mgmt_evq);
771 
772 	efx_ev_fini(sa->nic);
773 }
774 
775 int
776 sfc_ev_qinit(struct sfc_adapter *sa,
777 	     enum sfc_evq_type type, unsigned int type_index,
778 	     unsigned int entries, int socket_id, struct sfc_evq **evqp)
779 {
780 	struct sfc_evq *evq;
781 	int rc;
782 
783 	sfc_log_init(sa, "type=%s type_index=%u",
784 		     sfc_evq_type2str(type), type_index);
785 
786 	SFC_ASSERT(rte_is_power_of_2(entries));
787 
788 	rc = ENOMEM;
789 	evq = rte_zmalloc_socket("sfc-evq", sizeof(*evq), RTE_CACHE_LINE_SIZE,
790 				 socket_id);
791 	if (evq == NULL)
792 		goto fail_evq_alloc;
793 
794 	evq->sa = sa;
795 	evq->type = type;
796 	evq->entries = entries;
797 
798 	/* Allocate DMA space */
799 	rc = sfc_dma_alloc(sa, sfc_evq_type2str(type), type_index,
800 			   EFX_EVQ_SIZE(evq->entries), socket_id, &evq->mem);
801 	if (rc != 0)
802 		goto fail_dma_alloc;
803 
804 	evq->init_state = SFC_EVQ_INITIALIZED;
805 
806 	sa->evq_count++;
807 
808 	*evqp = evq;
809 
810 	return 0;
811 
812 fail_dma_alloc:
813 	rte_free(evq);
814 
815 fail_evq_alloc:
816 
817 	sfc_log_init(sa, "failed %d", rc);
818 	return rc;
819 }
820 
821 void
822 sfc_ev_qfini(struct sfc_evq *evq)
823 {
824 	struct sfc_adapter *sa = evq->sa;
825 
826 	SFC_ASSERT(evq->init_state == SFC_EVQ_INITIALIZED);
827 
828 	sfc_dma_free(sa, &evq->mem);
829 
830 	rte_free(evq);
831 
832 	SFC_ASSERT(sa->evq_count > 0);
833 	sa->evq_count--;
834 }
835 
836 static int
837 sfc_kvarg_perf_profile_handler(__rte_unused const char *key,
838 			       const char *value_str, void *opaque)
839 {
840 	uint64_t *value = opaque;
841 
842 	if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_THROUGHPUT) == 0)
843 		*value = EFX_EVQ_FLAGS_TYPE_THROUGHPUT;
844 	else if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_LOW_LATENCY) == 0)
845 		*value = EFX_EVQ_FLAGS_TYPE_LOW_LATENCY;
846 	else if (strcasecmp(value_str, SFC_KVARG_PERF_PROFILE_AUTO) == 0)
847 		*value = EFX_EVQ_FLAGS_TYPE_AUTO;
848 	else
849 		return -EINVAL;
850 
851 	return 0;
852 }
853 
854 int
855 sfc_ev_attach(struct sfc_adapter *sa)
856 {
857 	int rc;
858 
859 	sfc_log_init(sa, "entry");
860 
861 	sa->evq_flags = EFX_EVQ_FLAGS_TYPE_THROUGHPUT;
862 	rc = sfc_kvargs_process(sa, SFC_KVARG_PERF_PROFILE,
863 				sfc_kvarg_perf_profile_handler,
864 				&sa->evq_flags);
865 	if (rc != 0) {
866 		sfc_err(sa, "invalid %s parameter value",
867 			SFC_KVARG_PERF_PROFILE);
868 		goto fail_kvarg_perf_profile;
869 	}
870 
871 	sa->mgmt_evq_index = 0;
872 	rte_spinlock_init(&sa->mgmt_evq_lock);
873 
874 	rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_MGMT, 0, SFC_MGMT_EVQ_ENTRIES,
875 			  sa->socket_id, &sa->mgmt_evq);
876 	if (rc != 0)
877 		goto fail_mgmt_evq_init;
878 
879 	/*
880 	 * Rx/Tx event queues are created/destroyed when corresponding
881 	 * Rx/Tx queue is created/destroyed.
882 	 */
883 
884 	return 0;
885 
886 fail_mgmt_evq_init:
887 
888 fail_kvarg_perf_profile:
889 	sfc_log_init(sa, "failed %d", rc);
890 	return rc;
891 }
892 
893 void
894 sfc_ev_detach(struct sfc_adapter *sa)
895 {
896 	sfc_log_init(sa, "entry");
897 
898 	sfc_ev_qfini(sa->mgmt_evq);
899 
900 	if (sa->evq_count != 0)
901 		sfc_err(sa, "%u EvQs are not destroyed before detach",
902 			sa->evq_count);
903 }
904