xref: /linux/drivers/infiniband/sw/rxe/rxe_comp.c (revision 23bc06af)
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 	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 
retransmit_timer(struct timer_list * t)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->send_task);
126 	}
127 	spin_unlock_irqrestore(&qp->state_lock, flags);
128 }
129 
rxe_comp_queue_pkt(struct rxe_qp * qp,struct sk_buff * skb)130 void rxe_comp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
131 {
132 	rxe_counter_inc(SKB_TO_PKT(skb)->rxe, RXE_CNT_SENDER_SCHED);
133 	skb_queue_tail(&qp->resp_pkts, skb);
134 	rxe_sched_task(&qp->send_task);
135 }
136 
get_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe ** wqe_p)137 static inline enum comp_state get_wqe(struct rxe_qp *qp,
138 				      struct rxe_pkt_info *pkt,
139 				      struct rxe_send_wqe **wqe_p)
140 {
141 	struct rxe_send_wqe *wqe;
142 
143 	/* we come here whether or not we found a response packet to see if
144 	 * there are any posted WQEs
145 	 */
146 	wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
147 	*wqe_p = wqe;
148 
149 	/* no WQE or requester has not started it yet */
150 	if (!wqe || wqe->state == wqe_state_posted)
151 		return pkt ? COMPST_DONE : COMPST_EXIT;
152 
153 	/* WQE does not require an ack */
154 	if (wqe->state == wqe_state_done)
155 		return COMPST_COMP_WQE;
156 
157 	/* WQE caused an error */
158 	if (wqe->state == wqe_state_error)
159 		return COMPST_ERROR;
160 
161 	/* we have a WQE, if we also have an ack check its PSN */
162 	return pkt ? COMPST_CHECK_PSN : COMPST_EXIT;
163 }
164 
reset_retry_counters(struct rxe_qp * qp)165 static inline void reset_retry_counters(struct rxe_qp *qp)
166 {
167 	qp->comp.retry_cnt = qp->attr.retry_cnt;
168 	qp->comp.rnr_retry = qp->attr.rnr_retry;
169 	qp->comp.started_retry = 0;
170 }
171 
check_psn(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)172 static inline enum comp_state check_psn(struct rxe_qp *qp,
173 					struct rxe_pkt_info *pkt,
174 					struct rxe_send_wqe *wqe)
175 {
176 	s32 diff;
177 
178 	/* check to see if response is past the oldest WQE. if it is, complete
179 	 * send/write or error read/atomic
180 	 */
181 	diff = psn_compare(pkt->psn, wqe->last_psn);
182 	if (diff > 0) {
183 		if (wqe->state == wqe_state_pending) {
184 			if (wqe->mask & WR_ATOMIC_OR_READ_MASK)
185 				return COMPST_ERROR_RETRY;
186 
187 			reset_retry_counters(qp);
188 			return COMPST_COMP_WQE;
189 		} else {
190 			return COMPST_DONE;
191 		}
192 	}
193 
194 	/* compare response packet to expected response */
195 	diff = psn_compare(pkt->psn, qp->comp.psn);
196 	if (diff < 0) {
197 		/* response is most likely a retried packet if it matches an
198 		 * uncompleted WQE go complete it else ignore it
199 		 */
200 		if (pkt->psn == wqe->last_psn)
201 			return COMPST_COMP_ACK;
202 		else if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE &&
203 			 (qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST ||
204 			  qp->comp.opcode == IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE))
205 			return COMPST_CHECK_ACK;
206 		else
207 			return COMPST_DONE;
208 	} else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) {
209 		return COMPST_DONE;
210 	} else {
211 		return COMPST_CHECK_ACK;
212 	}
213 }
214 
check_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)215 static inline enum comp_state check_ack(struct rxe_qp *qp,
216 					struct rxe_pkt_info *pkt,
217 					struct rxe_send_wqe *wqe)
218 {
219 	unsigned int mask = pkt->mask;
220 	u8 syn;
221 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
222 
223 	/* Check the sequence only */
224 	switch (qp->comp.opcode) {
225 	case -1:
226 		/* Will catch all *_ONLY cases. */
227 		if (!(mask & RXE_START_MASK))
228 			return COMPST_ERROR;
229 
230 		break;
231 
232 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
233 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
234 		/* Check NAK code to handle a remote error */
235 		if (pkt->opcode == IB_OPCODE_RC_ACKNOWLEDGE)
236 			break;
237 
238 		if (pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE &&
239 		    pkt->opcode != IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST) {
240 			/* read retries of partial data may restart from
241 			 * read response first or response only.
242 			 */
243 			if ((pkt->psn == wqe->first_psn &&
244 			     pkt->opcode ==
245 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) ||
246 			    (wqe->first_psn == wqe->last_psn &&
247 			     pkt->opcode ==
248 			     IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY))
249 				break;
250 
251 			return COMPST_ERROR;
252 		}
253 		break;
254 	default:
255 		WARN_ON_ONCE(1);
256 	}
257 
258 	/* Check operation validity. */
259 	switch (pkt->opcode) {
260 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST:
261 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST:
262 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY:
263 		syn = aeth_syn(pkt);
264 
265 		if ((syn & AETH_TYPE_MASK) != AETH_ACK)
266 			return COMPST_ERROR;
267 
268 		if (wqe->wr.opcode == IB_WR_ATOMIC_WRITE)
269 			return COMPST_WRITE_SEND;
270 
271 		fallthrough;
272 		/* (IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE doesn't have an AETH)
273 		 */
274 	case IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE:
275 		if (wqe->wr.opcode != IB_WR_RDMA_READ &&
276 		    wqe->wr.opcode != IB_WR_RDMA_READ_WITH_INV &&
277 		    wqe->wr.opcode != IB_WR_FLUSH) {
278 			wqe->status = IB_WC_FATAL_ERR;
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 						qp->req.again = 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 				rxe_dbg_qp(qp, "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 		rxe_dbg_qp(qp, "unexpected opcode\n");
349 	}
350 
351 	return COMPST_ERROR;
352 }
353 
do_read(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)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), RXE_TO_MR_OBJ);
363 	if (ret) {
364 		wqe->status = IB_WC_LOC_PROT_ERR;
365 		return COMPST_ERROR;
366 	}
367 
368 	if (wqe->dma.resid == 0 && (pkt->mask & RXE_END_MASK))
369 		return COMPST_COMP_ACK;
370 
371 	return COMPST_UPDATE_COMP;
372 }
373 
do_atomic(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)374 static inline enum comp_state do_atomic(struct rxe_qp *qp,
375 					struct rxe_pkt_info *pkt,
376 					struct rxe_send_wqe *wqe)
377 {
378 	int ret;
379 
380 	u64 atomic_orig = atmack_orig(pkt);
381 
382 	ret = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE,
383 			&wqe->dma, &atomic_orig,
384 			sizeof(u64), RXE_TO_MR_OBJ);
385 	if (ret) {
386 		wqe->status = IB_WC_LOC_PROT_ERR;
387 		return COMPST_ERROR;
388 	}
389 
390 	return COMPST_COMP_ACK;
391 }
392 
make_send_cqe(struct rxe_qp * qp,struct rxe_send_wqe * wqe,struct rxe_cqe * cqe)393 static void make_send_cqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
394 			  struct rxe_cqe *cqe)
395 {
396 	struct ib_wc *wc = &cqe->ibwc;
397 	struct ib_uverbs_wc *uwc = &cqe->uibwc;
398 
399 	memset(cqe, 0, sizeof(*cqe));
400 
401 	if (!qp->is_user) {
402 		wc->wr_id = wqe->wr.wr_id;
403 		wc->status = wqe->status;
404 		wc->qp = &qp->ibqp;
405 	} else {
406 		uwc->wr_id = wqe->wr.wr_id;
407 		uwc->status = wqe->status;
408 		uwc->qp_num = qp->ibqp.qp_num;
409 	}
410 
411 	if (wqe->status == IB_WC_SUCCESS) {
412 		if (!qp->is_user) {
413 			wc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
414 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
415 			    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
416 				wc->wc_flags = IB_WC_WITH_IMM;
417 			wc->byte_len = wqe->dma.length;
418 		} else {
419 			uwc->opcode = wr_to_wc_opcode(wqe->wr.opcode);
420 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
421 			    wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
422 				uwc->wc_flags = IB_WC_WITH_IMM;
423 			uwc->byte_len = wqe->dma.length;
424 		}
425 	} else {
426 		if (wqe->status != IB_WC_WR_FLUSH_ERR)
427 			rxe_err_qp(qp, "non-flush error status = %d\n",
428 				wqe->status);
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  */
do_complete(struct rxe_qp * qp,struct rxe_send_wqe * wqe)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 	bool post;
445 
446 	/* do we need to post a completion */
447 	post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
448 			(wqe->wr.send_flags & IB_SEND_SIGNALED) ||
449 			wqe->status != IB_WC_SUCCESS);
450 
451 	if (post)
452 		make_send_cqe(qp, wqe, &cqe);
453 
454 	queue_advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_CLIENT);
455 
456 	if (post)
457 		rxe_cq_post(qp->scq, &cqe, 0);
458 
459 	if (wqe->wr.opcode == IB_WR_SEND ||
460 	    wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
461 	    wqe->wr.opcode == IB_WR_SEND_WITH_INV)
462 		rxe_counter_inc(rxe, RXE_CNT_RDMA_SEND);
463 
464 	/*
465 	 * we completed something so let req run again
466 	 * if it is trying to fence
467 	 */
468 	if (qp->req.wait_fence) {
469 		qp->req.wait_fence = 0;
470 		qp->req.again = 1;
471 	}
472 }
473 
comp_check_sq_drain_done(struct rxe_qp * qp)474 static void comp_check_sq_drain_done(struct rxe_qp *qp)
475 {
476 	unsigned long flags;
477 
478 	spin_lock_irqsave(&qp->state_lock, flags);
479 	if (unlikely(qp_state(qp) == IB_QPS_SQD)) {
480 		if (qp->attr.sq_draining && qp->comp.psn == qp->req.psn) {
481 			qp->attr.sq_draining = 0;
482 			spin_unlock_irqrestore(&qp->state_lock, flags);
483 
484 			if (qp->ibqp.event_handler) {
485 				struct ib_event ev;
486 
487 				ev.device = qp->ibqp.device;
488 				ev.element.qp = &qp->ibqp;
489 				ev.event = IB_EVENT_SQ_DRAINED;
490 				qp->ibqp.event_handler(&ev,
491 					qp->ibqp.qp_context);
492 			}
493 			return;
494 		}
495 	}
496 	spin_unlock_irqrestore(&qp->state_lock, flags);
497 }
498 
complete_ack(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)499 static inline enum comp_state complete_ack(struct rxe_qp *qp,
500 					   struct rxe_pkt_info *pkt,
501 					   struct rxe_send_wqe *wqe)
502 {
503 	if (wqe->has_rd_atomic) {
504 		wqe->has_rd_atomic = 0;
505 		atomic_inc(&qp->req.rd_atomic);
506 		if (qp->req.need_rd_atomic) {
507 			qp->comp.timeout_retry = 0;
508 			qp->req.need_rd_atomic = 0;
509 			qp->req.again = 1;
510 		}
511 	}
512 
513 	comp_check_sq_drain_done(qp);
514 
515 	do_complete(qp, wqe);
516 
517 	if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
518 		return COMPST_UPDATE_COMP;
519 	else
520 		return COMPST_DONE;
521 }
522 
complete_wqe(struct rxe_qp * qp,struct rxe_pkt_info * pkt,struct rxe_send_wqe * wqe)523 static inline enum comp_state complete_wqe(struct rxe_qp *qp,
524 					   struct rxe_pkt_info *pkt,
525 					   struct rxe_send_wqe *wqe)
526 {
527 	if (pkt && wqe->state == wqe_state_pending) {
528 		if (psn_compare(wqe->last_psn, qp->comp.psn) >= 0) {
529 			qp->comp.psn = (wqe->last_psn + 1) & BTH_PSN_MASK;
530 			qp->comp.opcode = -1;
531 		}
532 
533 		if (qp->req.wait_psn) {
534 			qp->req.wait_psn = 0;
535 			qp->req.again = 1;
536 		}
537 	}
538 
539 	do_complete(qp, wqe);
540 
541 	return COMPST_GET_WQE;
542 }
543 
544 /* drain incoming response packet queue */
drain_resp_pkts(struct rxe_qp * qp)545 static void drain_resp_pkts(struct rxe_qp *qp)
546 {
547 	struct sk_buff *skb;
548 
549 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
550 		rxe_put(qp);
551 		kfree_skb(skb);
552 		ib_device_put(qp->ibqp.device);
553 	}
554 }
555 
556 /* complete send wqe with flush error */
flush_send_wqe(struct rxe_qp * qp,struct rxe_send_wqe * wqe)557 static int flush_send_wqe(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
558 {
559 	struct rxe_cqe cqe = {};
560 	struct ib_wc *wc = &cqe.ibwc;
561 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
562 	int err;
563 
564 	if (qp->is_user) {
565 		uwc->wr_id = wqe->wr.wr_id;
566 		uwc->status = IB_WC_WR_FLUSH_ERR;
567 		uwc->qp_num = qp->ibqp.qp_num;
568 	} else {
569 		wc->wr_id = wqe->wr.wr_id;
570 		wc->status = IB_WC_WR_FLUSH_ERR;
571 		wc->qp = &qp->ibqp;
572 	}
573 
574 	err = rxe_cq_post(qp->scq, &cqe, 0);
575 	if (err)
576 		rxe_dbg_cq(qp->scq, "post cq failed, err = %d\n", err);
577 
578 	return err;
579 }
580 
581 /* drain and optionally complete the send queue
582  * if unable to complete a wqe, i.e. cq is full, stop
583  * completing and flush the remaining wqes
584  */
flush_send_queue(struct rxe_qp * qp,bool notify)585 static void flush_send_queue(struct rxe_qp *qp, bool notify)
586 {
587 	struct rxe_send_wqe *wqe;
588 	struct rxe_queue *q = qp->sq.queue;
589 	int err;
590 
591 	/* send queue never got created. nothing to do. */
592 	if (!qp->sq.queue)
593 		return;
594 
595 	while ((wqe = queue_head(q, q->type))) {
596 		if (notify) {
597 			err = flush_send_wqe(qp, wqe);
598 			if (err)
599 				notify = 0;
600 		}
601 		queue_advance_consumer(q, q->type);
602 	}
603 }
604 
free_pkt(struct rxe_pkt_info * pkt)605 static void free_pkt(struct rxe_pkt_info *pkt)
606 {
607 	struct sk_buff *skb = PKT_TO_SKB(pkt);
608 	struct rxe_qp *qp = pkt->qp;
609 	struct ib_device *dev = qp->ibqp.device;
610 
611 	kfree_skb(skb);
612 	rxe_put(qp);
613 	ib_device_put(dev);
614 }
615 
616 /* reset the retry timer if
617  * - QP is type RC
618  * - there is a packet sent by the requester that
619  *   might be acked (we still might get spurious
620  *   timeouts but try to keep them as few as possible)
621  * - the timeout parameter is set
622  * - the QP is alive
623  */
reset_retry_timer(struct rxe_qp * qp)624 static void reset_retry_timer(struct rxe_qp *qp)
625 {
626 	unsigned long flags;
627 
628 	if (qp_type(qp) == IB_QPT_RC && qp->qp_timeout_jiffies) {
629 		spin_lock_irqsave(&qp->state_lock, flags);
630 		if (qp_state(qp) >= IB_QPS_RTS &&
631 		    psn_compare(qp->req.psn, qp->comp.psn) > 0)
632 			mod_timer(&qp->retrans_timer,
633 				  jiffies + qp->qp_timeout_jiffies);
634 		spin_unlock_irqrestore(&qp->state_lock, flags);
635 	}
636 }
637 
rxe_completer(struct rxe_qp * qp)638 int rxe_completer(struct rxe_qp *qp)
639 {
640 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
641 	struct rxe_send_wqe *wqe = NULL;
642 	struct sk_buff *skb = NULL;
643 	struct rxe_pkt_info *pkt = NULL;
644 	enum comp_state state;
645 	int ret;
646 	unsigned long flags;
647 
648 	qp->req.again = 0;
649 
650 	spin_lock_irqsave(&qp->state_lock, flags);
651 	if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
652 			  qp_state(qp) == IB_QPS_RESET) {
653 		bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
654 
655 		drain_resp_pkts(qp);
656 		flush_send_queue(qp, notify);
657 		spin_unlock_irqrestore(&qp->state_lock, flags);
658 		goto exit;
659 	}
660 	spin_unlock_irqrestore(&qp->state_lock, flags);
661 
662 	if (qp->comp.timeout) {
663 		qp->comp.timeout_retry = 1;
664 		qp->comp.timeout = 0;
665 	} else {
666 		qp->comp.timeout_retry = 0;
667 	}
668 
669 	if (qp->req.need_retry)
670 		goto exit;
671 
672 	state = COMPST_GET_ACK;
673 
674 	while (1) {
675 		rxe_dbg_qp(qp, "state = %s\n", comp_state_name[state]);
676 		switch (state) {
677 		case COMPST_GET_ACK:
678 			skb = skb_dequeue(&qp->resp_pkts);
679 			if (skb) {
680 				pkt = SKB_TO_PKT(skb);
681 				qp->comp.timeout_retry = 0;
682 			}
683 			state = COMPST_GET_WQE;
684 			break;
685 
686 		case COMPST_GET_WQE:
687 			state = get_wqe(qp, pkt, &wqe);
688 			break;
689 
690 		case COMPST_CHECK_PSN:
691 			state = check_psn(qp, pkt, wqe);
692 			break;
693 
694 		case COMPST_CHECK_ACK:
695 			state = check_ack(qp, pkt, wqe);
696 			break;
697 
698 		case COMPST_READ:
699 			state = do_read(qp, pkt, wqe);
700 			break;
701 
702 		case COMPST_ATOMIC:
703 			state = do_atomic(qp, pkt, wqe);
704 			break;
705 
706 		case COMPST_WRITE_SEND:
707 			if (wqe->state == wqe_state_pending &&
708 			    wqe->last_psn == pkt->psn)
709 				state = COMPST_COMP_ACK;
710 			else
711 				state = COMPST_UPDATE_COMP;
712 			break;
713 
714 		case COMPST_COMP_ACK:
715 			state = complete_ack(qp, pkt, wqe);
716 			break;
717 
718 		case COMPST_COMP_WQE:
719 			state = complete_wqe(qp, pkt, wqe);
720 			break;
721 
722 		case COMPST_UPDATE_COMP:
723 			if (pkt->mask & RXE_END_MASK)
724 				qp->comp.opcode = -1;
725 			else
726 				qp->comp.opcode = pkt->opcode;
727 
728 			if (psn_compare(pkt->psn, qp->comp.psn) >= 0)
729 				qp->comp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
730 
731 			if (qp->req.wait_psn) {
732 				qp->req.wait_psn = 0;
733 				qp->req.again = 1;
734 			}
735 
736 			state = COMPST_DONE;
737 			break;
738 
739 		case COMPST_DONE:
740 			goto done;
741 
742 		case COMPST_EXIT:
743 			if (qp->comp.timeout_retry && wqe) {
744 				state = COMPST_ERROR_RETRY;
745 				break;
746 			}
747 
748 			reset_retry_timer(qp);
749 			goto exit;
750 
751 		case COMPST_ERROR_RETRY:
752 			/* we come here if the retry timer fired and we did
753 			 * not receive a response packet. try to retry the send
754 			 * queue if that makes sense and the limits have not
755 			 * been exceeded. remember that some timeouts are
756 			 * spurious since we do not reset the timer but kick
757 			 * it down the road or let it expire
758 			 */
759 
760 			/* there is nothing to retry in this case */
761 			if (!wqe || (wqe->state == wqe_state_posted))
762 				goto exit;
763 
764 			/* if we've started a retry, don't start another
765 			 * retry sequence, unless this is a timeout.
766 			 */
767 			if (qp->comp.started_retry &&
768 			    !qp->comp.timeout_retry)
769 				goto done;
770 
771 			if (qp->comp.retry_cnt > 0) {
772 				if (qp->comp.retry_cnt != 7)
773 					qp->comp.retry_cnt--;
774 
775 				/* no point in retrying if we have already
776 				 * seen the last ack that the requester could
777 				 * have caused
778 				 */
779 				if (psn_compare(qp->req.psn,
780 						qp->comp.psn) > 0) {
781 					/* tell the requester to retry the
782 					 * send queue next time around
783 					 */
784 					rxe_counter_inc(rxe,
785 							RXE_CNT_COMP_RETRY);
786 					qp->req.need_retry = 1;
787 					qp->comp.started_retry = 1;
788 					qp->req.again = 1;
789 				}
790 				goto done;
791 
792 			} else {
793 				rxe_counter_inc(rxe, RXE_CNT_RETRY_EXCEEDED);
794 				wqe->status = IB_WC_RETRY_EXC_ERR;
795 				state = COMPST_ERROR;
796 			}
797 			break;
798 
799 		case COMPST_RNR_RETRY:
800 			/* we come here if we received an RNR NAK */
801 			if (qp->comp.rnr_retry > 0) {
802 				if (qp->comp.rnr_retry != 7)
803 					qp->comp.rnr_retry--;
804 
805 				/* don't start a retry flow until the
806 				 * rnr timer has fired
807 				 */
808 				qp->req.wait_for_rnr_timer = 1;
809 				rxe_dbg_qp(qp, "set rnr nak timer\n");
810 				// TODO who protects from destroy_qp??
811 				mod_timer(&qp->rnr_nak_timer,
812 					  jiffies + rnrnak_jiffies(aeth_syn(pkt)
813 						& ~AETH_TYPE_MASK));
814 				goto exit;
815 			} else {
816 				rxe_counter_inc(rxe,
817 						RXE_CNT_RNR_RETRY_EXCEEDED);
818 				wqe->status = IB_WC_RNR_RETRY_EXC_ERR;
819 				state = COMPST_ERROR;
820 			}
821 			break;
822 
823 		case COMPST_ERROR:
824 			WARN_ON_ONCE(wqe->status == IB_WC_SUCCESS);
825 			do_complete(qp, wqe);
826 			rxe_qp_error(qp);
827 			goto exit;
828 		}
829 	}
830 
831 	/* A non-zero return value will cause rxe_do_task to
832 	 * exit its loop and end the work item. A zero return
833 	 * will continue looping and return to rxe_completer
834 	 */
835 done:
836 	ret = 0;
837 	goto out;
838 exit:
839 	ret = (qp->req.again) ? 0 : -EAGAIN;
840 out:
841 	qp->req.again = 0;
842 	if (pkt)
843 		free_pkt(pkt);
844 	return ret;
845 }
846