xref: /linux/drivers/infiniband/hw/hfi1/ruc.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 /*
2  * Copyright(c) 2015 - 2018 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 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
57 {
58 	return (gid->global.interface_id == id &&
59 		(gid->global.subnet_prefix == gid_prefix ||
60 		 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
61 }
62 
63 /*
64  *
65  * This should be called with the QP r_lock held.
66  *
67  * The s_lock will be acquired around the hfi1_migrate_qp() call.
68  */
69 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_packet *packet)
70 {
71 	__be64 guid;
72 	unsigned long flags;
73 	struct rvt_qp *qp = packet->qp;
74 	u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
75 	u32 dlid = packet->dlid;
76 	u32 slid = packet->slid;
77 	u32 sl = packet->sl;
78 	bool migrated = packet->migrated;
79 	u16 pkey = packet->pkey;
80 
81 	if (qp->s_mig_state == IB_MIG_ARMED && migrated) {
82 		if (!packet->grh) {
83 			if ((rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
84 			     IB_AH_GRH) &&
85 			    (packet->etype != RHF_RCV_TYPE_BYPASS))
86 				return 1;
87 		} else {
88 			const struct ib_global_route *grh;
89 
90 			if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
91 			      IB_AH_GRH))
92 				return 1;
93 			grh = rdma_ah_read_grh(&qp->alt_ah_attr);
94 			guid = get_sguid(ibp, grh->sgid_index);
95 			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
96 				    guid))
97 				return 1;
98 			if (!gid_ok(
99 				&packet->grh->sgid,
100 				grh->dgid.global.subnet_prefix,
101 				grh->dgid.global.interface_id))
102 				return 1;
103 		}
104 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
105 					    sc5, slid))) {
106 			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
107 				      slid, dlid);
108 			return 1;
109 		}
110 		/* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
111 		if (slid != rdma_ah_get_dlid(&qp->alt_ah_attr) ||
112 		    ppd_from_ibp(ibp)->port !=
113 			rdma_ah_get_port_num(&qp->alt_ah_attr))
114 			return 1;
115 		spin_lock_irqsave(&qp->s_lock, flags);
116 		hfi1_migrate_qp(qp);
117 		spin_unlock_irqrestore(&qp->s_lock, flags);
118 	} else {
119 		if (!packet->grh) {
120 			if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
121 			     IB_AH_GRH) &&
122 			    (packet->etype != RHF_RCV_TYPE_BYPASS))
123 				return 1;
124 		} else {
125 			const struct ib_global_route *grh;
126 
127 			if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
128 						   IB_AH_GRH))
129 				return 1;
130 			grh = rdma_ah_read_grh(&qp->remote_ah_attr);
131 			guid = get_sguid(ibp, grh->sgid_index);
132 			if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
133 				    guid))
134 				return 1;
135 			if (!gid_ok(
136 			     &packet->grh->sgid,
137 			     grh->dgid.global.subnet_prefix,
138 			     grh->dgid.global.interface_id))
139 				return 1;
140 		}
141 		if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
142 					    sc5, slid))) {
143 			hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
144 				      slid, dlid);
145 			return 1;
146 		}
147 		/* Validate the SLID. See Ch. 9.6.1.5 */
148 		if ((slid != rdma_ah_get_dlid(&qp->remote_ah_attr)) ||
149 		    ppd_from_ibp(ibp)->port != qp->port_num)
150 			return 1;
151 		if (qp->s_mig_state == IB_MIG_REARM && !migrated)
152 			qp->s_mig_state = IB_MIG_ARMED;
153 	}
154 
155 	return 0;
156 }
157 
158 /**
159  * ruc_loopback - handle UC and RC loopback requests
160  * @sqp: the sending QP
161  *
162  * This is called from hfi1_do_send() to
163  * forward a WQE addressed to the same HFI.
164  * Note that although we are single threaded due to the send engine, we still
165  * have to protect against post_send().  We don't have to worry about
166  * receive interrupts since this is a connected protocol and all packets
167  * will pass through here.
168  */
169 static void ruc_loopback(struct rvt_qp *sqp)
170 {
171 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
172 	struct rvt_qp *qp;
173 	struct rvt_swqe *wqe;
174 	struct rvt_sge *sge;
175 	unsigned long flags;
176 	struct ib_wc wc;
177 	u64 sdata;
178 	atomic64_t *maddr;
179 	enum ib_wc_status send_status;
180 	bool release;
181 	int ret;
182 	bool copy_last = false;
183 	int local_ops = 0;
184 
185 	rcu_read_lock();
186 
187 	/*
188 	 * Note that we check the responder QP state after
189 	 * checking the requester's state.
190 	 */
191 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
192 			    sqp->remote_qpn);
193 
194 	spin_lock_irqsave(&sqp->s_lock, flags);
195 
196 	/* Return if we are already busy processing a work request. */
197 	if ((sqp->s_flags & (RVT_S_BUSY | HFI1_S_ANY_WAIT)) ||
198 	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
199 		goto unlock;
200 
201 	sqp->s_flags |= RVT_S_BUSY;
202 
203 again:
204 	if (sqp->s_last == READ_ONCE(sqp->s_head))
205 		goto clr_busy;
206 	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
207 
208 	/* Return if it is not OK to start a new work request. */
209 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
210 		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
211 			goto clr_busy;
212 		/* We are in the error state, flush the work request. */
213 		send_status = IB_WC_WR_FLUSH_ERR;
214 		goto flush_send;
215 	}
216 
217 	/*
218 	 * We can rely on the entry not changing without the s_lock
219 	 * being held until we update s_last.
220 	 * We increment s_cur to indicate s_last is in progress.
221 	 */
222 	if (sqp->s_last == sqp->s_cur) {
223 		if (++sqp->s_cur >= sqp->s_size)
224 			sqp->s_cur = 0;
225 	}
226 	spin_unlock_irqrestore(&sqp->s_lock, flags);
227 
228 	if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
229 	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
230 		ibp->rvp.n_pkt_drops++;
231 		/*
232 		 * For RC, the requester would timeout and retry so
233 		 * shortcut the timeouts and just signal too many retries.
234 		 */
235 		if (sqp->ibqp.qp_type == IB_QPT_RC)
236 			send_status = IB_WC_RETRY_EXC_ERR;
237 		else
238 			send_status = IB_WC_SUCCESS;
239 		goto serr;
240 	}
241 
242 	memset(&wc, 0, sizeof(wc));
243 	send_status = IB_WC_SUCCESS;
244 
245 	release = true;
246 	sqp->s_sge.sge = wqe->sg_list[0];
247 	sqp->s_sge.sg_list = wqe->sg_list + 1;
248 	sqp->s_sge.num_sge = wqe->wr.num_sge;
249 	sqp->s_len = wqe->length;
250 	switch (wqe->wr.opcode) {
251 	case IB_WR_REG_MR:
252 		goto send_comp;
253 
254 	case IB_WR_LOCAL_INV:
255 		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
256 			if (rvt_invalidate_rkey(sqp,
257 						wqe->wr.ex.invalidate_rkey))
258 				send_status = IB_WC_LOC_PROT_ERR;
259 			local_ops = 1;
260 		}
261 		goto send_comp;
262 
263 	case IB_WR_SEND_WITH_INV:
264 		if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
265 			wc.wc_flags = IB_WC_WITH_INVALIDATE;
266 			wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
267 		}
268 		goto send;
269 
270 	case IB_WR_SEND_WITH_IMM:
271 		wc.wc_flags = IB_WC_WITH_IMM;
272 		wc.ex.imm_data = wqe->wr.ex.imm_data;
273 		/* FALLTHROUGH */
274 	case IB_WR_SEND:
275 send:
276 		ret = rvt_get_rwqe(qp, false);
277 		if (ret < 0)
278 			goto op_err;
279 		if (!ret)
280 			goto rnr_nak;
281 		break;
282 
283 	case IB_WR_RDMA_WRITE_WITH_IMM:
284 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
285 			goto inv_err;
286 		wc.wc_flags = IB_WC_WITH_IMM;
287 		wc.ex.imm_data = wqe->wr.ex.imm_data;
288 		ret = rvt_get_rwqe(qp, true);
289 		if (ret < 0)
290 			goto op_err;
291 		if (!ret)
292 			goto rnr_nak;
293 		/* skip copy_last set and qp_access_flags recheck */
294 		goto do_write;
295 	case IB_WR_RDMA_WRITE:
296 		copy_last = rvt_is_user_qp(qp);
297 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
298 			goto inv_err;
299 do_write:
300 		if (wqe->length == 0)
301 			break;
302 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
303 					  wqe->rdma_wr.remote_addr,
304 					  wqe->rdma_wr.rkey,
305 					  IB_ACCESS_REMOTE_WRITE)))
306 			goto acc_err;
307 		qp->r_sge.sg_list = NULL;
308 		qp->r_sge.num_sge = 1;
309 		qp->r_sge.total_len = wqe->length;
310 		break;
311 
312 	case IB_WR_RDMA_READ:
313 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
314 			goto inv_err;
315 		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
316 					  wqe->rdma_wr.remote_addr,
317 					  wqe->rdma_wr.rkey,
318 					  IB_ACCESS_REMOTE_READ)))
319 			goto acc_err;
320 		release = false;
321 		sqp->s_sge.sg_list = NULL;
322 		sqp->s_sge.num_sge = 1;
323 		qp->r_sge.sge = wqe->sg_list[0];
324 		qp->r_sge.sg_list = wqe->sg_list + 1;
325 		qp->r_sge.num_sge = wqe->wr.num_sge;
326 		qp->r_sge.total_len = wqe->length;
327 		break;
328 
329 	case IB_WR_ATOMIC_CMP_AND_SWP:
330 	case IB_WR_ATOMIC_FETCH_AND_ADD:
331 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
332 			goto inv_err;
333 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
334 					  wqe->atomic_wr.remote_addr,
335 					  wqe->atomic_wr.rkey,
336 					  IB_ACCESS_REMOTE_ATOMIC)))
337 			goto acc_err;
338 		/* Perform atomic OP and save result. */
339 		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
340 		sdata = wqe->atomic_wr.compare_add;
341 		*(u64 *)sqp->s_sge.sge.vaddr =
342 			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
343 			(u64)atomic64_add_return(sdata, maddr) - sdata :
344 			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
345 				      sdata, wqe->atomic_wr.swap);
346 		rvt_put_mr(qp->r_sge.sge.mr);
347 		qp->r_sge.num_sge = 0;
348 		goto send_comp;
349 
350 	default:
351 		send_status = IB_WC_LOC_QP_OP_ERR;
352 		goto serr;
353 	}
354 
355 	sge = &sqp->s_sge.sge;
356 	while (sqp->s_len) {
357 		u32 len = sqp->s_len;
358 
359 		if (len > sge->length)
360 			len = sge->length;
361 		if (len > sge->sge_length)
362 			len = sge->sge_length;
363 		WARN_ON_ONCE(len == 0);
364 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
365 		sge->vaddr += len;
366 		sge->length -= len;
367 		sge->sge_length -= len;
368 		if (sge->sge_length == 0) {
369 			if (!release)
370 				rvt_put_mr(sge->mr);
371 			if (--sqp->s_sge.num_sge)
372 				*sge = *sqp->s_sge.sg_list++;
373 		} else if (sge->length == 0 && sge->mr->lkey) {
374 			if (++sge->n >= RVT_SEGSZ) {
375 				if (++sge->m >= sge->mr->mapsz)
376 					break;
377 				sge->n = 0;
378 			}
379 			sge->vaddr =
380 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
381 			sge->length =
382 				sge->mr->map[sge->m]->segs[sge->n].length;
383 		}
384 		sqp->s_len -= len;
385 	}
386 	if (release)
387 		rvt_put_ss(&qp->r_sge);
388 
389 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
390 		goto send_comp;
391 
392 	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
393 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
394 	else
395 		wc.opcode = IB_WC_RECV;
396 	wc.wr_id = qp->r_wr_id;
397 	wc.status = IB_WC_SUCCESS;
398 	wc.byte_len = wqe->length;
399 	wc.qp = &qp->ibqp;
400 	wc.src_qp = qp->remote_qpn;
401 	wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
402 	wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
403 	wc.port_num = 1;
404 	/* Signal completion event if the solicited bit is set. */
405 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
406 		     wqe->wr.send_flags & IB_SEND_SOLICITED);
407 
408 send_comp:
409 	spin_lock_irqsave(&sqp->s_lock, flags);
410 	ibp->rvp.n_loop_pkts++;
411 flush_send:
412 	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
413 	hfi1_send_complete(sqp, wqe, send_status);
414 	if (local_ops) {
415 		atomic_dec(&sqp->local_ops_pending);
416 		local_ops = 0;
417 	}
418 	goto again;
419 
420 rnr_nak:
421 	/* Handle RNR NAK */
422 	if (qp->ibqp.qp_type == IB_QPT_UC)
423 		goto send_comp;
424 	ibp->rvp.n_rnr_naks++;
425 	/*
426 	 * Note: we don't need the s_lock held since the BUSY flag
427 	 * makes this single threaded.
428 	 */
429 	if (sqp->s_rnr_retry == 0) {
430 		send_status = IB_WC_RNR_RETRY_EXC_ERR;
431 		goto serr;
432 	}
433 	if (sqp->s_rnr_retry_cnt < 7)
434 		sqp->s_rnr_retry--;
435 	spin_lock_irqsave(&sqp->s_lock, flags);
436 	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
437 		goto clr_busy;
438 	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
439 				IB_AETH_CREDIT_SHIFT);
440 	goto clr_busy;
441 
442 op_err:
443 	send_status = IB_WC_REM_OP_ERR;
444 	wc.status = IB_WC_LOC_QP_OP_ERR;
445 	goto err;
446 
447 inv_err:
448 	send_status = IB_WC_REM_INV_REQ_ERR;
449 	wc.status = IB_WC_LOC_QP_OP_ERR;
450 	goto err;
451 
452 acc_err:
453 	send_status = IB_WC_REM_ACCESS_ERR;
454 	wc.status = IB_WC_LOC_PROT_ERR;
455 err:
456 	/* responder goes to error state */
457 	rvt_rc_error(qp, wc.status);
458 
459 serr:
460 	spin_lock_irqsave(&sqp->s_lock, flags);
461 	hfi1_send_complete(sqp, wqe, send_status);
462 	if (sqp->ibqp.qp_type == IB_QPT_RC) {
463 		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
464 
465 		sqp->s_flags &= ~RVT_S_BUSY;
466 		spin_unlock_irqrestore(&sqp->s_lock, flags);
467 		if (lastwqe) {
468 			struct ib_event ev;
469 
470 			ev.device = sqp->ibqp.device;
471 			ev.element.qp = &sqp->ibqp;
472 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
473 			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
474 		}
475 		goto done;
476 	}
477 clr_busy:
478 	sqp->s_flags &= ~RVT_S_BUSY;
479 unlock:
480 	spin_unlock_irqrestore(&sqp->s_lock, flags);
481 done:
482 	rcu_read_unlock();
483 }
484 
485 /**
486  * hfi1_make_grh - construct a GRH header
487  * @ibp: a pointer to the IB port
488  * @hdr: a pointer to the GRH header being constructed
489  * @grh: the global route address to send to
490  * @hwords: size of header after grh being sent in dwords
491  * @nwords: the number of 32 bit words of data being sent
492  *
493  * Return the size of the header in 32 bit words.
494  */
495 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
496 		  const struct ib_global_route *grh, u32 hwords, u32 nwords)
497 {
498 	hdr->version_tclass_flow =
499 		cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
500 			    (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
501 			    (grh->flow_label << IB_GRH_FLOW_SHIFT));
502 	hdr->paylen = cpu_to_be16((hwords + nwords) << 2);
503 	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
504 	hdr->next_hdr = IB_GRH_NEXT_HDR;
505 	hdr->hop_limit = grh->hop_limit;
506 	/* The SGID is 32-bit aligned. */
507 	hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
508 	hdr->sgid.global.interface_id =
509 		grh->sgid_index < HFI1_GUIDS_PER_PORT ?
510 		get_sguid(ibp, grh->sgid_index) :
511 		get_sguid(ibp, HFI1_PORT_GUID_INDEX);
512 	hdr->dgid = grh->dgid;
513 
514 	/* GRH header size in 32-bit words. */
515 	return sizeof(struct ib_grh) / sizeof(u32);
516 }
517 
518 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, \
519 			      hdr.ibh.u.oth.bth[2]) / 4)
520 
521 /**
522  * build_ahg - create ahg in s_ahg
523  * @qp: a pointer to QP
524  * @npsn: the next PSN for the request/response
525  *
526  * This routine handles the AHG by allocating an ahg entry and causing the
527  * copy of the first middle.
528  *
529  * Subsequent middles use the copied entry, editing the
530  * PSN with 1 or 2 edits.
531  */
532 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
533 {
534 	struct hfi1_qp_priv *priv = qp->priv;
535 
536 	if (unlikely(qp->s_flags & HFI1_S_AHG_CLEAR))
537 		clear_ahg(qp);
538 	if (!(qp->s_flags & HFI1_S_AHG_VALID)) {
539 		/* first middle that needs copy  */
540 		if (qp->s_ahgidx < 0)
541 			qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
542 		if (qp->s_ahgidx >= 0) {
543 			qp->s_ahgpsn = npsn;
544 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
545 			/* save to protect a change in another thread */
546 			priv->s_ahg->ahgidx = qp->s_ahgidx;
547 			qp->s_flags |= HFI1_S_AHG_VALID;
548 		}
549 	} else {
550 		/* subsequent middle after valid */
551 		if (qp->s_ahgidx >= 0) {
552 			priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
553 			priv->s_ahg->ahgidx = qp->s_ahgidx;
554 			priv->s_ahg->ahgcount++;
555 			priv->s_ahg->ahgdesc[0] =
556 				sdma_build_ahg_descriptor(
557 					(__force u16)cpu_to_be16((u16)npsn),
558 					BTH2_OFFSET,
559 					16,
560 					16);
561 			if ((npsn & 0xffff0000) !=
562 					(qp->s_ahgpsn & 0xffff0000)) {
563 				priv->s_ahg->ahgcount++;
564 				priv->s_ahg->ahgdesc[1] =
565 					sdma_build_ahg_descriptor(
566 						(__force u16)cpu_to_be16(
567 							(u16)(npsn >> 16)),
568 						BTH2_OFFSET,
569 						0,
570 						16);
571 			}
572 		}
573 	}
574 }
575 
576 static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
577 				     struct ib_other_headers *ohdr,
578 				     u32 bth0, u32 bth1, u32 bth2)
579 {
580 	bth1 |= qp->remote_qpn;
581 	ohdr->bth[0] = cpu_to_be32(bth0);
582 	ohdr->bth[1] = cpu_to_be32(bth1);
583 	ohdr->bth[2] = cpu_to_be32(bth2);
584 }
585 
586 /**
587  * hfi1_make_ruc_header_16B - build a 16B header
588  * @qp: the queue pair
589  * @ohdr: a pointer to the destination header memory
590  * @bth0: bth0 passed in from the RC/UC builder
591  * @bth2: bth2 passed in from the RC/UC builder
592  * @middle: non zero implies indicates ahg "could" be used
593  * @ps: the current packet state
594  *
595  * This routine may disarm ahg under these situations:
596  * - packet needs a GRH
597  * - BECN needed
598  * - migration state not IB_MIG_MIGRATED
599  */
600 static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
601 					    struct ib_other_headers *ohdr,
602 					    u32 bth0, u32 bth2, int middle,
603 					    struct hfi1_pkt_state *ps)
604 {
605 	struct hfi1_qp_priv *priv = qp->priv;
606 	struct hfi1_ibport *ibp = ps->ibp;
607 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
608 	u32 bth1 = 0;
609 	u32 slid;
610 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
611 	u8 l4 = OPA_16B_L4_IB_LOCAL;
612 	u8 extra_bytes = hfi1_get_16b_padding(
613 				(ps->s_txreq->hdr_dwords << 2),
614 				ps->s_txreq->s_cur_size);
615 	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
616 				 extra_bytes + SIZE_OF_LT) >> 2);
617 	bool becn = false;
618 
619 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
620 	    hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
621 		struct ib_grh *grh;
622 		struct ib_global_route *grd =
623 			rdma_ah_retrieve_grh(&qp->remote_ah_attr);
624 		/*
625 		 * Ensure OPA GIDs are transformed to IB gids
626 		 * before creating the GRH.
627 		 */
628 		if (grd->sgid_index == OPA_GID_INDEX)
629 			grd->sgid_index = 0;
630 		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
631 		l4 = OPA_16B_L4_IB_GLOBAL;
632 		ps->s_txreq->hdr_dwords +=
633 			hfi1_make_grh(ibp, grh, grd,
634 				      ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
635 				      nwords);
636 		middle = 0;
637 	}
638 
639 	if (qp->s_mig_state == IB_MIG_MIGRATED)
640 		bth1 |= OPA_BTH_MIG_REQ;
641 	else
642 		middle = 0;
643 
644 	if (qp->s_flags & RVT_S_ECN) {
645 		qp->s_flags &= ~RVT_S_ECN;
646 		/* we recently received a FECN, so return a BECN */
647 		becn = true;
648 		middle = 0;
649 	}
650 	if (middle)
651 		build_ahg(qp, bth2);
652 	else
653 		qp->s_flags &= ~HFI1_S_AHG_VALID;
654 
655 	bth0 |= pkey;
656 	bth0 |= extra_bytes << 20;
657 	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
658 
659 	if (!ppd->lid)
660 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
661 	else
662 		slid = ppd->lid |
663 			(rdma_ah_get_path_bits(&qp->remote_ah_attr) &
664 			((1 << ppd->lmc) - 1));
665 
666 	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
667 			  slid,
668 			  opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
669 				      16B),
670 			  (ps->s_txreq->hdr_dwords + nwords) >> 1,
671 			  pkey, becn, 0, l4, priv->s_sc);
672 }
673 
674 /**
675  * hfi1_make_ruc_header_9B - build a 9B header
676  * @qp: the queue pair
677  * @ohdr: a pointer to the destination header memory
678  * @bth0: bth0 passed in from the RC/UC builder
679  * @bth2: bth2 passed in from the RC/UC builder
680  * @middle: non zero implies indicates ahg "could" be used
681  * @ps: the current packet state
682  *
683  * This routine may disarm ahg under these situations:
684  * - packet needs a GRH
685  * - BECN needed
686  * - migration state not IB_MIG_MIGRATED
687  */
688 static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
689 					   struct ib_other_headers *ohdr,
690 					   u32 bth0, u32 bth2, int middle,
691 					   struct hfi1_pkt_state *ps)
692 {
693 	struct hfi1_qp_priv *priv = qp->priv;
694 	struct hfi1_ibport *ibp = ps->ibp;
695 	u32 bth1 = 0;
696 	u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
697 	u16 lrh0 = HFI1_LRH_BTH;
698 	u8 extra_bytes = -ps->s_txreq->s_cur_size & 3;
699 	u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
700 					 extra_bytes) >> 2);
701 
702 	if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
703 		struct ib_grh *grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
704 
705 		lrh0 = HFI1_LRH_GRH;
706 		ps->s_txreq->hdr_dwords +=
707 			hfi1_make_grh(ibp, grh,
708 				      rdma_ah_read_grh(&qp->remote_ah_attr),
709 				      ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
710 				      nwords);
711 		middle = 0;
712 	}
713 	lrh0 |= (priv->s_sc & 0xf) << 12 |
714 		(rdma_ah_get_sl(&qp->remote_ah_attr) & 0xf) << 4;
715 
716 	if (qp->s_mig_state == IB_MIG_MIGRATED)
717 		bth0 |= IB_BTH_MIG_REQ;
718 	else
719 		middle = 0;
720 
721 	if (qp->s_flags & RVT_S_ECN) {
722 		qp->s_flags &= ~RVT_S_ECN;
723 		/* we recently received a FECN, so return a BECN */
724 		bth1 |= (IB_BECN_MASK << IB_BECN_SHIFT);
725 		middle = 0;
726 	}
727 	if (middle)
728 		build_ahg(qp, bth2);
729 	else
730 		qp->s_flags &= ~HFI1_S_AHG_VALID;
731 
732 	bth0 |= pkey;
733 	bth0 |= extra_bytes << 20;
734 	hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
735 	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
736 			 lrh0,
737 			 ps->s_txreq->hdr_dwords + nwords,
738 			 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
739 			 ppd_from_ibp(ibp)->lid |
740 				rdma_ah_get_path_bits(&qp->remote_ah_attr));
741 }
742 
743 typedef void (*hfi1_make_ruc_hdr)(struct rvt_qp *qp,
744 				  struct ib_other_headers *ohdr,
745 				  u32 bth0, u32 bth2, int middle,
746 				  struct hfi1_pkt_state *ps);
747 
748 /* We support only two types - 9B and 16B for now */
749 static const hfi1_make_ruc_hdr hfi1_ruc_header_tbl[2] = {
750 	[HFI1_PKT_TYPE_9B] = &hfi1_make_ruc_header_9B,
751 	[HFI1_PKT_TYPE_16B] = &hfi1_make_ruc_header_16B
752 };
753 
754 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
755 			  u32 bth0, u32 bth2, int middle,
756 			  struct hfi1_pkt_state *ps)
757 {
758 	struct hfi1_qp_priv *priv = qp->priv;
759 
760 	/*
761 	 * reset s_ahg/AHG fields
762 	 *
763 	 * This insures that the ahgentry/ahgcount
764 	 * are at a non-AHG default to protect
765 	 * build_verbs_tx_desc() from using
766 	 * an include ahgidx.
767 	 *
768 	 * build_ahg() will modify as appropriate
769 	 * to use the AHG feature.
770 	 */
771 	priv->s_ahg->tx_flags = 0;
772 	priv->s_ahg->ahgcount = 0;
773 	priv->s_ahg->ahgidx = 0;
774 
775 	/* Make the appropriate header */
776 	hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth2, middle, ps);
777 }
778 
779 /* when sending, force a reschedule every one of these periods */
780 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
781 
782 /**
783  * schedule_send_yield - test for a yield required for QP send engine
784  * @timeout: Final time for timeout slice for jiffies
785  * @qp: a pointer to QP
786  * @ps: a pointer to a structure with commonly lookup values for
787  *      the the send engine progress
788  *
789  * This routine checks if the time slice for the QP has expired
790  * for RC QPs, if so an additional work entry is queued. At this
791  * point, other QPs have an opportunity to be scheduled. It
792  * returns true if a yield is required, otherwise, false
793  * is returned.
794  */
795 static bool schedule_send_yield(struct rvt_qp *qp,
796 				struct hfi1_pkt_state *ps)
797 {
798 	ps->pkts_sent = true;
799 
800 	if (unlikely(time_after(jiffies, ps->timeout))) {
801 		if (!ps->in_thread ||
802 		    workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
803 			spin_lock_irqsave(&qp->s_lock, ps->flags);
804 			qp->s_flags &= ~RVT_S_BUSY;
805 			hfi1_schedule_send(qp);
806 			spin_unlock_irqrestore(&qp->s_lock, ps->flags);
807 			this_cpu_inc(*ps->ppd->dd->send_schedule);
808 			trace_hfi1_rc_expired_time_slice(qp, true);
809 			return true;
810 		}
811 
812 		cond_resched();
813 		this_cpu_inc(*ps->ppd->dd->send_schedule);
814 		ps->timeout = jiffies + ps->timeout_int;
815 	}
816 
817 	trace_hfi1_rc_expired_time_slice(qp, false);
818 	return false;
819 }
820 
821 void hfi1_do_send_from_rvt(struct rvt_qp *qp)
822 {
823 	hfi1_do_send(qp, false);
824 }
825 
826 void _hfi1_do_send(struct work_struct *work)
827 {
828 	struct iowait *wait = container_of(work, struct iowait, iowork);
829 	struct rvt_qp *qp = iowait_to_qp(wait);
830 
831 	hfi1_do_send(qp, true);
832 }
833 
834 /**
835  * hfi1_do_send - perform a send on a QP
836  * @work: contains a pointer to the QP
837  * @in_thread: true if in a workqueue thread
838  *
839  * Process entries in the send work queue until credit or queue is
840  * exhausted.  Only allow one CPU to send a packet per QP.
841  * Otherwise, two threads could send packets out of order.
842  */
843 void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
844 {
845 	struct hfi1_pkt_state ps;
846 	struct hfi1_qp_priv *priv = qp->priv;
847 	int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
848 
849 	ps.dev = to_idev(qp->ibqp.device);
850 	ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
851 	ps.ppd = ppd_from_ibp(ps.ibp);
852 	ps.in_thread = in_thread;
853 
854 	trace_hfi1_rc_do_send(qp, in_thread);
855 
856 	switch (qp->ibqp.qp_type) {
857 	case IB_QPT_RC:
858 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
859 				   ~((1 << ps.ppd->lmc) - 1)) ==
860 				  ps.ppd->lid)) {
861 			ruc_loopback(qp);
862 			return;
863 		}
864 		make_req = hfi1_make_rc_req;
865 		ps.timeout_int = qp->timeout_jiffies;
866 		break;
867 	case IB_QPT_UC:
868 		if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
869 				   ~((1 << ps.ppd->lmc) - 1)) ==
870 				  ps.ppd->lid)) {
871 			ruc_loopback(qp);
872 			return;
873 		}
874 		make_req = hfi1_make_uc_req;
875 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
876 		break;
877 	default:
878 		make_req = hfi1_make_ud_req;
879 		ps.timeout_int = SEND_RESCHED_TIMEOUT;
880 	}
881 
882 	spin_lock_irqsave(&qp->s_lock, ps.flags);
883 
884 	/* Return if we are already busy processing a work request. */
885 	if (!hfi1_send_ok(qp)) {
886 		spin_unlock_irqrestore(&qp->s_lock, ps.flags);
887 		return;
888 	}
889 
890 	qp->s_flags |= RVT_S_BUSY;
891 
892 	ps.timeout_int = ps.timeout_int / 8;
893 	ps.timeout = jiffies + ps.timeout_int;
894 	ps.cpu = priv->s_sde ? priv->s_sde->cpu :
895 			cpumask_first(cpumask_of_node(ps.ppd->dd->node));
896 	ps.pkts_sent = false;
897 
898 	/* insure a pre-built packet is handled  */
899 	ps.s_txreq = get_waiting_verbs_txreq(qp);
900 	do {
901 		/* Check for a constructed packet to be sent. */
902 		if (ps.s_txreq) {
903 			spin_unlock_irqrestore(&qp->s_lock, ps.flags);
904 			/*
905 			 * If the packet cannot be sent now, return and
906 			 * the send engine will be woken up later.
907 			 */
908 			if (hfi1_verbs_send(qp, &ps))
909 				return;
910 			/* allow other tasks to run */
911 			if (schedule_send_yield(qp, &ps))
912 				return;
913 
914 			spin_lock_irqsave(&qp->s_lock, ps.flags);
915 		}
916 	} while (make_req(qp, &ps));
917 	iowait_starve_clear(ps.pkts_sent, &priv->s_iowait);
918 	spin_unlock_irqrestore(&qp->s_lock, ps.flags);
919 }
920 
921 /*
922  * This should be called with s_lock held.
923  */
924 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
925 			enum ib_wc_status status)
926 {
927 	u32 old_last, last;
928 
929 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
930 		return;
931 
932 	last = qp->s_last;
933 	old_last = last;
934 	trace_hfi1_qp_send_completion(qp, wqe, last);
935 	if (++last >= qp->s_size)
936 		last = 0;
937 	trace_hfi1_qp_send_completion(qp, wqe, last);
938 	qp->s_last = last;
939 	/* See post_send() */
940 	barrier();
941 	rvt_put_swqe(wqe);
942 	if (qp->ibqp.qp_type == IB_QPT_UD ||
943 	    qp->ibqp.qp_type == IB_QPT_SMI ||
944 	    qp->ibqp.qp_type == IB_QPT_GSI)
945 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
946 
947 	rvt_qp_swqe_complete(qp,
948 			     wqe,
949 			     ib_hfi1_wc_opcode[wqe->wr.opcode],
950 			     status);
951 
952 	if (qp->s_acked == old_last)
953 		qp->s_acked = last;
954 	if (qp->s_cur == old_last)
955 		qp->s_cur = last;
956 	if (qp->s_tail == old_last)
957 		qp->s_tail = last;
958 	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
959 		qp->s_draining = 0;
960 }
961