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 #include <crypto/hash.h>
36 
37 #include "rxe.h"
38 #include "rxe_loc.h"
39 #include "rxe_queue.h"
40 
41 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
42 		       u32 opcode);
43 
44 static inline void retry_first_write_send(struct rxe_qp *qp,
45 					  struct rxe_send_wqe *wqe,
46 					  unsigned int mask, int npsn)
47 {
48 	int i;
49 
50 	for (i = 0; i < npsn; i++) {
51 		int to_send = (wqe->dma.resid > qp->mtu) ?
52 				qp->mtu : wqe->dma.resid;
53 
54 		qp->req.opcode = next_opcode(qp, wqe,
55 					     wqe->wr.opcode);
56 
57 		if (wqe->wr.send_flags & IB_SEND_INLINE) {
58 			wqe->dma.resid -= to_send;
59 			wqe->dma.sge_offset += to_send;
60 		} else {
61 			advance_dma_data(&wqe->dma, to_send);
62 		}
63 		if (mask & WR_WRITE_MASK)
64 			wqe->iova += qp->mtu;
65 	}
66 }
67 
68 static void req_retry(struct rxe_qp *qp)
69 {
70 	struct rxe_send_wqe *wqe;
71 	unsigned int wqe_index;
72 	unsigned int mask;
73 	int npsn;
74 	int first = 1;
75 
76 	qp->req.wqe_index	= consumer_index(qp->sq.queue);
77 	qp->req.psn		= qp->comp.psn;
78 	qp->req.opcode		= -1;
79 
80 	for (wqe_index = consumer_index(qp->sq.queue);
81 		wqe_index != producer_index(qp->sq.queue);
82 		wqe_index = next_index(qp->sq.queue, wqe_index)) {
83 		wqe = addr_from_index(qp->sq.queue, wqe_index);
84 		mask = wr_opcode_mask(wqe->wr.opcode, qp);
85 
86 		if (wqe->state == wqe_state_posted)
87 			break;
88 
89 		if (wqe->state == wqe_state_done)
90 			continue;
91 
92 		wqe->iova = (mask & WR_ATOMIC_MASK) ?
93 			     wqe->wr.wr.atomic.remote_addr :
94 			     (mask & WR_READ_OR_WRITE_MASK) ?
95 			     wqe->wr.wr.rdma.remote_addr :
96 			     0;
97 
98 		if (!first || (mask & WR_READ_MASK) == 0) {
99 			wqe->dma.resid = wqe->dma.length;
100 			wqe->dma.cur_sge = 0;
101 			wqe->dma.sge_offset = 0;
102 		}
103 
104 		if (first) {
105 			first = 0;
106 
107 			if (mask & WR_WRITE_OR_SEND_MASK) {
108 				npsn = (qp->comp.psn - wqe->first_psn) &
109 					BTH_PSN_MASK;
110 				retry_first_write_send(qp, wqe, mask, npsn);
111 			}
112 
113 			if (mask & WR_READ_MASK) {
114 				npsn = (wqe->dma.length - wqe->dma.resid) /
115 					qp->mtu;
116 				wqe->iova += npsn * qp->mtu;
117 			}
118 		}
119 
120 		wqe->state = wqe_state_posted;
121 	}
122 }
123 
124 void rnr_nak_timer(struct timer_list *t)
125 {
126 	struct rxe_qp *qp = from_timer(qp, t, rnr_nak_timer);
127 
128 	pr_debug("qp#%d rnr nak timer fired\n", qp_num(qp));
129 	rxe_run_task(&qp->req.task, 1);
130 }
131 
132 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
133 {
134 	struct rxe_send_wqe *wqe = queue_head(qp->sq.queue);
135 	unsigned long flags;
136 
137 	if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
138 		/* check to see if we are drained;
139 		 * state_lock used by requester and completer
140 		 */
141 		spin_lock_irqsave(&qp->state_lock, flags);
142 		do {
143 			if (qp->req.state != QP_STATE_DRAIN) {
144 				/* comp just finished */
145 				spin_unlock_irqrestore(&qp->state_lock,
146 						       flags);
147 				break;
148 			}
149 
150 			if (wqe && ((qp->req.wqe_index !=
151 				consumer_index(qp->sq.queue)) ||
152 				(wqe->state != wqe_state_posted))) {
153 				/* comp not done yet */
154 				spin_unlock_irqrestore(&qp->state_lock,
155 						       flags);
156 				break;
157 			}
158 
159 			qp->req.state = QP_STATE_DRAINED;
160 			spin_unlock_irqrestore(&qp->state_lock, flags);
161 
162 			if (qp->ibqp.event_handler) {
163 				struct ib_event ev;
164 
165 				ev.device = qp->ibqp.device;
166 				ev.element.qp = &qp->ibqp;
167 				ev.event = IB_EVENT_SQ_DRAINED;
168 				qp->ibqp.event_handler(&ev,
169 					qp->ibqp.qp_context);
170 			}
171 		} while (0);
172 	}
173 
174 	if (qp->req.wqe_index == producer_index(qp->sq.queue))
175 		return NULL;
176 
177 	wqe = addr_from_index(qp->sq.queue, qp->req.wqe_index);
178 
179 	if (unlikely((qp->req.state == QP_STATE_DRAIN ||
180 		      qp->req.state == QP_STATE_DRAINED) &&
181 		     (wqe->state != wqe_state_processing)))
182 		return NULL;
183 
184 	if (unlikely((wqe->wr.send_flags & IB_SEND_FENCE) &&
185 		     (qp->req.wqe_index != consumer_index(qp->sq.queue)))) {
186 		qp->req.wait_fence = 1;
187 		return NULL;
188 	}
189 
190 	wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
191 	return wqe;
192 }
193 
194 static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
195 {
196 	switch (opcode) {
197 	case IB_WR_RDMA_WRITE:
198 		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
199 		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
200 			return fits ?
201 				IB_OPCODE_RC_RDMA_WRITE_LAST :
202 				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
203 		else
204 			return fits ?
205 				IB_OPCODE_RC_RDMA_WRITE_ONLY :
206 				IB_OPCODE_RC_RDMA_WRITE_FIRST;
207 
208 	case IB_WR_RDMA_WRITE_WITH_IMM:
209 		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
210 		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
211 			return fits ?
212 				IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
213 				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
214 		else
215 			return fits ?
216 				IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
217 				IB_OPCODE_RC_RDMA_WRITE_FIRST;
218 
219 	case IB_WR_SEND:
220 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
221 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
222 			return fits ?
223 				IB_OPCODE_RC_SEND_LAST :
224 				IB_OPCODE_RC_SEND_MIDDLE;
225 		else
226 			return fits ?
227 				IB_OPCODE_RC_SEND_ONLY :
228 				IB_OPCODE_RC_SEND_FIRST;
229 
230 	case IB_WR_SEND_WITH_IMM:
231 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
232 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
233 			return fits ?
234 				IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
235 				IB_OPCODE_RC_SEND_MIDDLE;
236 		else
237 			return fits ?
238 				IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
239 				IB_OPCODE_RC_SEND_FIRST;
240 
241 	case IB_WR_RDMA_READ:
242 		return IB_OPCODE_RC_RDMA_READ_REQUEST;
243 
244 	case IB_WR_ATOMIC_CMP_AND_SWP:
245 		return IB_OPCODE_RC_COMPARE_SWAP;
246 
247 	case IB_WR_ATOMIC_FETCH_AND_ADD:
248 		return IB_OPCODE_RC_FETCH_ADD;
249 
250 	case IB_WR_SEND_WITH_INV:
251 		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
252 		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
253 			return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
254 				IB_OPCODE_RC_SEND_MIDDLE;
255 		else
256 			return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
257 				IB_OPCODE_RC_SEND_FIRST;
258 	case IB_WR_REG_MR:
259 	case IB_WR_LOCAL_INV:
260 		return opcode;
261 	}
262 
263 	return -EINVAL;
264 }
265 
266 static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits)
267 {
268 	switch (opcode) {
269 	case IB_WR_RDMA_WRITE:
270 		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
271 		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
272 			return fits ?
273 				IB_OPCODE_UC_RDMA_WRITE_LAST :
274 				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
275 		else
276 			return fits ?
277 				IB_OPCODE_UC_RDMA_WRITE_ONLY :
278 				IB_OPCODE_UC_RDMA_WRITE_FIRST;
279 
280 	case IB_WR_RDMA_WRITE_WITH_IMM:
281 		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
282 		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
283 			return fits ?
284 				IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
285 				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
286 		else
287 			return fits ?
288 				IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
289 				IB_OPCODE_UC_RDMA_WRITE_FIRST;
290 
291 	case IB_WR_SEND:
292 		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
293 		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
294 			return fits ?
295 				IB_OPCODE_UC_SEND_LAST :
296 				IB_OPCODE_UC_SEND_MIDDLE;
297 		else
298 			return fits ?
299 				IB_OPCODE_UC_SEND_ONLY :
300 				IB_OPCODE_UC_SEND_FIRST;
301 
302 	case IB_WR_SEND_WITH_IMM:
303 		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
304 		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
305 			return fits ?
306 				IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
307 				IB_OPCODE_UC_SEND_MIDDLE;
308 		else
309 			return fits ?
310 				IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
311 				IB_OPCODE_UC_SEND_FIRST;
312 	}
313 
314 	return -EINVAL;
315 }
316 
317 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
318 		       u32 opcode)
319 {
320 	int fits = (wqe->dma.resid <= qp->mtu);
321 
322 	switch (qp_type(qp)) {
323 	case IB_QPT_RC:
324 		return next_opcode_rc(qp, opcode, fits);
325 
326 	case IB_QPT_UC:
327 		return next_opcode_uc(qp, opcode, fits);
328 
329 	case IB_QPT_SMI:
330 	case IB_QPT_UD:
331 	case IB_QPT_GSI:
332 		switch (opcode) {
333 		case IB_WR_SEND:
334 			return IB_OPCODE_UD_SEND_ONLY;
335 
336 		case IB_WR_SEND_WITH_IMM:
337 			return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
338 		}
339 		break;
340 
341 	default:
342 		break;
343 	}
344 
345 	return -EINVAL;
346 }
347 
348 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
349 {
350 	int depth;
351 
352 	if (wqe->has_rd_atomic)
353 		return 0;
354 
355 	qp->req.need_rd_atomic = 1;
356 	depth = atomic_dec_return(&qp->req.rd_atomic);
357 
358 	if (depth >= 0) {
359 		qp->req.need_rd_atomic = 0;
360 		wqe->has_rd_atomic = 1;
361 		return 0;
362 	}
363 
364 	atomic_inc(&qp->req.rd_atomic);
365 	return -EAGAIN;
366 }
367 
368 static inline int get_mtu(struct rxe_qp *qp)
369 {
370 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
371 
372 	if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
373 		return qp->mtu;
374 
375 	return rxe->port.mtu_cap;
376 }
377 
378 static struct sk_buff *init_req_packet(struct rxe_qp *qp,
379 				       struct rxe_send_wqe *wqe,
380 				       int opcode, int payload,
381 				       struct rxe_pkt_info *pkt)
382 {
383 	struct rxe_dev		*rxe = to_rdev(qp->ibqp.device);
384 	struct rxe_port		*port = &rxe->port;
385 	struct sk_buff		*skb;
386 	struct rxe_send_wr	*ibwr = &wqe->wr;
387 	struct rxe_av		*av;
388 	int			pad = (-payload) & 0x3;
389 	int			paylen;
390 	int			solicited;
391 	u16			pkey;
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 
398 	/* pkt->hdr, rxe, port_num and mask are initialized in ifc
399 	 * layer
400 	 */
401 	pkt->opcode	= opcode;
402 	pkt->qp		= qp;
403 	pkt->psn	= qp->req.psn;
404 	pkt->mask	= rxe_opcode[opcode].mask;
405 	pkt->paylen	= paylen;
406 	pkt->offset	= 0;
407 	pkt->wqe	= wqe;
408 
409 	/* init skb */
410 	av = rxe_get_av(pkt);
411 	skb = rxe_init_packet(rxe, av, paylen, pkt);
412 	if (unlikely(!skb))
413 		return NULL;
414 
415 	/* init bth */
416 	solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
417 			(pkt->mask & RXE_END_MASK) &&
418 			((pkt->mask & (RXE_SEND_MASK)) ||
419 			(pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
420 			(RXE_WRITE_MASK | RXE_IMMDT_MASK));
421 
422 	pkey = (qp_type(qp) == IB_QPT_GSI) ?
423 		 port->pkey_tbl[ibwr->wr.ud.pkey_index] :
424 		 port->pkey_tbl[qp->attr.pkey_index];
425 
426 	qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
427 					 qp->attr.dest_qp_num;
428 
429 	ack_req = ((pkt->mask & RXE_END_MASK) ||
430 		(qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
431 	if (ack_req)
432 		qp->req.noack_pkts = 0;
433 
434 	bth_init(pkt, pkt->opcode, solicited, 0, pad, pkey, qp_num,
435 		 ack_req, pkt->psn);
436 
437 	/* init optional headers */
438 	if (pkt->mask & RXE_RETH_MASK) {
439 		reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
440 		reth_set_va(pkt, wqe->iova);
441 		reth_set_len(pkt, wqe->dma.resid);
442 	}
443 
444 	if (pkt->mask & RXE_IMMDT_MASK)
445 		immdt_set_imm(pkt, ibwr->ex.imm_data);
446 
447 	if (pkt->mask & RXE_IETH_MASK)
448 		ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
449 
450 	if (pkt->mask & RXE_ATMETH_MASK) {
451 		atmeth_set_va(pkt, wqe->iova);
452 		if (opcode == IB_OPCODE_RC_COMPARE_SWAP ||
453 		    opcode == IB_OPCODE_RD_COMPARE_SWAP) {
454 			atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
455 			atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
456 		} else {
457 			atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
458 		}
459 		atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
460 	}
461 
462 	if (pkt->mask & RXE_DETH_MASK) {
463 		if (qp->ibqp.qp_num == 1)
464 			deth_set_qkey(pkt, GSI_QKEY);
465 		else
466 			deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
467 		deth_set_sqp(pkt, qp->ibqp.qp_num);
468 	}
469 
470 	return skb;
471 }
472 
473 static int fill_packet(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
474 		       struct rxe_pkt_info *pkt, struct sk_buff *skb,
475 		       int paylen)
476 {
477 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
478 	u32 crc = 0;
479 	u32 *p;
480 	int err;
481 
482 	err = rxe_prepare(pkt, skb, &crc);
483 	if (err)
484 		return err;
485 
486 	if (pkt->mask & RXE_WRITE_OR_SEND) {
487 		if (wqe->wr.send_flags & IB_SEND_INLINE) {
488 			u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
489 
490 			crc = rxe_crc32(rxe, crc, tmp, paylen);
491 			memcpy(payload_addr(pkt), tmp, paylen);
492 
493 			wqe->dma.resid -= paylen;
494 			wqe->dma.sge_offset += paylen;
495 		} else {
496 			err = copy_data(qp->pd, 0, &wqe->dma,
497 					payload_addr(pkt), paylen,
498 					from_mem_obj,
499 					&crc);
500 			if (err)
501 				return err;
502 		}
503 	}
504 	p = payload_addr(pkt) + paylen + bth_pad(pkt);
505 
506 	*p = ~crc;
507 
508 	return 0;
509 }
510 
511 static void update_wqe_state(struct rxe_qp *qp,
512 		struct rxe_send_wqe *wqe,
513 		struct rxe_pkt_info *pkt)
514 {
515 	if (pkt->mask & RXE_END_MASK) {
516 		if (qp_type(qp) == IB_QPT_RC)
517 			wqe->state = wqe_state_pending;
518 	} else {
519 		wqe->state = wqe_state_processing;
520 	}
521 }
522 
523 static void update_wqe_psn(struct rxe_qp *qp,
524 			   struct rxe_send_wqe *wqe,
525 			   struct rxe_pkt_info *pkt,
526 			   int payload)
527 {
528 	/* number of packets left to send including current one */
529 	int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
530 
531 	/* handle zero length packet case */
532 	if (num_pkt == 0)
533 		num_pkt = 1;
534 
535 	if (pkt->mask & RXE_START_MASK) {
536 		wqe->first_psn = qp->req.psn;
537 		wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
538 	}
539 
540 	if (pkt->mask & RXE_READ_MASK)
541 		qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
542 	else
543 		qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
544 }
545 
546 static void save_state(struct rxe_send_wqe *wqe,
547 		       struct rxe_qp *qp,
548 		       struct rxe_send_wqe *rollback_wqe,
549 		       u32 *rollback_psn)
550 {
551 	rollback_wqe->state     = wqe->state;
552 	rollback_wqe->first_psn = wqe->first_psn;
553 	rollback_wqe->last_psn  = wqe->last_psn;
554 	*rollback_psn		= qp->req.psn;
555 }
556 
557 static void rollback_state(struct rxe_send_wqe *wqe,
558 			   struct rxe_qp *qp,
559 			   struct rxe_send_wqe *rollback_wqe,
560 			   u32 rollback_psn)
561 {
562 	wqe->state     = rollback_wqe->state;
563 	wqe->first_psn = rollback_wqe->first_psn;
564 	wqe->last_psn  = rollback_wqe->last_psn;
565 	qp->req.psn    = rollback_psn;
566 }
567 
568 static void update_state(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
569 			 struct rxe_pkt_info *pkt, int payload)
570 {
571 	qp->req.opcode = pkt->opcode;
572 
573 	if (pkt->mask & RXE_END_MASK)
574 		qp->req.wqe_index = next_index(qp->sq.queue, qp->req.wqe_index);
575 
576 	qp->need_req_skb = 0;
577 
578 	if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
579 		mod_timer(&qp->retrans_timer,
580 			  jiffies + qp->qp_timeout_jiffies);
581 }
582 
583 int rxe_requester(void *arg)
584 {
585 	struct rxe_qp *qp = (struct rxe_qp *)arg;
586 	struct rxe_pkt_info pkt;
587 	struct sk_buff *skb;
588 	struct rxe_send_wqe *wqe;
589 	enum rxe_hdr_mask mask;
590 	int payload;
591 	int mtu;
592 	int opcode;
593 	int ret;
594 	struct rxe_send_wqe rollback_wqe;
595 	u32 rollback_psn;
596 
597 	rxe_add_ref(qp);
598 
599 next_wqe:
600 	if (unlikely(!qp->valid || qp->req.state == QP_STATE_ERROR))
601 		goto exit;
602 
603 	if (unlikely(qp->req.state == QP_STATE_RESET)) {
604 		qp->req.wqe_index = consumer_index(qp->sq.queue);
605 		qp->req.opcode = -1;
606 		qp->req.need_rd_atomic = 0;
607 		qp->req.wait_psn = 0;
608 		qp->req.need_retry = 0;
609 		goto exit;
610 	}
611 
612 	if (unlikely(qp->req.need_retry)) {
613 		req_retry(qp);
614 		qp->req.need_retry = 0;
615 	}
616 
617 	wqe = req_next_wqe(qp);
618 	if (unlikely(!wqe))
619 		goto exit;
620 
621 	if (wqe->mask & WR_REG_MASK) {
622 		if (wqe->wr.opcode == IB_WR_LOCAL_INV) {
623 			struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
624 			struct rxe_mem *rmr;
625 
626 			rmr = rxe_pool_get_index(&rxe->mr_pool,
627 						 wqe->wr.ex.invalidate_rkey >> 8);
628 			if (!rmr) {
629 				pr_err("No mr for key %#x\n",
630 				       wqe->wr.ex.invalidate_rkey);
631 				wqe->state = wqe_state_error;
632 				wqe->status = IB_WC_MW_BIND_ERR;
633 				goto exit;
634 			}
635 			rmr->state = RXE_MEM_STATE_FREE;
636 			rxe_drop_ref(rmr);
637 			wqe->state = wqe_state_done;
638 			wqe->status = IB_WC_SUCCESS;
639 		} else if (wqe->wr.opcode == IB_WR_REG_MR) {
640 			struct rxe_mem *rmr = to_rmr(wqe->wr.wr.reg.mr);
641 
642 			rmr->state = RXE_MEM_STATE_VALID;
643 			rmr->access = wqe->wr.wr.reg.access;
644 			rmr->lkey = wqe->wr.wr.reg.key;
645 			rmr->rkey = wqe->wr.wr.reg.key;
646 			rmr->iova = wqe->wr.wr.reg.mr->iova;
647 			wqe->state = wqe_state_done;
648 			wqe->status = IB_WC_SUCCESS;
649 		} else {
650 			goto exit;
651 		}
652 		if ((wqe->wr.send_flags & IB_SEND_SIGNALED) ||
653 		    qp->sq_sig_type == IB_SIGNAL_ALL_WR)
654 			rxe_run_task(&qp->comp.task, 1);
655 		qp->req.wqe_index = next_index(qp->sq.queue,
656 						qp->req.wqe_index);
657 		goto next_wqe;
658 	}
659 
660 	if (unlikely(qp_type(qp) == IB_QPT_RC &&
661 		     qp->req.psn > (qp->comp.psn + RXE_MAX_UNACKED_PSNS))) {
662 		qp->req.wait_psn = 1;
663 		goto exit;
664 	}
665 
666 	/* Limit the number of inflight SKBs per QP */
667 	if (unlikely(atomic_read(&qp->skb_out) >
668 		     RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
669 		qp->need_req_skb = 1;
670 		goto exit;
671 	}
672 
673 	opcode = next_opcode(qp, wqe, wqe->wr.opcode);
674 	if (unlikely(opcode < 0)) {
675 		wqe->status = IB_WC_LOC_QP_OP_ERR;
676 		goto exit;
677 	}
678 
679 	mask = rxe_opcode[opcode].mask;
680 	if (unlikely(mask & RXE_READ_OR_ATOMIC)) {
681 		if (check_init_depth(qp, wqe))
682 			goto exit;
683 	}
684 
685 	mtu = get_mtu(qp);
686 	payload = (mask & RXE_WRITE_OR_SEND) ? wqe->dma.resid : 0;
687 	if (payload > mtu) {
688 		if (qp_type(qp) == IB_QPT_UD) {
689 			/* C10-93.1.1: If the total sum of all the buffer lengths specified for a
690 			 * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
691 			 * shall not emit any packets for this message. Further, the CI shall not
692 			 * generate an error due to this condition.
693 			 */
694 
695 			/* fake a successful UD send */
696 			wqe->first_psn = qp->req.psn;
697 			wqe->last_psn = qp->req.psn;
698 			qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
699 			qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
700 			qp->req.wqe_index = next_index(qp->sq.queue,
701 						       qp->req.wqe_index);
702 			wqe->state = wqe_state_done;
703 			wqe->status = IB_WC_SUCCESS;
704 			__rxe_do_task(&qp->comp.task);
705 			rxe_drop_ref(qp);
706 			return 0;
707 		}
708 		payload = mtu;
709 	}
710 
711 	skb = init_req_packet(qp, wqe, opcode, payload, &pkt);
712 	if (unlikely(!skb)) {
713 		pr_err("qp#%d Failed allocating skb\n", qp_num(qp));
714 		goto err;
715 	}
716 
717 	if (fill_packet(qp, wqe, &pkt, skb, payload)) {
718 		pr_debug("qp#%d Error during fill packet\n", qp_num(qp));
719 		kfree_skb(skb);
720 		goto err;
721 	}
722 
723 	/*
724 	 * To prevent a race on wqe access between requester and completer,
725 	 * wqe members state and psn need to be set before calling
726 	 * rxe_xmit_packet().
727 	 * Otherwise, completer might initiate an unjustified retry flow.
728 	 */
729 	save_state(wqe, qp, &rollback_wqe, &rollback_psn);
730 	update_wqe_state(qp, wqe, &pkt);
731 	update_wqe_psn(qp, wqe, &pkt, payload);
732 	ret = rxe_xmit_packet(qp, &pkt, skb);
733 	if (ret) {
734 		qp->need_req_skb = 1;
735 
736 		rollback_state(wqe, qp, &rollback_wqe, rollback_psn);
737 
738 		if (ret == -EAGAIN) {
739 			rxe_run_task(&qp->req.task, 1);
740 			goto exit;
741 		}
742 
743 		goto err;
744 	}
745 
746 	update_state(qp, wqe, &pkt, payload);
747 
748 	goto next_wqe;
749 
750 err:
751 	wqe->status = IB_WC_LOC_PROT_ERR;
752 	wqe->state = wqe_state_error;
753 	__rxe_do_task(&qp->comp.task);
754 
755 exit:
756 	rxe_drop_ref(qp);
757 	return -EAGAIN;
758 }
759