xref: /openbmc/linux/drivers/infiniband/sw/rxe/rxe_comp.c (revision 4ed91d48259d9ddd378424d008f2e6559f7e78f8)
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(unsigned long data)
140 {
141 	struct rxe_qp *qp = (struct rxe_qp *)data;
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 	rxe_run_task(&qp->comp.task, must_sched);
158 }
159 
160 static inline enum comp_state get_wqe(struct rxe_qp *qp,
161 				      struct rxe_pkt_info *pkt,
162 				      struct rxe_send_wqe **wqe_p)
163 {
164 	struct rxe_send_wqe *wqe;
165 
166 	/* we come here whether or not we found a response packet to see if
167 	 * there are any posted WQEs
168 	 */
169 	wqe = queue_head(qp->sq.queue);
170 	*wqe_p = wqe;
171 
172 	/* no WQE or requester has not started it yet */
173 	if (!wqe || wqe->state == wqe_state_posted)
174 		return pkt ? COMPST_DONE : COMPST_EXIT;
175 
176 	/* WQE does not require an ack */
177 	if (wqe->state == wqe_state_done)
178 		return COMPST_COMP_WQE;
179 
180 	/* WQE caused an error */
181 	if (wqe->state == wqe_state_error)
182 		return COMPST_ERROR;
183 
184 	/* we have a WQE, if we also have an ack check its PSN */
185 	return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
186 }
187 
188 static inline void reset_retry_counters(struct rxe_qp *qp)
189 {
190 	qp->comp.retry_cnt = qp->attr.retry_cnt;
191 	qp->comp.rnr_retry = qp->attr.rnr_retry;
192 }
193 
194 static inline enum comp_state check_psn(struct rxe_qp *qp,
195 					struct rxe_pkt_info *pkt,
196 					struct rxe_send_wqe *wqe)
197 {
198 	s32 diff;
199 
200 	/* check to see if response is past the oldest WQE. if it is, complete
201 	 * send/write or error read/atomic
202 	 */
203 	diff = psn_compare(pkt->psn, wqe->last_psn);
204 	if (diff > 0) {
205 		if (wqe->state == wqe_state_pending) {
206 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
207 				return COMPST_ERROR_RETRY;
208 
209 			reset_retry_counters(qp);
210 			return COMPST_COMP_WQE;
211 		} else {
212 			return COMPST_DONE;
213 		}
214 	}
215 
216 	/* compare response packet to expected response */
217 	diff = psn_compare(pkt->psn, qp->comp.psn);
218 	if (diff < 0) {
219 		/* response is most likely a retried packet if it matches an
220 		 * uncompleted WQE go complete it else ignore it
221 		 */
222 		if (pkt->psn == wqe->last_psn)
223 			return COMPST_COMP_ACK;
224 		else
225 			return COMPST_DONE;
226 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
227 		return COMPST_DONE;
228 	} else {
229 		return COMPST_CHECK_ACK;
230 	}
231 }
232 
233 static inline enum comp_state check_ack(struct rxe_qp *qp,
234 					struct rxe_pkt_info *pkt,
235 					struct rxe_send_wqe *wqe)
236 {
237 	unsigned int mask = pkt->mask;
238 	u8 syn;
239 
240 	/* Check the sequence only */
241 	switch (qp->comp.opcode) {
242 	case -1:
243 		/* Will catch all *_ONLY cases. */
244 		if (!(mask & RXE_START_MASK))
245 			return COMPST_ERROR;
246 
247 		break;
248 
249 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
250 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
251 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
252 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
253 			return COMPST_ERROR;
254 		}
255 		break;
256 	default:
257 		WARN_ON_ONCE(1);
258 	}
259 
260 	/* Check operation validity. */
261 	switch (pkt->opcode) {
262 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
263 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
264 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
265 		syn = aeth_syn(pkt);
266 
267 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
268 			return COMPST_ERROR;
269 
270 		/* Fall through (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE
271 		 * doesn't have an AETH)
272 		 */
273 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
274 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
275 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV) {
276 			return COMPST_ERROR;
277 		}
278 		reset_retry_counters(qp);
279 		return COMPST_READ;
280 
281 	case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
282 		syn = aeth_syn(pkt);
283 
284 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
285 			return COMPST_ERROR;
286 
287 		if (wqe->wr.opcode != IB_WR_ATOMIC_CMP_AND_SWP &&
288 		    wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
289 			return COMPST_ERROR;
290 		reset_retry_counters(qp);
291 		return COMPST_ATOMIC;
292 
293 	case IB_OPCODE_RC_ACKNOWLEDGE:
294 		syn = aeth_syn(pkt);
295 		switch (syn & AETH_TYPE_MASK) {
296 		case AETH_ACK:
297 			reset_retry_counters(qp);
298 			return COMPST_WRITE_SEND;
299 
300 		case AETH_RNR_NAK:
301 			return COMPST_RNR_RETRY;
302 
303 		case AETH_NAK:
304 			switch (syn) {
305 			case AETH_NAK_PSN_SEQ_ERROR:
306 				/* a nak implicitly acks all packets with psns
307 				 * before
308 				 */
309 				if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
310 					qp->comp.psn = pkt->psn;
311 					if (qp->req.wait_psn) {
312 						qp->req.wait_psn = 0;
313 						rxe_run_task(&qp->req.task, 1);
314 					}
315 				}
316 				return COMPST_ERROR_RETRY;
317 
318 			case AETH_NAK_INVALID_REQ:
319 				wqe->status = IB_WC_REM_INV_REQ_ERR;
320 				return COMPST_ERROR;
321 
322 			case AETH_NAK_REM_ACC_ERR:
323 				wqe->status = IB_WC_REM_ACCESS_ERR;
324 				return COMPST_ERROR;
325 
326 			case AETH_NAK_REM_OP_ERR:
327 				wqe->status = IB_WC_REM_OP_ERR;
328 				return COMPST_ERROR;
329 
330 			default:
331 				pr_warn("unexpected nak %x\n", syn);
332 				wqe->status = IB_WC_REM_OP_ERR;
333 				return COMPST_ERROR;
334 			}
335 
336 		default:
337 			return COMPST_ERROR;
338 		}
339 		break;
340 
341 	default:
342 		pr_warn("unexpected opcode\n");
343 	}
344 
345 	return COMPST_ERROR;
346 }
347 
348 static inline enum comp_state do_read(struct rxe_qp *qp,
349 				      struct rxe_pkt_info *pkt,
350 				      struct rxe_send_wqe *wqe)
351 {
352 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
353 	int ret;
354 
355 	ret = copy_data(rxe, qp->pd, IB_ACCESS_LOCAL_WRITE,
356 			&wqe->dma, payload_addr(pkt),
357 			payload_size(pkt), to_mem_obj, NULL);
358 	if (ret)
359 		return COMPST_ERROR;
360 
361 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
362 		return COMPST_COMP_ACK;
363 	else
364 		return COMPST_UPDATE_COMP;
365 }
366 
367 static inline enum comp_state do_atomic(struct rxe_qp *qp,
368 					struct rxe_pkt_info *pkt,
369 					struct rxe_send_wqe *wqe)
370 {
371 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
372 	int ret;
373 
374 	u64 atomic_orig = atmack_orig(pkt);
375 
376 	ret = copy_data(rxe, qp->pd, IB_ACCESS_LOCAL_WRITE,
377 			&wqe->dma, &atomic_orig,
378 			sizeof(u64), to_mem_obj, NULL);
379 	if (ret)
380 		return COMPST_ERROR;
381 	else
382 		return COMPST_COMP_ACK;
383 }
384 
385 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
386 			  struct rxe_cqe *cqe)
387 {
388 	memset(cqe, 0, sizeof(*cqe));
389 
390 	if (!qp->is_user) {
391 		struct ib_wc		*wc	= &cqe->ibwc;
392 
393 		wc->wr_id		= wqe->wr.wr_id;
394 		wc->status		= wqe->status;
395 		wc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
396 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
397 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
398 			wc->wc_flags = IB_WC_WITH_IMM;
399 		wc->byte_len		= wqe->dma.length;
400 		wc->qp			= &qp->ibqp;
401 	} else {
402 		struct ib_uverbs_wc	*uwc	= &cqe->uibwc;
403 
404 		uwc->wr_id		= wqe->wr.wr_id;
405 		uwc->status		= wqe->status;
406 		uwc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
407 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
408 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
409 			uwc->wc_flags = IB_WC_WITH_IMM;
410 		uwc->byte_len		= wqe->dma.length;
411 		uwc->qp_num		= qp->ibqp.qp_num;
412 	}
413 }
414 
415 /*
416  * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
417  * ---------8<---------8<-------------
418  * ...Note that if a completion error occurs, a Work Completion
419  * will always be generated, even if the signaling
420  * indicator requests an Unsignaled Completion.
421  * ---------8<---------8<-------------
422  */
423 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
424 {
425 	struct rxe_cqe cqe;
426 
427 	if ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
428 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
429 	    wqe->status != IB_WC_SUCCESS) {
430 		make_send_cqe(qp, wqe, &cqe);
431 		advance_consumer(qp->sq.queue);
432 		rxe_cq_post(qp->scq, &cqe, 0);
433 	} else {
434 		advance_consumer(qp->sq.queue);
435 	}
436 
437 	/*
438 	 * we completed something so let req run again
439 	 * if it is trying to fence
440 	 */
441 	if (qp->req.wait_fence) {
442 		qp->req.wait_fence = 0;
443 		rxe_run_task(&qp->req.task, 1);
444 	}
445 }
446 
447 static inline enum comp_state complete_ack(struct rxe_qp *qp,
448 					   struct rxe_pkt_info *pkt,
449 					   struct rxe_send_wqe *wqe)
450 {
451 	unsigned long flags;
452 
453 	if (wqe->has_rd_atomic) {
454 		wqe->has_rd_atomic = 0;
455 		atomic_inc(&qp->req.rd_atomic);
456 		if (qp->req.need_rd_atomic) {
457 			qp->comp.timeout_retry = 0;
458 			qp->req.need_rd_atomic = 0;
459 			rxe_run_task(&qp->req.task, 1);
460 		}
461 	}
462 
463 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
464 		/* state_lock used by requester & completer */
465 		spin_lock_irqsave(&qp->state_lock, flags);
466 		if ((qp->req.state == QP_STATE_DRAIN) &&
467 		    (qp->comp.psn == qp->req.psn)) {
468 			qp->req.state = QP_STATE_DRAINED;
469 			spin_unlock_irqrestore(&qp->state_lock, flags);
470 
471 			if (qp->ibqp.event_handler) {
472 				struct ib_event ev;
473 
474 				ev.device = qp->ibqp.device;
475 				ev.element.qp = &qp->ibqp;
476 				ev.event = IB_EVENT_SQ_DRAINED;
477 				qp->ibqp.event_handler(&ev,
478 					qp->ibqp.qp_context);
479 			}
480 		} else {
481 			spin_unlock_irqrestore(&qp->state_lock, flags);
482 		}
483 	}
484 
485 	do_complete(qp, wqe);
486 
487 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
488 		return COMPST_UPDATE_COMP;
489 	else
490 		return COMPST_DONE;
491 }
492 
493 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
494 					   struct rxe_pkt_info *pkt,
495 					   struct rxe_send_wqe *wqe)
496 {
497 	qp->comp.opcode = -1;
498 
499 	if (pkt) {
500 		if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
501 			qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
502 
503 		if (qp->req.wait_psn) {
504 			qp->req.wait_psn = 0;
505 			rxe_run_task(&qp->req.task, 1);
506 		}
507 	}
508 
509 	do_complete(qp, wqe);
510 
511 	return COMPST_GET_WQE;
512 }
513 
514 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
515 {
516 	struct sk_buff *skb;
517 	struct rxe_send_wqe *wqe;
518 
519 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
520 		rxe_drop_ref(qp);
521 		kfree_skb(skb);
522 	}
523 
524 	while ((wqe = queue_head(qp->sq.queue))) {
525 		if (notify) {
526 			wqe->status = IB_WC_WR_FLUSH_ERR;
527 			do_complete(qp, wqe);
528 		} else {
529 			advance_consumer(qp->sq.queue);
530 		}
531 	}
532 }
533 
534 int rxe_completer(void *arg)
535 {
536 	struct rxe_qp *qp = (struct rxe_qp *)arg;
537 	struct rxe_send_wqe *wqe = wqe;
538 	struct sk_buff *skb = NULL;
539 	struct rxe_pkt_info *pkt = NULL;
540 	enum comp_state state;
541 
542 	rxe_add_ref(qp);
543 
544 	if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
545 	    qp->req.state == QP_STATE_RESET) {
546 		rxe_drain_resp_pkts(qp, qp->valid &&
547 				    qp->req.state == QP_STATE_ERROR);
548 		goto exit;
549 	}
550 
551 	if (qp->comp.timeout) {
552 		qp->comp.timeout_retry = 1;
553 		qp->comp.timeout = 0;
554 	} else {
555 		qp->comp.timeout_retry = 0;
556 	}
557 
558 	if (qp->req.need_retry)
559 		goto exit;
560 
561 	state = COMPST_GET_ACK;
562 
563 	while (1) {
564 		pr_debug("qp#%d state = %s\n", qp_num(qp),
565 			 comp_state_name[state]);
566 		switch (state) {
567 		case COMPST_GET_ACK:
568 			skb = skb_dequeue(&qp->resp_pkts);
569 			if (skb) {
570 				pkt = SKB_TO_PKT(skb);
571 				qp->comp.timeout_retry = 0;
572 			}
573 			state = COMPST_GET_WQE;
574 			break;
575 
576 		case COMPST_GET_WQE:
577 			state = get_wqe(qp, pkt, &wqe);
578 			break;
579 
580 		case COMPST_CHECK_PSN:
581 			state = check_psn(qp, pkt, wqe);
582 			break;
583 
584 		case COMPST_CHECK_ACK:
585 			state = check_ack(qp, pkt, wqe);
586 			break;
587 
588 		case COMPST_READ:
589 			state = do_read(qp, pkt, wqe);
590 			break;
591 
592 		case COMPST_ATOMIC:
593 			state = do_atomic(qp, pkt, wqe);
594 			break;
595 
596 		case COMPST_WRITE_SEND:
597 			if (wqe->state == wqe_state_pending &&
598 			    wqe->last_psn == pkt->psn)
599 				state = COMPST_COMP_ACK;
600 			else
601 				state = COMPST_UPDATE_COMP;
602 			break;
603 
604 		case COMPST_COMP_ACK:
605 			state = complete_ack(qp, pkt, wqe);
606 			break;
607 
608 		case COMPST_COMP_WQE:
609 			state = complete_wqe(qp, pkt, wqe);
610 			break;
611 
612 		case COMPST_UPDATE_COMP:
613 			if (pkt->mask & RXE_END_MASK)
614 				qp->comp.opcode = -1;
615 			else
616 				qp->comp.opcode = pkt->opcode;
617 
618 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
619 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
620 
621 			if (qp->req.wait_psn) {
622 				qp->req.wait_psn = 0;
623 				rxe_run_task(&qp->req.task, 1);
624 			}
625 
626 			state = COMPST_DONE;
627 			break;
628 
629 		case COMPST_DONE:
630 			if (pkt) {
631 				rxe_drop_ref(pkt->qp);
632 				kfree_skb(skb);
633 				skb = NULL;
634 			}
635 			goto done;
636 
637 		case COMPST_EXIT:
638 			if (qp->comp.timeout_retry && wqe) {
639 				state = COMPST_ERROR_RETRY;
640 				break;
641 			}
642 
643 			/* re reset the timeout counter if
644 			 * (1) QP is type RC
645 			 * (2) the QP is alive
646 			 * (3) there is a packet sent by the requester that
647 			 *     might be acked (we still might get spurious
648 			 *     timeouts but try to keep them as few as possible)
649 			 * (4) the timeout parameter is set
650 			 */
651 			if ((qp_type(qp) == IB_QPT_RC) &&
652 			    (qp->req.state == QP_STATE_READY) &&
653 			    (psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
654 			    qp->qp_timeout_jiffies)
655 				mod_timer(&qp->retrans_timer,
656 					  jiffies + qp->qp_timeout_jiffies);
657 			WARN_ON_ONCE(skb);
658 			goto exit;
659 
660 		case COMPST_ERROR_RETRY:
661 			/* we come here if the retry timer fired and we did
662 			 * not receive a response packet. try to retry the send
663 			 * queue if that makes sense and the limits have not
664 			 * been exceeded. remember that some timeouts are
665 			 * spurious since we do not reset the timer but kick
666 			 * it down the road or let it expire
667 			 */
668 
669 			/* there is nothing to retry in this case */
670 			if (!wqe || (wqe->state == wqe_state_posted)) {
671 				WARN_ON_ONCE(skb);
672 				goto exit;
673 			}
674 
675 			if (qp->comp.retry_cnt > 0) {
676 				if (qp->comp.retry_cnt != 7)
677 					qp->comp.retry_cnt--;
678 
679 				/* no point in retrying if we have already
680 				 * seen the last ack that the requester could
681 				 * have caused
682 				 */
683 				if (psn_compare(qp->req.psn,
684 						qp->comp.psn) > 0) {
685 					/* tell the requester to retry the
686 					 * send send queue next time around
687 					 */
688 					qp->req.need_retry = 1;
689 					rxe_run_task(&qp->req.task, 1);
690 				}
691 
692 				if (pkt) {
693 					rxe_drop_ref(pkt->qp);
694 					kfree_skb(skb);
695 					skb = NULL;
696 				}
697 
698 				WARN_ON_ONCE(skb);
699 				goto exit;
700 
701 			} else {
702 				wqe->status = IB_WC_RETRY_EXC_ERR;
703 				state = COMPST_ERROR;
704 			}
705 			break;
706 
707 		case COMPST_RNR_RETRY:
708 			if (qp->comp.rnr_retry > 0) {
709 				if (qp->comp.rnr_retry != 7)
710 					qp->comp.rnr_retry--;
711 
712 				qp->req.need_retry = 1;
713 				pr_debug("qp#%d set rnr nak timer\n",
714 					 qp_num(qp));
715 				mod_timer(&qp->rnr_nak_timer,
716 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
717 						& ~AETH_TYPE_MASK));
718 				rxe_drop_ref(pkt->qp);
719 				kfree_skb(skb);
720 				skb = NULL;
721 				goto exit;
722 			} else {
723 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
724 				state = COMPST_ERROR;
725 			}
726 			break;
727 
728 		case COMPST_ERROR:
729 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
730 			do_complete(qp, wqe);
731 			rxe_qp_error(qp);
732 
733 			if (pkt) {
734 				rxe_drop_ref(pkt->qp);
735 				kfree_skb(skb);
736 				skb = NULL;
737 			}
738 
739 			WARN_ON_ONCE(skb);
740 			goto exit;
741 		}
742 	}
743 
744 exit:
745 	/* we come here if we are done with processing and want the task to
746 	 * exit from the loop calling us
747 	 */
748 	WARN_ON_ONCE(skb);
749 	rxe_drop_ref(qp);
750 	return -EAGAIN;
751 
752 done:
753 	/* we come here if we have processed a packet we want the task to call
754 	 * us again to see if there is anything else to do
755 	 */
756 	WARN_ON_ONCE(skb);
757 	rxe_drop_ref(qp);
758 	return 0;
759 }
760