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 #include <crypto/hash.h>
9 
10 #include "rxe.h"
11 #include "rxe_loc.h"
12 #include "rxe_queue.h"
13 
14 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
15 		       u32 opcode);
16 
17 static inline void retry_first_write_send(struct rxe_qp *qp,
18 					  struct rxe_send_wqe *wqe, int npsn)
19 {
20 	int i;
21 
22 	for (i = 0; i < npsn; i++) {
23 		int to_send = (wqe->dma.resid > qp->mtu) ?
24 				qp->mtu : wqe->dma.resid;
25 
26 		qp->req.opcode = next_opcode(qp, wqe,
27 					     wqe->wr.opcode);
28 
29 		if (wqe->wr.send_flags & IB_SEND_INLINE) {
30 			wqe->dma.resid -= to_send;
31 			wqe->dma.sge_offset += to_send;
32 		} else {
33 			advance_dma_data(&wqe->dma, to_send);
34 		}
35 	}
36 }
37 
38 static void req_retry(struct rxe_qp *qp)
39 {
40 	struct rxe_send_wqe *wqe;
41 	unsigned int wqe_index;
42 	unsigned int mask;
43 	int npsn;
44 	int first = 1;
45 	struct rxe_queue *q = qp->sq.queue;
46 	unsigned int cons;
47 	unsigned int prod;
48 
49 	cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
50 	prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
51 
52 	qp->req.wqe_index	= cons;
53 	qp->req.psn		= qp->comp.psn;
54 	qp->req.opcode		= -1;
55 
56 	for (wqe_index = cons; wqe_index != prod;
57 			wqe_index = queue_next_index(q, wqe_index)) {
58 		wqe = queue_addr_from_index(qp->sq.queue, wqe_index);
59 		mask = wr_opcode_mask(wqe->wr.opcode, qp);
60 
61 		if (wqe->state == wqe_state_posted)
62 			break;
63 
64 		if (wqe->state == wqe_state_done)
65 			continue;
66 
67 		wqe->iova = (mask & WR_ATOMIC_MASK) ?
68 			     wqe->wr.wr.atomic.remote_addr :
69 			     (mask & WR_READ_OR_WRITE_MASK) ?
70 			     wqe->wr.wr.rdma.remote_addr :
71 			     0;
72 
73 		if (!first || (mask & WR_READ_MASK) == 0) {
74 			wqe->dma.resid = wqe->dma.length;
75 			wqe->dma.cur_sge = 0;
76 			wqe->dma.sge_offset = 0;
77 		}
78 
79 		if (first) {
80 			first = 0;
81 
82 			if (mask & WR_WRITE_OR_SEND_MASK) {
83 				npsn = (qp->comp.psn - wqe->first_psn) &
84 					BTH_PSN_MASK;
85 				retry_first_write_send(qp, wqe, npsn);
86 			}
87 
88 			if (mask & WR_READ_MASK) {
89 				npsn = (wqe->dma.length - wqe->dma.resid) /
90 					qp->mtu;
91 				wqe->iova += npsn * qp->mtu;
92 			}
93 		}
94 
95 		wqe->state = wqe_state_posted;
96 	}
97 }
98 
99 void rnr_nak_timer(struct timer_list *t)
100 {
101 	struct rxe_qp *qp = from_timer(qp, t, rnr_nak_timer);
102 
103 	pr_debug("%s: fired for qp#%d\n", __func__, qp_num(qp));
104 
105 	/* request a send queue retry */
106 	qp->req.need_retry = 1;
107 	qp->req.wait_for_rnr_timer = 0;
108 	rxe_run_task(&qp->req.task, 1);
109 }
110 
111 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
112 {
113 	struct rxe_send_wqe *wqe;
114 	struct rxe_queue *q = qp->sq.queue;
115 	unsigned int index = qp->req.wqe_index;
116 	unsigned int cons;
117 	unsigned int prod;
118 
119 	wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
120 	cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
121 	prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
122 
123 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
124 		/* check to see if we are drained;
125 		 * state_lock used by requester and completer
126 		 */
127 		spin_lock_bh(&qp->state_lock);
128 		do {
129 			if (qp->req.state != QP_STATE_DRAIN) {
130 				/* comp just finished */
131 				spin_unlock_bh(&qp->state_lock);
132 				break;
133 			}
134 
135 			if (wqe && ((index != cons) ||
136 				(wqe->state != wqe_state_posted))) {
137 				/* comp not done yet */
138 				spin_unlock_bh(&qp->state_lock);
139 				break;
140 			}
141 
142 			qp->req.state = QP_STATE_DRAINED;
143 			spin_unlock_bh(&qp->state_lock);
144 
145 			if (qp->ibqp.event_handler) {
146 				struct ib_event ev;
147 
148 				ev.device = qp->ibqp.device;
149 				ev.element.qp = &qp->ibqp;
150 				ev.event = IB_EVENT_SQ_DRAINED;
151 				qp->ibqp.event_handler(&ev,
152 					qp->ibqp.qp_context);
153 			}
154 		} while (0);
155 	}
156 
157 	if (index == prod)
158 		return NULL;
159 
160 	wqe = queue_addr_from_index(q, index);
161 
162 	if (unlikely((qp->req.state == QP_STATE_DRAIN ||
163 		      qp->req.state == QP_STATE_DRAINED) &&
164 		     (wqe->state != wqe_state_processing)))
165 		return NULL;
166 
167 	wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
168 	return wqe;
169 }
170 
171 /**
172  * rxe_wqe_is_fenced - check if next wqe is fenced
173  * @qp: the queue pair
174  * @wqe: the next wqe
175  *
176  * Returns: 1 if wqe needs to wait
177  *	    0 if wqe is ready to go
178  */
179 static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
180 {
181 	/* Local invalidate fence (LIF) see IBA 10.6.5.1
182 	 * Requires ALL previous operations on the send queue
183 	 * are complete. Make mandatory for the rxe driver.
184 	 */
185 	if (wqe->wr.opcode == IB_WR_LOCAL_INV)
186 		return qp->req.wqe_index != queue_get_consumer(qp->sq.queue,
187 						QUEUE_TYPE_FROM_CLIENT);
188 
189 	/* Fence see IBA 10.8.3.3
190 	 * Requires that all previous read and atomic operations
191 	 * are complete.
192 	 */
193 	return (wqe->wr.send_flags & IB_SEND_FENCE) &&
194 		atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic;
195 }
196 
197 static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
198 {
199 	switch (opcode) {
200 	case IB_WR_RDMA_WRITE:
201 		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
202 		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
203 			return fits ?
204 				IB_OPCODE_RC_RDMA_WRITE_LAST :
205 				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
206 		else
207 			return fits ?
208 				IB_OPCODE_RC_RDMA_WRITE_ONLY :
209 				IB_OPCODE_RC_RDMA_WRITE_FIRST;
210 
211 	case IB_WR_RDMA_WRITE_WITH_IMM:
212 		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
213 		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
214 			return fits ?
215 				IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
216 				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
217 		else
218 			return fits ?
219 				IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
220 				IB_OPCODE_RC_RDMA_WRITE_FIRST;
221 
222 	case IB_WR_SEND:
223 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
224 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
225 			return fits ?
226 				IB_OPCODE_RC_SEND_LAST :
227 				IB_OPCODE_RC_SEND_MIDDLE;
228 		else
229 			return fits ?
230 				IB_OPCODE_RC_SEND_ONLY :
231 				IB_OPCODE_RC_SEND_FIRST;
232 
233 	case IB_WR_SEND_WITH_IMM:
234 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
235 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
236 			return fits ?
237 				IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
238 				IB_OPCODE_RC_SEND_MIDDLE;
239 		else
240 			return fits ?
241 				IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
242 				IB_OPCODE_RC_SEND_FIRST;
243 
244 	case IB_WR_RDMA_READ:
245 		return IB_OPCODE_RC_RDMA_READ_REQUEST;
246 
247 	case IB_WR_ATOMIC_CMP_AND_SWP:
248 		return IB_OPCODE_RC_COMPARE_SWAP;
249 
250 	case IB_WR_ATOMIC_FETCH_AND_ADD:
251 		return IB_OPCODE_RC_FETCH_ADD;
252 
253 	case IB_WR_SEND_WITH_INV:
254 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
255 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
256 			return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
257 				IB_OPCODE_RC_SEND_MIDDLE;
258 		else
259 			return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
260 				IB_OPCODE_RC_SEND_FIRST;
261 	case IB_WR_REG_MR:
262 	case IB_WR_LOCAL_INV:
263 		return opcode;
264 	}
265 
266 	return -EINVAL;
267 }
268 
269 static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits)
270 {
271 	switch (opcode) {
272 	case IB_WR_RDMA_WRITE:
273 		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
274 		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
275 			return fits ?
276 				IB_OPCODE_UC_RDMA_WRITE_LAST :
277 				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
278 		else
279 			return fits ?
280 				IB_OPCODE_UC_RDMA_WRITE_ONLY :
281 				IB_OPCODE_UC_RDMA_WRITE_FIRST;
282 
283 	case IB_WR_RDMA_WRITE_WITH_IMM:
284 		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
285 		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
286 			return fits ?
287 				IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
288 				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
289 		else
290 			return fits ?
291 				IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
292 				IB_OPCODE_UC_RDMA_WRITE_FIRST;
293 
294 	case IB_WR_SEND:
295 		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
296 		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
297 			return fits ?
298 				IB_OPCODE_UC_SEND_LAST :
299 				IB_OPCODE_UC_SEND_MIDDLE;
300 		else
301 			return fits ?
302 				IB_OPCODE_UC_SEND_ONLY :
303 				IB_OPCODE_UC_SEND_FIRST;
304 
305 	case IB_WR_SEND_WITH_IMM:
306 		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
307 		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
308 			return fits ?
309 				IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
310 				IB_OPCODE_UC_SEND_MIDDLE;
311 		else
312 			return fits ?
313 				IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
314 				IB_OPCODE_UC_SEND_FIRST;
315 	}
316 
317 	return -EINVAL;
318 }
319 
320 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
321 		       u32 opcode)
322 {
323 	int fits = (wqe->dma.resid <= qp->mtu);
324 
325 	switch (qp_type(qp)) {
326 	case IB_QPT_RC:
327 		return next_opcode_rc(qp, opcode, fits);
328 
329 	case IB_QPT_UC:
330 		return next_opcode_uc(qp, opcode, fits);
331 
332 	case IB_QPT_UD:
333 	case IB_QPT_GSI:
334 		switch (opcode) {
335 		case IB_WR_SEND:
336 			return IB_OPCODE_UD_SEND_ONLY;
337 
338 		case IB_WR_SEND_WITH_IMM:
339 			return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
340 		}
341 		break;
342 
343 	default:
344 		break;
345 	}
346 
347 	return -EINVAL;
348 }
349 
350 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
351 {
352 	int depth;
353 
354 	if (wqe->has_rd_atomic)
355 		return 0;
356 
357 	qp->req.need_rd_atomic = 1;
358 	depth = atomic_dec_return(&qp->req.rd_atomic);
359 
360 	if (depth >= 0) {
361 		qp->req.need_rd_atomic = 0;
362 		wqe->has_rd_atomic = 1;
363 		return 0;
364 	}
365 
366 	atomic_inc(&qp->req.rd_atomic);
367 	return -EAGAIN;
368 }
369 
370 static inline int get_mtu(struct rxe_qp *qp)
371 {
372 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
373 
374 	if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
375 		return qp->mtu;
376 
377 	return rxe->port.mtu_cap;
378 }
379 
380 static struct sk_buff *init_req_packet(struct rxe_qp *qp,
381 				       struct rxe_av *av,
382 				       struct rxe_send_wqe *wqe,
383 				       int opcode, u32 payload,
384 				       struct rxe_pkt_info *pkt)
385 {
386 	struct rxe_dev		*rxe = to_rdev(qp->ibqp.device);
387 	struct sk_buff		*skb;
388 	struct rxe_send_wr	*ibwr = &wqe->wr;
389 	int			pad = (-payload) & 0x3;
390 	int			paylen;
391 	int			solicited;
392 	u32			qp_num;
393 	int			ack_req;
394 
395 	/* length from start of bth to end of icrc */
396 	paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
397 	pkt->paylen = paylen;
398 
399 	/* init skb */
400 	skb = rxe_init_packet(rxe, av, paylen, pkt);
401 	if (unlikely(!skb))
402 		return NULL;
403 
404 	/* init bth */
405 	solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
406 			(pkt->mask & RXE_END_MASK) &&
407 			((pkt->mask & (RXE_SEND_MASK)) ||
408 			(pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
409 			(RXE_WRITE_MASK | RXE_IMMDT_MASK));
410 
411 	qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
412 					 qp->attr.dest_qp_num;
413 
414 	ack_req = ((pkt->mask & RXE_END_MASK) ||
415 		(qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
416 	if (ack_req)
417 		qp->req.noack_pkts = 0;
418 
419 	bth_init(pkt, pkt->opcode, solicited, 0, pad, IB_DEFAULT_PKEY_FULL, qp_num,
420 		 ack_req, pkt->psn);
421 
422 	/* init optional headers */
423 	if (pkt->mask & RXE_RETH_MASK) {
424 		reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
425 		reth_set_va(pkt, wqe->iova);
426 		reth_set_len(pkt, wqe->dma.resid);
427 	}
428 
429 	if (pkt->mask & RXE_IMMDT_MASK)
430 		immdt_set_imm(pkt, ibwr->ex.imm_data);
431 
432 	if (pkt->mask & RXE_IETH_MASK)
433 		ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
434 
435 	if (pkt->mask & RXE_ATMETH_MASK) {
436 		atmeth_set_va(pkt, wqe->iova);
437 		if (opcode == IB_OPCODE_RC_COMPARE_SWAP) {
438 			atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
439 			atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
440 		} else {
441 			atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
442 		}
443 		atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
444 	}
445 
446 	if (pkt->mask & RXE_DETH_MASK) {
447 		if (qp->ibqp.qp_num == 1)
448 			deth_set_qkey(pkt, GSI_QKEY);
449 		else
450 			deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
451 		deth_set_sqp(pkt, qp->ibqp.qp_num);
452 	}
453 
454 	return skb;
455 }
456 
457 static int finish_packet(struct rxe_qp *qp, struct rxe_av *av,
458 			 struct rxe_send_wqe *wqe, struct rxe_pkt_info *pkt,
459 			 struct sk_buff *skb, u32 payload)
460 {
461 	int err;
462 
463 	err = rxe_prepare(av, pkt, skb);
464 	if (err)
465 		return err;
466 
467 	if (pkt->mask & RXE_WRITE_OR_SEND_MASK) {
468 		if (wqe->wr.send_flags & IB_SEND_INLINE) {
469 			u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
470 
471 			memcpy(payload_addr(pkt), tmp, payload);
472 
473 			wqe->dma.resid -= payload;
474 			wqe->dma.sge_offset += payload;
475 		} else {
476 			err = copy_data(qp->pd, 0, &wqe->dma,
477 					payload_addr(pkt), payload,
478 					RXE_FROM_MR_OBJ);
479 			if (err)
480 				return err;
481 		}
482 		if (bth_pad(pkt)) {
483 			u8 *pad = payload_addr(pkt) + payload;
484 
485 			memset(pad, 0, bth_pad(pkt));
486 		}
487 	}
488 
489 	return 0;
490 }
491 
492 static void update_wqe_state(struct rxe_qp *qp,
493 		struct rxe_send_wqe *wqe,
494 		struct rxe_pkt_info *pkt)
495 {
496 	if (pkt->mask & RXE_END_MASK) {
497 		if (qp_type(qp) == IB_QPT_RC)
498 			wqe->state = wqe_state_pending;
499 	} else {
500 		wqe->state = wqe_state_processing;
501 	}
502 }
503 
504 static void update_wqe_psn(struct rxe_qp *qp,
505 			   struct rxe_send_wqe *wqe,
506 			   struct rxe_pkt_info *pkt,
507 			   u32 payload)
508 {
509 	/* number of packets left to send including current one */
510 	int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
511 
512 	/* handle zero length packet case */
513 	if (num_pkt == 0)
514 		num_pkt = 1;
515 
516 	if (pkt->mask & RXE_START_MASK) {
517 		wqe->first_psn = qp->req.psn;
518 		wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
519 	}
520 
521 	if (pkt->mask & RXE_READ_MASK)
522 		qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
523 	else
524 		qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
525 }
526 
527 static void save_state(struct rxe_send_wqe *wqe,
528 		       struct rxe_qp *qp,
529 		       struct rxe_send_wqe *rollback_wqe,
530 		       u32 *rollback_psn)
531 {
532 	rollback_wqe->state     = wqe->state;
533 	rollback_wqe->first_psn = wqe->first_psn;
534 	rollback_wqe->last_psn  = wqe->last_psn;
535 	*rollback_psn		= qp->req.psn;
536 }
537 
538 static void rollback_state(struct rxe_send_wqe *wqe,
539 			   struct rxe_qp *qp,
540 			   struct rxe_send_wqe *rollback_wqe,
541 			   u32 rollback_psn)
542 {
543 	wqe->state     = rollback_wqe->state;
544 	wqe->first_psn = rollback_wqe->first_psn;
545 	wqe->last_psn  = rollback_wqe->last_psn;
546 	qp->req.psn    = rollback_psn;
547 }
548 
549 static void update_state(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
550 {
551 	qp->req.opcode = pkt->opcode;
552 
553 	if (pkt->mask & RXE_END_MASK)
554 		qp->req.wqe_index = queue_next_index(qp->sq.queue,
555 						     qp->req.wqe_index);
556 
557 	qp->need_req_skb = 0;
558 
559 	if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
560 		mod_timer(&qp->retrans_timer,
561 			  jiffies + qp->qp_timeout_jiffies);
562 }
563 
564 static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
565 {
566 	u8 opcode = wqe->wr.opcode;
567 	u32 rkey;
568 	int ret;
569 
570 	switch (opcode) {
571 	case IB_WR_LOCAL_INV:
572 		rkey = wqe->wr.ex.invalidate_rkey;
573 		if (rkey_is_mw(rkey))
574 			ret = rxe_invalidate_mw(qp, rkey);
575 		else
576 			ret = rxe_invalidate_mr(qp, rkey);
577 
578 		if (unlikely(ret)) {
579 			wqe->status = IB_WC_LOC_QP_OP_ERR;
580 			return ret;
581 		}
582 		break;
583 	case IB_WR_REG_MR:
584 		ret = rxe_reg_fast_mr(qp, wqe);
585 		if (unlikely(ret)) {
586 			wqe->status = IB_WC_LOC_QP_OP_ERR;
587 			return ret;
588 		}
589 		break;
590 	case IB_WR_BIND_MW:
591 		ret = rxe_bind_mw(qp, wqe);
592 		if (unlikely(ret)) {
593 			wqe->status = IB_WC_MW_BIND_ERR;
594 			return ret;
595 		}
596 		break;
597 	default:
598 		pr_err("Unexpected send wqe opcode %d\n", opcode);
599 		wqe->status = IB_WC_LOC_QP_OP_ERR;
600 		return -EINVAL;
601 	}
602 
603 	wqe->state = wqe_state_done;
604 	wqe->status = IB_WC_SUCCESS;
605 	qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
606 
607 	/* There is no ack coming for local work requests
608 	 * which can lead to a deadlock. So go ahead and complete
609 	 * it now.
610 	 */
611 	rxe_run_task(&qp->comp.task, 1);
612 
613 	return 0;
614 }
615 
616 int rxe_requester(void *arg)
617 {
618 	struct rxe_qp *qp = (struct rxe_qp *)arg;
619 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
620 	struct rxe_pkt_info pkt;
621 	struct sk_buff *skb;
622 	struct rxe_send_wqe *wqe;
623 	enum rxe_hdr_mask mask;
624 	u32 payload;
625 	int mtu;
626 	int opcode;
627 	int err;
628 	int ret;
629 	struct rxe_send_wqe rollback_wqe;
630 	u32 rollback_psn;
631 	struct rxe_queue *q = qp->sq.queue;
632 	struct rxe_ah *ah;
633 	struct rxe_av *av;
634 
635 	if (!rxe_get(qp))
636 		return -EAGAIN;
637 
638 	if (unlikely(!qp->valid))
639 		goto exit;
640 
641 	if (unlikely(qp->req.state == QP_STATE_ERROR)) {
642 		wqe = req_next_wqe(qp);
643 		if (wqe)
644 			/*
645 			 * Generate an error completion for error qp state
646 			 */
647 			goto err;
648 		else
649 			goto exit;
650 	}
651 
652 	if (unlikely(qp->req.state == QP_STATE_RESET)) {
653 		qp->req.wqe_index = queue_get_consumer(q,
654 						QUEUE_TYPE_FROM_CLIENT);
655 		qp->req.opcode = -1;
656 		qp->req.need_rd_atomic = 0;
657 		qp->req.wait_psn = 0;
658 		qp->req.need_retry = 0;
659 		qp->req.wait_for_rnr_timer = 0;
660 		goto exit;
661 	}
662 
663 	/* we come here if the retransmit timer has fired
664 	 * or if the rnr timer has fired. If the retransmit
665 	 * timer fires while we are processing an RNR NAK wait
666 	 * until the rnr timer has fired before starting the
667 	 * retry flow
668 	 */
669 	if (unlikely(qp->req.need_retry && !qp->req.wait_for_rnr_timer)) {
670 		req_retry(qp);
671 		qp->req.need_retry = 0;
672 	}
673 
674 	wqe = req_next_wqe(qp);
675 	if (unlikely(!wqe))
676 		goto exit;
677 
678 	if (rxe_wqe_is_fenced(qp, wqe)) {
679 		qp->req.wait_fence = 1;
680 		goto exit;
681 	}
682 
683 	if (wqe->mask & WR_LOCAL_OP_MASK) {
684 		err = rxe_do_local_ops(qp, wqe);
685 		if (unlikely(err))
686 			goto err;
687 		else
688 			goto done;
689 	}
690 
691 	if (unlikely(qp_type(qp) == IB_QPT_RC &&
692 		psn_compare(qp->req.psn, (qp->comp.psn +
693 				RXE_MAX_UNACKED_PSNS)) > 0)) {
694 		qp->req.wait_psn = 1;
695 		goto exit;
696 	}
697 
698 	/* Limit the number of inflight SKBs per QP */
699 	if (unlikely(atomic_read(&qp->skb_out) >
700 		     RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
701 		qp->need_req_skb = 1;
702 		goto exit;
703 	}
704 
705 	opcode = next_opcode(qp, wqe, wqe->wr.opcode);
706 	if (unlikely(opcode < 0)) {
707 		wqe->status = IB_WC_LOC_QP_OP_ERR;
708 		goto err;
709 	}
710 
711 	mask = rxe_opcode[opcode].mask;
712 	if (unlikely(mask & RXE_READ_OR_ATOMIC_MASK)) {
713 		if (check_init_depth(qp, wqe))
714 			goto exit;
715 	}
716 
717 	mtu = get_mtu(qp);
718 	payload = (mask & RXE_WRITE_OR_SEND_MASK) ? wqe->dma.resid : 0;
719 	if (payload > mtu) {
720 		if (qp_type(qp) == IB_QPT_UD) {
721 			/* C10-93.1.1: If the total sum of all the buffer lengths specified for a
722 			 * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
723 			 * shall not emit any packets for this message. Further, the CI shall not
724 			 * generate an error due to this condition.
725 			 */
726 
727 			/* fake a successful UD send */
728 			wqe->first_psn = qp->req.psn;
729 			wqe->last_psn = qp->req.psn;
730 			qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
731 			qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
732 			qp->req.wqe_index = queue_next_index(qp->sq.queue,
733 						       qp->req.wqe_index);
734 			wqe->state = wqe_state_done;
735 			wqe->status = IB_WC_SUCCESS;
736 			rxe_run_task(&qp->comp.task, 0);
737 			goto done;
738 		}
739 		payload = mtu;
740 	}
741 
742 	pkt.rxe = rxe;
743 	pkt.opcode = opcode;
744 	pkt.qp = qp;
745 	pkt.psn = qp->req.psn;
746 	pkt.mask = rxe_opcode[opcode].mask;
747 	pkt.wqe = wqe;
748 
749 	av = rxe_get_av(&pkt, &ah);
750 	if (unlikely(!av)) {
751 		pr_err("qp#%d Failed no address vector\n", qp_num(qp));
752 		wqe->status = IB_WC_LOC_QP_OP_ERR;
753 		goto err;
754 	}
755 
756 	skb = init_req_packet(qp, av, wqe, opcode, payload, &pkt);
757 	if (unlikely(!skb)) {
758 		pr_err("qp#%d Failed allocating skb\n", qp_num(qp));
759 		wqe->status = IB_WC_LOC_QP_OP_ERR;
760 		if (ah)
761 			rxe_put(ah);
762 		goto err;
763 	}
764 
765 	err = finish_packet(qp, av, wqe, &pkt, skb, payload);
766 	if (unlikely(err)) {
767 		pr_debug("qp#%d Error during finish packet\n", qp_num(qp));
768 		if (err == -EFAULT)
769 			wqe->status = IB_WC_LOC_PROT_ERR;
770 		else
771 			wqe->status = IB_WC_LOC_QP_OP_ERR;
772 		kfree_skb(skb);
773 		if (ah)
774 			rxe_put(ah);
775 		goto err;
776 	}
777 
778 	if (ah)
779 		rxe_put(ah);
780 
781 	/*
782 	 * To prevent a race on wqe access between requester and completer,
783 	 * wqe members state and psn need to be set before calling
784 	 * rxe_xmit_packet().
785 	 * Otherwise, completer might initiate an unjustified retry flow.
786 	 */
787 	save_state(wqe, qp, &rollback_wqe, &rollback_psn);
788 	update_wqe_state(qp, wqe, &pkt);
789 	update_wqe_psn(qp, wqe, &pkt, payload);
790 
791 	err = rxe_xmit_packet(qp, &pkt, skb);
792 	if (err) {
793 		qp->need_req_skb = 1;
794 
795 		rollback_state(wqe, qp, &rollback_wqe, rollback_psn);
796 
797 		if (err == -EAGAIN) {
798 			rxe_run_task(&qp->req.task, 1);
799 			goto exit;
800 		}
801 
802 		wqe->status = IB_WC_LOC_QP_OP_ERR;
803 		goto err;
804 	}
805 
806 	update_state(qp, &pkt);
807 
808 	/* A non-zero return value will cause rxe_do_task to
809 	 * exit its loop and end the tasklet. A zero return
810 	 * will continue looping and return to rxe_requester
811 	 */
812 done:
813 	ret = 0;
814 	goto out;
815 err:
816 	/* update wqe_index for each wqe completion */
817 	qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
818 	wqe->state = wqe_state_error;
819 	qp->req.state = QP_STATE_ERROR;
820 	rxe_run_task(&qp->comp.task, 0);
821 exit:
822 	ret = -EAGAIN;
823 out:
824 	rxe_put(qp);
825 
826 	return ret;
827 }
828