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 }
195 
196 static inline enum comp_state check_psn(struct rxe_qp *qp,
197 					struct rxe_pkt_info *pkt,
198 					struct rxe_send_wqe *wqe)
199 {
200 	s32 diff;
201 
202 	/* check to see if response is past the oldest WQE. if it is, complete
203 	 * send/write or error read/atomic
204 	 */
205 	diff = psn_compare(pkt->psn, wqe->last_psn);
206 	if (diff > 0) {
207 		if (wqe->state == wqe_state_pending) {
208 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
209 				return COMPST_ERROR_RETRY;
210 
211 			reset_retry_counters(qp);
212 			return COMPST_COMP_WQE;
213 		} else {
214 			return COMPST_DONE;
215 		}
216 	}
217 
218 	/* compare response packet to expected response */
219 	diff = psn_compare(pkt->psn, qp->comp.psn);
220 	if (diff < 0) {
221 		/* response is most likely a retried packet if it matches an
222 		 * uncompleted WQE go complete it else ignore it
223 		 */
224 		if (pkt->psn == wqe->last_psn)
225 			return COMPST_COMP_ACK;
226 		else
227 			return COMPST_DONE;
228 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
229 		return COMPST_DONE;
230 	} else {
231 		return COMPST_CHECK_ACK;
232 	}
233 }
234 
235 static inline enum comp_state check_ack(struct rxe_qp *qp,
236 					struct rxe_pkt_info *pkt,
237 					struct rxe_send_wqe *wqe)
238 {
239 	unsigned int mask = pkt->mask;
240 	u8 syn;
241 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
242 
243 	/* Check the sequence only */
244 	switch (qp->comp.opcode) {
245 	case -1:
246 		/* Will catch all *_ONLY cases. */
247 		if (!(mask & RXE_START_MASK))
248 			return COMPST_ERROR;
249 
250 		break;
251 
252 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
253 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
254 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
255 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
256 			return COMPST_ERROR;
257 		}
258 		break;
259 	default:
260 		WARN_ON_ONCE(1);
261 	}
262 
263 	/* Check operation validity. */
264 	switch (pkt->opcode) {
265 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
266 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
267 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
268 		syn = aeth_syn(pkt);
269 
270 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
271 			return COMPST_ERROR;
272 
273 		/* fall through */
274 		/* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
275 		 */
276 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
277 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
278 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV) {
279 			return COMPST_ERROR;
280 		}
281 		reset_retry_counters(qp);
282 		return COMPST_READ;
283 
284 	case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
285 		syn = aeth_syn(pkt);
286 
287 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
288 			return COMPST_ERROR;
289 
290 		if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP &&
291 		    wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
292 			return COMPST_ERROR;
293 		reset_retry_counters(qp);
294 		return COMPST_ATOMIC;
295 
296 	case IB_OPCODE_RC_ACKNOWLEDGE:
297 		syn = aeth_syn(pkt);
298 		switch (syn & AETH_TYPE_MASK) {
299 		case AETH_ACK:
300 			reset_retry_counters(qp);
301 			return COMPST_WRITE_SEND;
302 
303 		case AETH_RNR_NAK:
304 			rxe_counter_inc(rxe, RXE_CNT_RCV_RNR);
305 			return COMPST_RNR_RETRY;
306 
307 		case AETH_NAK:
308 			switch (syn) {
309 			case AETH_NAK_PSN_SEQ_ERROR:
310 				/* a nak implicitly acks all packets with psns
311 				 * before
312 				 */
313 				if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
314 					rxe_counter_inc(rxe,
315 							RXE_CNT_RCV_SEQ_ERR);
316 					qp->comp.psn = pkt->psn;
317 					if (qp->req.wait_psn) {
318 						qp->req.wait_psn = 0;
319 						rxe_run_task(&qp->req.task, 1);
320 					}
321 				}
322 				return COMPST_ERROR_RETRY;
323 
324 			case AETH_NAK_INVALID_REQ:
325 				wqe->status = IB_WC_REM_INV_REQ_ERR;
326 				return COMPST_ERROR;
327 
328 			case AETH_NAK_REM_ACC_ERR:
329 				wqe->status = IB_WC_REM_ACCESS_ERR;
330 				return COMPST_ERROR;
331 
332 			case AETH_NAK_REM_OP_ERR:
333 				wqe->status = IB_WC_REM_OP_ERR;
334 				return COMPST_ERROR;
335 
336 			default:
337 				pr_warn("unexpected nak %x\n", syn);
338 				wqe->status = IB_WC_REM_OP_ERR;
339 				return COMPST_ERROR;
340 			}
341 
342 		default:
343 			return COMPST_ERROR;
344 		}
345 		break;
346 
347 	default:
348 		pr_warn("unexpected opcode\n");
349 	}
350 
351 	return COMPST_ERROR;
352 }
353 
354 static inline enum comp_state do_read(struct rxe_qp *qp,
355 				      struct rxe_pkt_info *pkt,
356 				      struct rxe_send_wqe *wqe)
357 {
358 	int ret;
359 
360 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
361 			&wqe->dma, payload_addr(pkt),
362 			payload_size(pkt), to_mem_obj, NULL);
363 	if (ret)
364 		return COMPST_ERROR;
365 
366 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
367 		return COMPST_COMP_ACK;
368 	else
369 		return COMPST_UPDATE_COMP;
370 }
371 
372 static inline enum comp_state do_atomic(struct rxe_qp *qp,
373 					struct rxe_pkt_info *pkt,
374 					struct rxe_send_wqe *wqe)
375 {
376 	int ret;
377 
378 	u64 atomic_orig = atmack_orig(pkt);
379 
380 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
381 			&wqe->dma, &atomic_orig,
382 			sizeof(u64), to_mem_obj, NULL);
383 	if (ret)
384 		return COMPST_ERROR;
385 	else
386 		return COMPST_COMP_ACK;
387 }
388 
389 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
390 			  struct rxe_cqe *cqe)
391 {
392 	memset(cqe, 0, sizeof(*cqe));
393 
394 	if (!qp->is_user) {
395 		struct ib_wc		*wc	= &cqe->ibwc;
396 
397 		wc->wr_id		= wqe->wr.wr_id;
398 		wc->status		= wqe->status;
399 		wc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
400 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
401 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
402 			wc->wc_flags = IB_WC_WITH_IMM;
403 		wc->byte_len		= wqe->dma.length;
404 		wc->qp			= &qp->ibqp;
405 	} else {
406 		struct ib_uverbs_wc	*uwc	= &cqe->uibwc;
407 
408 		uwc->wr_id		= wqe->wr.wr_id;
409 		uwc->status		= wqe->status;
410 		uwc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
411 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
412 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
413 			uwc->wc_flags = IB_WC_WITH_IMM;
414 		uwc->byte_len		= wqe->dma.length;
415 		uwc->qp_num		= qp->ibqp.qp_num;
416 	}
417 }
418 
419 /*
420  * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
421  * ---------8<---------8<-------------
422  * ...Note that if a completion error occurs, a Work Completion
423  * will always be generated, even if the signaling
424  * indicator requests an Unsignaled Completion.
425  * ---------8<---------8<-------------
426  */
427 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
428 {
429 	struct rxe_cqe cqe;
430 
431 	if ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
432 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
433 	    wqe->status != IB_WC_SUCCESS) {
434 		make_send_cqe(qp, wqe, &cqe);
435 		advance_consumer(qp->sq.queue);
436 		rxe_cq_post(qp->scq, &cqe, 0);
437 	} else {
438 		advance_consumer(qp->sq.queue);
439 	}
440 
441 	/*
442 	 * we completed something so let req run again
443 	 * if it is trying to fence
444 	 */
445 	if (qp->req.wait_fence) {
446 		qp->req.wait_fence = 0;
447 		rxe_run_task(&qp->req.task, 1);
448 	}
449 }
450 
451 static inline enum comp_state complete_ack(struct rxe_qp *qp,
452 					   struct rxe_pkt_info *pkt,
453 					   struct rxe_send_wqe *wqe)
454 {
455 	unsigned long flags;
456 
457 	if (wqe->has_rd_atomic) {
458 		wqe->has_rd_atomic = 0;
459 		atomic_inc(&qp->req.rd_atomic);
460 		if (qp->req.need_rd_atomic) {
461 			qp->comp.timeout_retry = 0;
462 			qp->req.need_rd_atomic = 0;
463 			rxe_run_task(&qp->req.task, 1);
464 		}
465 	}
466 
467 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
468 		/* state_lock used by requester & completer */
469 		spin_lock_irqsave(&qp->state_lock, flags);
470 		if ((qp->req.state == QP_STATE_DRAIN) &&
471 		    (qp->comp.psn == qp->req.psn)) {
472 			qp->req.state = QP_STATE_DRAINED;
473 			spin_unlock_irqrestore(&qp->state_lock, flags);
474 
475 			if (qp->ibqp.event_handler) {
476 				struct ib_event ev;
477 
478 				ev.device = qp->ibqp.device;
479 				ev.element.qp = &qp->ibqp;
480 				ev.event = IB_EVENT_SQ_DRAINED;
481 				qp->ibqp.event_handler(&ev,
482 					qp->ibqp.qp_context);
483 			}
484 		} else {
485 			spin_unlock_irqrestore(&qp->state_lock, flags);
486 		}
487 	}
488 
489 	do_complete(qp, wqe);
490 
491 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
492 		return COMPST_UPDATE_COMP;
493 	else
494 		return COMPST_DONE;
495 }
496 
497 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
498 					   struct rxe_pkt_info *pkt,
499 					   struct rxe_send_wqe *wqe)
500 {
501 	qp->comp.opcode = -1;
502 
503 	if (pkt) {
504 		if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
505 			qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
506 
507 		if (qp->req.wait_psn) {
508 			qp->req.wait_psn = 0;
509 			rxe_run_task(&qp->req.task, 1);
510 		}
511 	}
512 
513 	do_complete(qp, wqe);
514 
515 	return COMPST_GET_WQE;
516 }
517 
518 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
519 {
520 	struct sk_buff *skb;
521 	struct rxe_send_wqe *wqe;
522 
523 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
524 		rxe_drop_ref(qp);
525 		kfree_skb(skb);
526 	}
527 
528 	while ((wqe = queue_head(qp->sq.queue))) {
529 		if (notify) {
530 			wqe->status = IB_WC_WR_FLUSH_ERR;
531 			do_complete(qp, wqe);
532 		} else {
533 			advance_consumer(qp->sq.queue);
534 		}
535 	}
536 }
537 
538 int rxe_completer(void *arg)
539 {
540 	struct rxe_qp *qp = (struct rxe_qp *)arg;
541 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
542 	struct rxe_send_wqe *wqe = wqe;
543 	struct sk_buff *skb = NULL;
544 	struct rxe_pkt_info *pkt = NULL;
545 	enum comp_state state;
546 
547 	rxe_add_ref(qp);
548 
549 	if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
550 	    qp->req.state == QP_STATE_RESET) {
551 		rxe_drain_resp_pkts(qp, qp->valid &&
552 				    qp->req.state == QP_STATE_ERROR);
553 		goto exit;
554 	}
555 
556 	if (qp->comp.timeout) {
557 		qp->comp.timeout_retry = 1;
558 		qp->comp.timeout = 0;
559 	} else {
560 		qp->comp.timeout_retry = 0;
561 	}
562 
563 	if (qp->req.need_retry)
564 		goto exit;
565 
566 	state = COMPST_GET_ACK;
567 
568 	while (1) {
569 		pr_debug("qp#%d state = %s\n", qp_num(qp),
570 			 comp_state_name[state]);
571 		switch (state) {
572 		case COMPST_GET_ACK:
573 			skb = skb_dequeue(&qp->resp_pkts);
574 			if (skb) {
575 				pkt = SKB_TO_PKT(skb);
576 				qp->comp.timeout_retry = 0;
577 			}
578 			state = COMPST_GET_WQE;
579 			break;
580 
581 		case COMPST_GET_WQE:
582 			state = get_wqe(qp, pkt, &wqe);
583 			break;
584 
585 		case COMPST_CHECK_PSN:
586 			state = check_psn(qp, pkt, wqe);
587 			break;
588 
589 		case COMPST_CHECK_ACK:
590 			state = check_ack(qp, pkt, wqe);
591 			break;
592 
593 		case COMPST_READ:
594 			state = do_read(qp, pkt, wqe);
595 			break;
596 
597 		case COMPST_ATOMIC:
598 			state = do_atomic(qp, pkt, wqe);
599 			break;
600 
601 		case COMPST_WRITE_SEND:
602 			if (wqe->state == wqe_state_pending &&
603 			    wqe->last_psn == pkt->psn)
604 				state = COMPST_COMP_ACK;
605 			else
606 				state = COMPST_UPDATE_COMP;
607 			break;
608 
609 		case COMPST_COMP_ACK:
610 			state = complete_ack(qp, pkt, wqe);
611 			break;
612 
613 		case COMPST_COMP_WQE:
614 			state = complete_wqe(qp, pkt, wqe);
615 			break;
616 
617 		case COMPST_UPDATE_COMP:
618 			if (pkt->mask & RXE_END_MASK)
619 				qp->comp.opcode = -1;
620 			else
621 				qp->comp.opcode = pkt->opcode;
622 
623 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
624 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
625 
626 			if (qp->req.wait_psn) {
627 				qp->req.wait_psn = 0;
628 				rxe_run_task(&qp->req.task, 1);
629 			}
630 
631 			state = COMPST_DONE;
632 			break;
633 
634 		case COMPST_DONE:
635 			if (pkt) {
636 				rxe_drop_ref(pkt->qp);
637 				kfree_skb(skb);
638 				skb = NULL;
639 			}
640 			goto done;
641 
642 		case COMPST_EXIT:
643 			if (qp->comp.timeout_retry && wqe) {
644 				state = COMPST_ERROR_RETRY;
645 				break;
646 			}
647 
648 			/* re reset the timeout counter if
649 			 * (1) QP is type RC
650 			 * (2) the QP is alive
651 			 * (3) there is a packet sent by the requester that
652 			 *     might be acked (we still might get spurious
653 			 *     timeouts but try to keep them as few as possible)
654 			 * (4) the timeout parameter is set
655 			 */
656 			if ((qp_type(qp) == IB_QPT_RC) &&
657 			    (qp->req.state == QP_STATE_READY) &&
658 			    (psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
659 			    qp->qp_timeout_jiffies)
660 				mod_timer(&qp->retrans_timer,
661 					  jiffies + qp->qp_timeout_jiffies);
662 			goto exit;
663 
664 		case COMPST_ERROR_RETRY:
665 			/* we come here if the retry timer fired and we did
666 			 * not receive a response packet. try to retry the send
667 			 * queue if that makes sense and the limits have not
668 			 * been exceeded. remember that some timeouts are
669 			 * spurious since we do not reset the timer but kick
670 			 * it down the road or let it expire
671 			 */
672 
673 			/* there is nothing to retry in this case */
674 			if (!wqe || (wqe->state == wqe_state_posted)) {
675 				goto exit;
676 			}
677 
678 			if (qp->comp.retry_cnt > 0) {
679 				if (qp->comp.retry_cnt != 7)
680 					qp->comp.retry_cnt--;
681 
682 				/* no point in retrying if we have already
683 				 * seen the last ack that the requester could
684 				 * have caused
685 				 */
686 				if (psn_compare(qp->req.psn,
687 						qp->comp.psn) > 0) {
688 					/* tell the requester to retry the
689 					 * send queue next time around
690 					 */
691 					rxe_counter_inc(rxe,
692 							RXE_CNT_COMP_RETRY);
693 					qp->req.need_retry = 1;
694 					rxe_run_task(&qp->req.task, 1);
695 				}
696 
697 				if (pkt) {
698 					rxe_drop_ref(pkt->qp);
699 					kfree_skb(skb);
700 					skb = NULL;
701 				}
702 
703 				goto exit;
704 
705 			} else {
706 				rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
707 				wqe->status = IB_WC_RETRY_EXC_ERR;
708 				state = COMPST_ERROR;
709 			}
710 			break;
711 
712 		case COMPST_RNR_RETRY:
713 			if (qp->comp.rnr_retry > 0) {
714 				if (qp->comp.rnr_retry != 7)
715 					qp->comp.rnr_retry--;
716 
717 				qp->req.need_retry = 1;
718 				pr_debug("qp#%d set rnr nak timer\n",
719 					 qp_num(qp));
720 				mod_timer(&qp->rnr_nak_timer,
721 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
722 						& ~AETH_TYPE_MASK));
723 				rxe_drop_ref(pkt->qp);
724 				kfree_skb(skb);
725 				skb = NULL;
726 				goto exit;
727 			} else {
728 				rxe_counter_inc(rxe,
729 						RXE_CNT_RNR_RETRY_EXCEEDED);
730 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
731 				state = COMPST_ERROR;
732 			}
733 			break;
734 
735 		case COMPST_ERROR:
736 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
737 			do_complete(qp, wqe);
738 			rxe_qp_error(qp);
739 
740 			if (pkt) {
741 				rxe_drop_ref(pkt->qp);
742 				kfree_skb(skb);
743 				skb = NULL;
744 			}
745 
746 			goto exit;
747 		}
748 	}
749 
750 exit:
751 	/* we come here if we are done with processing and want the task to
752 	 * exit from the loop calling us
753 	 */
754 	WARN_ON_ONCE(skb);
755 	rxe_drop_ref(qp);
756 	return -EAGAIN;
757 
758 done:
759 	/* we come here if we have processed a packet we want the task to call
760 	 * us again to see if there is anything else to do
761 	 */
762 	WARN_ON_ONCE(skb);
763 	rxe_drop_ref(qp);
764 	return 0;
765 }
766