xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ruc.c (revision d236d361)
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 		if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
78 				 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
79 			goto bad_lkey;
80 		qp->r_len += wqe->sg_list[i].length;
81 		j++;
82 	}
83 	ss->num_sge = j;
84 	ss->total_len = qp->r_len;
85 	ret = 1;
86 	goto bail;
87 
88 bad_lkey:
89 	while (j) {
90 		struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
91 
92 		rvt_put_mr(sge->mr);
93 	}
94 	ss->num_sge = 0;
95 	memset(&wc, 0, sizeof(wc));
96 	wc.wr_id = wqe->wr_id;
97 	wc.status = IB_WC_LOC_PROT_ERR;
98 	wc.opcode = IB_WC_RECV;
99 	wc.qp = &qp->ibqp;
100 	/* Signal solicited completion event. */
101 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
102 	ret = 0;
103 bail:
104 	return ret;
105 }
106 
107 /**
108  * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE
109  * @qp: the QP
110  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
111  *
112  * Return -1 if there is a local error, 0 if no RWQE is available,
113  * otherwise return 1.
114  *
115  * Can be called from interrupt level.
116  */
117 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only)
118 {
119 	unsigned long flags;
120 	struct rvt_rq *rq;
121 	struct rvt_rwq *wq;
122 	struct rvt_srq *srq;
123 	struct rvt_rwqe *wqe;
124 	void (*handler)(struct ib_event *, void *);
125 	u32 tail;
126 	int ret;
127 
128 	if (qp->ibqp.srq) {
129 		srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
130 		handler = srq->ibsrq.event_handler;
131 		rq = &srq->rq;
132 	} else {
133 		srq = NULL;
134 		handler = NULL;
135 		rq = &qp->r_rq;
136 	}
137 
138 	spin_lock_irqsave(&rq->lock, flags);
139 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
140 		ret = 0;
141 		goto unlock;
142 	}
143 
144 	wq = rq->wq;
145 	tail = wq->tail;
146 	/* Validate tail before using it since it is user writable. */
147 	if (tail >= rq->size)
148 		tail = 0;
149 	if (unlikely(tail == wq->head)) {
150 		ret = 0;
151 		goto unlock;
152 	}
153 	/* Make sure entry is read after head index is read. */
154 	smp_rmb();
155 	wqe = rvt_get_rwqe_ptr(rq, tail);
156 	/*
157 	 * Even though we update the tail index in memory, the verbs
158 	 * consumer is not supposed to post more entries until a
159 	 * completion is generated.
160 	 */
161 	if (++tail >= rq->size)
162 		tail = 0;
163 	wq->tail = tail;
164 	if (!wr_id_only && !init_sge(qp, wqe)) {
165 		ret = -1;
166 		goto unlock;
167 	}
168 	qp->r_wr_id = wqe->wr_id;
169 
170 	ret = 1;
171 	set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
172 	if (handler) {
173 		u32 n;
174 
175 		/*
176 		 * Validate head pointer value and compute
177 		 * the number of remaining WQEs.
178 		 */
179 		n = wq->head;
180 		if (n >= rq->size)
181 			n = 0;
182 		if (n < tail)
183 			n += rq->size - tail;
184 		else
185 			n -= tail;
186 		if (n < srq->limit) {
187 			struct ib_event ev;
188 
189 			srq->limit = 0;
190 			spin_unlock_irqrestore(&rq->lock, flags);
191 			ev.device = qp->ibqp.device;
192 			ev.element.srq = qp->ibqp.srq;
193 			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
194 			handler(&ev, srq->ibsrq.srq_context);
195 			goto bail;
196 		}
197 	}
198 unlock:
199 	spin_unlock_irqrestore(&rq->lock, flags);
200 bail:
201 	return ret;
202 }
203 
204 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
205 {
206 	return (gid->global.interface_id == id &&
207 		(gid->global.subnet_prefix == gid_prefix ||
208 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
209 }
210 
211 /*
212  *
213  * This should be called with the QP r_lock held.
214  *
215  * The s_lock will be acquired around the hfi1_migrate_qp() call.
216  */
217 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct ib_header *hdr,
218 		       int has_grh, struct rvt_qp *qp, u32 bth0)
219 {
220 	__be64 guid;
221 	unsigned long flags;
222 	u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
223 
224 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
225 		if (!has_grh) {
226 			if (rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
227 			    IB_AH_GRH)
228 				goto err;
229 		} else {
230 			const struct ib_global_route *grh;
231 
232 			if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
233 			      IB_AH_GRH))
234 				goto err;
235 			grh = rdma_ah_read_grh(&qp->alt_ah_attr);
236 			guid = get_sguid(ibp, grh->sgid_index);
237 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
238 				    guid))
239 				goto err;
240 			if (!gid_ok(
241 				&hdr->u.l.grh.sgid,
242 				grh->dgid.global.subnet_prefix,
243 				grh->dgid.global.interface_id))
244 				goto err;
245 		}
246 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0, sc5,
247 					    ib_get_slid(hdr)))) {
248 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
249 				       (u16)bth0,
250 				       ib_get_sl(hdr),
251 				       0, qp->ibqp.qp_num,
252 				       ib_get_slid(hdr),
253 				       ib_get_dlid(hdr));
254 			goto err;
255 		}
256 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
257 		if (ib_get_slid(hdr) !=
258 			rdma_ah_get_dlid(&qp->alt_ah_attr) ||
259 		    ppd_from_ibp(ibp)->port !=
260 			rdma_ah_get_port_num(&qp->alt_ah_attr))
261 			goto err;
262 		spin_lock_irqsave(&qp->s_lock, flags);
263 		hfi1_migrate_qp(qp);
264 		spin_unlock_irqrestore(&qp->s_lock, flags);
265 	} else {
266 		if (!has_grh) {
267 			if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
268 						 IB_AH_GRH)
269 				goto err;
270 		} else {
271 			const struct ib_global_route *grh;
272 
273 			if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
274 						   IB_AH_GRH))
275 				goto err;
276 			grh = rdma_ah_read_grh(&qp->remote_ah_attr);
277 			guid = get_sguid(ibp, grh->sgid_index);
278 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
279 				    guid))
280 				goto err;
281 			if (!gid_ok(
282 			     &hdr->u.l.grh.sgid,
283 			     grh->dgid.global.subnet_prefix,
284 			     grh->dgid.global.interface_id))
285 				goto err;
286 		}
287 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0, sc5,
288 					    ib_get_slid(hdr)))) {
289 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
290 				       (u16)bth0,
291 				       ib_get_sl(hdr),
292 				       0, qp->ibqp.qp_num,
293 				       ib_get_slid(hdr),
294 				       ib_get_dlid(hdr));
295 			goto err;
296 		}
297 		/* Validate the SLID. See Ch. 9.6.1.5 */
298 		if (ib_get_slid(hdr) !=
299 			rdma_ah_get_dlid(&qp->remote_ah_attr) ||
300 		    ppd_from_ibp(ibp)->port != qp->port_num)
301 			goto err;
302 		if (qp->s_mig_state == IB_MIG_REARM &&
303 		    !(bth0 & IB_BTH_MIG_REQ))
304 			qp->s_mig_state = IB_MIG_ARMED;
305 	}
306 
307 	return 0;
308 
309 err:
310 	return 1;
311 }
312 
313 /**
314  * ruc_loopback - handle UC and RC loopback requests
315  * @sqp: the sending QP
316  *
317  * This is called from hfi1_do_send() to
318  * forward a WQE addressed to the same HFI.
319  * Note that although we are single threaded due to the send engine, we still
320  * have to protect against post_send().  We don't have to worry about
321  * receive interrupts since this is a connected protocol and all packets
322  * will pass through here.
323  */
324 static void ruc_loopback(struct rvt_qp *sqp)
325 {
326 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
327 	struct rvt_qp *qp;
328 	struct rvt_swqe *wqe;
329 	struct rvt_sge *sge;
330 	unsigned long flags;
331 	struct ib_wc wc;
332 	u64 sdata;
333 	atomic64_t *maddr;
334 	enum ib_wc_status send_status;
335 	bool release;
336 	int ret;
337 	bool copy_last = false;
338 	int local_ops = 0;
339 
340 	rcu_read_lock();
341 
342 	/*
343 	 * Note that we check the responder QP state after
344 	 * checking the requester's state.
345 	 */
346 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
347 			    sqp->remote_qpn);
348 
349 	spin_lock_irqsave(&sqp->s_lock, flags);
350 
351 	/* Return if we are already busy processing a work request. */
352 	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
353 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
354 		goto unlock;
355 
356 	sqp->s_flags |= RVT_S_BUSY;
357 
358 again:
359 	smp_read_barrier_depends(); /* see post_one_send() */
360 	if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
361 		goto clr_busy;
362 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
363 
364 	/* Return if it is not OK to start a new work request. */
365 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
366 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
367 			goto clr_busy;
368 		/* We are in the error state, flush the work request. */
369 		send_status = IB_WC_WR_FLUSH_ERR;
370 		goto flush_send;
371 	}
372 
373 	/*
374 	 * We can rely on the entry not changing without the s_lock
375 	 * being held until we update s_last.
376 	 * We increment s_cur to indicate s_last is in progress.
377 	 */
378 	if (sqp->s_last == sqp->s_cur) {
379 		if (++sqp->s_cur >= sqp->s_size)
380 			sqp->s_cur = 0;
381 	}
382 	spin_unlock_irqrestore(&sqp->s_lock, flags);
383 
384 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
385 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
386 		ibp->rvp.n_pkt_drops++;
387 		/*
388 		 * For RC, the requester would timeout and retry so
389 		 * shortcut the timeouts and just signal too many retries.
390 		 */
391 		if (sqp->ibqp.qp_type == IB_QPT_RC)
392 			send_status = IB_WC_RETRY_EXC_ERR;
393 		else
394 			send_status = IB_WC_SUCCESS;
395 		goto serr;
396 	}
397 
398 	memset(&wc, 0, sizeof(wc));
399 	send_status = IB_WC_SUCCESS;
400 
401 	release = true;
402 	sqp->s_sge.sge = wqe->sg_list[0];
403 	sqp->s_sge.sg_list = wqe->sg_list + 1;
404 	sqp->s_sge.num_sge = wqe->wr.num_sge;
405 	sqp->s_len = wqe->length;
406 	switch (wqe->wr.opcode) {
407 	case IB_WR_REG_MR:
408 		goto send_comp;
409 
410 	case IB_WR_LOCAL_INV:
411 		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
412 			if (rvt_invalidate_rkey(sqp,
413 						wqe->wr.ex.invalidate_rkey))
414 				send_status = IB_WC_LOC_PROT_ERR;
415 			local_ops = 1;
416 		}
417 		goto send_comp;
418 
419 	case IB_WR_SEND_WITH_INV:
420 		if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
421 			wc.wc_flags = IB_WC_WITH_INVALIDATE;
422 			wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
423 		}
424 		goto send;
425 
426 	case IB_WR_SEND_WITH_IMM:
427 		wc.wc_flags = IB_WC_WITH_IMM;
428 		wc.ex.imm_data = wqe->wr.ex.imm_data;
429 		/* FALLTHROUGH */
430 	case IB_WR_SEND:
431 send:
432 		ret = hfi1_rvt_get_rwqe(qp, 0);
433 		if (ret < 0)
434 			goto op_err;
435 		if (!ret)
436 			goto rnr_nak;
437 		break;
438 
439 	case IB_WR_RDMA_WRITE_WITH_IMM:
440 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
441 			goto inv_err;
442 		wc.wc_flags = IB_WC_WITH_IMM;
443 		wc.ex.imm_data = wqe->wr.ex.imm_data;
444 		ret = hfi1_rvt_get_rwqe(qp, 1);
445 		if (ret < 0)
446 			goto op_err;
447 		if (!ret)
448 			goto rnr_nak;
449 		/* skip copy_last set and qp_access_flags recheck */
450 		goto do_write;
451 	case IB_WR_RDMA_WRITE:
452 		copy_last = rvt_is_user_qp(qp);
453 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
454 			goto inv_err;
455 do_write:
456 		if (wqe->length == 0)
457 			break;
458 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
459 					  wqe->rdma_wr.remote_addr,
460 					  wqe->rdma_wr.rkey,
461 					  IB_ACCESS_REMOTE_WRITE)))
462 			goto acc_err;
463 		qp->r_sge.sg_list = NULL;
464 		qp->r_sge.num_sge = 1;
465 		qp->r_sge.total_len = wqe->length;
466 		break;
467 
468 	case IB_WR_RDMA_READ:
469 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
470 			goto inv_err;
471 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
472 					  wqe->rdma_wr.remote_addr,
473 					  wqe->rdma_wr.rkey,
474 					  IB_ACCESS_REMOTE_READ)))
475 			goto acc_err;
476 		release = false;
477 		sqp->s_sge.sg_list = NULL;
478 		sqp->s_sge.num_sge = 1;
479 		qp->r_sge.sge = wqe->sg_list[0];
480 		qp->r_sge.sg_list = wqe->sg_list + 1;
481 		qp->r_sge.num_sge = wqe->wr.num_sge;
482 		qp->r_sge.total_len = wqe->length;
483 		break;
484 
485 	case IB_WR_ATOMIC_CMP_AND_SWP:
486 	case IB_WR_ATOMIC_FETCH_AND_ADD:
487 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
488 			goto inv_err;
489 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
490 					  wqe->atomic_wr.remote_addr,
491 					  wqe->atomic_wr.rkey,
492 					  IB_ACCESS_REMOTE_ATOMIC)))
493 			goto acc_err;
494 		/* Perform atomic OP and save result. */
495 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
496 		sdata = wqe->atomic_wr.compare_add;
497 		*(u64 *)sqp->s_sge.sge.vaddr =
498 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
499 			(u64)atomic64_add_return(sdata, maddr) - sdata :
500 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
501 				      sdata, wqe->atomic_wr.swap);
502 		rvt_put_mr(qp->r_sge.sge.mr);
503 		qp->r_sge.num_sge = 0;
504 		goto send_comp;
505 
506 	default:
507 		send_status = IB_WC_LOC_QP_OP_ERR;
508 		goto serr;
509 	}
510 
511 	sge = &sqp->s_sge.sge;
512 	while (sqp->s_len) {
513 		u32 len = sqp->s_len;
514 
515 		if (len > sge->length)
516 			len = sge->length;
517 		if (len > sge->sge_length)
518 			len = sge->sge_length;
519 		WARN_ON_ONCE(len == 0);
520 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
521 		sge->vaddr += len;
522 		sge->length -= len;
523 		sge->sge_length -= len;
524 		if (sge->sge_length == 0) {
525 			if (!release)
526 				rvt_put_mr(sge->mr);
527 			if (--sqp->s_sge.num_sge)
528 				*sge = *sqp->s_sge.sg_list++;
529 		} else if (sge->length == 0 && sge->mr->lkey) {
530 			if (++sge->n >= RVT_SEGSZ) {
531 				if (++sge->m >= sge->mr->mapsz)
532 					break;
533 				sge->n = 0;
534 			}
535 			sge->vaddr =
536 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
537 			sge->length =
538 				sge->mr->map[sge->m]->segs[sge->n].length;
539 		}
540 		sqp->s_len -= len;
541 	}
542 	if (release)
543 		rvt_put_ss(&qp->r_sge);
544 
545 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
546 		goto send_comp;
547 
548 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
549 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
550 	else
551 		wc.opcode = IB_WC_RECV;
552 	wc.wr_id = qp->r_wr_id;
553 	wc.status = IB_WC_SUCCESS;
554 	wc.byte_len = wqe->length;
555 	wc.qp = &qp->ibqp;
556 	wc.src_qp = qp->remote_qpn;
557 	wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
558 	wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
559 	wc.port_num = 1;
560 	/* Signal completion event if the solicited bit is set. */
561 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
562 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
563 
564 send_comp:
565 	spin_lock_irqsave(&sqp->s_lock, flags);
566 	ibp->rvp.n_loop_pkts++;
567 flush_send:
568 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
569 	hfi1_send_complete(sqp, wqe, send_status);
570 	if (local_ops) {
571 		atomic_dec(&sqp->local_ops_pending);
572 		local_ops = 0;
573 	}
574 	goto again;
575 
576 rnr_nak:
577 	/* Handle RNR NAK */
578 	if (qp->ibqp.qp_type == IB_QPT_UC)
579 		goto send_comp;
580 	ibp->rvp.n_rnr_naks++;
581 	/*
582 	 * Note: we don't need the s_lock held since the BUSY flag
583 	 * makes this single threaded.
584 	 */
585 	if (sqp->s_rnr_retry == 0) {
586 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
587 		goto serr;
588 	}
589 	if (sqp->s_rnr_retry_cnt < 7)
590 		sqp->s_rnr_retry--;
591 	spin_lock_irqsave(&sqp->s_lock, flags);
592 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
593 		goto clr_busy;
594 	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
595 				IB_AETH_CREDIT_SHIFT);
596 	goto clr_busy;
597 
598 op_err:
599 	send_status = IB_WC_REM_OP_ERR;
600 	wc.status = IB_WC_LOC_QP_OP_ERR;
601 	goto err;
602 
603 inv_err:
604 	send_status = IB_WC_REM_INV_REQ_ERR;
605 	wc.status = IB_WC_LOC_QP_OP_ERR;
606 	goto err;
607 
608 acc_err:
609 	send_status = IB_WC_REM_ACCESS_ERR;
610 	wc.status = IB_WC_LOC_PROT_ERR;
611 err:
612 	/* responder goes to error state */
613 	rvt_rc_error(qp, wc.status);
614 
615 serr:
616 	spin_lock_irqsave(&sqp->s_lock, flags);
617 	hfi1_send_complete(sqp, wqe, send_status);
618 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
619 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
620 
621 		sqp->s_flags &= ~RVT_S_BUSY;
622 		spin_unlock_irqrestore(&sqp->s_lock, flags);
623 		if (lastwqe) {
624 			struct ib_event ev;
625 
626 			ev.device = sqp->ibqp.device;
627 			ev.element.qp = &sqp->ibqp;
628 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
629 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
630 		}
631 		goto done;
632 	}
633 clr_busy:
634 	sqp->s_flags &= ~RVT_S_BUSY;
635 unlock:
636 	spin_unlock_irqrestore(&sqp->s_lock, flags);
637 done:
638 	rcu_read_unlock();
639 }
640 
641 /**
642  * hfi1_make_grh - construct a GRH header
643  * @ibp: a pointer to the IB port
644  * @hdr: a pointer to the GRH header being constructed
645  * @grh: the global route address to send to
646  * @hwords: the number of 32 bit words of header being sent
647  * @nwords: the number of 32 bit words of data being sent
648  *
649  * Return the size of the header in 32 bit words.
650  */
651 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
652 		  const struct ib_global_route *grh, u32 hwords, u32 nwords)
653 {
654 	hdr->version_tclass_flow =
655 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
656 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
657 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
658 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
659 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
660 	hdr->next_hdr = IB_GRH_NEXT_HDR;
661 	hdr->hop_limit = grh->hop_limit;
662 	/* The SGID is 32-bit aligned. */
663 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
664 	hdr->sgid.global.interface_id =
665 		grh->sgid_index < HFI1_GUIDS_PER_PORT ?
666 		get_sguid(ibp, grh->sgid_index) :
667 		get_sguid(ibp, HFI1_PORT_GUID_INDEX);
668 	hdr->dgid = grh->dgid;
669 
670 	/* GRH header size in 32-bit words. */
671 	return sizeof(struct ib_grh) / sizeof(u32);
672 }
673 
674 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, hdr.u.oth.bth[2]) / 4)
675 
676 /**
677  * build_ahg - create ahg in s_ahg
678  * @qp: a pointer to QP
679  * @npsn: the next PSN for the request/response
680  *
681  * This routine handles the AHG by allocating an ahg entry and causing the
682  * copy of the first middle.
683  *
684  * Subsequent middles use the copied entry, editing the
685  * PSN with 1 or 2 edits.
686  */
687 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
688 {
689 	struct hfi1_qp_priv *priv = qp->priv;
690 
691 	if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
692 		clear_ahg(qp);
693 	if (!(qp->s_flags & RVT_S_AHG_VALID)) {
694 		/* first middle that needs copy  */
695 		if (qp->s_ahgidx < 0)
696 			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
697 		if (qp->s_ahgidx >= 0) {
698 			qp->s_ahgpsn = npsn;
699 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
700 			/* save to protect a change in another thread */
701 			priv->s_ahg->ahgidx = qp->s_ahgidx;
702 			qp->s_flags |= RVT_S_AHG_VALID;
703 		}
704 	} else {
705 		/* subsequent middle after valid */
706 		if (qp->s_ahgidx >= 0) {
707 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
708 			priv->s_ahg->ahgidx = qp->s_ahgidx;
709 			priv->s_ahg->ahgcount++;
710 			priv->s_ahg->ahgdesc[0] =
711 				sdma_build_ahg_descriptor(
712 					(__force u16)cpu_to_be16((u16)npsn),
713 					BTH2_OFFSET,
714 					16,
715 					16);
716 			if ((npsn & 0xffff0000) !=
717 					(qp->s_ahgpsn & 0xffff0000)) {
718 				priv->s_ahg->ahgcount++;
719 				priv->s_ahg->ahgdesc[1] =
720 					sdma_build_ahg_descriptor(
721 						(__force u16)cpu_to_be16(
722 							(u16)(npsn >> 16)),
723 						BTH2_OFFSET,
724 						0,
725 						16);
726 			}
727 		}
728 	}
729 }
730 
731 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
732 			  u32 bth0, u32 bth2, int middle,
733 			  struct hfi1_pkt_state *ps)
734 {
735 	struct hfi1_qp_priv *priv = qp->priv;
736 	struct hfi1_ibport *ibp = ps->ibp;
737 	u16 lrh0;
738 	u32 nwords;
739 	u32 extra_bytes;
740 	u32 bth1;
741 
742 	/* Construct the header. */
743 	extra_bytes = -ps->s_txreq->s_cur_size & 3;
744 	nwords = (ps->s_txreq->s_cur_size + extra_bytes) >> 2;
745 	lrh0 = HFI1_LRH_BTH;
746 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
747 		qp->s_hdrwords +=
748 			hfi1_make_grh(ibp,
749 				      &ps->s_txreq->phdr.hdr.u.l.grh,
750 				      rdma_ah_read_grh(&qp->remote_ah_attr),
751 				      qp->s_hdrwords, nwords);
752 		lrh0 = HFI1_LRH_GRH;
753 		middle = 0;
754 	}
755 	lrh0 |= (priv->s_sc & 0xf) << 12 |
756 		(rdma_ah_get_sl(&qp->remote_ah_attr) & 0xf) << 4;
757 	/*
758 	 * reset s_ahg/AHG fields
759 	 *
760 	 * This insures that the ahgentry/ahgcount
761 	 * are at a non-AHG default to protect
762 	 * build_verbs_tx_desc() from using
763 	 * an include ahgidx.
764 	 *
765 	 * build_ahg() will modify as appropriate
766 	 * to use the AHG feature.
767 	 */
768 	priv->s_ahg->tx_flags = 0;
769 	priv->s_ahg->ahgcount = 0;
770 	priv->s_ahg->ahgidx = 0;
771 	if (qp->s_mig_state == IB_MIG_MIGRATED)
772 		bth0 |= IB_BTH_MIG_REQ;
773 	else
774 		middle = 0;
775 	if (middle)
776 		build_ahg(qp, bth2);
777 	else
778 		qp->s_flags &= ~RVT_S_AHG_VALID;
779 	ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
780 	ps->s_txreq->phdr.hdr.lrh[1] =
781 		cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr));
782 	ps->s_txreq->phdr.hdr.lrh[2] =
783 		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
784 	ps->s_txreq->phdr.hdr.lrh[3] =
785 		cpu_to_be16(ppd_from_ibp(ibp)->lid |
786 			    rdma_ah_get_path_bits(&qp->remote_ah_attr));
787 	bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
788 	bth0 |= extra_bytes << 20;
789 	ohdr->bth[0] = cpu_to_be32(bth0);
790 	bth1 = qp->remote_qpn;
791 	if (qp->s_flags & RVT_S_ECN) {
792 		qp->s_flags &= ~RVT_S_ECN;
793 		/* we recently received a FECN, so return a BECN */
794 		bth1 |= (IB_BECN_MASK << IB_BECN_SHIFT);
795 	}
796 	ohdr->bth[1] = cpu_to_be32(bth1);
797 	ohdr->bth[2] = cpu_to_be32(bth2);
798 }
799 
800 /* when sending, force a reschedule every one of these periods */
801 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
802 
803 /**
804  * schedule_send_yield - test for a yield required for QP send engine
805  * @timeout: Final time for timeout slice for jiffies
806  * @qp: a pointer to QP
807  * @ps: a pointer to a structure with commonly lookup values for
808  *      the the send engine progress
809  *
810  * This routine checks if the time slice for the QP has expired
811  * for RC QPs, if so an additional work entry is queued. At this
812  * point, other QPs have an opportunity to be scheduled. It
813  * returns true if a yield is required, otherwise, false
814  * is returned.
815  */
816 static bool schedule_send_yield(struct rvt_qp *qp,
817 				struct hfi1_pkt_state *ps)
818 {
819 	if (unlikely(time_after(jiffies, ps->timeout))) {
820 		if (!ps->in_thread ||
821 		    workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
822 			spin_lock_irqsave(&qp->s_lock, ps->flags);
823 			qp->s_flags &= ~RVT_S_BUSY;
824 			hfi1_schedule_send(qp);
825 			spin_unlock_irqrestore(&qp->s_lock, ps->flags);
826 			this_cpu_inc(*ps->ppd->dd->send_schedule);
827 			trace_hfi1_rc_expired_time_slice(qp, true);
828 			return true;
829 		}
830 
831 		cond_resched();
832 		this_cpu_inc(*ps->ppd->dd->send_schedule);
833 		ps->timeout = jiffies + ps->timeout_int;
834 	}
835 
836 	trace_hfi1_rc_expired_time_slice(qp, false);
837 	return false;
838 }
839 
840 void hfi1_do_send_from_rvt(struct rvt_qp *qp)
841 {
842 	hfi1_do_send(qp, false);
843 }
844 
845 void _hfi1_do_send(struct work_struct *work)
846 {
847 	struct iowait *wait = container_of(work, struct iowait, iowork);
848 	struct rvt_qp *qp = iowait_to_qp(wait);
849 
850 	hfi1_do_send(qp, true);
851 }
852 
853 /**
854  * hfi1_do_send - perform a send on a QP
855  * @work: contains a pointer to the QP
856  * @in_thread: true if in a workqueue thread
857  *
858  * Process entries in the send work queue until credit or queue is
859  * exhausted.  Only allow one CPU to send a packet per QP.
860  * Otherwise, two threads could send packets out of order.
861  */
862 void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
863 {
864 	struct hfi1_pkt_state ps;
865 	struct hfi1_qp_priv *priv = qp->priv;
866 	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
867 
868 	ps.dev = to_idev(qp->ibqp.device);
869 	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
870 	ps.ppd = ppd_from_ibp(ps.ibp);
871 	ps.in_thread = in_thread;
872 
873 	trace_hfi1_rc_do_send(qp, in_thread);
874 
875 	switch (qp->ibqp.qp_type) {
876 	case IB_QPT_RC:
877 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
878 				   ~((1 << ps.ppd->lmc) - 1)) ==
879 				  ps.ppd->lid)) {
880 			ruc_loopback(qp);
881 			return;
882 		}
883 		make_req = hfi1_make_rc_req;
884 		ps.timeout_int = qp->timeout_jiffies;
885 		break;
886 	case IB_QPT_UC:
887 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
888 				   ~((1 << ps.ppd->lmc) - 1)) ==
889 				  ps.ppd->lid)) {
890 			ruc_loopback(qp);
891 			return;
892 		}
893 		make_req = hfi1_make_uc_req;
894 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
895 		break;
896 	default:
897 		make_req = hfi1_make_ud_req;
898 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
899 	}
900 
901 	spin_lock_irqsave(&qp->s_lock, ps.flags);
902 
903 	/* Return if we are already busy processing a work request. */
904 	if (!hfi1_send_ok(qp)) {
905 		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
906 		return;
907 	}
908 
909 	qp->s_flags |= RVT_S_BUSY;
910 
911 	ps.timeout_int = ps.timeout_int / 8;
912 	ps.timeout = jiffies + ps.timeout_int;
913 	ps.cpu = priv->s_sde ? priv->s_sde->cpu :
914 			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
915 
916 	/* insure a pre-built packet is handled  */
917 	ps.s_txreq = get_waiting_verbs_txreq(qp);
918 	do {
919 		/* Check for a constructed packet to be sent. */
920 		if (qp->s_hdrwords != 0) {
921 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
922 			/*
923 			 * If the packet cannot be sent now, return and
924 			 * the send engine will be woken up later.
925 			 */
926 			if (hfi1_verbs_send(qp, &ps))
927 				return;
928 			/* Record that s_ahg is empty. */
929 			qp->s_hdrwords = 0;
930 			/* allow other tasks to run */
931 			if (schedule_send_yield(qp, &ps))
932 				return;
933 
934 			spin_lock_irqsave(&qp->s_lock, ps.flags);
935 		}
936 	} while (make_req(qp, &ps));
937 
938 	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
939 }
940 
941 /*
942  * This should be called with s_lock held.
943  */
944 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
945 			enum ib_wc_status status)
946 {
947 	u32 old_last, last;
948 
949 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
950 		return;
951 
952 	last = qp->s_last;
953 	old_last = last;
954 	trace_hfi1_qp_send_completion(qp, wqe, last);
955 	if (++last >= qp->s_size)
956 		last = 0;
957 	trace_hfi1_qp_send_completion(qp, wqe, last);
958 	qp->s_last = last;
959 	/* See post_send() */
960 	barrier();
961 	rvt_put_swqe(wqe);
962 	if (qp->ibqp.qp_type == IB_QPT_UD ||
963 	    qp->ibqp.qp_type == IB_QPT_SMI ||
964 	    qp->ibqp.qp_type == IB_QPT_GSI)
965 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
966 
967 	rvt_qp_swqe_complete(qp,
968 			     wqe,
969 			     ib_hfi1_wc_opcode[wqe->wr.opcode],
970 			     status);
971 
972 	if (qp->s_acked == old_last)
973 		qp->s_acked = last;
974 	if (qp->s_cur == old_last)
975 		qp->s_cur = last;
976 	if (qp->s_tail == old_last)
977 		qp->s_tail = last;
978 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
979 		qp->s_draining = 0;
980 }
981