rxe_req.c

  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
 13static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
 14		       u32 opcode);
 15
 16static inline void retry_first_write_send(struct rxe_qp *qp,
 17					  struct rxe_send_wqe *wqe, int npsn)
 18{
 19	int i;
 20
 21	for (i = 0; i < npsn; i++) {
 22		int to_send = (wqe->dma.resid > qp->mtu) ?
 23				qp->mtu : wqe->dma.resid;
 24
 25		qp->req.opcode = next_opcode(qp, wqe,
 26					     wqe->wr.opcode);
 27
 28		if (wqe->wr.send_flags & IB_SEND_INLINE) {
 29			wqe->dma.resid -= to_send;
 30			wqe->dma.sge_offset += to_send;
 31		} else {
 32			advance_dma_data(&wqe->dma, to_send);
 33		}
 34	}
 35}
 36
 37static void req_retry(struct rxe_qp *qp)
 38{
 39	struct rxe_send_wqe *wqe;
 40	unsigned int wqe_index;
 41	unsigned int mask;
 42	int npsn;
 43	int first = 1;
 44	struct rxe_queue *q = qp->sq.queue;
 45	unsigned int cons;
 46	unsigned int prod;
 47
 48	cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
 49	prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
 50
 51	qp->req.wqe_index	= cons;
 52	qp->req.psn		= qp->comp.psn;
 53	qp->req.opcode		= -1;
 54
 55	for (wqe_index = cons; wqe_index != prod;
 56			wqe_index = queue_next_index(q, wqe_index)) {
 57		wqe = queue_addr_from_index(qp->sq.queue, wqe_index);
 58		mask = wr_opcode_mask(wqe->wr.opcode, qp);
 59
 60		if (wqe->state == wqe_state_posted)
 61			break;
 62
 63		if (wqe->state == wqe_state_done)
 64			continue;
 65
 66		wqe->iova = (mask & WR_ATOMIC_MASK) ?
 67			     wqe->wr.wr.atomic.remote_addr :
 68			     (mask & WR_READ_OR_WRITE_MASK) ?
 69			     wqe->wr.wr.rdma.remote_addr :
 70			     0;
 71
 72		if (!first || (mask & WR_READ_MASK) == 0) {
 73			wqe->dma.resid = wqe->dma.length;
 74			wqe->dma.cur_sge = 0;
 75			wqe->dma.sge_offset = 0;
 76		}
 77
 78		if (first) {
 79			first = 0;
 80
 81			if (mask & WR_WRITE_OR_SEND_MASK) {
 82				npsn = (qp->comp.psn - wqe->first_psn) &
 83					BTH_PSN_MASK;
 84				retry_first_write_send(qp, wqe, npsn);
 85			}
 86
 87			if (mask & WR_READ_MASK) {
 88				npsn = (wqe->dma.length - wqe->dma.resid) /
 89					qp->mtu;
 90				wqe->iova += npsn * qp->mtu;
 91			}
 92		}
 93
 94		wqe->state = wqe_state_posted;
 95	}
 96}
 97
 98void rnr_nak_timer(struct timer_list *t)
 99{
100	struct rxe_qp *qp = timer_container_of(qp, t, rnr_nak_timer);
101	unsigned long flags;
102
103	rxe_dbg_qp(qp, "nak timer fired\n");
104
105	if (!rxe_get(qp))
106		return;
107	spin_lock_irqsave(&qp->state_lock, flags);
108	if (qp->valid) {
109		/* request a send queue retry */
110		qp->req.need_retry = 1;
111		qp->req.wait_for_rnr_timer = 0;
112		rxe_sched_task(&qp->send_task);
113	}
114	spin_unlock_irqrestore(&qp->state_lock, flags);
115	rxe_put(qp);
116}
117
118static void req_check_sq_drain_done(struct rxe_qp *qp)
119{
120	struct rxe_queue *q;
121	unsigned int index;
122	unsigned int cons;
123	struct rxe_send_wqe *wqe;
124	unsigned long flags;
125
126	spin_lock_irqsave(&qp->state_lock, flags);
127	if (qp_state(qp) == IB_QPS_SQD) {
128		q = qp->sq.queue;
129		index = qp->req.wqe_index;
130		cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
131		wqe = queue_addr_from_index(q, cons);
132
133		/* check to see if we are drained;
134		 * state_lock used by requester and completer
135		 */
136		do {
137			if (!qp->attr.sq_draining)
138				/* comp just finished */
139				break;
140
141			if (wqe && ((index != cons) ||
142				(wqe->state != wqe_state_posted)))
143				/* comp not done yet */
144				break;
145
146			qp->attr.sq_draining = 0;
147			spin_unlock_irqrestore(&qp->state_lock, flags);
148
149			if (qp->ibqp.event_handler) {
150				struct ib_event ev;
151
152				ev.device = qp->ibqp.device;
153				ev.element.qp = &qp->ibqp;
154				ev.event = IB_EVENT_SQ_DRAINED;
155				qp->ibqp.event_handler(&ev,
156					qp->ibqp.qp_context);
157			}
158			return;
159		} while (0);
160	}
161	spin_unlock_irqrestore(&qp->state_lock, flags);
162}
163
164static struct rxe_send_wqe *__req_next_wqe(struct rxe_qp *qp)
165{
166	struct rxe_queue *q = qp->sq.queue;
167	unsigned int index = qp->req.wqe_index;
168	unsigned int prod;
169
170	prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
171	if (index == prod)
172		return NULL;
173	else
174		return queue_addr_from_index(q, index);
175}
176
177static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
178{
179	struct rxe_send_wqe *wqe;
180	unsigned long flags;
181
182	req_check_sq_drain_done(qp);
183
184	wqe = __req_next_wqe(qp);
185	if (wqe == NULL)
186		return NULL;
187
188	spin_lock_irqsave(&qp->state_lock, flags);
189	if (unlikely((qp_state(qp) == IB_QPS_SQD) &&
190		     (wqe->state != wqe_state_processing))) {
191		spin_unlock_irqrestore(&qp->state_lock, flags);
192		return NULL;
193	}
194	spin_unlock_irqrestore(&qp->state_lock, flags);
195
196	wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
197	return wqe;
198}
199
200/**
201 * rxe_wqe_is_fenced - check if next wqe is fenced
202 * @qp: the queue pair
203 * @wqe: the next wqe
204 *
205 * Returns: 1 if wqe needs to wait
206 *	    0 if wqe is ready to go
207 */
208static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
209{
210	/* Local invalidate fence (LIF) see IBA 10.6.5.1
211	 * Requires ALL previous operations on the send queue
212	 * are complete. Make mandatory for the rxe driver.
213	 */
214	if (wqe->wr.opcode == IB_WR_LOCAL_INV)
215		return qp->req.wqe_index != queue_get_consumer(qp->sq.queue,
216						QUEUE_TYPE_FROM_CLIENT);
217
218	/* Fence see IBA 10.8.3.3
219	 * Requires that all previous read and atomic operations
220	 * are complete.
221	 */
222	return (wqe->wr.send_flags & IB_SEND_FENCE) &&
223		atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic;
224}
225
226static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
227{
228	switch (opcode) {
229	case IB_WR_RDMA_WRITE:
230		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
231		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
232			return fits ?
233				IB_OPCODE_RC_RDMA_WRITE_LAST :
234				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
235		else
236			return fits ?
237				IB_OPCODE_RC_RDMA_WRITE_ONLY :
238				IB_OPCODE_RC_RDMA_WRITE_FIRST;
239
240	case IB_WR_RDMA_WRITE_WITH_IMM:
241		if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
242		    qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
243			return fits ?
244				IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
245				IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
246		else
247			return fits ?
248				IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
249				IB_OPCODE_RC_RDMA_WRITE_FIRST;
250
251	case IB_WR_SEND:
252		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
253		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
254			return fits ?
255				IB_OPCODE_RC_SEND_LAST :
256				IB_OPCODE_RC_SEND_MIDDLE;
257		else
258			return fits ?
259				IB_OPCODE_RC_SEND_ONLY :
260				IB_OPCODE_RC_SEND_FIRST;
261
262	case IB_WR_SEND_WITH_IMM:
263		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
264		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
265			return fits ?
266				IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
267				IB_OPCODE_RC_SEND_MIDDLE;
268		else
269			return fits ?
270				IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
271				IB_OPCODE_RC_SEND_FIRST;
272
273	case IB_WR_FLUSH:
274		return IB_OPCODE_RC_FLUSH;
275
276	case IB_WR_RDMA_READ:
277		return IB_OPCODE_RC_RDMA_READ_REQUEST;
278
279	case IB_WR_ATOMIC_CMP_AND_SWP:
280		return IB_OPCODE_RC_COMPARE_SWAP;
281
282	case IB_WR_ATOMIC_FETCH_AND_ADD:
283		return IB_OPCODE_RC_FETCH_ADD;
284
285	case IB_WR_SEND_WITH_INV:
286		if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
287		    qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
288			return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
289				IB_OPCODE_RC_SEND_MIDDLE;
290		else
291			return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
292				IB_OPCODE_RC_SEND_FIRST;
293
294	case IB_WR_ATOMIC_WRITE:
295		return IB_OPCODE_RC_ATOMIC_WRITE;
296
297	case IB_WR_REG_MR:
298	case IB_WR_LOCAL_INV:
299		return opcode;
300	}
301
302	return -EINVAL;
303}
304
305static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits)
306{
307	switch (opcode) {
308	case IB_WR_RDMA_WRITE:
309		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
310		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
311			return fits ?
312				IB_OPCODE_UC_RDMA_WRITE_LAST :
313				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
314		else
315			return fits ?
316				IB_OPCODE_UC_RDMA_WRITE_ONLY :
317				IB_OPCODE_UC_RDMA_WRITE_FIRST;
318
319	case IB_WR_RDMA_WRITE_WITH_IMM:
320		if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
321		    qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
322			return fits ?
323				IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
324				IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
325		else
326			return fits ?
327				IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
328				IB_OPCODE_UC_RDMA_WRITE_FIRST;
329
330	case IB_WR_SEND:
331		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
332		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
333			return fits ?
334				IB_OPCODE_UC_SEND_LAST :
335				IB_OPCODE_UC_SEND_MIDDLE;
336		else
337			return fits ?
338				IB_OPCODE_UC_SEND_ONLY :
339				IB_OPCODE_UC_SEND_FIRST;
340
341	case IB_WR_SEND_WITH_IMM:
342		if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
343		    qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
344			return fits ?
345				IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
346				IB_OPCODE_UC_SEND_MIDDLE;
347		else
348			return fits ?
349				IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
350				IB_OPCODE_UC_SEND_FIRST;
351	}
352
353	return -EINVAL;
354}
355
356static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
357		       u32 opcode)
358{
359	int fits = (wqe->dma.resid <= qp->mtu);
360
361	switch (qp_type(qp)) {
362	case IB_QPT_RC:
363		return next_opcode_rc(qp, opcode, fits);
364
365	case IB_QPT_UC:
366		return next_opcode_uc(qp, opcode, fits);
367
368	case IB_QPT_UD:
369	case IB_QPT_GSI:
370		switch (opcode) {
371		case IB_WR_SEND:
372			return IB_OPCODE_UD_SEND_ONLY;
373
374		case IB_WR_SEND_WITH_IMM:
375			return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
376		}
377		break;
378
379	default:
380		break;
381	}
382
383	return -EINVAL;
384}
385
386static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
387{
388	int depth;
389
390	if (wqe->has_rd_atomic)
391		return 0;
392
393	qp->req.need_rd_atomic = 1;
394	depth = atomic_dec_return(&qp->req.rd_atomic);
395
396	if (depth >= 0) {
397		qp->req.need_rd_atomic = 0;
398		wqe->has_rd_atomic = 1;
399		return 0;
400	}
401
402	atomic_inc(&qp->req.rd_atomic);
403	return -EAGAIN;
404}
405
406static inline int get_mtu(struct rxe_qp *qp)
407{
408	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
409
410	if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
411		return qp->mtu;
412
413	return rxe->port.mtu_cap;
414}
415
416static struct sk_buff *init_req_packet(struct rxe_qp *qp,
417				       struct rxe_av *av,
418				       struct rxe_send_wqe *wqe,
419				       int opcode, u32 payload,
420				       struct rxe_pkt_info *pkt)
421{
422	struct rxe_dev		*rxe = to_rdev(qp->ibqp.device);
423	struct sk_buff		*skb;
424	struct rxe_send_wr	*ibwr = &wqe->wr;
425	int			pad = (-payload) & 0x3;
426	int			paylen;
427	int			solicited;
428	u32			qp_num;
429	int			ack_req = 0;
430
431	/* length from start of bth to end of icrc */
432	paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
433	pkt->paylen = paylen;
434
435	/* init skb */
436	skb = rxe_init_packet(rxe, av, paylen, pkt);
437	if (unlikely(!skb))
438		return NULL;
439
440	/* init bth */
441	solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
442			(pkt->mask & RXE_END_MASK) &&
443			((pkt->mask & (RXE_SEND_MASK)) ||
444			(pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
445			(RXE_WRITE_MASK | RXE_IMMDT_MASK));
446
447	qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
448					 qp->attr.dest_qp_num;
449
450	if (qp_type(qp) != IB_QPT_UD && qp_type(qp) != IB_QPT_UC)
451		ack_req = ((pkt->mask & RXE_END_MASK) ||
452			   (qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
453	if (ack_req)
454		qp->req.noack_pkts = 0;
455
456	bth_init(pkt, pkt->opcode, solicited, 0, pad, IB_DEFAULT_PKEY_FULL, qp_num,
457		 ack_req, pkt->psn);
458
459	/* init optional headers */
460	if (pkt->mask & RXE_RETH_MASK) {
461		if (pkt->mask & RXE_FETH_MASK)
462			reth_set_rkey(pkt, ibwr->wr.flush.rkey);
463		else
464			reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
465		reth_set_va(pkt, wqe->iova);
466		reth_set_len(pkt, wqe->dma.resid);
467	}
468
469	/* Fill Flush Extension Transport Header */
470	if (pkt->mask & RXE_FETH_MASK)
471		feth_init(pkt, ibwr->wr.flush.type, ibwr->wr.flush.level);
472
473	if (pkt->mask & RXE_IMMDT_MASK)
474		immdt_set_imm(pkt, ibwr->ex.imm_data);
475
476	if (pkt->mask & RXE_IETH_MASK)
477		ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
478
479	if (pkt->mask & RXE_ATMETH_MASK) {
480		atmeth_set_va(pkt, wqe->iova);
481		if (opcode == IB_OPCODE_RC_COMPARE_SWAP) {
482			atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
483			atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
484		} else {
485			atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
486		}
487		atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
488	}
489
490	if (pkt->mask & RXE_DETH_MASK) {
491		if (qp->ibqp.qp_num == 1)
492			deth_set_qkey(pkt, GSI_QKEY);
493		else
494			deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
495		deth_set_sqp(pkt, qp->ibqp.qp_num);
496	}
497
498	return skb;
499}
500
501static int finish_packet(struct rxe_qp *qp, struct rxe_av *av,
502			 struct rxe_send_wqe *wqe, struct rxe_pkt_info *pkt,
503			 struct sk_buff *skb, u32 payload)
504{
505	int err;
506
507	err = rxe_prepare(av, pkt, skb);
508	if (err)
509		return err;
510
511	if (pkt->mask & RXE_WRITE_OR_SEND_MASK) {
512		if (wqe->wr.send_flags & IB_SEND_INLINE) {
513			u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
514
515			memcpy(payload_addr(pkt), tmp, payload);
516
517			wqe->dma.resid -= payload;
518			wqe->dma.sge_offset += payload;
519		} else {
520			err = copy_data(qp->pd, 0, &wqe->dma,
521					payload_addr(pkt), payload,
522					RXE_FROM_MR_OBJ);
523			if (err)
524				return err;
525		}
526		if (bth_pad(pkt)) {
527			u8 *pad = payload_addr(pkt) + payload;
528
529			memset(pad, 0, bth_pad(pkt));
530		}
531	} else if (pkt->mask & RXE_FLUSH_MASK) {
532		/* oA19-2: shall have no payload. */
533		wqe->dma.resid = 0;
534	}
535
536	if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
537		memcpy(payload_addr(pkt), wqe->dma.atomic_wr, payload);
538		wqe->dma.resid -= payload;
539	}
540
541	return 0;
542}
543
544static void update_wqe_state(struct rxe_qp *qp,
545		struct rxe_send_wqe *wqe,
546		struct rxe_pkt_info *pkt)
547{
548	if (pkt->mask & RXE_END_MASK) {
549		if (qp_type(qp) == IB_QPT_RC)
550			wqe->state = wqe_state_pending;
551		else
552			wqe->state = wqe_state_done;
553	} else {
554		wqe->state = wqe_state_processing;
555	}
556}
557
558static void update_wqe_psn(struct rxe_qp *qp,
559			   struct rxe_send_wqe *wqe,
560			   struct rxe_pkt_info *pkt,
561			   u32 payload)
562{
563	/* number of packets left to send including current one */
564	int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
565
566	/* handle zero length packet case */
567	if (num_pkt == 0)
568		num_pkt = 1;
569
570	if (pkt->mask & RXE_START_MASK) {
571		wqe->first_psn = qp->req.psn;
572		wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
573	}
574
575	if (pkt->mask & RXE_READ_MASK)
576		qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
577	else
578		qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
579}
580
581static void update_state(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
582{
583	qp->req.opcode = pkt->opcode;
584
585	if (pkt->mask & RXE_END_MASK)
586		qp->req.wqe_index = queue_next_index(qp->sq.queue,
587						     qp->req.wqe_index);
588
589	qp->need_req_skb = 0;
590
591	if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
592		mod_timer(&qp->retrans_timer,
593			  jiffies + qp->qp_timeout_jiffies);
594}
595
596static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
597{
598	u8 opcode = wqe->wr.opcode;
599	u32 rkey;
600	int ret;
601
602	switch (opcode) {
603	case IB_WR_LOCAL_INV:
604		rkey = wqe->wr.ex.invalidate_rkey;
605		if (rkey_is_mw(rkey))
606			ret = rxe_invalidate_mw(qp, rkey);
607		else
608			ret = rxe_invalidate_mr(qp, rkey);
609
610		if (unlikely(ret)) {
611			wqe->status = IB_WC_LOC_QP_OP_ERR;
612			return ret;
613		}
614		break;
615	case IB_WR_REG_MR:
616		ret = rxe_reg_fast_mr(qp, wqe);
617		if (unlikely(ret)) {
618			wqe->status = IB_WC_LOC_QP_OP_ERR;
619			return ret;
620		}
621		break;
622	case IB_WR_BIND_MW:
623		ret = rxe_bind_mw(qp, wqe);
624		if (unlikely(ret)) {
625			wqe->status = IB_WC_MW_BIND_ERR;
626			return ret;
627		}
628		break;
629	default:
630		rxe_dbg_qp(qp, "Unexpected send wqe opcode %d\n", opcode);
631		wqe->status = IB_WC_LOC_QP_OP_ERR;
632		return -EINVAL;
633	}
634
635	wqe->state = wqe_state_done;
636	wqe->status = IB_WC_SUCCESS;
637	qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
638
639	return 0;
640}
641
642int rxe_requester(struct rxe_qp *qp)
643{
644	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
645	struct rxe_pkt_info pkt;
646	struct sk_buff *skb;
647	struct rxe_send_wqe *wqe;
648	enum rxe_hdr_mask mask;
649	u32 payload;
650	int mtu;
651	int opcode;
652	int err;
653	int ret;
654	struct rxe_queue *q = qp->sq.queue;
655	struct rxe_ah *ah;
656	struct rxe_av *av;
657	unsigned long flags;
658
659	spin_lock_irqsave(&qp->state_lock, flags);
660	if (unlikely(!qp->valid)) {
661		spin_unlock_irqrestore(&qp->state_lock, flags);
662		goto exit;
663	}
664
665	if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
666		wqe = __req_next_wqe(qp);
667		spin_unlock_irqrestore(&qp->state_lock, flags);
668		if (wqe) {
669			wqe->status = IB_WC_WR_FLUSH_ERR;
670			goto err;
671		} else {
672			goto exit;
673		}
674	}
675
676	if (unlikely(qp_state(qp) == IB_QPS_RESET)) {
677		qp->req.wqe_index = queue_get_consumer(q,
678						QUEUE_TYPE_FROM_CLIENT);
679		qp->req.opcode = -1;
680		qp->req.need_rd_atomic = 0;
681		qp->req.wait_psn = 0;
682		qp->req.need_retry = 0;
683		qp->req.wait_for_rnr_timer = 0;
684		spin_unlock_irqrestore(&qp->state_lock, flags);
685		goto exit;
686	}
687	spin_unlock_irqrestore(&qp->state_lock, flags);
688
689	/* we come here if the retransmit timer has fired
690	 * or if the rnr timer has fired. If the retransmit
691	 * timer fires while we are processing an RNR NAK wait
692	 * until the rnr timer has fired before starting the
693	 * retry flow
694	 */
695	if (unlikely(qp->req.need_retry && !qp->req.wait_for_rnr_timer)) {
696		req_retry(qp);
697		qp->req.need_retry = 0;
698	}
699
700	wqe = req_next_wqe(qp);
701	if (unlikely(!wqe))
702		goto exit;
703
704	if (rxe_wqe_is_fenced(qp, wqe)) {
705		qp->req.wait_fence = 1;
706		goto exit;
707	}
708
709	if (wqe->mask & WR_LOCAL_OP_MASK) {
710		err = rxe_do_local_ops(qp, wqe);
711		if (unlikely(err))
712			goto err;
713		else
714			goto done;
715	}
716
717	if (unlikely(qp_type(qp) == IB_QPT_RC &&
718		psn_compare(qp->req.psn, (qp->comp.psn +
719				RXE_MAX_UNACKED_PSNS)) > 0)) {
720		qp->req.wait_psn = 1;
721		goto exit;
722	}
723
724	/* Limit the number of inflight SKBs per QP */
725	if (unlikely(atomic_read(&qp->skb_out) >
726		     RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
727		qp->need_req_skb = 1;
728		goto exit;
729	}
730
731	opcode = next_opcode(qp, wqe, wqe->wr.opcode);
732	if (unlikely(opcode < 0)) {
733		wqe->status = IB_WC_LOC_QP_OP_ERR;
734		goto err;
735	}
736
737	mask = rxe_opcode[opcode].mask;
738	if (unlikely(mask & (RXE_READ_OR_ATOMIC_MASK |
739			RXE_ATOMIC_WRITE_MASK))) {
740		if (check_init_depth(qp, wqe))
741			goto exit;
742	}
743
744	mtu = get_mtu(qp);
745	payload = (mask & (RXE_WRITE_OR_SEND_MASK | RXE_ATOMIC_WRITE_MASK)) ?
746			wqe->dma.resid : 0;
747	if (payload > mtu) {
748		if (qp_type(qp) == IB_QPT_UD) {
749			/* C10-93.1.1: If the total sum of all the buffer lengths specified for a
750			 * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
751			 * shall not emit any packets for this message. Further, the CI shall not
752			 * generate an error due to this condition.
753			 */
754
755			/* fake a successful UD send */
756			wqe->first_psn = qp->req.psn;
757			wqe->last_psn = qp->req.psn;
758			qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
759			qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
760			qp->req.wqe_index = queue_next_index(qp->sq.queue,
761						       qp->req.wqe_index);
762			wqe->state = wqe_state_done;
763			wqe->status = IB_WC_SUCCESS;
764			goto done;
765		}
766		payload = mtu;
767	}
768
769	pkt.rxe = rxe;
770	pkt.opcode = opcode;
771	pkt.qp = qp;
772	pkt.psn = qp->req.psn;
773	pkt.mask = rxe_opcode[opcode].mask;
774	pkt.wqe = wqe;
775
776	av = rxe_get_av(&pkt, &ah);
777	if (unlikely(!av)) {
778		rxe_dbg_qp(qp, "Failed no address vector\n");
779		wqe->status = IB_WC_LOC_QP_OP_ERR;
780		goto err;
781	}
782
783	skb = init_req_packet(qp, av, wqe, opcode, payload, &pkt);
784	if (unlikely(!skb)) {
785		rxe_dbg_qp(qp, "Failed allocating skb\n");
786		wqe->status = IB_WC_LOC_QP_OP_ERR;
787		if (ah)
788			rxe_put(ah);
789		goto err;
790	}
791
792	err = finish_packet(qp, av, wqe, &pkt, skb, payload);
793	if (unlikely(err)) {
794		rxe_dbg_qp(qp, "Error during finish packet\n");
795		if (err == -EFAULT)
796			wqe->status = IB_WC_LOC_PROT_ERR;
797		else
798			wqe->status = IB_WC_LOC_QP_OP_ERR;
799		kfree_skb(skb);
800		if (ah)
801			rxe_put(ah);
802		goto err;
803	}
804
805	if (ah)
806		rxe_put(ah);
807
808	err = rxe_xmit_packet(qp, &pkt, skb);
809	if (err) {
810		wqe->status = IB_WC_LOC_QP_OP_ERR;
811		goto err;
812	}
813
814	update_wqe_state(qp, wqe, &pkt);
815	update_wqe_psn(qp, wqe, &pkt, payload);
816	update_state(qp, &pkt);
817
818	/* A non-zero return value will cause rxe_do_task to
819	 * exit its loop and end the work item. A zero return
820	 * will continue looping and return to rxe_requester
821	 */
822done:
823	ret = 0;
824	goto out;
825err:
826	/* update wqe_index for each wqe completion */
827	qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
828	wqe->state = wqe_state_error;
829	rxe_qp_error(qp);
830exit:
831	ret = -EAGAIN;
832out:
833	return ret;
834}
835
836int rxe_sender(struct rxe_qp *qp)
837{
838	int req_ret;
839	int comp_ret;
840
841	/* process the send queue */
842	req_ret = rxe_requester(qp);
843
844	/* process the response queue */
845	comp_ret = rxe_completer(qp);
846
847	/* exit the task loop if both requester and completer
848	 * are ready
849	 */
850	return (req_ret && comp_ret) ? -EAGAIN : 0;
851}