• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
rnrnak_jiffies(u8 timeout)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 
wr_to_wc_opcode(enum ib_wr_opcode opcode)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 
108 	default:
109 		return 0xff;
110 	}
111 }
112 
retransmit_timer(struct timer_list * t)113 void retransmit_timer(struct timer_list *t)
114 {
115 	struct rxe_qp *qp = from_timer(qp, t, retrans_timer);
116 
117 	if (qp->valid) {
118 		qp->comp.timeout = 1;
119 		rxe_run_task(&qp->comp.task, 1);
120 	}
121 }
122 
rxe_comp_queue_pkt(struct rxe_qp * qp,struct sk_buff * skb)123 void rxe_comp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
124 {
125 	int must_sched;
126 
127 	skb_queue_tail(&qp->resp_pkts, skb);
128 
129 	must_sched = skb_queue_len(&qp->resp_pkts) > 1;
130 	if (must_sched != 0)
131 		rxe_counter_inc(SKB_TO_PKT(skb)->rxe, RXE_CNT_COMPLETER_SCHED);
132 
133 	rxe_run_task(&qp->comp.task, must_sched);
134 }
135 
get_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe ** wqe_p)136 static inline enum comp_state get_wqe(struct rxe_qp *qp,
137 				      struct rxe_pkt_info *pkt,
138 				      struct rxe_send_wqe **wqe_p)
139 {
140 	struct rxe_send_wqe *wqe;
141 
142 	/* we come here whether or not we found a response packet to see if
143 	 * there are any posted WQEs
144 	 */
145 	wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
146 	*wqe_p = wqe;
147 
148 	/* no WQE or requester has not started it yet */
149 	if (!wqe || wqe->state == wqe_state_posted)
150 		return pkt ? COMPST_DONE : COMPST_EXIT;
151 
152 	/* WQE does not require an ack */
153 	if (wqe->state == wqe_state_done)
154 		return COMPST_COMP_WQE;
155 
156 	/* WQE caused an error */
157 	if (wqe->state == wqe_state_error)
158 		return COMPST_ERROR;
159 
160 	/* we have a WQE, if we also have an ack check its PSN */
161 	return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
162 }
163 
reset_retry_counters(struct rxe_qp * qp)164 static inline void reset_retry_counters(struct rxe_qp *qp)
165 {
166 	qp->comp.retry_cnt = qp->attr.retry_cnt;
167 	qp->comp.rnr_retry = qp->attr.rnr_retry;
168 	qp->comp.started_retry = 0;
169 }
170 
check_psn(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)171 static inline enum comp_state check_psn(struct rxe_qp *qp,
172 					struct rxe_pkt_info *pkt,
173 					struct rxe_send_wqe *wqe)
174 {
175 	s32 diff;
176 
177 	/* check to see if response is past the oldest WQE. if it is, complete
178 	 * send/write or error read/atomic
179 	 */
180 	diff = psn_compare(pkt->psn, wqe->last_psn);
181 	if (diff > 0) {
182 		if (wqe->state == wqe_state_pending) {
183 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
184 				return COMPST_ERROR_RETRY;
185 
186 			reset_retry_counters(qp);
187 			return COMPST_COMP_WQE;
188 		} else {
189 			return COMPST_DONE;
190 		}
191 	}
192 
193 	/* compare response packet to expected response */
194 	diff = psn_compare(pkt->psn, qp->comp.psn);
195 	if (diff < 0) {
196 		/* response is most likely a retried packet if it matches an
197 		 * uncompleted WQE go complete it else ignore it
198 		 */
199 		if (pkt->psn == wqe->last_psn)
200 			return COMPST_COMP_ACK;
201 		else
202 			return COMPST_DONE;
203 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
204 		return COMPST_DONE;
205 	} else {
206 		return COMPST_CHECK_ACK;
207 	}
208 }
209 
check_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)210 static inline enum comp_state check_ack(struct rxe_qp *qp,
211 					struct rxe_pkt_info *pkt,
212 					struct rxe_send_wqe *wqe)
213 {
214 	unsigned int mask = pkt->mask;
215 	u8 syn;
216 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
217 
218 	/* Check the sequence only */
219 	switch (qp->comp.opcode) {
220 	case -1:
221 		/* Will catch all *_ONLY cases. */
222 		if (!(mask & RXE_START_MASK))
223 			return COMPST_ERROR;
224 
225 		break;
226 
227 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
228 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
229 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
230 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
231 			/* read retries of partial data may restart from
232 			 * read response first or response only.
233 			 */
234 			if ((pkt->psn == wqe->first_psn &&
235 			     pkt->opcode ==
236 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) ||
237 			    (wqe->first_psn == wqe->last_psn &&
238 			     pkt->opcode ==
239 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY))
240 				break;
241 
242 			return COMPST_ERROR;
243 		}
244 		break;
245 	default:
246 		WARN_ON_ONCE(1);
247 	}
248 
249 	/* Check operation validity. */
250 	switch (pkt->opcode) {
251 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
252 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
253 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
254 		syn = aeth_syn(pkt);
255 
256 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
257 			return COMPST_ERROR;
258 
259 		fallthrough;
260 		/* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
261 		 */
262 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
263 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
264 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV) {
265 			wqe->status = IB_WC_FATAL_ERR;
266 			return COMPST_ERROR;
267 		}
268 		reset_retry_counters(qp);
269 		return COMPST_READ;
270 
271 	case IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE:
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_CMP_AND_SWP &&
278 		    wqe->wr.opcode != IB_WR_ATOMIC_FETCH_AND_ADD)
279 			return COMPST_ERROR;
280 		reset_retry_counters(qp);
281 		return COMPST_ATOMIC;
282 
283 	case IB_OPCODE_RC_ACKNOWLEDGE:
284 		syn = aeth_syn(pkt);
285 		switch (syn & AETH_TYPE_MASK) {
286 		case AETH_ACK:
287 			reset_retry_counters(qp);
288 			return COMPST_WRITE_SEND;
289 
290 		case AETH_RNR_NAK:
291 			rxe_counter_inc(rxe, RXE_CNT_RCV_RNR);
292 			return COMPST_RNR_RETRY;
293 
294 		case AETH_NAK:
295 			switch (syn) {
296 			case AETH_NAK_PSN_SEQ_ERROR:
297 				/* a nak implicitly acks all packets with psns
298 				 * before
299 				 */
300 				if (psn_compare(pkt->psn, qp->comp.psn) > 0) {
301 					rxe_counter_inc(rxe,
302 							RXE_CNT_RCV_SEQ_ERR);
303 					qp->comp.psn = pkt->psn;
304 					if (qp->req.wait_psn) {
305 						qp->req.wait_psn = 0;
306 						rxe_run_task(&qp->req.task, 0);
307 					}
308 				}
309 				return COMPST_ERROR_RETRY;
310 
311 			case AETH_NAK_INVALID_REQ:
312 				wqe->status = IB_WC_REM_INV_REQ_ERR;
313 				return COMPST_ERROR;
314 
315 			case AETH_NAK_REM_ACC_ERR:
316 				wqe->status = IB_WC_REM_ACCESS_ERR;
317 				return COMPST_ERROR;
318 
319 			case AETH_NAK_REM_OP_ERR:
320 				wqe->status = IB_WC_REM_OP_ERR;
321 				return COMPST_ERROR;
322 
323 			default:
324 				pr_warn("unexpected nak %x\n", syn);
325 				wqe->status = IB_WC_REM_OP_ERR;
326 				return COMPST_ERROR;
327 			}
328 
329 		default:
330 			return COMPST_ERROR;
331 		}
332 		break;
333 
334 	default:
335 		pr_warn("unexpected opcode\n");
336 	}
337 
338 	return COMPST_ERROR;
339 }
340 
do_read(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)341 static inline enum comp_state do_read(struct rxe_qp *qp,
342 				      struct rxe_pkt_info *pkt,
343 				      struct rxe_send_wqe *wqe)
344 {
345 	int ret;
346 
347 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
348 			&wqe->dma, payload_addr(pkt),
349 			payload_size(pkt), RXE_TO_MR_OBJ);
350 	if (ret) {
351 		wqe->status = IB_WC_LOC_PROT_ERR;
352 		return COMPST_ERROR;
353 	}
354 
355 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
356 		return COMPST_COMP_ACK;
357 
358 	return COMPST_UPDATE_COMP;
359 }
360 
do_atomic(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)361 static inline enum comp_state do_atomic(struct rxe_qp *qp,
362 					struct rxe_pkt_info *pkt,
363 					struct rxe_send_wqe *wqe)
364 {
365 	int ret;
366 
367 	u64 atomic_orig = atmack_orig(pkt);
368 
369 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
370 			&wqe->dma, &atomic_orig,
371 			sizeof(u64), RXE_TO_MR_OBJ);
372 	if (ret) {
373 		wqe->status = IB_WC_LOC_PROT_ERR;
374 		return COMPST_ERROR;
375 	}
376 
377 	return COMPST_COMP_ACK;
378 }
379 
make_send_cqe(struct rxe_qp * qp,struct rxe_send_wqe * wqe,struct rxe_cqe * cqe)380 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
381 			  struct rxe_cqe *cqe)
382 {
383 	memset(cqe, 0, sizeof(*cqe));
384 
385 	if (!qp->is_user) {
386 		struct ib_wc		*wc	= &cqe->ibwc;
387 
388 		wc->wr_id		= wqe->wr.wr_id;
389 		wc->status		= wqe->status;
390 		wc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
391 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
392 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
393 			wc->wc_flags = IB_WC_WITH_IMM;
394 		wc->byte_len		= wqe->dma.length;
395 		wc->qp			= &qp->ibqp;
396 	} else {
397 		struct ib_uverbs_wc	*uwc	= &cqe->uibwc;
398 
399 		uwc->wr_id		= wqe->wr.wr_id;
400 		uwc->status		= wqe->status;
401 		uwc->opcode		= wr_to_wc_opcode(wqe->wr.opcode);
402 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
403 		    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
404 			uwc->wc_flags = IB_WC_WITH_IMM;
405 		uwc->byte_len		= wqe->dma.length;
406 		uwc->qp_num		= qp->ibqp.qp_num;
407 	}
408 }
409 
410 /*
411  * IBA Spec. Section 10.7.3.1 SIGNALED COMPLETIONS
412  * ---------8<---------8<-------------
413  * ...Note that if a completion error occurs, a Work Completion
414  * will always be generated, even if the signaling
415  * indicator requests an Unsignaled Completion.
416  * ---------8<---------8<-------------
417  */
do_complete(struct rxe_qp * qp,struct rxe_send_wqe * wqe)418 static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
419 {
420 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
421 	struct rxe_cqe cqe;
422 	bool post;
423 
424 	/* do we need to post a completion */
425 	post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
426 			(wqe->wr.send_flags & IB_SEND_SIGNALED) ||
427 			wqe->status != IB_WC_SUCCESS);
428 
429 	if (post)
430 		make_send_cqe(qp, wqe, &cqe);
431 
432 	queue_advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
433 
434 	if (post)
435 		rxe_cq_post(qp->scq, &cqe, 0);
436 
437 	if (wqe->wr.opcode == IB_WR_SEND ||
438 	    wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
439 	    wqe->wr.opcode == IB_WR_SEND_WITH_INV)
440 		rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND);
441 
442 	/*
443 	 * we completed something so let req run again
444 	 * if it is trying to fence
445 	 */
446 	if (qp->req.wait_fence) {
447 		qp->req.wait_fence = 0;
448 		rxe_run_task(&qp->req.task, 0);
449 	}
450 }
451 
complete_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)452 static inline enum comp_state complete_ack(struct rxe_qp *qp,
453 					   struct rxe_pkt_info *pkt,
454 					   struct rxe_send_wqe *wqe)
455 {
456 	unsigned long flags;
457 
458 	if (wqe->has_rd_atomic) {
459 		wqe->has_rd_atomic = 0;
460 		atomic_inc(&qp->req.rd_atomic);
461 		if (qp->req.need_rd_atomic) {
462 			qp->comp.timeout_retry = 0;
463 			qp->req.need_rd_atomic = 0;
464 			rxe_run_task(&qp->req.task, 0);
465 		}
466 	}
467 
468 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
469 		/* state_lock used by requester & completer */
470 		spin_lock_irqsave(&qp->state_lock, flags);
471 		if ((qp->req.state == QP_STATE_DRAIN) &&
472 		    (qp->comp.psn == qp->req.psn)) {
473 			qp->req.state = QP_STATE_DRAINED;
474 			spin_unlock_irqrestore(&qp->state_lock, flags);
475 
476 			if (qp->ibqp.event_handler) {
477 				struct ib_event ev;
478 
479 				ev.device = qp->ibqp.device;
480 				ev.element.qp = &qp->ibqp;
481 				ev.event = IB_EVENT_SQ_DRAINED;
482 				qp->ibqp.event_handler(&ev,
483 					qp->ibqp.qp_context);
484 			}
485 		} else {
486 			spin_unlock_irqrestore(&qp->state_lock, flags);
487 		}
488 	}
489 
490 	do_complete(qp, wqe);
491 
492 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
493 		return COMPST_UPDATE_COMP;
494 	else
495 		return COMPST_DONE;
496 }
497 
complete_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)498 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
499 					   struct rxe_pkt_info *pkt,
500 					   struct rxe_send_wqe *wqe)
501 {
502 	if (pkt && wqe->state == wqe_state_pending) {
503 		if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) {
504 			qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK;
505 			qp->comp.opcode = -1;
506 		}
507 
508 		if (qp->req.wait_psn) {
509 			qp->req.wait_psn = 0;
510 			rxe_run_task(&qp->req.task, 1);
511 		}
512 	}
513 
514 	do_complete(qp, wqe);
515 
516 	return COMPST_GET_WQE;
517 }
518 
rxe_drain_resp_pkts(struct rxe_qp * qp,bool notify)519 static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
520 {
521 	struct sk_buff *skb;
522 	struct rxe_send_wqe *wqe;
523 	struct rxe_queue *q = qp->sq.queue;
524 
525 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
526 		rxe_drop_ref(qp);
527 		kfree_skb(skb);
528 		ib_device_put(qp->ibqp.device);
529 	}
530 
531 	while ((wqe = queue_head(q, q->type))) {
532 		if (notify) {
533 			wqe->status = IB_WC_WR_FLUSH_ERR;
534 			do_complete(qp, wqe);
535 		} else {
536 			queue_advance_consumer(q, q->type);
537 		}
538 	}
539 }
540 
free_pkt(struct rxe_pkt_info * pkt)541 static void free_pkt(struct rxe_pkt_info *pkt)
542 {
543 	struct sk_buff *skb = PKT_TO_SKB(pkt);
544 	struct rxe_qp *qp = pkt->qp;
545 	struct ib_device *dev = qp->ibqp.device;
546 
547 	kfree_skb(skb);
548 	rxe_drop_ref(qp);
549 	ib_device_put(dev);
550 }
551 
rxe_completer(void * arg)552 int rxe_completer(void *arg)
553 {
554 	struct rxe_qp *qp = (struct rxe_qp *)arg;
555 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
556 	struct rxe_send_wqe *wqe = NULL;
557 	struct sk_buff *skb = NULL;
558 	struct rxe_pkt_info *pkt = NULL;
559 	enum comp_state state;
560 	int ret = 0;
561 
562 	rxe_add_ref(qp);
563 
564 	if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
565 	    qp->req.state == QP_STATE_RESET) {
566 		rxe_drain_resp_pkts(qp, qp->valid &&
567 				    qp->req.state == QP_STATE_ERROR);
568 		ret = -EAGAIN;
569 		goto done;
570 	}
571 
572 	if (qp->comp.timeout) {
573 		qp->comp.timeout_retry = 1;
574 		qp->comp.timeout = 0;
575 	} else {
576 		qp->comp.timeout_retry = 0;
577 	}
578 
579 	if (qp->req.need_retry) {
580 		ret = -EAGAIN;
581 		goto done;
582 	}
583 
584 	state = COMPST_GET_ACK;
585 
586 	while (1) {
587 		pr_debug("qp#%d state = %s\n", qp_num(qp),
588 			 comp_state_name[state]);
589 		switch (state) {
590 		case COMPST_GET_ACK:
591 			skb = skb_dequeue(&qp->resp_pkts);
592 			if (skb) {
593 				pkt = SKB_TO_PKT(skb);
594 				qp->comp.timeout_retry = 0;
595 			}
596 			state = COMPST_GET_WQE;
597 			break;
598 
599 		case COMPST_GET_WQE:
600 			state = get_wqe(qp, pkt, &wqe);
601 			break;
602 
603 		case COMPST_CHECK_PSN:
604 			state = check_psn(qp, pkt, wqe);
605 			break;
606 
607 		case COMPST_CHECK_ACK:
608 			state = check_ack(qp, pkt, wqe);
609 			break;
610 
611 		case COMPST_READ:
612 			state = do_read(qp, pkt, wqe);
613 			break;
614 
615 		case COMPST_ATOMIC:
616 			state = do_atomic(qp, pkt, wqe);
617 			break;
618 
619 		case COMPST_WRITE_SEND:
620 			if (wqe->state == wqe_state_pending &&
621 			    wqe->last_psn == pkt->psn)
622 				state = COMPST_COMP_ACK;
623 			else
624 				state = COMPST_UPDATE_COMP;
625 			break;
626 
627 		case COMPST_COMP_ACK:
628 			state = complete_ack(qp, pkt, wqe);
629 			break;
630 
631 		case COMPST_COMP_WQE:
632 			state = complete_wqe(qp, pkt, wqe);
633 			break;
634 
635 		case COMPST_UPDATE_COMP:
636 			if (pkt->mask & RXE_END_MASK)
637 				qp->comp.opcode = -1;
638 			else
639 				qp->comp.opcode = pkt->opcode;
640 
641 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
642 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
643 
644 			if (qp->req.wait_psn) {
645 				qp->req.wait_psn = 0;
646 				rxe_run_task(&qp->req.task, 1);
647 			}
648 
649 			state = COMPST_DONE;
650 			break;
651 
652 		case COMPST_DONE:
653 			goto done;
654 
655 		case COMPST_EXIT:
656 			if (qp->comp.timeout_retry && wqe) {
657 				state = COMPST_ERROR_RETRY;
658 				break;
659 			}
660 
661 			/* re reset the timeout counter if
662 			 * (1) QP is type RC
663 			 * (2) the QP is alive
664 			 * (3) there is a packet sent by the requester that
665 			 *     might be acked (we still might get spurious
666 			 *     timeouts but try to keep them as few as possible)
667 			 * (4) the timeout parameter is set
668 			 */
669 			if ((qp_type(qp) == IB_QPT_RC) &&
670 			    (qp->req.state == QP_STATE_READY) &&
671 			    (psn_compare(qp->req.psn, qp->comp.psn) > 0) &&
672 			    qp->qp_timeout_jiffies)
673 				mod_timer(&qp->retrans_timer,
674 					  jiffies + qp->qp_timeout_jiffies);
675 			ret = -EAGAIN;
676 			goto done;
677 
678 		case COMPST_ERROR_RETRY:
679 			/* we come here if the retry timer fired and we did
680 			 * not receive a response packet. try to retry the send
681 			 * queue if that makes sense and the limits have not
682 			 * been exceeded. remember that some timeouts are
683 			 * spurious since we do not reset the timer but kick
684 			 * it down the road or let it expire
685 			 */
686 
687 			/* there is nothing to retry in this case */
688 			if (!wqe || (wqe->state == wqe_state_posted)) {
689 				ret = -EAGAIN;
690 				goto done;
691 			}
692 
693 			/* if we've started a retry, don't start another
694 			 * retry sequence, unless this is a timeout.
695 			 */
696 			if (qp->comp.started_retry &&
697 			    !qp->comp.timeout_retry)
698 				goto done;
699 
700 			if (qp->comp.retry_cnt > 0) {
701 				if (qp->comp.retry_cnt != 7)
702 					qp->comp.retry_cnt--;
703 
704 				/* no point in retrying if we have already
705 				 * seen the last ack that the requester could
706 				 * have caused
707 				 */
708 				if (psn_compare(qp->req.psn,
709 						qp->comp.psn) > 0) {
710 					/* tell the requester to retry the
711 					 * send queue next time around
712 					 */
713 					rxe_counter_inc(rxe,
714 							RXE_CNT_COMP_RETRY);
715 					qp->req.need_retry = 1;
716 					qp->comp.started_retry = 1;
717 					rxe_run_task(&qp->req.task, 0);
718 				}
719 				goto done;
720 
721 			} else {
722 				rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
723 				wqe->status = IB_WC_RETRY_EXC_ERR;
724 				state = COMPST_ERROR;
725 			}
726 			break;
727 
728 		case COMPST_RNR_RETRY:
729 			if (qp->comp.rnr_retry > 0) {
730 				if (qp->comp.rnr_retry != 7)
731 					qp->comp.rnr_retry--;
732 
733 				qp->req.need_retry = 1;
734 				pr_debug("qp#%d set rnr nak timer\n",
735 					 qp_num(qp));
736 				mod_timer(&qp->rnr_nak_timer,
737 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
738 						& ~AETH_TYPE_MASK));
739 				ret = -EAGAIN;
740 				goto done;
741 			} else {
742 				rxe_counter_inc(rxe,
743 						RXE_CNT_RNR_RETRY_EXCEEDED);
744 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
745 				state = COMPST_ERROR;
746 			}
747 			break;
748 
749 		case COMPST_ERROR:
750 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
751 			do_complete(qp, wqe);
752 			rxe_qp_error(qp);
753 			ret = -EAGAIN;
754 			goto done;
755 		}
756 	}
757 
758 done:
759 	if (pkt)
760 		free_pkt(pkt);
761 	rxe_drop_ref(qp);
762 
763 	return ret;
764 }
765