xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ruc.c (revision 4ed91d48259d9ddd378424d008f2e6559f7e78f8)
1 /*
2  * Copyright(c) 2015, 2016 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[qp->remote_ah_attr.sl];
223 
224 	if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
225 		if (!has_grh) {
226 			if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
227 				goto err;
228 		} else {
229 			if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
230 				goto err;
231 			guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
232 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
233 				    guid))
234 				goto err;
235 			if (!gid_ok(
236 				&hdr->u.l.grh.sgid,
237 				qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
238 				qp->alt_ah_attr.grh.dgid.global.interface_id))
239 				goto err;
240 		}
241 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
242 					    sc5, be16_to_cpu(hdr->lrh[3])))) {
243 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
244 				       (u16)bth0,
245 				       (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
246 				       0, qp->ibqp.qp_num,
247 				       be16_to_cpu(hdr->lrh[3]),
248 				       be16_to_cpu(hdr->lrh[1]));
249 			goto err;
250 		}
251 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
252 		if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
253 		    ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
254 			goto err;
255 		spin_lock_irqsave(&qp->s_lock, flags);
256 		hfi1_migrate_qp(qp);
257 		spin_unlock_irqrestore(&qp->s_lock, flags);
258 	} else {
259 		if (!has_grh) {
260 			if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
261 				goto err;
262 		} else {
263 			if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
264 				goto err;
265 			guid = get_sguid(ibp,
266 					 qp->remote_ah_attr.grh.sgid_index);
267 			if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
268 				    guid))
269 				goto err;
270 			if (!gid_ok(
271 			     &hdr->u.l.grh.sgid,
272 			     qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
273 			     qp->remote_ah_attr.grh.dgid.global.interface_id))
274 				goto err;
275 		}
276 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
277 					    sc5, be16_to_cpu(hdr->lrh[3])))) {
278 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
279 				       (u16)bth0,
280 				       (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
281 				       0, qp->ibqp.qp_num,
282 				       be16_to_cpu(hdr->lrh[3]),
283 				       be16_to_cpu(hdr->lrh[1]));
284 			goto err;
285 		}
286 		/* Validate the SLID. See Ch. 9.6.1.5 */
287 		if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
288 		    ppd_from_ibp(ibp)->port != qp->port_num)
289 			goto err;
290 		if (qp->s_mig_state == IB_MIG_REARM &&
291 		    !(bth0 & IB_BTH_MIG_REQ))
292 			qp->s_mig_state = IB_MIG_ARMED;
293 	}
294 
295 	return 0;
296 
297 err:
298 	return 1;
299 }
300 
301 /**
302  * ruc_loopback - handle UC and RC loopback requests
303  * @sqp: the sending QP
304  *
305  * This is called from hfi1_do_send() to
306  * forward a WQE addressed to the same HFI.
307  * Note that although we are single threaded due to the send engine, we still
308  * have to protect against post_send().  We don't have to worry about
309  * receive interrupts since this is a connected protocol and all packets
310  * will pass through here.
311  */
312 static void ruc_loopback(struct rvt_qp *sqp)
313 {
314 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
315 	struct rvt_qp *qp;
316 	struct rvt_swqe *wqe;
317 	struct rvt_sge *sge;
318 	unsigned long flags;
319 	struct ib_wc wc;
320 	u64 sdata;
321 	atomic64_t *maddr;
322 	enum ib_wc_status send_status;
323 	bool release;
324 	int ret;
325 	bool copy_last = false;
326 	int local_ops = 0;
327 
328 	rcu_read_lock();
329 
330 	/*
331 	 * Note that we check the responder QP state after
332 	 * checking the requester's state.
333 	 */
334 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
335 			    sqp->remote_qpn);
336 
337 	spin_lock_irqsave(&sqp->s_lock, flags);
338 
339 	/* Return if we are already busy processing a work request. */
340 	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
341 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
342 		goto unlock;
343 
344 	sqp->s_flags |= RVT_S_BUSY;
345 
346 again:
347 	smp_read_barrier_depends(); /* see post_one_send() */
348 	if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
349 		goto clr_busy;
350 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
351 
352 	/* Return if it is not OK to start a new work request. */
353 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
354 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
355 			goto clr_busy;
356 		/* We are in the error state, flush the work request. */
357 		send_status = IB_WC_WR_FLUSH_ERR;
358 		goto flush_send;
359 	}
360 
361 	/*
362 	 * We can rely on the entry not changing without the s_lock
363 	 * being held until we update s_last.
364 	 * We increment s_cur to indicate s_last is in progress.
365 	 */
366 	if (sqp->s_last == sqp->s_cur) {
367 		if (++sqp->s_cur >= sqp->s_size)
368 			sqp->s_cur = 0;
369 	}
370 	spin_unlock_irqrestore(&sqp->s_lock, flags);
371 
372 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
373 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
374 		ibp->rvp.n_pkt_drops++;
375 		/*
376 		 * For RC, the requester would timeout and retry so
377 		 * shortcut the timeouts and just signal too many retries.
378 		 */
379 		if (sqp->ibqp.qp_type == IB_QPT_RC)
380 			send_status = IB_WC_RETRY_EXC_ERR;
381 		else
382 			send_status = IB_WC_SUCCESS;
383 		goto serr;
384 	}
385 
386 	memset(&wc, 0, sizeof(wc));
387 	send_status = IB_WC_SUCCESS;
388 
389 	release = true;
390 	sqp->s_sge.sge = wqe->sg_list[0];
391 	sqp->s_sge.sg_list = wqe->sg_list + 1;
392 	sqp->s_sge.num_sge = wqe->wr.num_sge;
393 	sqp->s_len = wqe->length;
394 	switch (wqe->wr.opcode) {
395 	case IB_WR_REG_MR:
396 		goto send_comp;
397 
398 	case IB_WR_LOCAL_INV:
399 		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
400 			if (rvt_invalidate_rkey(sqp,
401 						wqe->wr.ex.invalidate_rkey))
402 				send_status = IB_WC_LOC_PROT_ERR;
403 			local_ops = 1;
404 		}
405 		goto send_comp;
406 
407 	case IB_WR_SEND_WITH_INV:
408 		if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
409 			wc.wc_flags = IB_WC_WITH_INVALIDATE;
410 			wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
411 		}
412 		goto send;
413 
414 	case IB_WR_SEND_WITH_IMM:
415 		wc.wc_flags = IB_WC_WITH_IMM;
416 		wc.ex.imm_data = wqe->wr.ex.imm_data;
417 		/* FALLTHROUGH */
418 	case IB_WR_SEND:
419 send:
420 		ret = hfi1_rvt_get_rwqe(qp, 0);
421 		if (ret < 0)
422 			goto op_err;
423 		if (!ret)
424 			goto rnr_nak;
425 		break;
426 
427 	case IB_WR_RDMA_WRITE_WITH_IMM:
428 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
429 			goto inv_err;
430 		wc.wc_flags = IB_WC_WITH_IMM;
431 		wc.ex.imm_data = wqe->wr.ex.imm_data;
432 		ret = hfi1_rvt_get_rwqe(qp, 1);
433 		if (ret < 0)
434 			goto op_err;
435 		if (!ret)
436 			goto rnr_nak;
437 		/* skip copy_last set and qp_access_flags recheck */
438 		goto do_write;
439 	case IB_WR_RDMA_WRITE:
440 		copy_last = rvt_is_user_qp(qp);
441 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
442 			goto inv_err;
443 do_write:
444 		if (wqe->length == 0)
445 			break;
446 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
447 					  wqe->rdma_wr.remote_addr,
448 					  wqe->rdma_wr.rkey,
449 					  IB_ACCESS_REMOTE_WRITE)))
450 			goto acc_err;
451 		qp->r_sge.sg_list = NULL;
452 		qp->r_sge.num_sge = 1;
453 		qp->r_sge.total_len = wqe->length;
454 		break;
455 
456 	case IB_WR_RDMA_READ:
457 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
458 			goto inv_err;
459 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
460 					  wqe->rdma_wr.remote_addr,
461 					  wqe->rdma_wr.rkey,
462 					  IB_ACCESS_REMOTE_READ)))
463 			goto acc_err;
464 		release = false;
465 		sqp->s_sge.sg_list = NULL;
466 		sqp->s_sge.num_sge = 1;
467 		qp->r_sge.sge = wqe->sg_list[0];
468 		qp->r_sge.sg_list = wqe->sg_list + 1;
469 		qp->r_sge.num_sge = wqe->wr.num_sge;
470 		qp->r_sge.total_len = wqe->length;
471 		break;
472 
473 	case IB_WR_ATOMIC_CMP_AND_SWP:
474 	case IB_WR_ATOMIC_FETCH_AND_ADD:
475 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
476 			goto inv_err;
477 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
478 					  wqe->atomic_wr.remote_addr,
479 					  wqe->atomic_wr.rkey,
480 					  IB_ACCESS_REMOTE_ATOMIC)))
481 			goto acc_err;
482 		/* Perform atomic OP and save result. */
483 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
484 		sdata = wqe->atomic_wr.compare_add;
485 		*(u64 *)sqp->s_sge.sge.vaddr =
486 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
487 			(u64)atomic64_add_return(sdata, maddr) - sdata :
488 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
489 				      sdata, wqe->atomic_wr.swap);
490 		rvt_put_mr(qp->r_sge.sge.mr);
491 		qp->r_sge.num_sge = 0;
492 		goto send_comp;
493 
494 	default:
495 		send_status = IB_WC_LOC_QP_OP_ERR;
496 		goto serr;
497 	}
498 
499 	sge = &sqp->s_sge.sge;
500 	while (sqp->s_len) {
501 		u32 len = sqp->s_len;
502 
503 		if (len > sge->length)
504 			len = sge->length;
505 		if (len > sge->sge_length)
506 			len = sge->sge_length;
507 		WARN_ON_ONCE(len == 0);
508 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
509 		sge->vaddr += len;
510 		sge->length -= len;
511 		sge->sge_length -= len;
512 		if (sge->sge_length == 0) {
513 			if (!release)
514 				rvt_put_mr(sge->mr);
515 			if (--sqp->s_sge.num_sge)
516 				*sge = *sqp->s_sge.sg_list++;
517 		} else if (sge->length == 0 && sge->mr->lkey) {
518 			if (++sge->n >= RVT_SEGSZ) {
519 				if (++sge->m >= sge->mr->mapsz)
520 					break;
521 				sge->n = 0;
522 			}
523 			sge->vaddr =
524 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
525 			sge->length =
526 				sge->mr->map[sge->m]->segs[sge->n].length;
527 		}
528 		sqp->s_len -= len;
529 	}
530 	if (release)
531 		rvt_put_ss(&qp->r_sge);
532 
533 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
534 		goto send_comp;
535 
536 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
537 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
538 	else
539 		wc.opcode = IB_WC_RECV;
540 	wc.wr_id = qp->r_wr_id;
541 	wc.status = IB_WC_SUCCESS;
542 	wc.byte_len = wqe->length;
543 	wc.qp = &qp->ibqp;
544 	wc.src_qp = qp->remote_qpn;
545 	wc.slid = qp->remote_ah_attr.dlid;
546 	wc.sl = qp->remote_ah_attr.sl;
547 	wc.port_num = 1;
548 	/* Signal completion event if the solicited bit is set. */
549 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
550 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
551 
552 send_comp:
553 	spin_lock_irqsave(&sqp->s_lock, flags);
554 	ibp->rvp.n_loop_pkts++;
555 flush_send:
556 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
557 	hfi1_send_complete(sqp, wqe, send_status);
558 	if (local_ops) {
559 		atomic_dec(&sqp->local_ops_pending);
560 		local_ops = 0;
561 	}
562 	goto again;
563 
564 rnr_nak:
565 	/* Handle RNR NAK */
566 	if (qp->ibqp.qp_type == IB_QPT_UC)
567 		goto send_comp;
568 	ibp->rvp.n_rnr_naks++;
569 	/*
570 	 * Note: we don't need the s_lock held since the BUSY flag
571 	 * makes this single threaded.
572 	 */
573 	if (sqp->s_rnr_retry == 0) {
574 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
575 		goto serr;
576 	}
577 	if (sqp->s_rnr_retry_cnt < 7)
578 		sqp->s_rnr_retry--;
579 	spin_lock_irqsave(&sqp->s_lock, flags);
580 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
581 		goto clr_busy;
582 	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
583 				IB_AETH_CREDIT_SHIFT);
584 	goto clr_busy;
585 
586 op_err:
587 	send_status = IB_WC_REM_OP_ERR;
588 	wc.status = IB_WC_LOC_QP_OP_ERR;
589 	goto err;
590 
591 inv_err:
592 	send_status = IB_WC_REM_INV_REQ_ERR;
593 	wc.status = IB_WC_LOC_QP_OP_ERR;
594 	goto err;
595 
596 acc_err:
597 	send_status = IB_WC_REM_ACCESS_ERR;
598 	wc.status = IB_WC_LOC_PROT_ERR;
599 err:
600 	/* responder goes to error state */
601 	rvt_rc_error(qp, wc.status);
602 
603 serr:
604 	spin_lock_irqsave(&sqp->s_lock, flags);
605 	hfi1_send_complete(sqp, wqe, send_status);
606 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
607 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
608 
609 		sqp->s_flags &= ~RVT_S_BUSY;
610 		spin_unlock_irqrestore(&sqp->s_lock, flags);
611 		if (lastwqe) {
612 			struct ib_event ev;
613 
614 			ev.device = sqp->ibqp.device;
615 			ev.element.qp = &sqp->ibqp;
616 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
617 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
618 		}
619 		goto done;
620 	}
621 clr_busy:
622 	sqp->s_flags &= ~RVT_S_BUSY;
623 unlock:
624 	spin_unlock_irqrestore(&sqp->s_lock, flags);
625 done:
626 	rcu_read_unlock();
627 }
628 
629 /**
630  * hfi1_make_grh - construct a GRH header
631  * @ibp: a pointer to the IB port
632  * @hdr: a pointer to the GRH header being constructed
633  * @grh: the global route address to send to
634  * @hwords: the number of 32 bit words of header being sent
635  * @nwords: the number of 32 bit words of data being sent
636  *
637  * Return the size of the header in 32 bit words.
638  */
639 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
640 		  struct ib_global_route *grh, u32 hwords, u32 nwords)
641 {
642 	hdr->version_tclass_flow =
643 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
644 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
645 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
646 	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
647 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
648 	hdr->next_hdr = IB_GRH_NEXT_HDR;
649 	hdr->hop_limit = grh->hop_limit;
650 	/* The SGID is 32-bit aligned. */
651 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
652 	hdr->sgid.global.interface_id =
653 		grh->sgid_index < HFI1_GUIDS_PER_PORT ?
654 		get_sguid(ibp, grh->sgid_index) :
655 		get_sguid(ibp, HFI1_PORT_GUID_INDEX);
656 	hdr->dgid = grh->dgid;
657 
658 	/* GRH header size in 32-bit words. */
659 	return sizeof(struct ib_grh) / sizeof(u32);
660 }
661 
662 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, hdr.u.oth.bth[2]) / 4)
663 
664 /**
665  * build_ahg - create ahg in s_ahg
666  * @qp: a pointer to QP
667  * @npsn: the next PSN for the request/response
668  *
669  * This routine handles the AHG by allocating an ahg entry and causing the
670  * copy of the first middle.
671  *
672  * Subsequent middles use the copied entry, editing the
673  * PSN with 1 or 2 edits.
674  */
675 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
676 {
677 	struct hfi1_qp_priv *priv = qp->priv;
678 
679 	if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
680 		clear_ahg(qp);
681 	if (!(qp->s_flags & RVT_S_AHG_VALID)) {
682 		/* first middle that needs copy  */
683 		if (qp->s_ahgidx < 0)
684 			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
685 		if (qp->s_ahgidx >= 0) {
686 			qp->s_ahgpsn = npsn;
687 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
688 			/* save to protect a change in another thread */
689 			priv->s_ahg->ahgidx = qp->s_ahgidx;
690 			qp->s_flags |= RVT_S_AHG_VALID;
691 		}
692 	} else {
693 		/* subsequent middle after valid */
694 		if (qp->s_ahgidx >= 0) {
695 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
696 			priv->s_ahg->ahgidx = qp->s_ahgidx;
697 			priv->s_ahg->ahgcount++;
698 			priv->s_ahg->ahgdesc[0] =
699 				sdma_build_ahg_descriptor(
700 					(__force u16)cpu_to_be16((u16)npsn),
701 					BTH2_OFFSET,
702 					16,
703 					16);
704 			if ((npsn & 0xffff0000) !=
705 					(qp->s_ahgpsn & 0xffff0000)) {
706 				priv->s_ahg->ahgcount++;
707 				priv->s_ahg->ahgdesc[1] =
708 					sdma_build_ahg_descriptor(
709 						(__force u16)cpu_to_be16(
710 							(u16)(npsn >> 16)),
711 						BTH2_OFFSET,
712 						0,
713 						16);
714 			}
715 		}
716 	}
717 }
718 
719 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
720 			  u32 bth0, u32 bth2, int middle,
721 			  struct hfi1_pkt_state *ps)
722 {
723 	struct hfi1_qp_priv *priv = qp->priv;
724 	struct hfi1_ibport *ibp = ps->ibp;
725 	u16 lrh0;
726 	u32 nwords;
727 	u32 extra_bytes;
728 	u32 bth1;
729 
730 	/* Construct the header. */
731 	extra_bytes = -ps->s_txreq->s_cur_size & 3;
732 	nwords = (ps->s_txreq->s_cur_size + extra_bytes) >> 2;
733 	lrh0 = HFI1_LRH_BTH;
734 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
735 		qp->s_hdrwords += hfi1_make_grh(ibp,
736 						&ps->s_txreq->phdr.hdr.u.l.grh,
737 						&qp->remote_ah_attr.grh,
738 						qp->s_hdrwords, nwords);
739 		lrh0 = HFI1_LRH_GRH;
740 		middle = 0;
741 	}
742 	lrh0 |= (priv->s_sc & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
743 	/*
744 	 * reset s_ahg/AHG fields
745 	 *
746 	 * This insures that the ahgentry/ahgcount
747 	 * are at a non-AHG default to protect
748 	 * build_verbs_tx_desc() from using
749 	 * an include ahgidx.
750 	 *
751 	 * build_ahg() will modify as appropriate
752 	 * to use the AHG feature.
753 	 */
754 	priv->s_ahg->tx_flags = 0;
755 	priv->s_ahg->ahgcount = 0;
756 	priv->s_ahg->ahgidx = 0;
757 	if (qp->s_mig_state == IB_MIG_MIGRATED)
758 		bth0 |= IB_BTH_MIG_REQ;
759 	else
760 		middle = 0;
761 	if (middle)
762 		build_ahg(qp, bth2);
763 	else
764 		qp->s_flags &= ~RVT_S_AHG_VALID;
765 	ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
766 	ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
767 	ps->s_txreq->phdr.hdr.lrh[2] =
768 		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
769 	ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
770 				       qp->remote_ah_attr.src_path_bits);
771 	bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
772 	bth0 |= extra_bytes << 20;
773 	ohdr->bth[0] = cpu_to_be32(bth0);
774 	bth1 = qp->remote_qpn;
775 	if (qp->s_flags & RVT_S_ECN) {
776 		qp->s_flags &= ~RVT_S_ECN;
777 		/* we recently received a FECN, so return a BECN */
778 		bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT);
779 	}
780 	ohdr->bth[1] = cpu_to_be32(bth1);
781 	ohdr->bth[2] = cpu_to_be32(bth2);
782 }
783 
784 /* when sending, force a reschedule every one of these periods */
785 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
786 
787 void _hfi1_do_send(struct work_struct *work)
788 {
789 	struct iowait *wait = container_of(work, struct iowait, iowork);
790 	struct rvt_qp *qp = iowait_to_qp(wait);
791 
792 	hfi1_do_send(qp);
793 }
794 
795 /**
796  * hfi1_do_send - perform a send on a QP
797  * @work: contains a pointer to the QP
798  *
799  * Process entries in the send work queue until credit or queue is
800  * exhausted.  Only allow one CPU to send a packet per QP.
801  * Otherwise, two threads could send packets out of order.
802  */
803 void hfi1_do_send(struct rvt_qp *qp)
804 {
805 	struct hfi1_pkt_state ps;
806 	struct hfi1_qp_priv *priv = qp->priv;
807 	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
808 	unsigned long timeout;
809 	unsigned long timeout_int;
810 	int cpu;
811 
812 	ps.dev = to_idev(qp->ibqp.device);
813 	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
814 	ps.ppd = ppd_from_ibp(ps.ibp);
815 
816 	switch (qp->ibqp.qp_type) {
817 	case IB_QPT_RC:
818 		if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
819 								) - 1)) ==
820 				 ps.ppd->lid)) {
821 			ruc_loopback(qp);
822 			return;
823 		}
824 		make_req = hfi1_make_rc_req;
825 		timeout_int = (qp->timeout_jiffies);
826 		break;
827 	case IB_QPT_UC:
828 		if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
829 								) - 1)) ==
830 				 ps.ppd->lid)) {
831 			ruc_loopback(qp);
832 			return;
833 		}
834 		make_req = hfi1_make_uc_req;
835 		timeout_int = SEND_RESCHED_TIMEOUT;
836 		break;
837 	default:
838 		make_req = hfi1_make_ud_req;
839 		timeout_int = SEND_RESCHED_TIMEOUT;
840 	}
841 
842 	spin_lock_irqsave(&qp->s_lock, ps.flags);
843 
844 	/* Return if we are already busy processing a work request. */
845 	if (!hfi1_send_ok(qp)) {
846 		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
847 		return;
848 	}
849 
850 	qp->s_flags |= RVT_S_BUSY;
851 
852 	timeout = jiffies + (timeout_int) / 8;
853 	cpu = priv->s_sde ? priv->s_sde->cpu :
854 			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
855 	/* insure a pre-built packet is handled  */
856 	ps.s_txreq = get_waiting_verbs_txreq(qp);
857 	do {
858 		/* Check for a constructed packet to be sent. */
859 		if (qp->s_hdrwords != 0) {
860 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
861 			/*
862 			 * If the packet cannot be sent now, return and
863 			 * the send engine will be woken up later.
864 			 */
865 			if (hfi1_verbs_send(qp, &ps))
866 				return;
867 			/* Record that s_ahg is empty. */
868 			qp->s_hdrwords = 0;
869 			/* allow other tasks to run */
870 			if (unlikely(time_after(jiffies, timeout))) {
871 				if (workqueue_congested(cpu,
872 							ps.ppd->hfi1_wq)) {
873 					spin_lock_irqsave(
874 						&qp->s_lock,
875 						ps.flags);
876 					qp->s_flags &= ~RVT_S_BUSY;
877 					hfi1_schedule_send(qp);
878 					spin_unlock_irqrestore(
879 						&qp->s_lock,
880 						ps.flags);
881 					this_cpu_inc(
882 						*ps.ppd->dd->send_schedule);
883 					return;
884 				}
885 				if (!irqs_disabled()) {
886 					cond_resched();
887 					this_cpu_inc(
888 					   *ps.ppd->dd->send_schedule);
889 				}
890 				timeout = jiffies + (timeout_int) / 8;
891 			}
892 			spin_lock_irqsave(&qp->s_lock, ps.flags);
893 		}
894 	} while (make_req(qp, &ps));
895 
896 	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
897 }
898 
899 /*
900  * This should be called with s_lock held.
901  */
902 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
903 			enum ib_wc_status status)
904 {
905 	u32 old_last, last;
906 
907 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
908 		return;
909 
910 	last = qp->s_last;
911 	old_last = last;
912 	if (++last >= qp->s_size)
913 		last = 0;
914 	qp->s_last = last;
915 	/* See post_send() */
916 	barrier();
917 	rvt_put_swqe(wqe);
918 	if (qp->ibqp.qp_type == IB_QPT_UD ||
919 	    qp->ibqp.qp_type == IB_QPT_SMI ||
920 	    qp->ibqp.qp_type == IB_QPT_GSI)
921 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
922 
923 	rvt_qp_swqe_complete(qp, wqe, status);
924 
925 	if (qp->s_acked == old_last)
926 		qp->s_acked = last;
927 	if (qp->s_cur == old_last)
928 		qp->s_cur = last;
929 	if (qp->s_tail == old_last)
930 		qp->s_tail = last;
931 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
932 		qp->s_draining = 0;
933 }
934