1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #include <linux/skbuff.h>
8 
9 #include "rxe.h"
10 #include "rxe_loc.h"
11 #include "rxe_queue.h"
12 #include "rxe_task.h"
13 
14 enum comp_state {
15 	COMPST_GET_ACK,
16 	COMPST_GET_WQE,
17 	COMPST_COMP_WQE,
18 	COMPST_COMP_ACK,
19 	COMPST_CHECK_PSN,
20 	COMPST_CHECK_ACK,
21 	COMPST_READ,
22 	COMPST_ATOMIC,
23 	COMPST_WRITE_SEND,
24 	COMPST_UPDATE_COMP,
25 	COMPST_ERROR_RETRY,
26 	COMPST_RNR_RETRY,
27 	COMPST_ERROR,
28 	COMPST_EXIT, /* We have an issue, and we want to rerun the completer */
29 	COMPST_DONE, /* The completer finished successflly */
30 };
31 
32 static char *comp_state_name[] =  {
33 	[COMPST_GET_ACK]		= "GET ACK",
34 	[COMPST_GET_WQE]		= "GET WQE",
35 	[COMPST_COMP_WQE]		= "COMP WQE",
36 	[COMPST_COMP_ACK]		= "COMP ACK",
37 	[COMPST_CHECK_PSN]		= "CHECK PSN",
38 	[COMPST_CHECK_ACK]		= "CHECK ACK",
39 	[COMPST_READ]			= "READ",
40 	[COMPST_ATOMIC]			= "ATOMIC",
41 	[COMPST_WRITE_SEND]		= "WRITE/SEND",
42 	[COMPST_UPDATE_COMP]		= "UPDATE COMP",
43 	[COMPST_ERROR_RETRY]		= "ERROR RETRY",
44 	[COMPST_RNR_RETRY]		= "RNR RETRY",
45 	[COMPST_ERROR]			= "ERROR",
46 	[COMPST_EXIT]			= "EXIT",
47 	[COMPST_DONE]			= "DONE",
48 };
49 
50 static unsigned long rnrnak_usec[32] = {
51 	[IB_RNR_TIMER_655_36] = 655360,
52 	[IB_RNR_TIMER_000_01] = 10,
53 	[IB_RNR_TIMER_000_02] = 20,
54 	[IB_RNR_TIMER_000_03] = 30,
55 	[IB_RNR_TIMER_000_04] = 40,
56 	[IB_RNR_TIMER_000_06] = 60,
57 	[IB_RNR_TIMER_000_08] = 80,
58 	[IB_RNR_TIMER_000_12] = 120,
59 	[IB_RNR_TIMER_000_16] = 160,
60 	[IB_RNR_TIMER_000_24] = 240,
61 	[IB_RNR_TIMER_000_32] = 320,
62 	[IB_RNR_TIMER_000_48] = 480,
63 	[IB_RNR_TIMER_000_64] = 640,
64 	[IB_RNR_TIMER_000_96] = 960,
65 	[IB_RNR_TIMER_001_28] = 1280,
66 	[IB_RNR_TIMER_001_92] = 1920,
67 	[IB_RNR_TIMER_002_56] = 2560,
68 	[IB_RNR_TIMER_003_84] = 3840,
69 	[IB_RNR_TIMER_005_12] = 5120,
70 	[IB_RNR_TIMER_007_68] = 7680,
71 	[IB_RNR_TIMER_010_24] = 10240,
72 	[IB_RNR_TIMER_015_36] = 15360,
73 	[IB_RNR_TIMER_020_48] = 20480,
74 	[IB_RNR_TIMER_030_72] = 30720,
75 	[IB_RNR_TIMER_040_96] = 40960,
76 	[IB_RNR_TIMER_061_44] = 61410,
77 	[IB_RNR_TIMER_081_92] = 81920,
78 	[IB_RNR_TIMER_122_88] = 122880,
79 	[IB_RNR_TIMER_163_84] = 163840,
80 	[IB_RNR_TIMER_245_76] = 245760,
81 	[IB_RNR_TIMER_327_68] = 327680,
82 	[IB_RNR_TIMER_491_52] = 491520,
83 };
84 
85 static inline unsigned long rnrnak_jiffies(u8 timeout)
86 {
87 	return max_t(unsigned long,
88 		usecs_to_jiffies(rnrnak_usec[timeout]), 1);
89 }
90 
91 static enum ib_wc_opcode wr_to_wc_opcode(enum ib_wr_opcode opcode)
92 {
93 	switch (opcode) {
94 	case IB_WR_RDMA_WRITE:			return IB_WC_RDMA_WRITE;
95 	case IB_WR_RDMA_WRITE_WITH_IMM:		return IB_WC_RDMA_WRITE;
96 	case IB_WR_SEND:			return IB_WC_SEND;
97 	case IB_WR_SEND_WITH_IMM:		return IB_WC_SEND;
98 	case IB_WR_RDMA_READ:			return IB_WC_RDMA_READ;
99 	case IB_WR_ATOMIC_CMP_AND_SWP:		return IB_WC_COMP_SWAP;
100 	case IB_WR_ATOMIC_FETCH_AND_ADD:	return IB_WC_FETCH_ADD;
101 	case IB_WR_LSO:				return IB_WC_LSO;
102 	case IB_WR_SEND_WITH_INV:		return IB_WC_SEND;
103 	case IB_WR_RDMA_READ_WITH_INV:		return IB_WC_RDMA_READ;
104 	case IB_WR_LOCAL_INV:			return IB_WC_LOCAL_INV;
105 	case IB_WR_REG_MR:			return IB_WC_REG_MR;
106 	case IB_WR_BIND_MW:			return IB_WC_BIND_MW;
107 	case IB_WR_ATOMIC_WRITE:		return IB_WC_ATOMIC_WRITE;
108 	case IB_WR_FLUSH:			return IB_WC_FLUSH;
109 
110 	default:
111 		return 0xff;
112 	}
113 }
114 
115 void retransmit_timer(struct timer_list *t)
116 {
117 	struct rxe_qp *qp = from_timer(qp, t, retrans_timer);
118 	unsigned long flags;
119 
120 	rxe_dbg_qp(qp, "retransmit timer fired\n");
121 
122 	spin_lock_irqsave(&qp->state_lock, flags);
123 	if (qp->valid) {
124 		qp->comp.timeout = 1;
125 		rxe_sched_task(&qp->comp.task);
126 	}
127 	spin_unlock_irqrestore(&qp->state_lock, flags);
128 }
129 
130 void rxe_comp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
131 {
132 	int must_sched;
133 
134 	skb_queue_tail(&qp->resp_pkts, skb);
135 
136 	must_sched = skb_queue_len(&qp->resp_pkts) > 1;
137 	if (must_sched != 0)
138 		rxe_counter_inc(SKB_TO_PKT(skb)->rxe, RXE_CNT_COMPLETER_SCHED);
139 
140 	if (must_sched)
141 		rxe_sched_task(&qp->comp.task);
142 	else
143 		rxe_run_task(&qp->comp.task);
144 }
145 
146 static inline enum comp_state get_wqe(struct rxe_qp *qp,
147 				      struct rxe_pkt_info *pkt,
148 				      struct rxe_send_wqe **wqe_p)
149 {
150 	struct rxe_send_wqe *wqe;
151 
152 	/* we come here whether or not we found a response packet to see if
153 	 * there are any posted WQEs
154 	 */
155 	wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
156 	*wqe_p = wqe;
157 
158 	/* no WQE or requester has not started it yet */
159 	if (!wqe || wqe->state == wqe_state_posted)
160 		return pkt ? COMPST_DONE : COMPST_EXIT;
161 
162 	/* WQE does not require an ack */
163 	if (wqe->state == wqe_state_done)
164 		return COMPST_COMP_WQE;
165 
166 	/* WQE caused an error */
167 	if (wqe->state == wqe_state_error)
168 		return COMPST_ERROR;
169 
170 	/* we have a WQE, if we also have an ack check its PSN */
171 	return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
172 }
173 
174 static inline void reset_retry_counters(struct rxe_qp *qp)
175 {
176 	qp->comp.retry_cnt = qp->attr.retry_cnt;
177 	qp->comp.rnr_retry = qp->attr.rnr_retry;
178 	qp->comp.started_retry = 0;
179 }
180 
181 static inline enum comp_state check_psn(struct rxe_qp *qp,
182 					struct rxe_pkt_info *pkt,
183 					struct rxe_send_wqe *wqe)
184 {
185 	s32 diff;
186 
187 	/* check to see if response is past the oldest WQE. if it is, complete
188 	 * send/write or error read/atomic
189 	 */
190 	diff = psn_compare(pkt->psn, wqe->last_psn);
191 	if (diff > 0) {
192 		if (wqe->state == wqe_state_pending) {
193 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
194 				return COMPST_ERROR_RETRY;
195 
196 			reset_retry_counters(qp);
197 			return COMPST_COMP_WQE;
198 		} else {
199 			return COMPST_DONE;
200 		}
201 	}
202 
203 	/* compare response packet to expected response */
204 	diff = psn_compare(pkt->psn, qp->comp.psn);
205 	if (diff < 0) {
206 		/* response is most likely a retried packet if it matches an
207 		 * uncompleted WQE go complete it else ignore it
208 		 */
209 		if (pkt->psn == wqe->last_psn)
210 			return COMPST_COMP_ACK;
211 		else if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE &&
212 			 (qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST ||
213 			  qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE))
214 			return COMPST_CHECK_ACK;
215 		else
216 			return COMPST_DONE;
217 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
218 		return COMPST_DONE;
219 	} else {
220 		return COMPST_CHECK_ACK;
221 	}
222 }
223 
224 static inline enum comp_state check_ack(struct rxe_qp *qp,
225 					struct rxe_pkt_info *pkt,
226 					struct rxe_send_wqe *wqe)
227 {
228 	unsigned int mask = pkt->mask;
229 	u8 syn;
230 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
231 
232 	/* Check the sequence only */
233 	switch (qp->comp.opcode) {
234 	case -1:
235 		/* Will catch all *_ONLY cases. */
236 		if (!(mask & RXE_START_MASK))
237 			return COMPST_ERROR;
238 
239 		break;
240 
241 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
242 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
243 		/* Check NAK code to handle a remote error */
244 		if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE)
245 			break;
246 
247 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
248 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
249 			/* read retries of partial data may restart from
250 			 * read response first or response only.
251 			 */
252 			if ((pkt->psn == wqe->first_psn &&
253 			     pkt->opcode ==
254 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) ||
255 			    (wqe->first_psn == wqe->last_psn &&
256 			     pkt->opcode ==
257 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY))
258 				break;
259 
260 			return COMPST_ERROR;
261 		}
262 		break;
263 	default:
264 		WARN_ON_ONCE(1);
265 	}
266 
267 	/* Check operation validity. */
268 	switch (pkt->opcode) {
269 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
270 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
271 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
272 		syn = aeth_syn(pkt);
273 
274 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
275 			return COMPST_ERROR;
276 
277 		if (wqe->wr.opcode == IB_WR_ATOMIC_WRITE)
278 			return COMPST_WRITE_SEND;
279 
280 		fallthrough;
281 		/* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
282 		 */
283 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
284 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
285 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV &&
286 		    wqe->wr.opcode != IB_WR_FLUSH) {
287 			wqe->status = IB_WC_FATAL_ERR;
288 			return COMPST_ERROR;
289 		}
290 		reset_retry_counters(qp);
291 		return COMPST_READ;
292 
293 	case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
294 		syn = aeth_syn(pkt);
295 
296 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
297 			return COMPST_ERROR;
298 
299 		if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP &&
300 		    wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
301 			return COMPST_ERROR;
302 		reset_retry_counters(qp);
303 		return COMPST_ATOMIC;
304 
305 	case IB_OPCODE_RC_ACKNOWLEDGE:
306 		syn = aeth_syn(pkt);
307 		switch (syn & AETH_TYPE_MASK) {
308 		case AETH_ACK:
309 			reset_retry_counters(qp);
310 			return COMPST_WRITE_SEND;
311 
312 		case AETH_RNR_NAK:
313 			rxe_counter_inc(rxe, RXE_CNT_RCV_RNR);
314 			return COMPST_RNR_RETRY;
315 
316 		case AETH_NAK:
317 			switch (syn) {
318 			case AETH_NAK_PSN_SEQ_ERROR:
319 				/* a nak implicitly acks all packets with psns
320 				 * before
321 				 */
322 				if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
323 					rxe_counter_inc(rxe,
324 							RXE_CNT_RCV_SEQ_ERR);
325 					qp->comp.psn = pkt->psn;
326 					if (qp->req.wait_psn) {
327 						qp->req.wait_psn = 0;
328 						rxe_sched_task(&qp->req.task);
329 					}
330 				}
331 				return COMPST_ERROR_RETRY;
332 
333 			case AETH_NAK_INVALID_REQ:
334 				wqe->status = IB_WC_REM_INV_REQ_ERR;
335 				return COMPST_ERROR;
336 
337 			case AETH_NAK_REM_ACC_ERR:
338 				wqe->status = IB_WC_REM_ACCESS_ERR;
339 				return COMPST_ERROR;
340 
341 			case AETH_NAK_REM_OP_ERR:
342 				wqe->status = IB_WC_REM_OP_ERR;
343 				return COMPST_ERROR;
344 
345 			default:
346 				rxe_dbg_qp(qp, "unexpected nak %x\n", syn);
347 				wqe->status = IB_WC_REM_OP_ERR;
348 				return COMPST_ERROR;
349 			}
350 
351 		default:
352 			return COMPST_ERROR;
353 		}
354 		break;
355 
356 	default:
357 		rxe_dbg_qp(qp, "unexpected opcode\n");
358 	}
359 
360 	return COMPST_ERROR;
361 }
362 
363 static inline enum comp_state do_read(struct rxe_qp *qp,
364 				      struct rxe_pkt_info *pkt,
365 				      struct rxe_send_wqe *wqe)
366 {
367 	int ret;
368 
369 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
370 			&wqe->dma, payload_addr(pkt),
371 			payload_size(pkt), RXE_TO_MR_OBJ);
372 	if (ret) {
373 		wqe->status = IB_WC_LOC_PROT_ERR;
374 		return COMPST_ERROR;
375 	}
376 
377 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
378 		return COMPST_COMP_ACK;
379 
380 	return COMPST_UPDATE_COMP;
381 }
382 
383 static inline enum comp_state do_atomic(struct rxe_qp *qp,
384 					struct rxe_pkt_info *pkt,
385 					struct rxe_send_wqe *wqe)
386 {
387 	int ret;
388 
389 	u64 atomic_orig = atmack_orig(pkt);
390 
391 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
392 			&wqe->dma, &atomic_orig,
393 			sizeof(u64), RXE_TO_MR_OBJ);
394 	if (ret) {
395 		wqe->status = IB_WC_LOC_PROT_ERR;
396 		return COMPST_ERROR;
397 	}
398 
399 	return COMPST_COMP_ACK;
400 }
401 
402 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
403 			  struct rxe_cqe *cqe)
404 {
405 	struct ib_wc *wc = &cqe->ibwc;
406 	struct ib_uverbs_wc *uwc = &cqe->uibwc;
407 
408 	memset(cqe, 0, sizeof(*cqe));
409 
410 	if (!qp->is_user) {
411 		wc->wr_id = wqe->wr.wr_id;
412 		wc->status = wqe->status;
413 		wc->qp = &qp->ibqp;
414 	} else {
415 		uwc->wr_id = wqe->wr.wr_id;
416 		uwc->status = wqe->status;
417 		uwc->qp_num = qp->ibqp.qp_num;
418 	}
419 
420 	if (wqe->status == IB_WC_SUCCESS) {
421 		if (!qp->is_user) {
422 			wc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
423 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
424 			    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
425 				wc->wc_flags = IB_WC_WITH_IMM;
426 			wc->byte_len = wqe->dma.length;
427 		} else {
428 			uwc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
429 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
430 			    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
431 				uwc->wc_flags = IB_WC_WITH_IMM;
432 			uwc->byte_len = wqe->dma.length;
433 		}
434 	} else {
435 		if (wqe->status != IB_WC_WR_FLUSH_ERR)
436 			rxe_err_qp(qp, "non-flush error status = %d",
437 				wqe->status);
438 	}
439 }
440 
441 /*
442  * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
443  * ---------8<---------8<-------------
444  * ...Note that if a completion error occurs, a Work Completion
445  * will always be generated, even if the signaling
446  * indicator requests an Unsignaled Completion.
447  * ---------8<---------8<-------------
448  */
449 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
450 {
451 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
452 	struct rxe_cqe cqe;
453 	bool post;
454 
455 	/* do we need to post a completion */
456 	post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
457 			(wqe->wr.send_flags & IB_SEND_SIGNALED) ||
458 			wqe->status != IB_WC_SUCCESS);
459 
460 	if (post)
461 		make_send_cqe(qp, wqe, &cqe);
462 
463 	queue_advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
464 
465 	if (post)
466 		rxe_cq_post(qp->scq, &cqe, 0);
467 
468 	if (wqe->wr.opcode == IB_WR_SEND ||
469 	    wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
470 	    wqe->wr.opcode == IB_WR_SEND_WITH_INV)
471 		rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND);
472 
473 	/*
474 	 * we completed something so let req run again
475 	 * if it is trying to fence
476 	 */
477 	if (qp->req.wait_fence) {
478 		qp->req.wait_fence = 0;
479 		rxe_sched_task(&qp->req.task);
480 	}
481 }
482 
483 static void comp_check_sq_drain_done(struct rxe_qp *qp)
484 {
485 	unsigned long flags;
486 
487 	spin_lock_irqsave(&qp->state_lock, flags);
488 	if (unlikely(qp_state(qp) == IB_QPS_SQD)) {
489 		if (qp->attr.sq_draining && qp->comp.psn == qp->req.psn) {
490 			qp->attr.sq_draining = 0;
491 			spin_unlock_irqrestore(&qp->state_lock, flags);
492 
493 			if (qp->ibqp.event_handler) {
494 				struct ib_event ev;
495 
496 				ev.device = qp->ibqp.device;
497 				ev.element.qp = &qp->ibqp;
498 				ev.event = IB_EVENT_SQ_DRAINED;
499 				qp->ibqp.event_handler(&ev,
500 					qp->ibqp.qp_context);
501 			}
502 			return;
503 		}
504 	}
505 	spin_unlock_irqrestore(&qp->state_lock, flags);
506 }
507 
508 static inline enum comp_state complete_ack(struct rxe_qp *qp,
509 					   struct rxe_pkt_info *pkt,
510 					   struct rxe_send_wqe *wqe)
511 {
512 	if (wqe->has_rd_atomic) {
513 		wqe->has_rd_atomic = 0;
514 		atomic_inc(&qp->req.rd_atomic);
515 		if (qp->req.need_rd_atomic) {
516 			qp->comp.timeout_retry = 0;
517 			qp->req.need_rd_atomic = 0;
518 			rxe_sched_task(&qp->req.task);
519 		}
520 	}
521 
522 	comp_check_sq_drain_done(qp);
523 
524 	do_complete(qp, wqe);
525 
526 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
527 		return COMPST_UPDATE_COMP;
528 	else
529 		return COMPST_DONE;
530 }
531 
532 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
533 					   struct rxe_pkt_info *pkt,
534 					   struct rxe_send_wqe *wqe)
535 {
536 	if (pkt && wqe->state == wqe_state_pending) {
537 		if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) {
538 			qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK;
539 			qp->comp.opcode = -1;
540 		}
541 
542 		if (qp->req.wait_psn) {
543 			qp->req.wait_psn = 0;
544 			rxe_sched_task(&qp->req.task);
545 		}
546 	}
547 
548 	do_complete(qp, wqe);
549 
550 	return COMPST_GET_WQE;
551 }
552 
553 /* drain incoming response packet queue */
554 static void drain_resp_pkts(struct rxe_qp *qp)
555 {
556 	struct sk_buff *skb;
557 
558 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
559 		rxe_put(qp);
560 		kfree_skb(skb);
561 		ib_device_put(qp->ibqp.device);
562 	}
563 }
564 
565 /* complete send wqe with flush error */
566 static int flush_send_wqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
567 {
568 	struct rxe_cqe cqe = {};
569 	struct ib_wc *wc = &cqe.ibwc;
570 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
571 	int err;
572 
573 	if (qp->is_user) {
574 		uwc->wr_id = wqe->wr.wr_id;
575 		uwc->status = IB_WC_WR_FLUSH_ERR;
576 		uwc->qp_num = qp->ibqp.qp_num;
577 	} else {
578 		wc->wr_id = wqe->wr.wr_id;
579 		wc->status = IB_WC_WR_FLUSH_ERR;
580 		wc->qp = &qp->ibqp;
581 	}
582 
583 	err = rxe_cq_post(qp->scq, &cqe, 0);
584 	if (err)
585 		rxe_dbg_cq(qp->scq, "post cq failed, err = %d", err);
586 
587 	return err;
588 }
589 
590 /* drain and optionally complete the send queue
591  * if unable to complete a wqe, i.e. cq is full, stop
592  * completing and flush the remaining wqes
593  */
594 static void flush_send_queue(struct rxe_qp *qp, bool notify)
595 {
596 	struct rxe_send_wqe *wqe;
597 	struct rxe_queue *q = qp->sq.queue;
598 	int err;
599 
600 	while ((wqe = queue_head(q, q->type))) {
601 		if (notify) {
602 			err = flush_send_wqe(qp, wqe);
603 			if (err)
604 				notify = 0;
605 		}
606 		queue_advance_consumer(q, q->type);
607 	}
608 }
609 
610 static void free_pkt(struct rxe_pkt_info *pkt)
611 {
612 	struct sk_buff *skb = PKT_TO_SKB(pkt);
613 	struct rxe_qp *qp = pkt->qp;
614 	struct ib_device *dev = qp->ibqp.device;
615 
616 	kfree_skb(skb);
617 	rxe_put(qp);
618 	ib_device_put(dev);
619 }
620 
621 /* reset the retry timer if
622  * - QP is type RC
623  * - there is a packet sent by the requester that
624  *   might be acked (we still might get spurious
625  *   timeouts but try to keep them as few as possible)
626  * - the timeout parameter is set
627  * - the QP is alive
628  */
629 static void reset_retry_timer(struct rxe_qp *qp)
630 {
631 	unsigned long flags;
632 
633 	if (qp_type(qp) == IB_QPT_RC && qp->qp_timeout_jiffies) {
634 		spin_lock_irqsave(&qp->state_lock, flags);
635 		if (qp_state(qp) >= IB_QPS_RTS &&
636 		    psn_compare(qp->req.psn, qp->comp.psn) > 0)
637 			mod_timer(&qp->retrans_timer,
638 				  jiffies + qp->qp_timeout_jiffies);
639 		spin_unlock_irqrestore(&qp->state_lock, flags);
640 	}
641 }
642 
643 int rxe_completer(struct rxe_qp *qp)
644 {
645 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
646 	struct rxe_send_wqe *wqe = NULL;
647 	struct sk_buff *skb = NULL;
648 	struct rxe_pkt_info *pkt = NULL;
649 	enum comp_state state;
650 	int ret;
651 	unsigned long flags;
652 
653 	spin_lock_irqsave(&qp->state_lock, flags);
654 	if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
655 			  qp_state(qp) == IB_QPS_RESET) {
656 		bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
657 
658 		drain_resp_pkts(qp);
659 		flush_send_queue(qp, notify);
660 		spin_unlock_irqrestore(&qp->state_lock, flags);
661 		goto exit;
662 	}
663 	spin_unlock_irqrestore(&qp->state_lock, flags);
664 
665 	if (qp->comp.timeout) {
666 		qp->comp.timeout_retry = 1;
667 		qp->comp.timeout = 0;
668 	} else {
669 		qp->comp.timeout_retry = 0;
670 	}
671 
672 	if (qp->req.need_retry)
673 		goto exit;
674 
675 	state = COMPST_GET_ACK;
676 
677 	while (1) {
678 		rxe_dbg_qp(qp, "state = %s\n", comp_state_name[state]);
679 		switch (state) {
680 		case COMPST_GET_ACK:
681 			skb = skb_dequeue(&qp->resp_pkts);
682 			if (skb) {
683 				pkt = SKB_TO_PKT(skb);
684 				qp->comp.timeout_retry = 0;
685 			}
686 			state = COMPST_GET_WQE;
687 			break;
688 
689 		case COMPST_GET_WQE:
690 			state = get_wqe(qp, pkt, &wqe);
691 			break;
692 
693 		case COMPST_CHECK_PSN:
694 			state = check_psn(qp, pkt, wqe);
695 			break;
696 
697 		case COMPST_CHECK_ACK:
698 			state = check_ack(qp, pkt, wqe);
699 			break;
700 
701 		case COMPST_READ:
702 			state = do_read(qp, pkt, wqe);
703 			break;
704 
705 		case COMPST_ATOMIC:
706 			state = do_atomic(qp, pkt, wqe);
707 			break;
708 
709 		case COMPST_WRITE_SEND:
710 			if (wqe->state == wqe_state_pending &&
711 			    wqe->last_psn == pkt->psn)
712 				state = COMPST_COMP_ACK;
713 			else
714 				state = COMPST_UPDATE_COMP;
715 			break;
716 
717 		case COMPST_COMP_ACK:
718 			state = complete_ack(qp, pkt, wqe);
719 			break;
720 
721 		case COMPST_COMP_WQE:
722 			state = complete_wqe(qp, pkt, wqe);
723 			break;
724 
725 		case COMPST_UPDATE_COMP:
726 			if (pkt->mask & RXE_END_MASK)
727 				qp->comp.opcode = -1;
728 			else
729 				qp->comp.opcode = pkt->opcode;
730 
731 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
732 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
733 
734 			if (qp->req.wait_psn) {
735 				qp->req.wait_psn = 0;
736 				rxe_sched_task(&qp->req.task);
737 			}
738 
739 			state = COMPST_DONE;
740 			break;
741 
742 		case COMPST_DONE:
743 			goto done;
744 
745 		case COMPST_EXIT:
746 			if (qp->comp.timeout_retry && wqe) {
747 				state = COMPST_ERROR_RETRY;
748 				break;
749 			}
750 
751 			reset_retry_timer(qp);
752 			goto exit;
753 
754 		case COMPST_ERROR_RETRY:
755 			/* we come here if the retry timer fired and we did
756 			 * not receive a response packet. try to retry the send
757 			 * queue if that makes sense and the limits have not
758 			 * been exceeded. remember that some timeouts are
759 			 * spurious since we do not reset the timer but kick
760 			 * it down the road or let it expire
761 			 */
762 
763 			/* there is nothing to retry in this case */
764 			if (!wqe || (wqe->state == wqe_state_posted))
765 				goto exit;
766 
767 			/* if we've started a retry, don't start another
768 			 * retry sequence, unless this is a timeout.
769 			 */
770 			if (qp->comp.started_retry &&
771 			    !qp->comp.timeout_retry)
772 				goto done;
773 
774 			if (qp->comp.retry_cnt > 0) {
775 				if (qp->comp.retry_cnt != 7)
776 					qp->comp.retry_cnt--;
777 
778 				/* no point in retrying if we have already
779 				 * seen the last ack that the requester could
780 				 * have caused
781 				 */
782 				if (psn_compare(qp->req.psn,
783 						qp->comp.psn) > 0) {
784 					/* tell the requester to retry the
785 					 * send queue next time around
786 					 */
787 					rxe_counter_inc(rxe,
788 							RXE_CNT_COMP_RETRY);
789 					qp->req.need_retry = 1;
790 					qp->comp.started_retry = 1;
791 					rxe_sched_task(&qp->req.task);
792 				}
793 				goto done;
794 
795 			} else {
796 				rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
797 				wqe->status = IB_WC_RETRY_EXC_ERR;
798 				state = COMPST_ERROR;
799 			}
800 			break;
801 
802 		case COMPST_RNR_RETRY:
803 			/* we come here if we received an RNR NAK */
804 			if (qp->comp.rnr_retry > 0) {
805 				if (qp->comp.rnr_retry != 7)
806 					qp->comp.rnr_retry--;
807 
808 				/* don't start a retry flow until the
809 				 * rnr timer has fired
810 				 */
811 				qp->req.wait_for_rnr_timer = 1;
812 				rxe_dbg_qp(qp, "set rnr nak timer\n");
813 				// TODO who protects from destroy_qp??
814 				mod_timer(&qp->rnr_nak_timer,
815 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
816 						& ~AETH_TYPE_MASK));
817 				goto exit;
818 			} else {
819 				rxe_counter_inc(rxe,
820 						RXE_CNT_RNR_RETRY_EXCEEDED);
821 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
822 				state = COMPST_ERROR;
823 			}
824 			break;
825 
826 		case COMPST_ERROR:
827 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
828 			do_complete(qp, wqe);
829 			rxe_qp_error(qp);
830 			goto exit;
831 		}
832 	}
833 
834 	/* A non-zero return value will cause rxe_do_task to
835 	 * exit its loop and end the work item. A zero return
836 	 * will continue looping and return to rxe_completer
837 	 */
838 done:
839 	ret = 0;
840 	goto out;
841 exit:
842 	ret = -EAGAIN;
843 out:
844 	if (pkt)
845 		free_pkt(pkt);
846 	return ret;
847 }
848