xref: /linux/drivers/infiniband/sw/rxe/rxe_comp.c (revision bd628c1bed7902ec1f24ba0fe70758949146abbe)
1 /*
2  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *	- Redistributions of source code must retain the above
16  *	  copyright notice, this list of conditions and the following
17  *	  disclaimer.
18  *
19  *	- Redistributions in binary form must reproduce the above
20  *	  copyright notice, this list of conditions and the following
21  *	  disclaimer in the documentation and/or other materials
22  *	  provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/skbuff.h>
35 
36 #include "rxe.h"
37 #include "rxe_loc.h"
38 #include "rxe_queue.h"
39 #include "rxe_task.h"
40 
41 enum comp_state {
42 	COMPST_GET_ACK,
43 	COMPST_GET_WQE,
44 	COMPST_COMP_WQE,
45 	COMPST_COMP_ACK,
46 	COMPST_CHECK_PSN,
47 	COMPST_CHECK_ACK,
48 	COMPST_READ,
49 	COMPST_ATOMIC,
50 	COMPST_WRITE_SEND,
51 	COMPST_UPDATE_COMP,
52 	COMPST_ERROR_RETRY,
53 	COMPST_RNR_RETRY,
54 	COMPST_ERROR,
55 	COMPST_EXIT, /* We have an issue, and we want to rerun the completer */
56 	COMPST_DONE, /* The completer finished successflly */
57 };
58 
59 static char *comp_state_name[] =  {
60 	[COMPST_GET_ACK]		= "GET ACK",
61 	[COMPST_GET_WQE]		= "GET WQE",
62 	[COMPST_COMP_WQE]		= "COMP WQE",
63 	[COMPST_COMP_ACK]		= "COMP ACK",
64 	[COMPST_CHECK_PSN]		= "CHECK PSN",
65 	[COMPST_CHECK_ACK]		= "CHECK ACK",
66 	[COMPST_READ]			= "READ",
67 	[COMPST_ATOMIC]			= "ATOMIC",
68 	[COMPST_WRITE_SEND]		= "WRITE/SEND",
69 	[COMPST_UPDATE_COMP]		= "UPDATE COMP",
70 	[COMPST_ERROR_RETRY]		= "ERROR RETRY",
71 	[COMPST_RNR_RETRY]		= "RNR RETRY",
72 	[COMPST_ERROR]			= "ERROR",
73 	[COMPST_EXIT]			= "EXIT",
74 	[COMPST_DONE]			= "DONE",
75 };
76 
77 static unsigned long rnrnak_usec[32] = {
78 	[IB_RNR_TIMER_655_36] = 655360,
79 	[IB_RNR_TIMER_000_01] = 10,
80 	[IB_RNR_TIMER_000_02] = 20,
81 	[IB_RNR_TIMER_000_03] = 30,
82 	[IB_RNR_TIMER_000_04] = 40,
83 	[IB_RNR_TIMER_000_06] = 60,
84 	[IB_RNR_TIMER_000_08] = 80,
85 	[IB_RNR_TIMER_000_12] = 120,
86 	[IB_RNR_TIMER_000_16] = 160,
87 	[IB_RNR_TIMER_000_24] = 240,
88 	[IB_RNR_TIMER_000_32] = 320,
89 	[IB_RNR_TIMER_000_48] = 480,
90 	[IB_RNR_TIMER_000_64] = 640,
91 	[IB_RNR_TIMER_000_96] = 960,
92 	[IB_RNR_TIMER_001_28] = 1280,
93 	[IB_RNR_TIMER_001_92] = 1920,
94 	[IB_RNR_TIMER_002_56] = 2560,
95 	[IB_RNR_TIMER_003_84] = 3840,
96 	[IB_RNR_TIMER_005_12] = 5120,
97 	[IB_RNR_TIMER_007_68] = 7680,
98 	[IB_RNR_TIMER_010_24] = 10240,
99 	[IB_RNR_TIMER_015_36] = 15360,
100 	[IB_RNR_TIMER_020_48] = 20480,
101 	[IB_RNR_TIMER_030_72] = 30720,
102 	[IB_RNR_TIMER_040_96] = 40960,
103 	[IB_RNR_TIMER_061_44] = 61410,
104 	[IB_RNR_TIMER_081_92] = 81920,
105 	[IB_RNR_TIMER_122_88] = 122880,
106 	[IB_RNR_TIMER_163_84] = 163840,
107 	[IB_RNR_TIMER_245_76] = 245760,
108 	[IB_RNR_TIMER_327_68] = 327680,
109 	[IB_RNR_TIMER_491_52] = 491520,
110 };
111 
112 static inline unsigned long rnrnak_jiffies(u8 timeout)
113 {
114 	return max_t(unsigned long,
115 		usecs_to_jiffies(rnrnak_usec[timeout]), 1);
116 }
117 
118 static enum ib_wc_opcode wr_to_wc_opcode(enum ib_wr_opcode opcode)
119 {
120 	switch (opcode) {
121 	case IB_WR_RDMA_WRITE:			return IB_WC_RDMA_WRITE;
122 	case IB_WR_RDMA_WRITE_WITH_IMM:		return IB_WC_RDMA_WRITE;
123 	case IB_WR_SEND:			return IB_WC_SEND;
124 	case IB_WR_SEND_WITH_IMM:		return IB_WC_SEND;
125 	case IB_WR_RDMA_READ:			return IB_WC_RDMA_READ;
126 	case IB_WR_ATOMIC_CMP_AND_SWP:		return IB_WC_COMP_SWAP;
127 	case IB_WR_ATOMIC_FETCH_AND_ADD:	return IB_WC_FETCH_ADD;
128 	case IB_WR_LSO:				return IB_WC_LSO;
129 	case IB_WR_SEND_WITH_INV:		return IB_WC_SEND;
130 	case IB_WR_RDMA_READ_WITH_INV:		return IB_WC_RDMA_READ;
131 	case IB_WR_LOCAL_INV:			return IB_WC_LOCAL_INV;
132 	case IB_WR_REG_MR:			return IB_WC_REG_MR;
133 
134 	default:
135 		return 0xff;
136 	}
137 }
138 
139 void retransmit_timer(struct timer_list *t)
140 {
141 	struct rxe_qp *qp = from_timer(qp, t, retrans_timer);
142 
143 	if (qp->valid) {
144 		qp->comp.timeout = 1;
145 		rxe_run_task(&qp->comp.task, 1);
146 	}
147 }
148 
149 void rxe_comp_queue_pkt(struct rxe_dev *rxe, struct rxe_qp *qp,
150 			struct sk_buff *skb)
151 {
152 	int must_sched;
153 
154 	skb_queue_tail(&qp->resp_pkts, skb);
155 
156 	must_sched = skb_queue_len(&qp->resp_pkts) > 1;
157 	if (must_sched != 0)
158 		rxe_counter_inc(rxe, RXE_CNT_COMPLETER_SCHED);
159 	rxe_run_task(&qp->comp.task, must_sched);
160 }
161 
162 static inline enum comp_state get_wqe(struct rxe_qp *qp,
163 				      struct rxe_pkt_info *pkt,
164 				      struct rxe_send_wqe **wqe_p)
165 {
166 	struct rxe_send_wqe *wqe;
167 
168 	/* we come here whether or not we found a response packet to see if
169 	 * there are any posted WQEs
170 	 */
171 	wqe = queue_head(qp->sq.queue);
172 	*wqe_p = wqe;
173 
174 	/* no WQE or requester has not started it yet */
175 	if (!wqe || wqe->state == wqe_state_posted)
176 		return pkt ? COMPST_DONE : COMPST_EXIT;
177 
178 	/* WQE does not require an ack */
179 	if (wqe->state == wqe_state_done)
180 		return COMPST_COMP_WQE;
181 
182 	/* WQE caused an error */
183 	if (wqe->state == wqe_state_error)
184 		return COMPST_ERROR;
185 
186 	/* we have a WQE, if we also have an ack check its PSN */
187 	return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
188 }
189 
190 static inline void reset_retry_counters(struct rxe_qp *qp)
191 {
192 	qp->comp.retry_cnt = qp->attr.retry_cnt;
193 	qp->comp.rnr_retry = qp->attr.rnr_retry;
194 	qp->comp.started_retry = 0;
195 }
196 
197 static inline enum comp_state check_psn(struct rxe_qp *qp,
198 					struct rxe_pkt_info *pkt,
199 					struct rxe_send_wqe *wqe)
200 {
201 	s32 diff;
202 
203 	/* check to see if response is past the oldest WQE. if it is, complete
204 	 * send/write or error read/atomic
205 	 */
206 	diff = psn_compare(pkt->psn, wqe->last_psn);
207 	if (diff > 0) {
208 		if (wqe->state == wqe_state_pending) {
209 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
210 				return COMPST_ERROR_RETRY;
211 
212 			reset_retry_counters(qp);
213 			return COMPST_COMP_WQE;
214 		} else {
215 			return COMPST_DONE;
216 		}
217 	}
218 
219 	/* compare response packet to expected response */
220 	diff = psn_compare(pkt->psn, qp->comp.psn);
221 	if (diff < 0) {
222 		/* response is most likely a retried packet if it matches an
223 		 * uncompleted WQE go complete it else ignore it
224 		 */
225 		if (pkt->psn == wqe->last_psn)
226 			return COMPST_COMP_ACK;
227 		else
228 			return COMPST_DONE;
229 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
230 		return COMPST_DONE;
231 	} else {
232 		return COMPST_CHECK_ACK;
233 	}
234 }
235 
236 static inline enum comp_state check_ack(struct rxe_qp *qp,
237 					struct rxe_pkt_info *pkt,
238 					struct rxe_send_wqe *wqe)
239 {
240 	unsigned int mask = pkt->mask;
241 	u8 syn;
242 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
243 
244 	/* Check the sequence only */
245 	switch (qp->comp.opcode) {
246 	case -1:
247 		/* Will catch all *_ONLY cases. */
248 		if (!(mask & RXE_START_MASK))
249 			return COMPST_ERROR;
250 
251 		break;
252 
253 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
254 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
255 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
256 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
257 			/* read retries of partial data may restart from
258 			 * read response first or response only.
259 			 */
260 			if ((pkt->psn == wqe->first_psn &&
261 			     pkt->opcode ==
262 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) ||
263 			    (wqe->first_psn == wqe->last_psn &&
264 			     pkt->opcode ==
265 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY))
266 				break;
267 
268 			return COMPST_ERROR;
269 		}
270 		break;
271 	default:
272 		WARN_ON_ONCE(1);
273 	}
274 
275 	/* Check operation validity. */
276 	switch (pkt->opcode) {
277 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
278 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
279 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
280 		syn = aeth_syn(pkt);
281 
282 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
283 			return COMPST_ERROR;
284 
285 		/* fall through */
286 		/* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
287 		 */
288 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
289 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
290 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV) {
291 			wqe->status = IB_WC_FATAL_ERR;
292 			return COMPST_ERROR;
293 		}
294 		reset_retry_counters(qp);
295 		return COMPST_READ;
296 
297 	case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
298 		syn = aeth_syn(pkt);
299 
300 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
301 			return COMPST_ERROR;
302 
303 		if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP &&
304 		    wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
305 			return COMPST_ERROR;
306 		reset_retry_counters(qp);
307 		return COMPST_ATOMIC;
308 
309 	case IB_OPCODE_RC_ACKNOWLEDGE:
310 		syn = aeth_syn(pkt);
311 		switch (syn & AETH_TYPE_MASK) {
312 		case AETH_ACK:
313 			reset_retry_counters(qp);
314 			return COMPST_WRITE_SEND;
315 
316 		case AETH_RNR_NAK:
317 			rxe_counter_inc(rxe, RXE_CNT_RCV_RNR);
318 			return COMPST_RNR_RETRY;
319 
320 		case AETH_NAK:
321 			switch (syn) {
322 			case AETH_NAK_PSN_SEQ_ERROR:
323 				/* a nak implicitly acks all packets with psns
324 				 * before
325 				 */
326 				if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
327 					rxe_counter_inc(rxe,
328 							RXE_CNT_RCV_SEQ_ERR);
329 					qp->comp.psn = pkt->psn;
330 					if (qp->req.wait_psn) {
331 						qp->req.wait_psn = 0;
332 						rxe_run_task(&qp->req.task, 1);
333 					}
334 				}
335 				return COMPST_ERROR_RETRY;
336 
337 			case AETH_NAK_INVALID_REQ:
338 				wqe->status = IB_WC_REM_INV_REQ_ERR;
339 				return COMPST_ERROR;
340 
341 			case AETH_NAK_REM_ACC_ERR:
342 				wqe->status = IB_WC_REM_ACCESS_ERR;
343 				return COMPST_ERROR;
344 
345 			case AETH_NAK_REM_OP_ERR:
346 				wqe->status = IB_WC_REM_OP_ERR;
347 				return COMPST_ERROR;
348 
349 			default:
350 				pr_warn("unexpected nak %x\n", syn);
351 				wqe->status = IB_WC_REM_OP_ERR;
352 				return COMPST_ERROR;
353 			}
354 
355 		default:
356 			return COMPST_ERROR;
357 		}
358 		break;
359 
360 	default:
361 		pr_warn("unexpected opcode\n");
362 	}
363 
364 	return COMPST_ERROR;
365 }
366 
367 static inline enum comp_state do_read(struct rxe_qp *qp,
368 				      struct rxe_pkt_info *pkt,
369 				      struct rxe_send_wqe *wqe)
370 {
371 	int ret;
372 
373 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
374 			&wqe->dma, payload_addr(pkt),
375 			payload_size(pkt), to_mem_obj, NULL);
376 	if (ret)
377 		return COMPST_ERROR;
378 
379 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
380 		return COMPST_COMP_ACK;
381 	else
382 		return COMPST_UPDATE_COMP;
383 }
384 
385 static inline enum comp_state do_atomic(struct rxe_qp *qp,
386 					struct rxe_pkt_info *pkt,
387 					struct rxe_send_wqe *wqe)
388 {
389 	int ret;
390 
391 	u64 atomic_orig = atmack_orig(pkt);
392 
393 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
394 			&wqe->dma, &atomic_orig,
395 			sizeof(u64), to_mem_obj, NULL);
396 	if (ret)
397 		return COMPST_ERROR;
398 	else
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 	memset(cqe, 0, sizeof(*cqe));
406 
407 	if (!qp->is_user) {
408 		struct ib_wc		*wc	= &cqe->ibwc;
409 
410 		wc->wr_id		= wqe->wr.wr_id;
411 		wc->status		= wqe->status;
412 		wc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
413 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
414 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
415 			wc->wc_flags = IB_WC_WITH_IMM;
416 		wc->byte_len		= wqe->dma.length;
417 		wc->qp			= &qp->ibqp;
418 	} else {
419 		struct ib_uverbs_wc	*uwc	= &cqe->uibwc;
420 
421 		uwc->wr_id		= wqe->wr.wr_id;
422 		uwc->status		= wqe->status;
423 		uwc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
424 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
425 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
426 			uwc->wc_flags = IB_WC_WITH_IMM;
427 		uwc->byte_len		= wqe->dma.length;
428 		uwc->qp_num		= qp->ibqp.qp_num;
429 	}
430 }
431 
432 /*
433  * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
434  * ---------8<---------8<-------------
435  * ...Note that if a completion error occurs, a Work Completion
436  * will always be generated, even if the signaling
437  * indicator requests an Unsignaled Completion.
438  * ---------8<---------8<-------------
439  */
440 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
441 {
442 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
443 	struct rxe_cqe cqe;
444 
445 	if ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
446 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
447 	    wqe->status != IB_WC_SUCCESS) {
448 		make_send_cqe(qp, wqe, &cqe);
449 		advance_consumer(qp->sq.queue);
450 		rxe_cq_post(qp->scq, &cqe, 0);
451 	} else {
452 		advance_consumer(qp->sq.queue);
453 	}
454 
455 	if (wqe->wr.opcode == IB_WR_SEND ||
456 	    wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
457 	    wqe->wr.opcode == IB_WR_SEND_WITH_INV)
458 		rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND);
459 
460 	/*
461 	 * we completed something so let req run again
462 	 * if it is trying to fence
463 	 */
464 	if (qp->req.wait_fence) {
465 		qp->req.wait_fence = 0;
466 		rxe_run_task(&qp->req.task, 1);
467 	}
468 }
469 
470 static inline enum comp_state complete_ack(struct rxe_qp *qp,
471 					   struct rxe_pkt_info *pkt,
472 					   struct rxe_send_wqe *wqe)
473 {
474 	unsigned long flags;
475 
476 	if (wqe->has_rd_atomic) {
477 		wqe->has_rd_atomic = 0;
478 		atomic_inc(&qp->req.rd_atomic);
479 		if (qp->req.need_rd_atomic) {
480 			qp->comp.timeout_retry = 0;
481 			qp->req.need_rd_atomic = 0;
482 			rxe_run_task(&qp->req.task, 1);
483 		}
484 	}
485 
486 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
487 		/* state_lock used by requester & completer */
488 		spin_lock_irqsave(&qp->state_lock, flags);
489 		if ((qp->req.state == QP_STATE_DRAIN) &&
490 		    (qp->comp.psn == qp->req.psn)) {
491 			qp->req.state = QP_STATE_DRAINED;
492 			spin_unlock_irqrestore(&qp->state_lock, flags);
493 
494 			if (qp->ibqp.event_handler) {
495 				struct ib_event ev;
496 
497 				ev.device = qp->ibqp.device;
498 				ev.element.qp = &qp->ibqp;
499 				ev.event = IB_EVENT_SQ_DRAINED;
500 				qp->ibqp.event_handler(&ev,
501 					qp->ibqp.qp_context);
502 			}
503 		} else {
504 			spin_unlock_irqrestore(&qp->state_lock, flags);
505 		}
506 	}
507 
508 	do_complete(qp, wqe);
509 
510 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
511 		return COMPST_UPDATE_COMP;
512 	else
513 		return COMPST_DONE;
514 }
515 
516 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
517 					   struct rxe_pkt_info *pkt,
518 					   struct rxe_send_wqe *wqe)
519 {
520 	if (pkt && wqe->state == wqe_state_pending) {
521 		if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) {
522 			qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK;
523 			qp->comp.opcode = -1;
524 		}
525 
526 		if (qp->req.wait_psn) {
527 			qp->req.wait_psn = 0;
528 			rxe_run_task(&qp->req.task, 1);
529 		}
530 	}
531 
532 	do_complete(qp, wqe);
533 
534 	return COMPST_GET_WQE;
535 }
536 
537 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
538 {
539 	struct sk_buff *skb;
540 	struct rxe_send_wqe *wqe;
541 
542 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
543 		rxe_drop_ref(qp);
544 		kfree_skb(skb);
545 	}
546 
547 	while ((wqe = queue_head(qp->sq.queue))) {
548 		if (notify) {
549 			wqe->status = IB_WC_WR_FLUSH_ERR;
550 			do_complete(qp, wqe);
551 		} else {
552 			advance_consumer(qp->sq.queue);
553 		}
554 	}
555 }
556 
557 int rxe_completer(void *arg)
558 {
559 	struct rxe_qp *qp = (struct rxe_qp *)arg;
560 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
561 	struct rxe_send_wqe *wqe = wqe;
562 	struct sk_buff *skb = NULL;
563 	struct rxe_pkt_info *pkt = NULL;
564 	enum comp_state state;
565 
566 	rxe_add_ref(qp);
567 
568 	if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
569 	    qp->req.state == QP_STATE_RESET) {
570 		rxe_drain_resp_pkts(qp, qp->valid &&
571 				    qp->req.state == QP_STATE_ERROR);
572 		goto exit;
573 	}
574 
575 	if (qp->comp.timeout) {
576 		qp->comp.timeout_retry = 1;
577 		qp->comp.timeout = 0;
578 	} else {
579 		qp->comp.timeout_retry = 0;
580 	}
581 
582 	if (qp->req.need_retry)
583 		goto exit;
584 
585 	state = COMPST_GET_ACK;
586 
587 	while (1) {
588 		pr_debug("qp#%d state = %s\n", qp_num(qp),
589 			 comp_state_name[state]);
590 		switch (state) {
591 		case COMPST_GET_ACK:
592 			skb = skb_dequeue(&qp->resp_pkts);
593 			if (skb) {
594 				pkt = SKB_TO_PKT(skb);
595 				qp->comp.timeout_retry = 0;
596 			}
597 			state = COMPST_GET_WQE;
598 			break;
599 
600 		case COMPST_GET_WQE:
601 			state = get_wqe(qp, pkt, &wqe);
602 			break;
603 
604 		case COMPST_CHECK_PSN:
605 			state = check_psn(qp, pkt, wqe);
606 			break;
607 
608 		case COMPST_CHECK_ACK:
609 			state = check_ack(qp, pkt, wqe);
610 			break;
611 
612 		case COMPST_READ:
613 			state = do_read(qp, pkt, wqe);
614 			break;
615 
616 		case COMPST_ATOMIC:
617 			state = do_atomic(qp, pkt, wqe);
618 			break;
619 
620 		case COMPST_WRITE_SEND:
621 			if (wqe->state == wqe_state_pending &&
622 			    wqe->last_psn == pkt->psn)
623 				state = COMPST_COMP_ACK;
624 			else
625 				state = COMPST_UPDATE_COMP;
626 			break;
627 
628 		case COMPST_COMP_ACK:
629 			state = complete_ack(qp, pkt, wqe);
630 			break;
631 
632 		case COMPST_COMP_WQE:
633 			state = complete_wqe(qp, pkt, wqe);
634 			break;
635 
636 		case COMPST_UPDATE_COMP:
637 			if (pkt->mask & RXE_END_MASK)
638 				qp->comp.opcode = -1;
639 			else
640 				qp->comp.opcode = pkt->opcode;
641 
642 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
643 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
644 
645 			if (qp->req.wait_psn) {
646 				qp->req.wait_psn = 0;
647 				rxe_run_task(&qp->req.task, 1);
648 			}
649 
650 			state = COMPST_DONE;
651 			break;
652 
653 		case COMPST_DONE:
654 			if (pkt) {
655 				rxe_drop_ref(pkt->qp);
656 				kfree_skb(skb);
657 				skb = NULL;
658 			}
659 			goto done;
660 
661 		case COMPST_EXIT:
662 			if (qp->comp.timeout_retry && wqe) {
663 				state = COMPST_ERROR_RETRY;
664 				break;
665 			}
666 
667 			/* re reset the timeout counter if
668 			 * (1) QP is type RC
669 			 * (2) the QP is alive
670 			 * (3) there is a packet sent by the requester that
671 			 *     might be acked (we still might get spurious
672 			 *     timeouts but try to keep them as few as possible)
673 			 * (4) the timeout parameter is set
674 			 */
675 			if ((qp_type(qp) == IB_QPT_RC) &&
676 			    (qp->req.state == QP_STATE_READY) &&
677 			    (psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
678 			    qp->qp_timeout_jiffies)
679 				mod_timer(&qp->retrans_timer,
680 					  jiffies + qp->qp_timeout_jiffies);
681 			goto exit;
682 
683 		case COMPST_ERROR_RETRY:
684 			/* we come here if the retry timer fired and we did
685 			 * not receive a response packet. try to retry the send
686 			 * queue if that makes sense and the limits have not
687 			 * been exceeded. remember that some timeouts are
688 			 * spurious since we do not reset the timer but kick
689 			 * it down the road or let it expire
690 			 */
691 
692 			/* there is nothing to retry in this case */
693 			if (!wqe || (wqe->state == wqe_state_posted)) {
694 				goto exit;
695 			}
696 
697 			/* if we've started a retry, don't start another
698 			 * retry sequence, unless this is a timeout.
699 			 */
700 			if (qp->comp.started_retry &&
701 			    !qp->comp.timeout_retry) {
702 				if (pkt) {
703 					rxe_drop_ref(pkt->qp);
704 					kfree_skb(skb);
705 					skb = NULL;
706 				}
707 
708 				goto done;
709 			}
710 
711 			if (qp->comp.retry_cnt > 0) {
712 				if (qp->comp.retry_cnt != 7)
713 					qp->comp.retry_cnt--;
714 
715 				/* no point in retrying if we have already
716 				 * seen the last ack that the requester could
717 				 * have caused
718 				 */
719 				if (psn_compare(qp->req.psn,
720 						qp->comp.psn) > 0) {
721 					/* tell the requester to retry the
722 					 * send queue next time around
723 					 */
724 					rxe_counter_inc(rxe,
725 							RXE_CNT_COMP_RETRY);
726 					qp->req.need_retry = 1;
727 					qp->comp.started_retry = 1;
728 					rxe_run_task(&qp->req.task, 1);
729 				}
730 
731 				if (pkt) {
732 					rxe_drop_ref(pkt->qp);
733 					kfree_skb(skb);
734 					skb = NULL;
735 				}
736 
737 				goto done;
738 
739 			} else {
740 				rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
741 				wqe->status = IB_WC_RETRY_EXC_ERR;
742 				state = COMPST_ERROR;
743 			}
744 			break;
745 
746 		case COMPST_RNR_RETRY:
747 			if (qp->comp.rnr_retry > 0) {
748 				if (qp->comp.rnr_retry != 7)
749 					qp->comp.rnr_retry--;
750 
751 				qp->req.need_retry = 1;
752 				pr_debug("qp#%d set rnr nak timer\n",
753 					 qp_num(qp));
754 				mod_timer(&qp->rnr_nak_timer,
755 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
756 						& ~AETH_TYPE_MASK));
757 				rxe_drop_ref(pkt->qp);
758 				kfree_skb(skb);
759 				skb = NULL;
760 				goto exit;
761 			} else {
762 				rxe_counter_inc(rxe,
763 						RXE_CNT_RNR_RETRY_EXCEEDED);
764 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
765 				state = COMPST_ERROR;
766 			}
767 			break;
768 
769 		case COMPST_ERROR:
770 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
771 			do_complete(qp, wqe);
772 			rxe_qp_error(qp);
773 
774 			if (pkt) {
775 				rxe_drop_ref(pkt->qp);
776 				kfree_skb(skb);
777 				skb = NULL;
778 			}
779 
780 			goto exit;
781 		}
782 	}
783 
784 exit:
785 	/* we come here if we are done with processing and want the task to
786 	 * exit from the loop calling us
787 	 */
788 	WARN_ON_ONCE(skb);
789 	rxe_drop_ref(qp);
790 	return -EAGAIN;
791 
792 done:
793 	/* we come here if we have processed a packet we want the task to call
794 	 * us again to see if there is anything else to do
795 	 */
796 	WARN_ON_ONCE(skb);
797 	rxe_drop_ref(qp);
798 	return 0;
799 }
800