xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ruc.c (revision 99fee508)
1 /*
2  * Copyright(c) 2015 - 2017 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/spinlock.h>
49 
50 #include "hfi.h"
51 #include "mad.h"
52 #include "qp.h"
53 #include "verbs_txreq.h"
54 #include "trace.h"
55 
56 /*
57  * Validate a RWQE and fill in the SGE state.
58  * Return 1 if OK.
59  */
60 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
61 {
62 	int i, j, ret;
63 	struct ib_wc wc;
64 	struct rvt_lkey_table *rkt;
65 	struct rvt_pd *pd;
66 	struct rvt_sge_state *ss;
67 
68 	rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
69 	pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
70 	ss = &qp->r_sge;
71 	ss->sg_list = qp->r_sg_list;
72 	qp->r_len = 0;
73 	for (i = j = 0; i < wqe->num_sge; i++) {
74 		if (wqe->sg_list[i].length == 0)
75 			continue;
76 		/* Check LKEY */
77 		ret = rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
78 				  NULL, &wqe->sg_list[i],
79 				  IB_ACCESS_LOCAL_WRITE);
80 		if (unlikely(ret <= 0))
81 			goto bad_lkey;
82 		qp->r_len += wqe->sg_list[i].length;
83 		j++;
84 	}
85 	ss->num_sge = j;
86 	ss->total_len = qp->r_len;
87 	ret = 1;
88 	goto bail;
89 
90 bad_lkey:
91 	while (j) {
92 		struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
93 
94 		rvt_put_mr(sge->mr);
95 	}
96 	ss->num_sge = 0;
97 	memset(&wc, 0, sizeof(wc));
98 	wc.wr_id = wqe->wr_id;
99 	wc.status = IB_WC_LOC_PROT_ERR;
100 	wc.opcode = IB_WC_RECV;
101 	wc.qp = &qp->ibqp;
102 	/* Signal solicited completion event. */
103 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
104 	ret = 0;
105 bail:
106 	return ret;
107 }
108 
109 /**
110  * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE
111  * @qp: the QP
112  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
113  *
114  * Return -1 if there is a local error, 0 if no RWQE is available,
115  * otherwise return 1.
116  *
117  * Can be called from interrupt level.
118  */
119 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only)
120 {
121 	unsigned long flags;
122 	struct rvt_rq *rq;
123 	struct rvt_rwq *wq;
124 	struct rvt_srq *srq;
125 	struct rvt_rwqe *wqe;
126 	void (*handler)(struct ib_event *, void *);
127 	u32 tail;
128 	int ret;
129 
130 	if (qp->ibqp.srq) {
131 		srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
132 		handler = srq->ibsrq.event_handler;
133 		rq = &srq->rq;
134 	} else {
135 		srq = NULL;
136 		handler = NULL;
137 		rq = &qp->r_rq;
138 	}
139 
140 	spin_lock_irqsave(&rq->lock, flags);
141 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
142 		ret = 0;
143 		goto unlock;
144 	}
145 
146 	wq = rq->wq;
147 	tail = wq->tail;
148 	/* Validate tail before using it since it is user writable. */
149 	if (tail >= rq->size)
150 		tail = 0;
151 	if (unlikely(tail == wq->head)) {
152 		ret = 0;
153 		goto unlock;
154 	}
155 	/* Make sure entry is read after head index is read. */
156 	smp_rmb();
157 	wqe = rvt_get_rwqe_ptr(rq, tail);
158 	/*
159 	 * Even though we update the tail index in memory, the verbs
160 	 * consumer is not supposed to post more entries until a
161 	 * completion is generated.
162 	 */
163 	if (++tail >= rq->size)
164 		tail = 0;
165 	wq->tail = tail;
166 	if (!wr_id_only && !init_sge(qp, wqe)) {
167 		ret = -1;
168 		goto unlock;
169 	}
170 	qp->r_wr_id = wqe->wr_id;
171 
172 	ret = 1;
173 	set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
174 	if (handler) {
175 		u32 n;
176 
177 		/*
178 		 * Validate head pointer value and compute
179 		 * the number of remaining WQEs.
180 		 */
181 		n = wq->head;
182 		if (n >= rq->size)
183 			n = 0;
184 		if (n < tail)
185 			n += rq->size - tail;
186 		else
187 			n -= tail;
188 		if (n < srq->limit) {
189 			struct ib_event ev;
190 
191 			srq->limit = 0;
192 			spin_unlock_irqrestore(&rq->lock, flags);
193 			ev.device = qp->ibqp.device;
194 			ev.element.srq = qp->ibqp.srq;
195 			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
196 			handler(&ev, srq->ibsrq.srq_context);
197 			goto bail;
198 		}
199 	}
200 unlock:
201 	spin_unlock_irqrestore(&rq->lock, flags);
202 bail:
203 	return ret;
204 }
205 
206 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
207 {
208 	return (gid->global.interface_id == id &&
209 		(gid->global.subnet_prefix == gid_prefix ||
210 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
211 }
212 
213 /*
214  *
215  * This should be called with the QP r_lock held.
216  *
217  * The s_lock will be acquired around the hfi1_migrate_qp() call.
218  */
219 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_packet *packet)
220 {
221 	__be64 guid;
222 	unsigned long flags;
223 	struct rvt_qp *qp = packet->qp;
224 	u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
225 	u32 dlid = packet->dlid;
226 	u32 slid = packet->slid;
227 	u32 sl = packet->sl;
228 	int migrated;
229 	u32 bth0, bth1;
230 	u16 pkey;
231 
232 	bth0 = be32_to_cpu(packet->ohdr->bth[0]);
233 	bth1 = be32_to_cpu(packet->ohdr->bth[1]);
234 	if (packet->etype == RHF_RCV_TYPE_BYPASS) {
235 		pkey = hfi1_16B_get_pkey(packet->hdr);
236 		migrated = bth1 & OPA_BTH_MIG_REQ;
237 	} else {
238 		pkey = ib_bth_get_pkey(packet->ohdr);
239 		migrated = bth0 & IB_BTH_MIG_REQ;
240 	}
241 
242 	if (qp->s_mig_state == IB_MIG_ARMED && migrated) {
243 		if (!packet->grh) {
244 			if ((rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
245 			     IB_AH_GRH) &&
246 			    (packet->etype != RHF_RCV_TYPE_BYPASS))
247 				return 1;
248 		} else {
249 			const struct ib_global_route *grh;
250 
251 			if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
252 			      IB_AH_GRH))
253 				return 1;
254 			grh = rdma_ah_read_grh(&qp->alt_ah_attr);
255 			guid = get_sguid(ibp, grh->sgid_index);
256 			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
257 				    guid))
258 				return 1;
259 			if (!gid_ok(
260 				&packet->grh->sgid,
261 				grh->dgid.global.subnet_prefix,
262 				grh->dgid.global.interface_id))
263 				return 1;
264 		}
265 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
266 					    sc5, slid))) {
267 			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
268 				      slid, dlid);
269 			return 1;
270 		}
271 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
272 		if (slid != rdma_ah_get_dlid(&qp->alt_ah_attr) ||
273 		    ppd_from_ibp(ibp)->port !=
274 			rdma_ah_get_port_num(&qp->alt_ah_attr))
275 			return 1;
276 		spin_lock_irqsave(&qp->s_lock, flags);
277 		hfi1_migrate_qp(qp);
278 		spin_unlock_irqrestore(&qp->s_lock, flags);
279 	} else {
280 		if (!packet->grh) {
281 			if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
282 			     IB_AH_GRH) &&
283 			    (packet->etype != RHF_RCV_TYPE_BYPASS))
284 				return 1;
285 		} else {
286 			const struct ib_global_route *grh;
287 
288 			if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
289 						   IB_AH_GRH))
290 				return 1;
291 			grh = rdma_ah_read_grh(&qp->remote_ah_attr);
292 			guid = get_sguid(ibp, grh->sgid_index);
293 			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
294 				    guid))
295 				return 1;
296 			if (!gid_ok(
297 			     &packet->grh->sgid,
298 			     grh->dgid.global.subnet_prefix,
299 			     grh->dgid.global.interface_id))
300 				return 1;
301 		}
302 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
303 					    sc5, slid))) {
304 			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
305 				      slid, dlid);
306 			return 1;
307 		}
308 		/* Validate the SLID. See Ch. 9.6.1.5 */
309 		if ((slid != rdma_ah_get_dlid(&qp->remote_ah_attr)) ||
310 		    ppd_from_ibp(ibp)->port != qp->port_num)
311 			return 1;
312 		if (qp->s_mig_state == IB_MIG_REARM && !migrated)
313 			qp->s_mig_state = IB_MIG_ARMED;
314 	}
315 
316 	return 0;
317 }
318 
319 /**
320  * ruc_loopback - handle UC and RC loopback requests
321  * @sqp: the sending QP
322  *
323  * This is called from hfi1_do_send() to
324  * forward a WQE addressed to the same HFI.
325  * Note that although we are single threaded due to the send engine, we still
326  * have to protect against post_send().  We don't have to worry about
327  * receive interrupts since this is a connected protocol and all packets
328  * will pass through here.
329  */
330 static void ruc_loopback(struct rvt_qp *sqp)
331 {
332 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
333 	struct rvt_qp *qp;
334 	struct rvt_swqe *wqe;
335 	struct rvt_sge *sge;
336 	unsigned long flags;
337 	struct ib_wc wc;
338 	u64 sdata;
339 	atomic64_t *maddr;
340 	enum ib_wc_status send_status;
341 	bool release;
342 	int ret;
343 	bool copy_last = false;
344 	int local_ops = 0;
345 
346 	rcu_read_lock();
347 
348 	/*
349 	 * Note that we check the responder QP state after
350 	 * checking the requester's state.
351 	 */
352 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
353 			    sqp->remote_qpn);
354 
355 	spin_lock_irqsave(&sqp->s_lock, flags);
356 
357 	/* Return if we are already busy processing a work request. */
358 	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
359 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
360 		goto unlock;
361 
362 	sqp->s_flags |= RVT_S_BUSY;
363 
364 again:
365 	smp_read_barrier_depends(); /* see post_one_send() */
366 	if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
367 		goto clr_busy;
368 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
369 
370 	/* Return if it is not OK to start a new work request. */
371 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
372 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
373 			goto clr_busy;
374 		/* We are in the error state, flush the work request. */
375 		send_status = IB_WC_WR_FLUSH_ERR;
376 		goto flush_send;
377 	}
378 
379 	/*
380 	 * We can rely on the entry not changing without the s_lock
381 	 * being held until we update s_last.
382 	 * We increment s_cur to indicate s_last is in progress.
383 	 */
384 	if (sqp->s_last == sqp->s_cur) {
385 		if (++sqp->s_cur >= sqp->s_size)
386 			sqp->s_cur = 0;
387 	}
388 	spin_unlock_irqrestore(&sqp->s_lock, flags);
389 
390 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
391 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
392 		ibp->rvp.n_pkt_drops++;
393 		/*
394 		 * For RC, the requester would timeout and retry so
395 		 * shortcut the timeouts and just signal too many retries.
396 		 */
397 		if (sqp->ibqp.qp_type == IB_QPT_RC)
398 			send_status = IB_WC_RETRY_EXC_ERR;
399 		else
400 			send_status = IB_WC_SUCCESS;
401 		goto serr;
402 	}
403 
404 	memset(&wc, 0, sizeof(wc));
405 	send_status = IB_WC_SUCCESS;
406 
407 	release = true;
408 	sqp->s_sge.sge = wqe->sg_list[0];
409 	sqp->s_sge.sg_list = wqe->sg_list + 1;
410 	sqp->s_sge.num_sge = wqe->wr.num_sge;
411 	sqp->s_len = wqe->length;
412 	switch (wqe->wr.opcode) {
413 	case IB_WR_REG_MR:
414 		goto send_comp;
415 
416 	case IB_WR_LOCAL_INV:
417 		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
418 			if (rvt_invalidate_rkey(sqp,
419 						wqe->wr.ex.invalidate_rkey))
420 				send_status = IB_WC_LOC_PROT_ERR;
421 			local_ops = 1;
422 		}
423 		goto send_comp;
424 
425 	case IB_WR_SEND_WITH_INV:
426 		if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
427 			wc.wc_flags = IB_WC_WITH_INVALIDATE;
428 			wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
429 		}
430 		goto send;
431 
432 	case IB_WR_SEND_WITH_IMM:
433 		wc.wc_flags = IB_WC_WITH_IMM;
434 		wc.ex.imm_data = wqe->wr.ex.imm_data;
435 		/* FALLTHROUGH */
436 	case IB_WR_SEND:
437 send:
438 		ret = hfi1_rvt_get_rwqe(qp, 0);
439 		if (ret < 0)
440 			goto op_err;
441 		if (!ret)
442 			goto rnr_nak;
443 		break;
444 
445 	case IB_WR_RDMA_WRITE_WITH_IMM:
446 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
447 			goto inv_err;
448 		wc.wc_flags = IB_WC_WITH_IMM;
449 		wc.ex.imm_data = wqe->wr.ex.imm_data;
450 		ret = hfi1_rvt_get_rwqe(qp, 1);
451 		if (ret < 0)
452 			goto op_err;
453 		if (!ret)
454 			goto rnr_nak;
455 		/* skip copy_last set and qp_access_flags recheck */
456 		goto do_write;
457 	case IB_WR_RDMA_WRITE:
458 		copy_last = rvt_is_user_qp(qp);
459 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
460 			goto inv_err;
461 do_write:
462 		if (wqe->length == 0)
463 			break;
464 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
465 					  wqe->rdma_wr.remote_addr,
466 					  wqe->rdma_wr.rkey,
467 					  IB_ACCESS_REMOTE_WRITE)))
468 			goto acc_err;
469 		qp->r_sge.sg_list = NULL;
470 		qp->r_sge.num_sge = 1;
471 		qp->r_sge.total_len = wqe->length;
472 		break;
473 
474 	case IB_WR_RDMA_READ:
475 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
476 			goto inv_err;
477 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
478 					  wqe->rdma_wr.remote_addr,
479 					  wqe->rdma_wr.rkey,
480 					  IB_ACCESS_REMOTE_READ)))
481 			goto acc_err;
482 		release = false;
483 		sqp->s_sge.sg_list = NULL;
484 		sqp->s_sge.num_sge = 1;
485 		qp->r_sge.sge = wqe->sg_list[0];
486 		qp->r_sge.sg_list = wqe->sg_list + 1;
487 		qp->r_sge.num_sge = wqe->wr.num_sge;
488 		qp->r_sge.total_len = wqe->length;
489 		break;
490 
491 	case IB_WR_ATOMIC_CMP_AND_SWP:
492 	case IB_WR_ATOMIC_FETCH_AND_ADD:
493 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
494 			goto inv_err;
495 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
496 					  wqe->atomic_wr.remote_addr,
497 					  wqe->atomic_wr.rkey,
498 					  IB_ACCESS_REMOTE_ATOMIC)))
499 			goto acc_err;
500 		/* Perform atomic OP and save result. */
501 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
502 		sdata = wqe->atomic_wr.compare_add;
503 		*(u64 *)sqp->s_sge.sge.vaddr =
504 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
505 			(u64)atomic64_add_return(sdata, maddr) - sdata :
506 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
507 				      sdata, wqe->atomic_wr.swap);
508 		rvt_put_mr(qp->r_sge.sge.mr);
509 		qp->r_sge.num_sge = 0;
510 		goto send_comp;
511 
512 	default:
513 		send_status = IB_WC_LOC_QP_OP_ERR;
514 		goto serr;
515 	}
516 
517 	sge = &sqp->s_sge.sge;
518 	while (sqp->s_len) {
519 		u32 len = sqp->s_len;
520 
521 		if (len > sge->length)
522 			len = sge->length;
523 		if (len > sge->sge_length)
524 			len = sge->sge_length;
525 		WARN_ON_ONCE(len == 0);
526 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
527 		sge->vaddr += len;
528 		sge->length -= len;
529 		sge->sge_length -= len;
530 		if (sge->sge_length == 0) {
531 			if (!release)
532 				rvt_put_mr(sge->mr);
533 			if (--sqp->s_sge.num_sge)
534 				*sge = *sqp->s_sge.sg_list++;
535 		} else if (sge->length == 0 && sge->mr->lkey) {
536 			if (++sge->n >= RVT_SEGSZ) {
537 				if (++sge->m >= sge->mr->mapsz)
538 					break;
539 				sge->n = 0;
540 			}
541 			sge->vaddr =
542 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
543 			sge->length =
544 				sge->mr->map[sge->m]->segs[sge->n].length;
545 		}
546 		sqp->s_len -= len;
547 	}
548 	if (release)
549 		rvt_put_ss(&qp->r_sge);
550 
551 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
552 		goto send_comp;
553 
554 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
555 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
556 	else
557 		wc.opcode = IB_WC_RECV;
558 	wc.wr_id = qp->r_wr_id;
559 	wc.status = IB_WC_SUCCESS;
560 	wc.byte_len = wqe->length;
561 	wc.qp = &qp->ibqp;
562 	wc.src_qp = qp->remote_qpn;
563 	wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
564 	wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
565 	wc.port_num = 1;
566 	/* Signal completion event if the solicited bit is set. */
567 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
568 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
569 
570 send_comp:
571 	spin_lock_irqsave(&sqp->s_lock, flags);
572 	ibp->rvp.n_loop_pkts++;
573 flush_send:
574 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
575 	hfi1_send_complete(sqp, wqe, send_status);
576 	if (local_ops) {
577 		atomic_dec(&sqp->local_ops_pending);
578 		local_ops = 0;
579 	}
580 	goto again;
581 
582 rnr_nak:
583 	/* Handle RNR NAK */
584 	if (qp->ibqp.qp_type == IB_QPT_UC)
585 		goto send_comp;
586 	ibp->rvp.n_rnr_naks++;
587 	/*
588 	 * Note: we don't need the s_lock held since the BUSY flag
589 	 * makes this single threaded.
590 	 */
591 	if (sqp->s_rnr_retry == 0) {
592 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
593 		goto serr;
594 	}
595 	if (sqp->s_rnr_retry_cnt < 7)
596 		sqp->s_rnr_retry--;
597 	spin_lock_irqsave(&sqp->s_lock, flags);
598 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
599 		goto clr_busy;
600 	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
601 				IB_AETH_CREDIT_SHIFT);
602 	goto clr_busy;
603 
604 op_err:
605 	send_status = IB_WC_REM_OP_ERR;
606 	wc.status = IB_WC_LOC_QP_OP_ERR;
607 	goto err;
608 
609 inv_err:
610 	send_status = IB_WC_REM_INV_REQ_ERR;
611 	wc.status = IB_WC_LOC_QP_OP_ERR;
612 	goto err;
613 
614 acc_err:
615 	send_status = IB_WC_REM_ACCESS_ERR;
616 	wc.status = IB_WC_LOC_PROT_ERR;
617 err:
618 	/* responder goes to error state */
619 	rvt_rc_error(qp, wc.status);
620 
621 serr:
622 	spin_lock_irqsave(&sqp->s_lock, flags);
623 	hfi1_send_complete(sqp, wqe, send_status);
624 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
625 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
626 
627 		sqp->s_flags &= ~RVT_S_BUSY;
628 		spin_unlock_irqrestore(&sqp->s_lock, flags);
629 		if (lastwqe) {
630 			struct ib_event ev;
631 
632 			ev.device = sqp->ibqp.device;
633 			ev.element.qp = &sqp->ibqp;
634 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
635 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
636 		}
637 		goto done;
638 	}
639 clr_busy:
640 	sqp->s_flags &= ~RVT_S_BUSY;
641 unlock:
642 	spin_unlock_irqrestore(&sqp->s_lock, flags);
643 done:
644 	rcu_read_unlock();
645 }
646 
647 /**
648  * hfi1_make_grh - construct a GRH header
649  * @ibp: a pointer to the IB port
650  * @hdr: a pointer to the GRH header being constructed
651  * @grh: the global route address to send to
652  * @hwords: size of header after grh being sent in dwords
653  * @nwords: the number of 32 bit words of data being sent
654  *
655  * Return the size of the header in 32 bit words.
656  */
657 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
658 		  const struct ib_global_route *grh, u32 hwords, u32 nwords)
659 {
660 	hdr->version_tclass_flow =
661 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
662 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
663 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
664 	hdr->paylen = cpu_to_be16((hwords + nwords) << 2);
665 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
666 	hdr->next_hdr = IB_GRH_NEXT_HDR;
667 	hdr->hop_limit = grh->hop_limit;
668 	/* The SGID is 32-bit aligned. */
669 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
670 	hdr->sgid.global.interface_id =
671 		grh->sgid_index < HFI1_GUIDS_PER_PORT ?
672 		get_sguid(ibp, grh->sgid_index) :
673 		get_sguid(ibp, HFI1_PORT_GUID_INDEX);
674 	hdr->dgid = grh->dgid;
675 
676 	/* GRH header size in 32-bit words. */
677 	return sizeof(struct ib_grh) / sizeof(u32);
678 }
679 
680 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, \
681 			      hdr.ibh.u.oth.bth[2]) / 4)
682 
683 /**
684  * build_ahg - create ahg in s_ahg
685  * @qp: a pointer to QP
686  * @npsn: the next PSN for the request/response
687  *
688  * This routine handles the AHG by allocating an ahg entry and causing the
689  * copy of the first middle.
690  *
691  * Subsequent middles use the copied entry, editing the
692  * PSN with 1 or 2 edits.
693  */
694 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
695 {
696 	struct hfi1_qp_priv *priv = qp->priv;
697 
698 	if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
699 		clear_ahg(qp);
700 	if (!(qp->s_flags & RVT_S_AHG_VALID)) {
701 		/* first middle that needs copy  */
702 		if (qp->s_ahgidx < 0)
703 			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
704 		if (qp->s_ahgidx >= 0) {
705 			qp->s_ahgpsn = npsn;
706 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
707 			/* save to protect a change in another thread */
708 			priv->s_ahg->ahgidx = qp->s_ahgidx;
709 			qp->s_flags |= RVT_S_AHG_VALID;
710 		}
711 	} else {
712 		/* subsequent middle after valid */
713 		if (qp->s_ahgidx >= 0) {
714 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
715 			priv->s_ahg->ahgidx = qp->s_ahgidx;
716 			priv->s_ahg->ahgcount++;
717 			priv->s_ahg->ahgdesc[0] =
718 				sdma_build_ahg_descriptor(
719 					(__force u16)cpu_to_be16((u16)npsn),
720 					BTH2_OFFSET,
721 					16,
722 					16);
723 			if ((npsn & 0xffff0000) !=
724 					(qp->s_ahgpsn & 0xffff0000)) {
725 				priv->s_ahg->ahgcount++;
726 				priv->s_ahg->ahgdesc[1] =
727 					sdma_build_ahg_descriptor(
728 						(__force u16)cpu_to_be16(
729 							(u16)(npsn >> 16)),
730 						BTH2_OFFSET,
731 						0,
732 						16);
733 			}
734 		}
735 	}
736 }
737 
738 static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
739 				     struct ib_other_headers *ohdr,
740 				     u32 bth0, u32 bth1, u32 bth2)
741 {
742 	bth1 |= qp->remote_qpn;
743 	ohdr->bth[0] = cpu_to_be32(bth0);
744 	ohdr->bth[1] = cpu_to_be32(bth1);
745 	ohdr->bth[2] = cpu_to_be32(bth2);
746 }
747 
748 static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
749 					    struct ib_other_headers *ohdr,
750 					    u32 bth0, u32 bth2, int middle,
751 					    struct hfi1_pkt_state *ps)
752 {
753 	struct hfi1_qp_priv *priv = qp->priv;
754 	struct hfi1_ibport *ibp = ps->ibp;
755 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
756 	u32 bth1 = 0;
757 	u32 slid;
758 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
759 	u8 l4 = OPA_16B_L4_IB_LOCAL;
760 	u8 extra_bytes = hfi1_get_16b_padding((qp->s_hdrwords << 2),
761 				   ps->s_txreq->s_cur_size);
762 	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
763 				 extra_bytes + SIZE_OF_LT) >> 2);
764 	u8 becn = 0;
765 
766 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
767 	    hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
768 		struct ib_grh *grh;
769 		struct ib_global_route *grd =
770 			rdma_ah_retrieve_grh(&qp->remote_ah_attr);
771 		int hdrwords;
772 
773 		/*
774 		 * Ensure OPA GIDs are transformed to IB gids
775 		 * before creating the GRH.
776 		 */
777 		if (grd->sgid_index == OPA_GID_INDEX)
778 			grd->sgid_index = 0;
779 		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
780 		l4 = OPA_16B_L4_IB_GLOBAL;
781 		hdrwords = qp->s_hdrwords - 4;
782 		qp->s_hdrwords += hfi1_make_grh(ibp, grh, grd,
783 						hdrwords, nwords);
784 		middle = 0;
785 	}
786 
787 	if (qp->s_mig_state == IB_MIG_MIGRATED)
788 		bth1 |= OPA_BTH_MIG_REQ;
789 	else
790 		middle = 0;
791 
792 	if (middle)
793 		build_ahg(qp, bth2);
794 	else
795 		qp->s_flags &= ~RVT_S_AHG_VALID;
796 
797 	bth0 |= pkey;
798 	bth0 |= extra_bytes << 20;
799 	if (qp->s_flags & RVT_S_ECN) {
800 		qp->s_flags &= ~RVT_S_ECN;
801 		/* we recently received a FECN, so return a BECN */
802 		becn = 1;
803 	}
804 	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
805 
806 	if (!ppd->lid)
807 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
808 	else
809 		slid = ppd->lid |
810 			(rdma_ah_get_path_bits(&qp->remote_ah_attr) &
811 			((1 << ppd->lmc) - 1));
812 
813 	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
814 			  slid,
815 			  opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
816 				      16B),
817 			  (qp->s_hdrwords + nwords) >> 1,
818 			  pkey, becn, 0, l4, priv->s_sc);
819 }
820 
821 static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
822 					   struct ib_other_headers *ohdr,
823 					   u32 bth0, u32 bth2, int middle,
824 					   struct hfi1_pkt_state *ps)
825 {
826 	struct hfi1_qp_priv *priv = qp->priv;
827 	struct hfi1_ibport *ibp = ps->ibp;
828 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
829 	u32 bth1 = 0;
830 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
831 	u16 lrh0 = HFI1_LRH_BTH;
832 	u16 slid;
833 	u8 extra_bytes = -ps->s_txreq->s_cur_size & 3;
834 	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
835 					 extra_bytes) >> 2);
836 
837 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
838 		struct ib_grh *grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
839 		int hdrwords = qp->s_hdrwords - 2;
840 
841 		lrh0 = HFI1_LRH_GRH;
842 		qp->s_hdrwords +=
843 			hfi1_make_grh(ibp, grh,
844 				      rdma_ah_read_grh(&qp->remote_ah_attr),
845 				      hdrwords, nwords);
846 		middle = 0;
847 	}
848 	lrh0 |= (priv->s_sc & 0xf) << 12 |
849 		(rdma_ah_get_sl(&qp->remote_ah_attr) & 0xf) << 4;
850 
851 	if (qp->s_mig_state == IB_MIG_MIGRATED)
852 		bth0 |= IB_BTH_MIG_REQ;
853 	else
854 		middle = 0;
855 
856 	if (middle)
857 		build_ahg(qp, bth2);
858 	else
859 		qp->s_flags &= ~RVT_S_AHG_VALID;
860 
861 	bth0 |= pkey;
862 	bth0 |= extra_bytes << 20;
863 	if (qp->s_flags & RVT_S_ECN) {
864 		qp->s_flags &= ~RVT_S_ECN;
865 		/* we recently received a FECN, so return a BECN */
866 		bth1 |= (IB_BECN_MASK << IB_BECN_SHIFT);
867 	}
868 	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
869 
870 	if (!ppd->lid)
871 		slid = be16_to_cpu(IB_LID_PERMISSIVE);
872 	else
873 		slid = ppd->lid |
874 			(rdma_ah_get_path_bits(&qp->remote_ah_attr) &
875 			((1 << ppd->lmc) - 1));
876 	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
877 			 lrh0,
878 			 qp->s_hdrwords + nwords,
879 			 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
880 			 ppd_from_ibp(ibp)->lid |
881 				rdma_ah_get_path_bits(&qp->remote_ah_attr));
882 }
883 
884 typedef void (*hfi1_make_ruc_hdr)(struct rvt_qp *qp,
885 				  struct ib_other_headers *ohdr,
886 				  u32 bth0, u32 bth2, int middle,
887 				  struct hfi1_pkt_state *ps);
888 
889 /* We support only two types - 9B and 16B for now */
890 static const hfi1_make_ruc_hdr hfi1_ruc_header_tbl[2] = {
891 	[HFI1_PKT_TYPE_9B] = &hfi1_make_ruc_header_9B,
892 	[HFI1_PKT_TYPE_16B] = &hfi1_make_ruc_header_16B
893 };
894 
895 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
896 			  u32 bth0, u32 bth2, int middle,
897 			  struct hfi1_pkt_state *ps)
898 {
899 	struct hfi1_qp_priv *priv = qp->priv;
900 
901 	/*
902 	 * reset s_ahg/AHG fields
903 	 *
904 	 * This insures that the ahgentry/ahgcount
905 	 * are at a non-AHG default to protect
906 	 * build_verbs_tx_desc() from using
907 	 * an include ahgidx.
908 	 *
909 	 * build_ahg() will modify as appropriate
910 	 * to use the AHG feature.
911 	 */
912 	priv->s_ahg->tx_flags = 0;
913 	priv->s_ahg->ahgcount = 0;
914 	priv->s_ahg->ahgidx = 0;
915 
916 	/* Make the appropriate header */
917 	hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth2, middle, ps);
918 }
919 
920 /* when sending, force a reschedule every one of these periods */
921 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
922 
923 /**
924  * schedule_send_yield - test for a yield required for QP send engine
925  * @timeout: Final time for timeout slice for jiffies
926  * @qp: a pointer to QP
927  * @ps: a pointer to a structure with commonly lookup values for
928  *      the the send engine progress
929  *
930  * This routine checks if the time slice for the QP has expired
931  * for RC QPs, if so an additional work entry is queued. At this
932  * point, other QPs have an opportunity to be scheduled. It
933  * returns true if a yield is required, otherwise, false
934  * is returned.
935  */
936 static bool schedule_send_yield(struct rvt_qp *qp,
937 				struct hfi1_pkt_state *ps)
938 {
939 	ps->pkts_sent = true;
940 
941 	if (unlikely(time_after(jiffies, ps->timeout))) {
942 		if (!ps->in_thread ||
943 		    workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
944 			spin_lock_irqsave(&qp->s_lock, ps->flags);
945 			qp->s_flags &= ~RVT_S_BUSY;
946 			hfi1_schedule_send(qp);
947 			spin_unlock_irqrestore(&qp->s_lock, ps->flags);
948 			this_cpu_inc(*ps->ppd->dd->send_schedule);
949 			trace_hfi1_rc_expired_time_slice(qp, true);
950 			return true;
951 		}
952 
953 		cond_resched();
954 		this_cpu_inc(*ps->ppd->dd->send_schedule);
955 		ps->timeout = jiffies + ps->timeout_int;
956 	}
957 
958 	trace_hfi1_rc_expired_time_slice(qp, false);
959 	return false;
960 }
961 
962 void hfi1_do_send_from_rvt(struct rvt_qp *qp)
963 {
964 	hfi1_do_send(qp, false);
965 }
966 
967 void _hfi1_do_send(struct work_struct *work)
968 {
969 	struct iowait *wait = container_of(work, struct iowait, iowork);
970 	struct rvt_qp *qp = iowait_to_qp(wait);
971 
972 	hfi1_do_send(qp, true);
973 }
974 
975 /**
976  * hfi1_do_send - perform a send on a QP
977  * @work: contains a pointer to the QP
978  * @in_thread: true if in a workqueue thread
979  *
980  * Process entries in the send work queue until credit or queue is
981  * exhausted.  Only allow one CPU to send a packet per QP.
982  * Otherwise, two threads could send packets out of order.
983  */
984 void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
985 {
986 	struct hfi1_pkt_state ps;
987 	struct hfi1_qp_priv *priv = qp->priv;
988 	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
989 
990 	ps.dev = to_idev(qp->ibqp.device);
991 	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
992 	ps.ppd = ppd_from_ibp(ps.ibp);
993 	ps.in_thread = in_thread;
994 
995 	trace_hfi1_rc_do_send(qp, in_thread);
996 
997 	switch (qp->ibqp.qp_type) {
998 	case IB_QPT_RC:
999 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
1000 				   ~((1 << ps.ppd->lmc) - 1)) ==
1001 				  ps.ppd->lid)) {
1002 			ruc_loopback(qp);
1003 			return;
1004 		}
1005 		make_req = hfi1_make_rc_req;
1006 		ps.timeout_int = qp->timeout_jiffies;
1007 		break;
1008 	case IB_QPT_UC:
1009 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
1010 				   ~((1 << ps.ppd->lmc) - 1)) ==
1011 				  ps.ppd->lid)) {
1012 			ruc_loopback(qp);
1013 			return;
1014 		}
1015 		make_req = hfi1_make_uc_req;
1016 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
1017 		break;
1018 	default:
1019 		make_req = hfi1_make_ud_req;
1020 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
1021 	}
1022 
1023 	spin_lock_irqsave(&qp->s_lock, ps.flags);
1024 
1025 	/* Return if we are already busy processing a work request. */
1026 	if (!hfi1_send_ok(qp)) {
1027 		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
1028 		return;
1029 	}
1030 
1031 	qp->s_flags |= RVT_S_BUSY;
1032 
1033 	ps.timeout_int = ps.timeout_int / 8;
1034 	ps.timeout = jiffies + ps.timeout_int;
1035 	ps.cpu = priv->s_sde ? priv->s_sde->cpu :
1036 			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
1037 	ps.pkts_sent = false;
1038 
1039 	/* insure a pre-built packet is handled  */
1040 	ps.s_txreq = get_waiting_verbs_txreq(qp);
1041 	do {
1042 		/* Check for a constructed packet to be sent. */
1043 		if (qp->s_hdrwords != 0) {
1044 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
1045 			/*
1046 			 * If the packet cannot be sent now, return and
1047 			 * the send engine will be woken up later.
1048 			 */
1049 			if (hfi1_verbs_send(qp, &ps))
1050 				return;
1051 			/* Record that s_ahg is empty. */
1052 			qp->s_hdrwords = 0;
1053 			/* allow other tasks to run */
1054 			if (schedule_send_yield(qp, &ps))
1055 				return;
1056 
1057 			spin_lock_irqsave(&qp->s_lock, ps.flags);
1058 		}
1059 	} while (make_req(qp, &ps));
1060 	iowait_starve_clear(ps.pkts_sent, &priv->s_iowait);
1061 	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
1062 }
1063 
1064 /*
1065  * This should be called with s_lock held.
1066  */
1067 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
1068 			enum ib_wc_status status)
1069 {
1070 	u32 old_last, last;
1071 
1072 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
1073 		return;
1074 
1075 	last = qp->s_last;
1076 	old_last = last;
1077 	trace_hfi1_qp_send_completion(qp, wqe, last);
1078 	if (++last >= qp->s_size)
1079 		last = 0;
1080 	trace_hfi1_qp_send_completion(qp, wqe, last);
1081 	qp->s_last = last;
1082 	/* See post_send() */
1083 	barrier();
1084 	rvt_put_swqe(wqe);
1085 	if (qp->ibqp.qp_type == IB_QPT_UD ||
1086 	    qp->ibqp.qp_type == IB_QPT_SMI ||
1087 	    qp->ibqp.qp_type == IB_QPT_GSI)
1088 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
1089 
1090 	rvt_qp_swqe_complete(qp,
1091 			     wqe,
1092 			     ib_hfi1_wc_opcode[wqe->wr.opcode],
1093 			     status);
1094 
1095 	if (qp->s_acked == old_last)
1096 		qp->s_acked = last;
1097 	if (qp->s_cur == old_last)
1098 		qp->s_cur = last;
1099 	if (qp->s_tail == old_last)
1100 		qp->s_tail = last;
1101 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
1102 		qp->s_draining = 0;
1103 }
1104