xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ud.c (revision a9a08845)
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/net.h>
49 #include <rdma/ib_smi.h>
50 
51 #include "hfi.h"
52 #include "mad.h"
53 #include "verbs_txreq.h"
54 #include "qp.h"
55 
56 /* We support only two types - 9B and 16B for now */
57 static const hfi1_make_req hfi1_make_ud_req_tbl[2] = {
58 	[HFI1_PKT_TYPE_9B] = &hfi1_make_ud_req_9B,
59 	[HFI1_PKT_TYPE_16B] = &hfi1_make_ud_req_16B
60 };
61 
62 /**
63  * ud_loopback - handle send on loopback QPs
64  * @sqp: the sending QP
65  * @swqe: the send work request
66  *
67  * This is called from hfi1_make_ud_req() to forward a WQE addressed
68  * to the same HFI.
69  * Note that the receive interrupt handler may be calling hfi1_ud_rcv()
70  * while this is being called.
71  */
72 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
73 {
74 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
75 	struct hfi1_pportdata *ppd;
76 	struct hfi1_qp_priv *priv = sqp->priv;
77 	struct rvt_qp *qp;
78 	struct rdma_ah_attr *ah_attr;
79 	unsigned long flags;
80 	struct rvt_sge_state ssge;
81 	struct rvt_sge *sge;
82 	struct ib_wc wc;
83 	u32 length;
84 	enum ib_qp_type sqptype, dqptype;
85 
86 	rcu_read_lock();
87 
88 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
89 			    swqe->ud_wr.remote_qpn);
90 	if (!qp) {
91 		ibp->rvp.n_pkt_drops++;
92 		rcu_read_unlock();
93 		return;
94 	}
95 
96 	sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
97 			IB_QPT_UD : sqp->ibqp.qp_type;
98 	dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
99 			IB_QPT_UD : qp->ibqp.qp_type;
100 
101 	if (dqptype != sqptype ||
102 	    !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
103 		ibp->rvp.n_pkt_drops++;
104 		goto drop;
105 	}
106 
107 	ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr;
108 	ppd = ppd_from_ibp(ibp);
109 
110 	if (qp->ibqp.qp_num > 1) {
111 		u16 pkey;
112 		u32 slid;
113 		u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
114 
115 		pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
116 		slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
117 				   ((1 << ppd->lmc) - 1));
118 		if (unlikely(ingress_pkey_check(ppd, pkey, sc5,
119 						qp->s_pkey_index,
120 						slid, false))) {
121 			hfi1_bad_pkey(ibp, pkey,
122 				      rdma_ah_get_sl(ah_attr),
123 				      sqp->ibqp.qp_num, qp->ibqp.qp_num,
124 				      slid, rdma_ah_get_dlid(ah_attr));
125 			goto drop;
126 		}
127 	}
128 
129 	/*
130 	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
131 	 * Qkeys with the high order bit set mean use the
132 	 * qkey from the QP context instead of the WR (see 10.2.5).
133 	 */
134 	if (qp->ibqp.qp_num) {
135 		u32 qkey;
136 
137 		qkey = (int)swqe->ud_wr.remote_qkey < 0 ?
138 			sqp->qkey : swqe->ud_wr.remote_qkey;
139 		if (unlikely(qkey != qp->qkey))
140 			goto drop; /* silently drop per IBTA spec */
141 	}
142 
143 	/*
144 	 * A GRH is expected to precede the data even if not
145 	 * present on the wire.
146 	 */
147 	length = swqe->length;
148 	memset(&wc, 0, sizeof(wc));
149 	wc.byte_len = length + sizeof(struct ib_grh);
150 
151 	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
152 		wc.wc_flags = IB_WC_WITH_IMM;
153 		wc.ex.imm_data = swqe->wr.ex.imm_data;
154 	}
155 
156 	spin_lock_irqsave(&qp->r_lock, flags);
157 
158 	/*
159 	 * Get the next work request entry to find where to put the data.
160 	 */
161 	if (qp->r_flags & RVT_R_REUSE_SGE) {
162 		qp->r_flags &= ~RVT_R_REUSE_SGE;
163 	} else {
164 		int ret;
165 
166 		ret = hfi1_rvt_get_rwqe(qp, 0);
167 		if (ret < 0) {
168 			rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
169 			goto bail_unlock;
170 		}
171 		if (!ret) {
172 			if (qp->ibqp.qp_num == 0)
173 				ibp->rvp.n_vl15_dropped++;
174 			goto bail_unlock;
175 		}
176 	}
177 	/* Silently drop packets which are too big. */
178 	if (unlikely(wc.byte_len > qp->r_len)) {
179 		qp->r_flags |= RVT_R_REUSE_SGE;
180 		ibp->rvp.n_pkt_drops++;
181 		goto bail_unlock;
182 	}
183 
184 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
185 		struct ib_grh grh;
186 		struct ib_global_route grd = *(rdma_ah_read_grh(ah_attr));
187 
188 		/*
189 		 * For loopback packets with extended LIDs, the
190 		 * sgid_index in the GRH is 0 and the dgid is
191 		 * OPA GID of the sender. While creating a response
192 		 * to the loopback packet, IB core creates the new
193 		 * sgid_index from the DGID and that will be the
194 		 * OPA_GID_INDEX. The new dgid is from the sgid
195 		 * index and that will be in the IB GID format.
196 		 *
197 		 * We now have a case where the sent packet had a
198 		 * different sgid_index and dgid compared to the
199 		 * one that was received in response.
200 		 *
201 		 * Fix this inconsistency.
202 		 */
203 		if (priv->hdr_type == HFI1_PKT_TYPE_16B) {
204 			if (grd.sgid_index == 0)
205 				grd.sgid_index = OPA_GID_INDEX;
206 
207 			if (ib_is_opa_gid(&grd.dgid))
208 				grd.dgid.global.interface_id =
209 				cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]);
210 		}
211 
212 		hfi1_make_grh(ibp, &grh, &grd, 0, 0);
213 		hfi1_copy_sge(&qp->r_sge, &grh,
214 			      sizeof(grh), true, false);
215 		wc.wc_flags |= IB_WC_GRH;
216 	} else {
217 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
218 	}
219 	ssge.sg_list = swqe->sg_list + 1;
220 	ssge.sge = *swqe->sg_list;
221 	ssge.num_sge = swqe->wr.num_sge;
222 	sge = &ssge.sge;
223 	while (length) {
224 		u32 len = sge->length;
225 
226 		if (len > length)
227 			len = length;
228 		if (len > sge->sge_length)
229 			len = sge->sge_length;
230 		WARN_ON_ONCE(len == 0);
231 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, true, false);
232 		sge->vaddr += len;
233 		sge->length -= len;
234 		sge->sge_length -= len;
235 		if (sge->sge_length == 0) {
236 			if (--ssge.num_sge)
237 				*sge = *ssge.sg_list++;
238 		} else if (sge->length == 0 && sge->mr->lkey) {
239 			if (++sge->n >= RVT_SEGSZ) {
240 				if (++sge->m >= sge->mr->mapsz)
241 					break;
242 				sge->n = 0;
243 			}
244 			sge->vaddr =
245 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
246 			sge->length =
247 				sge->mr->map[sge->m]->segs[sge->n].length;
248 		}
249 		length -= len;
250 	}
251 	rvt_put_ss(&qp->r_sge);
252 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
253 		goto bail_unlock;
254 	wc.wr_id = qp->r_wr_id;
255 	wc.status = IB_WC_SUCCESS;
256 	wc.opcode = IB_WC_RECV;
257 	wc.qp = &qp->ibqp;
258 	wc.src_qp = sqp->ibqp.qp_num;
259 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) {
260 		if (sqp->ibqp.qp_type == IB_QPT_GSI ||
261 		    sqp->ibqp.qp_type == IB_QPT_SMI)
262 			wc.pkey_index = swqe->ud_wr.pkey_index;
263 		else
264 			wc.pkey_index = sqp->s_pkey_index;
265 	} else {
266 		wc.pkey_index = 0;
267 	}
268 	wc.slid = (ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
269 				   ((1 << ppd->lmc) - 1))) & U16_MAX;
270 	/* Check for loopback when the port lid is not set */
271 	if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
272 		wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
273 	wc.sl = rdma_ah_get_sl(ah_attr);
274 	wc.dlid_path_bits = rdma_ah_get_dlid(ah_attr) & ((1 << ppd->lmc) - 1);
275 	wc.port_num = qp->port_num;
276 	/* Signal completion event if the solicited bit is set. */
277 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
278 		     swqe->wr.send_flags & IB_SEND_SOLICITED);
279 	ibp->rvp.n_loop_pkts++;
280 bail_unlock:
281 	spin_unlock_irqrestore(&qp->r_lock, flags);
282 drop:
283 	rcu_read_unlock();
284 }
285 
286 static void hfi1_make_bth_deth(struct rvt_qp *qp, struct rvt_swqe *wqe,
287 			       struct ib_other_headers *ohdr,
288 			       u16 *pkey, u32 extra_bytes, bool bypass)
289 {
290 	u32 bth0;
291 	struct hfi1_ibport *ibp;
292 
293 	ibp = to_iport(qp->ibqp.device, qp->port_num);
294 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
295 		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
296 		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
297 	} else {
298 		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
299 	}
300 
301 	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
302 		bth0 |= IB_BTH_SOLICITED;
303 	bth0 |= extra_bytes << 20;
304 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
305 		*pkey = hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index);
306 	else
307 		*pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
308 	if (!bypass)
309 		bth0 |= *pkey;
310 	ohdr->bth[0] = cpu_to_be32(bth0);
311 	ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn);
312 	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
313 	/*
314 	 * Qkeys with the high order bit set mean use the
315 	 * qkey from the QP context instead of the WR (see 10.2.5).
316 	 */
317 	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
318 					 qp->qkey : wqe->ud_wr.remote_qkey);
319 	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
320 }
321 
322 void hfi1_make_ud_req_9B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
323 			 struct rvt_swqe *wqe)
324 {
325 	u32 nwords, extra_bytes;
326 	u16 len, slid, dlid, pkey;
327 	u16 lrh0 = 0;
328 	u8 sc5;
329 	struct hfi1_qp_priv *priv = qp->priv;
330 	struct ib_other_headers *ohdr;
331 	struct rdma_ah_attr *ah_attr;
332 	struct hfi1_pportdata *ppd;
333 	struct hfi1_ibport *ibp;
334 	struct ib_grh *grh;
335 
336 	ibp = to_iport(qp->ibqp.device, qp->port_num);
337 	ppd = ppd_from_ibp(ibp);
338 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
339 
340 	extra_bytes = -wqe->length & 3;
341 	nwords = ((wqe->length + extra_bytes) >> 2) + SIZE_OF_CRC;
342 	/* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
343 	ps->s_txreq->hdr_dwords = 7;
344 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
345 		ps->s_txreq->hdr_dwords++;
346 
347 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
348 		grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
349 		ps->s_txreq->hdr_dwords +=
350 			hfi1_make_grh(ibp, grh, rdma_ah_read_grh(ah_attr),
351 				      ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
352 				      nwords);
353 		lrh0 = HFI1_LRH_GRH;
354 		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
355 	} else {
356 		lrh0 = HFI1_LRH_BTH;
357 		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
358 	}
359 
360 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
361 	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
362 	if (qp->ibqp.qp_type == IB_QPT_SMI) {
363 		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
364 		priv->s_sc = 0xf;
365 	} else {
366 		lrh0 |= (sc5 & 0xf) << 12;
367 		priv->s_sc = sc5;
368 	}
369 
370 	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
371 	if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
372 		slid = be16_to_cpu(IB_LID_PERMISSIVE);
373 	} else {
374 		u16 lid = (u16)ppd->lid;
375 
376 		if (lid) {
377 			lid |= rdma_ah_get_path_bits(ah_attr) &
378 				((1 << ppd->lmc) - 1);
379 			slid = lid;
380 		} else {
381 			slid = be16_to_cpu(IB_LID_PERMISSIVE);
382 		}
383 	}
384 	hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, false);
385 	len = ps->s_txreq->hdr_dwords + nwords;
386 
387 	/* Setup the packet */
388 	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_9B;
389 	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
390 			 lrh0, len, dlid, slid);
391 }
392 
393 void hfi1_make_ud_req_16B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
394 			  struct rvt_swqe *wqe)
395 {
396 	struct hfi1_qp_priv *priv = qp->priv;
397 	struct ib_other_headers *ohdr;
398 	struct rdma_ah_attr *ah_attr;
399 	struct hfi1_pportdata *ppd;
400 	struct hfi1_ibport *ibp;
401 	u32 dlid, slid, nwords, extra_bytes;
402 	u16 len, pkey;
403 	u8 l4, sc5;
404 
405 	ibp = to_iport(qp->ibqp.device, qp->port_num);
406 	ppd = ppd_from_ibp(ibp);
407 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
408 	/* header size in dwords 16B LRH+BTH+DETH = (16+12+8)/4. */
409 	ps->s_txreq->hdr_dwords = 9;
410 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
411 		ps->s_txreq->hdr_dwords++;
412 
413 	/* SW provides space for CRC and LT for bypass packets. */
414 	extra_bytes = hfi1_get_16b_padding((ps->s_txreq->hdr_dwords << 2),
415 					   wqe->length);
416 	nwords = ((wqe->length + extra_bytes + SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
417 
418 	if ((rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) &&
419 	    hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) {
420 		struct ib_grh *grh;
421 		struct ib_global_route *grd = rdma_ah_retrieve_grh(ah_attr);
422 		/*
423 		 * Ensure OPA GIDs are transformed to IB gids
424 		 * before creating the GRH.
425 		 */
426 		if (grd->sgid_index == OPA_GID_INDEX) {
427 			dd_dev_warn(ppd->dd, "Bad sgid_index. sgid_index: %d\n",
428 				    grd->sgid_index);
429 			grd->sgid_index = 0;
430 		}
431 		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
432 		ps->s_txreq->hdr_dwords += hfi1_make_grh(
433 			ibp, grh, grd,
434 			ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
435 			nwords);
436 		ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
437 		l4 = OPA_16B_L4_IB_GLOBAL;
438 	} else {
439 		ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
440 		l4 = OPA_16B_L4_IB_LOCAL;
441 	}
442 
443 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
444 	if (qp->ibqp.qp_type == IB_QPT_SMI)
445 		priv->s_sc = 0xf;
446 	else
447 		priv->s_sc = sc5;
448 
449 	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 16B);
450 	if (!ppd->lid)
451 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
452 	else
453 		slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
454 			   ((1 << ppd->lmc) - 1));
455 
456 	hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, true);
457 	/* Convert dwords to flits */
458 	len = (ps->s_txreq->hdr_dwords + nwords) >> 1;
459 
460 	/* Setup the packet */
461 	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_16B;
462 	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
463 			  slid, dlid, len, pkey, 0, 0, l4, priv->s_sc);
464 }
465 
466 /**
467  * hfi1_make_ud_req - construct a UD request packet
468  * @qp: the QP
469  *
470  * Assume s_lock is held.
471  *
472  * Return 1 if constructed; otherwise, return 0.
473  */
474 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
475 {
476 	struct hfi1_qp_priv *priv = qp->priv;
477 	struct rdma_ah_attr *ah_attr;
478 	struct hfi1_pportdata *ppd;
479 	struct hfi1_ibport *ibp;
480 	struct rvt_swqe *wqe;
481 	int next_cur;
482 	u32 lid;
483 
484 	ps->s_txreq = get_txreq(ps->dev, qp);
485 	if (IS_ERR(ps->s_txreq))
486 		goto bail_no_tx;
487 
488 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
489 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
490 			goto bail;
491 		/* We are in the error state, flush the work request. */
492 		if (qp->s_last == READ_ONCE(qp->s_head))
493 			goto bail;
494 		/* If DMAs are in progress, we can't flush immediately. */
495 		if (iowait_sdma_pending(&priv->s_iowait)) {
496 			qp->s_flags |= RVT_S_WAIT_DMA;
497 			goto bail;
498 		}
499 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
500 		hfi1_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
501 		goto done_free_tx;
502 	}
503 
504 	/* see post_one_send() */
505 	if (qp->s_cur == READ_ONCE(qp->s_head))
506 		goto bail;
507 
508 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
509 	next_cur = qp->s_cur + 1;
510 	if (next_cur >= qp->s_size)
511 		next_cur = 0;
512 
513 	/* Construct the header. */
514 	ibp = to_iport(qp->ibqp.device, qp->port_num);
515 	ppd = ppd_from_ibp(ibp);
516 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
517 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, ah_attr);
518 	if ((!hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) ||
519 	    (rdma_ah_get_dlid(ah_attr) == be32_to_cpu(OPA_LID_PERMISSIVE))) {
520 		lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1);
521 		if (unlikely(!loopback &&
522 			     ((lid == ppd->lid) ||
523 			      ((lid == be32_to_cpu(OPA_LID_PERMISSIVE)) &&
524 			       (qp->ibqp.qp_type == IB_QPT_GSI))))) {
525 			unsigned long tflags = ps->flags;
526 			/*
527 			 * If DMAs are in progress, we can't generate
528 			 * a completion for the loopback packet since
529 			 * it would be out of order.
530 			 * Instead of waiting, we could queue a
531 			 * zero length descriptor so we get a callback.
532 			 */
533 			if (iowait_sdma_pending(&priv->s_iowait)) {
534 				qp->s_flags |= RVT_S_WAIT_DMA;
535 				goto bail;
536 			}
537 			qp->s_cur = next_cur;
538 			spin_unlock_irqrestore(&qp->s_lock, tflags);
539 			ud_loopback(qp, wqe);
540 			spin_lock_irqsave(&qp->s_lock, tflags);
541 			ps->flags = tflags;
542 			hfi1_send_complete(qp, wqe, IB_WC_SUCCESS);
543 			goto done_free_tx;
544 		}
545 	}
546 
547 	qp->s_cur = next_cur;
548 	ps->s_txreq->s_cur_size = wqe->length;
549 	ps->s_txreq->ss = &qp->s_sge;
550 	qp->s_srate = rdma_ah_get_static_rate(ah_attr);
551 	qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
552 	qp->s_wqe = wqe;
553 	qp->s_sge.sge = wqe->sg_list[0];
554 	qp->s_sge.sg_list = wqe->sg_list + 1;
555 	qp->s_sge.num_sge = wqe->wr.num_sge;
556 	qp->s_sge.total_len = wqe->length;
557 
558 	/* Make the appropriate header */
559 	hfi1_make_ud_req_tbl[priv->hdr_type](qp, ps, qp->s_wqe);
560 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
561 	ps->s_txreq->sde = priv->s_sde;
562 	priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
563 	ps->s_txreq->psc = priv->s_sendcontext;
564 	/* disarm any ahg */
565 	priv->s_ahg->ahgcount = 0;
566 	priv->s_ahg->ahgidx = 0;
567 	priv->s_ahg->tx_flags = 0;
568 
569 	return 1;
570 
571 done_free_tx:
572 	hfi1_put_txreq(ps->s_txreq);
573 	ps->s_txreq = NULL;
574 	return 1;
575 
576 bail:
577 	hfi1_put_txreq(ps->s_txreq);
578 
579 bail_no_tx:
580 	ps->s_txreq = NULL;
581 	qp->s_flags &= ~RVT_S_BUSY;
582 	return 0;
583 }
584 
585 /*
586  * Hardware can't check this so we do it here.
587  *
588  * This is a slightly different algorithm than the standard pkey check.  It
589  * special cases the management keys and allows for 0x7fff and 0xffff to be in
590  * the table at the same time.
591  *
592  * @returns the index found or -1 if not found
593  */
594 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
595 {
596 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
597 	unsigned i;
598 
599 	if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) {
600 		unsigned lim_idx = -1;
601 
602 		for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) {
603 			/* here we look for an exact match */
604 			if (ppd->pkeys[i] == pkey)
605 				return i;
606 			if (ppd->pkeys[i] == LIM_MGMT_P_KEY)
607 				lim_idx = i;
608 		}
609 
610 		/* did not find 0xffff return 0x7fff idx if found */
611 		if (pkey == FULL_MGMT_P_KEY)
612 			return lim_idx;
613 
614 		/* no match...  */
615 		return -1;
616 	}
617 
618 	pkey &= 0x7fff; /* remove limited/full membership bit */
619 
620 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i)
621 		if ((ppd->pkeys[i] & 0x7fff) == pkey)
622 			return i;
623 
624 	/*
625 	 * Should not get here, this means hardware failed to validate pkeys.
626 	 */
627 	return -1;
628 }
629 
630 void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp,
631 		    u32 remote_qpn, u32 pkey, u32 slid, u32 dlid,
632 		    u8 sc5, const struct ib_grh *old_grh)
633 {
634 	u64 pbc, pbc_flags = 0;
635 	u32 bth0, plen, vl, hwords = 7;
636 	u16 len;
637 	u8 l4;
638 	struct hfi1_16b_header hdr;
639 	struct ib_other_headers *ohdr;
640 	struct pio_buf *pbuf;
641 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
642 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
643 	u32 nwords;
644 
645 	/* Populate length */
646 	nwords = ((hfi1_get_16b_padding(hwords << 2, 0) +
647 		   SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
648 	if (old_grh) {
649 		struct ib_grh *grh = &hdr.u.l.grh;
650 
651 		grh->version_tclass_flow = old_grh->version_tclass_flow;
652 		grh->paylen = cpu_to_be16(
653 			(hwords - LRH_16B_DWORDS + nwords) << 2);
654 		grh->hop_limit = 0xff;
655 		grh->sgid = old_grh->dgid;
656 		grh->dgid = old_grh->sgid;
657 		ohdr = &hdr.u.l.oth;
658 		l4 = OPA_16B_L4_IB_GLOBAL;
659 		hwords += sizeof(struct ib_grh) / sizeof(u32);
660 	} else {
661 		ohdr = &hdr.u.oth;
662 		l4 = OPA_16B_L4_IB_LOCAL;
663 	}
664 
665 	/* BIT 16 to 19 is TVER. Bit 20 to 22 is pad cnt */
666 	bth0 = (IB_OPCODE_CNP << 24) | (1 << 16) |
667 	       (hfi1_get_16b_padding(hwords << 2, 0) << 20);
668 	ohdr->bth[0] = cpu_to_be32(bth0);
669 
670 	ohdr->bth[1] = cpu_to_be32(remote_qpn);
671 	ohdr->bth[2] = 0; /* PSN 0 */
672 
673 	/* Convert dwords to flits */
674 	len = (hwords + nwords) >> 1;
675 	hfi1_make_16b_hdr(&hdr, slid, dlid, len, pkey, 1, 0, l4, sc5);
676 
677 	plen = 2 /* PBC */ + hwords + nwords;
678 	pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
679 	vl = sc_to_vlt(ppd->dd, sc5);
680 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
681 	if (ctxt) {
682 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
683 		if (pbuf)
684 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
685 						 &hdr, hwords);
686 	}
687 }
688 
689 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
690 		u32 pkey, u32 slid, u32 dlid, u8 sc5,
691 		const struct ib_grh *old_grh)
692 {
693 	u64 pbc, pbc_flags = 0;
694 	u32 bth0, plen, vl, hwords = 5;
695 	u16 lrh0;
696 	u8 sl = ibp->sc_to_sl[sc5];
697 	struct ib_header hdr;
698 	struct ib_other_headers *ohdr;
699 	struct pio_buf *pbuf;
700 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
701 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
702 
703 	if (old_grh) {
704 		struct ib_grh *grh = &hdr.u.l.grh;
705 
706 		grh->version_tclass_flow = old_grh->version_tclass_flow;
707 		grh->paylen = cpu_to_be16(
708 			(hwords - LRH_9B_DWORDS + SIZE_OF_CRC) << 2);
709 		grh->hop_limit = 0xff;
710 		grh->sgid = old_grh->dgid;
711 		grh->dgid = old_grh->sgid;
712 		ohdr = &hdr.u.l.oth;
713 		lrh0 = HFI1_LRH_GRH;
714 		hwords += sizeof(struct ib_grh) / sizeof(u32);
715 	} else {
716 		ohdr = &hdr.u.oth;
717 		lrh0 = HFI1_LRH_BTH;
718 	}
719 
720 	lrh0 |= (sc5 & 0xf) << 12 | sl << 4;
721 
722 	bth0 = pkey | (IB_OPCODE_CNP << 24);
723 	ohdr->bth[0] = cpu_to_be32(bth0);
724 
725 	ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT));
726 	ohdr->bth[2] = 0; /* PSN 0 */
727 
728 	hfi1_make_ib_hdr(&hdr, lrh0, hwords + SIZE_OF_CRC, dlid, slid);
729 	plen = 2 /* PBC */ + hwords;
730 	pbc_flags |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
731 	vl = sc_to_vlt(ppd->dd, sc5);
732 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
733 	if (ctxt) {
734 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
735 		if (pbuf)
736 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
737 						 &hdr, hwords);
738 	}
739 }
740 
741 /*
742  * opa_smp_check() - Do the regular pkey checking, and the additional
743  * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section
744  * 9.10.25 ("SMA Packet Checks").
745  *
746  * Note that:
747  *   - Checks are done using the pkey directly from the packet's BTH,
748  *     and specifically _not_ the pkey that we attach to the completion,
749  *     which may be different.
750  *   - These checks are specifically for "non-local" SMPs (i.e., SMPs
751  *     which originated on another node). SMPs which are sent from, and
752  *     destined to this node are checked in opa_local_smp_check().
753  *
754  * At the point where opa_smp_check() is called, we know:
755  *   - destination QP is QP0
756  *
757  * opa_smp_check() returns 0 if all checks succeed, 1 otherwise.
758  */
759 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5,
760 			 struct rvt_qp *qp, u16 slid, struct opa_smp *smp)
761 {
762 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
763 
764 	/*
765 	 * I don't think it's possible for us to get here with sc != 0xf,
766 	 * but check it to be certain.
767 	 */
768 	if (sc5 != 0xf)
769 		return 1;
770 
771 	if (rcv_pkey_check(ppd, pkey, sc5, slid))
772 		return 1;
773 
774 	/*
775 	 * At this point we know (and so don't need to check again) that
776 	 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY
777 	 * (see ingress_pkey_check).
778 	 */
779 	if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE &&
780 	    smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) {
781 		ingress_pkey_table_fail(ppd, pkey, slid);
782 		return 1;
783 	}
784 
785 	/*
786 	 * SMPs fall into one of four (disjoint) categories:
787 	 * SMA request, SMA response, SMA trap, or SMA trap repress.
788 	 * Our response depends, in part, on which type of SMP we're
789 	 * processing.
790 	 *
791 	 * If this is an SMA response, skip the check here.
792 	 *
793 	 * If this is an SMA request or SMA trap repress:
794 	 *   - pkey != FULL_MGMT_P_KEY =>
795 	 *       increment port recv constraint errors, drop MAD
796 	 *
797 	 * Otherwise:
798 	 *    - accept if the port is running an SM
799 	 *    - drop MAD if it's an SMA trap
800 	 *    - pkey == FULL_MGMT_P_KEY =>
801 	 *        reply with unsupported method
802 	 *    - pkey != FULL_MGMT_P_KEY =>
803 	 *	  increment port recv constraint errors, drop MAD
804 	 */
805 	switch (smp->method) {
806 	case IB_MGMT_METHOD_GET_RESP:
807 	case IB_MGMT_METHOD_REPORT_RESP:
808 		break;
809 	case IB_MGMT_METHOD_GET:
810 	case IB_MGMT_METHOD_SET:
811 	case IB_MGMT_METHOD_REPORT:
812 	case IB_MGMT_METHOD_TRAP_REPRESS:
813 		if (pkey != FULL_MGMT_P_KEY) {
814 			ingress_pkey_table_fail(ppd, pkey, slid);
815 			return 1;
816 		}
817 		break;
818 	default:
819 		if (ibp->rvp.port_cap_flags & IB_PORT_SM)
820 			return 0;
821 		if (smp->method == IB_MGMT_METHOD_TRAP)
822 			return 1;
823 		if (pkey == FULL_MGMT_P_KEY) {
824 			smp->status |= IB_SMP_UNSUP_METHOD;
825 			return 0;
826 		}
827 		ingress_pkey_table_fail(ppd, pkey, slid);
828 		return 1;
829 	}
830 	return 0;
831 }
832 
833 /**
834  * hfi1_ud_rcv - receive an incoming UD packet
835  * @ibp: the port the packet came in on
836  * @hdr: the packet header
837  * @rcv_flags: flags relevant to rcv processing
838  * @data: the packet data
839  * @tlen: the packet length
840  * @qp: the QP the packet came on
841  *
842  * This is called from qp_rcv() to process an incoming UD packet
843  * for the given QP.
844  * Called at interrupt level.
845  */
846 void hfi1_ud_rcv(struct hfi1_packet *packet)
847 {
848 	struct ib_other_headers *ohdr = packet->ohdr;
849 	u32 hdrsize = packet->hlen;
850 	struct ib_wc wc;
851 	u32 qkey;
852 	u32 src_qp;
853 	u16 pkey;
854 	int mgmt_pkey_idx = -1;
855 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
856 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
857 	void *data = packet->payload;
858 	u32 tlen = packet->tlen;
859 	struct rvt_qp *qp = packet->qp;
860 	u8 sc5 = packet->sc;
861 	u8 sl_from_sc;
862 	u8 opcode = packet->opcode;
863 	u8 sl = packet->sl;
864 	u32 dlid = packet->dlid;
865 	u32 slid = packet->slid;
866 	u8 extra_bytes;
867 	bool dlid_is_permissive;
868 	bool slid_is_permissive;
869 
870 	extra_bytes = packet->pad + packet->extra_byte + (SIZE_OF_CRC << 2);
871 	qkey = ib_get_qkey(ohdr);
872 	src_qp = ib_get_sqpn(ohdr);
873 
874 	if (packet->etype == RHF_RCV_TYPE_BYPASS) {
875 		u32 permissive_lid =
876 			opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B);
877 
878 		pkey = hfi1_16B_get_pkey(packet->hdr);
879 		dlid_is_permissive = (dlid == permissive_lid);
880 		slid_is_permissive = (slid == permissive_lid);
881 	} else {
882 		pkey = ib_bth_get_pkey(ohdr);
883 		dlid_is_permissive = (dlid == be16_to_cpu(IB_LID_PERMISSIVE));
884 		slid_is_permissive = (slid == be16_to_cpu(IB_LID_PERMISSIVE));
885 	}
886 	sl_from_sc = ibp->sc_to_sl[sc5];
887 
888 	process_ecn(qp, packet, (opcode != IB_OPCODE_CNP));
889 	/*
890 	 * Get the number of bytes the message was padded by
891 	 * and drop incomplete packets.
892 	 */
893 	if (unlikely(tlen < (hdrsize + extra_bytes)))
894 		goto drop;
895 
896 	tlen -= hdrsize + extra_bytes;
897 
898 	/*
899 	 * Check that the permissive LID is only used on QP0
900 	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
901 	 */
902 	if (qp->ibqp.qp_num) {
903 		if (unlikely(dlid_is_permissive || slid_is_permissive))
904 			goto drop;
905 		if (qp->ibqp.qp_num > 1) {
906 			if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
907 				/*
908 				 * Traps will not be sent for packets dropped
909 				 * by the HW. This is fine, as sending trap
910 				 * for invalid pkeys is optional according to
911 				 * IB spec (release 1.3, section 10.9.4)
912 				 */
913 				hfi1_bad_pkey(ibp,
914 					      pkey, sl,
915 					      src_qp, qp->ibqp.qp_num,
916 					      slid, dlid);
917 				return;
918 			}
919 		} else {
920 			/* GSI packet */
921 			mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
922 			if (mgmt_pkey_idx < 0)
923 				goto drop;
924 		}
925 		if (unlikely(qkey != qp->qkey)) /* Silent drop */
926 			return;
927 
928 		/* Drop invalid MAD packets (see 13.5.3.1). */
929 		if (unlikely(qp->ibqp.qp_num == 1 &&
930 			     (tlen > 2048 || (sc5 == 0xF))))
931 			goto drop;
932 	} else {
933 		/* Received on QP0, and so by definition, this is an SMP */
934 		struct opa_smp *smp = (struct opa_smp *)data;
935 
936 		if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp))
937 			goto drop;
938 
939 		if (tlen > 2048)
940 			goto drop;
941 		if ((dlid_is_permissive || slid_is_permissive) &&
942 		    smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
943 			goto drop;
944 
945 		/* look up SMI pkey */
946 		mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
947 		if (mgmt_pkey_idx < 0)
948 			goto drop;
949 	}
950 
951 	if (qp->ibqp.qp_num > 1 &&
952 	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
953 		wc.ex.imm_data = ohdr->u.ud.imm_data;
954 		wc.wc_flags = IB_WC_WITH_IMM;
955 		tlen -= sizeof(u32);
956 	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
957 		wc.ex.imm_data = 0;
958 		wc.wc_flags = 0;
959 	} else {
960 		goto drop;
961 	}
962 
963 	/*
964 	 * A GRH is expected to precede the data even if not
965 	 * present on the wire.
966 	 */
967 	wc.byte_len = tlen + sizeof(struct ib_grh);
968 
969 	/*
970 	 * Get the next work request entry to find where to put the data.
971 	 */
972 	if (qp->r_flags & RVT_R_REUSE_SGE) {
973 		qp->r_flags &= ~RVT_R_REUSE_SGE;
974 	} else {
975 		int ret;
976 
977 		ret = hfi1_rvt_get_rwqe(qp, 0);
978 		if (ret < 0) {
979 			rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
980 			return;
981 		}
982 		if (!ret) {
983 			if (qp->ibqp.qp_num == 0)
984 				ibp->rvp.n_vl15_dropped++;
985 			return;
986 		}
987 	}
988 	/* Silently drop packets which are too big. */
989 	if (unlikely(wc.byte_len > qp->r_len)) {
990 		qp->r_flags |= RVT_R_REUSE_SGE;
991 		goto drop;
992 	}
993 	if (packet->grh) {
994 		hfi1_copy_sge(&qp->r_sge, packet->grh,
995 			      sizeof(struct ib_grh), true, false);
996 		wc.wc_flags |= IB_WC_GRH;
997 	} else if (packet->etype == RHF_RCV_TYPE_BYPASS) {
998 		struct ib_grh grh;
999 		/*
1000 		 * Assuming we only created 16B on the send side
1001 		 * if we want to use large LIDs, since GRH was stripped
1002 		 * out when creating 16B, add back the GRH here.
1003 		 */
1004 		hfi1_make_ext_grh(packet, &grh, slid, dlid);
1005 		hfi1_copy_sge(&qp->r_sge, &grh,
1006 			      sizeof(struct ib_grh), true, false);
1007 		wc.wc_flags |= IB_WC_GRH;
1008 	} else {
1009 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
1010 	}
1011 	hfi1_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
1012 		      true, false);
1013 	rvt_put_ss(&qp->r_sge);
1014 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1015 		return;
1016 	wc.wr_id = qp->r_wr_id;
1017 	wc.status = IB_WC_SUCCESS;
1018 	wc.opcode = IB_WC_RECV;
1019 	wc.vendor_err = 0;
1020 	wc.qp = &qp->ibqp;
1021 	wc.src_qp = src_qp;
1022 
1023 	if (qp->ibqp.qp_type == IB_QPT_GSI ||
1024 	    qp->ibqp.qp_type == IB_QPT_SMI) {
1025 		if (mgmt_pkey_idx < 0) {
1026 			if (net_ratelimit()) {
1027 				struct hfi1_devdata *dd = ppd->dd;
1028 
1029 				dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n",
1030 					   qp->ibqp.qp_type);
1031 				mgmt_pkey_idx = 0;
1032 			}
1033 		}
1034 		wc.pkey_index = (unsigned)mgmt_pkey_idx;
1035 	} else {
1036 		wc.pkey_index = 0;
1037 	}
1038 	if (slid_is_permissive)
1039 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
1040 	wc.slid = slid & U16_MAX;
1041 	wc.sl = sl_from_sc;
1042 
1043 	/*
1044 	 * Save the LMC lower bits if the destination LID is a unicast LID.
1045 	 */
1046 	wc.dlid_path_bits = hfi1_check_mcast(dlid) ? 0 :
1047 		dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
1048 	wc.port_num = qp->port_num;
1049 	/* Signal completion event if the solicited bit is set. */
1050 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
1051 		     ib_bth_is_solicited(ohdr));
1052 	return;
1053 
1054 drop:
1055 	ibp->rvp.n_pkt_drops++;
1056 }
1057