xref: /linux/drivers/infiniband/hw/hfi1/ud.c (revision 72bea132f3680ee51e7ed2cee62892b6f5121909)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2019 Intel Corporation.
4  */
5 
6 #include <linux/net.h>
7 #include <rdma/ib_smi.h>
8 
9 #include "hfi.h"
10 #include "mad.h"
11 #include "verbs_txreq.h"
12 #include "trace_ibhdrs.h"
13 #include "qp.h"
14 
15 /* We support only two types - 9B and 16B for now */
16 static const hfi1_make_req hfi1_make_ud_req_tbl[2] = {
17 	[HFI1_PKT_TYPE_9B] = &hfi1_make_ud_req_9B,
18 	[HFI1_PKT_TYPE_16B] = &hfi1_make_ud_req_16B
19 };
20 
21 /**
22  * ud_loopback - handle send on loopback QPs
23  * @sqp: the sending QP
24  * @swqe: the send work request
25  *
26  * This is called from hfi1_make_ud_req() to forward a WQE addressed
27  * to the same HFI.
28  * Note that the receive interrupt handler may be calling hfi1_ud_rcv()
29  * while this is being called.
30  */
31 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
32 {
33 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
34 	struct hfi1_pportdata *ppd;
35 	struct hfi1_qp_priv *priv = sqp->priv;
36 	struct rvt_qp *qp;
37 	struct rdma_ah_attr *ah_attr;
38 	unsigned long flags;
39 	struct rvt_sge_state ssge;
40 	struct rvt_sge *sge;
41 	struct ib_wc wc;
42 	u32 length;
43 	enum ib_qp_type sqptype, dqptype;
44 
45 	rcu_read_lock();
46 
47 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
48 			    rvt_get_swqe_remote_qpn(swqe));
49 	if (!qp) {
50 		ibp->rvp.n_pkt_drops++;
51 		rcu_read_unlock();
52 		return;
53 	}
54 
55 	sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
56 			IB_QPT_UD : sqp->ibqp.qp_type;
57 	dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
58 			IB_QPT_UD : qp->ibqp.qp_type;
59 
60 	if (dqptype != sqptype ||
61 	    !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
62 		ibp->rvp.n_pkt_drops++;
63 		goto drop;
64 	}
65 
66 	ah_attr = rvt_get_swqe_ah_attr(swqe);
67 	ppd = ppd_from_ibp(ibp);
68 
69 	if (qp->ibqp.qp_num > 1) {
70 		u16 pkey;
71 		u32 slid;
72 		u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
73 
74 		pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
75 		slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
76 				   ((1 << ppd->lmc) - 1));
77 		if (unlikely(ingress_pkey_check(ppd, pkey, sc5,
78 						qp->s_pkey_index,
79 						slid, false))) {
80 			hfi1_bad_pkey(ibp, pkey,
81 				      rdma_ah_get_sl(ah_attr),
82 				      sqp->ibqp.qp_num, qp->ibqp.qp_num,
83 				      slid, rdma_ah_get_dlid(ah_attr));
84 			goto drop;
85 		}
86 	}
87 
88 	/*
89 	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
90 	 * Qkeys with the high order bit set mean use the
91 	 * qkey from the QP context instead of the WR (see 10.2.5).
92 	 */
93 	if (qp->ibqp.qp_num) {
94 		u32 qkey;
95 
96 		qkey = (int)rvt_get_swqe_remote_qkey(swqe) < 0 ?
97 			sqp->qkey : rvt_get_swqe_remote_qkey(swqe);
98 		if (unlikely(qkey != qp->qkey))
99 			goto drop; /* silently drop per IBTA spec */
100 	}
101 
102 	/*
103 	 * A GRH is expected to precede the data even if not
104 	 * present on the wire.
105 	 */
106 	length = swqe->length;
107 	memset(&wc, 0, sizeof(wc));
108 	wc.byte_len = length + sizeof(struct ib_grh);
109 
110 	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
111 		wc.wc_flags = IB_WC_WITH_IMM;
112 		wc.ex.imm_data = swqe->wr.ex.imm_data;
113 	}
114 
115 	spin_lock_irqsave(&qp->r_lock, flags);
116 
117 	/*
118 	 * Get the next work request entry to find where to put the data.
119 	 */
120 	if (qp->r_flags & RVT_R_REUSE_SGE) {
121 		qp->r_flags &= ~RVT_R_REUSE_SGE;
122 	} else {
123 		int ret;
124 
125 		ret = rvt_get_rwqe(qp, false);
126 		if (ret < 0) {
127 			rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
128 			goto bail_unlock;
129 		}
130 		if (!ret) {
131 			if (qp->ibqp.qp_num == 0)
132 				ibp->rvp.n_vl15_dropped++;
133 			goto bail_unlock;
134 		}
135 	}
136 	/* Silently drop packets which are too big. */
137 	if (unlikely(wc.byte_len > qp->r_len)) {
138 		qp->r_flags |= RVT_R_REUSE_SGE;
139 		ibp->rvp.n_pkt_drops++;
140 		goto bail_unlock;
141 	}
142 
143 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
144 		struct ib_grh grh;
145 		struct ib_global_route grd = *(rdma_ah_read_grh(ah_attr));
146 
147 		/*
148 		 * For loopback packets with extended LIDs, the
149 		 * sgid_index in the GRH is 0 and the dgid is
150 		 * OPA GID of the sender. While creating a response
151 		 * to the loopback packet, IB core creates the new
152 		 * sgid_index from the DGID and that will be the
153 		 * OPA_GID_INDEX. The new dgid is from the sgid
154 		 * index and that will be in the IB GID format.
155 		 *
156 		 * We now have a case where the sent packet had a
157 		 * different sgid_index and dgid compared to the
158 		 * one that was received in response.
159 		 *
160 		 * Fix this inconsistency.
161 		 */
162 		if (priv->hdr_type == HFI1_PKT_TYPE_16B) {
163 			if (grd.sgid_index == 0)
164 				grd.sgid_index = OPA_GID_INDEX;
165 
166 			if (ib_is_opa_gid(&grd.dgid))
167 				grd.dgid.global.interface_id =
168 				cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]);
169 		}
170 
171 		hfi1_make_grh(ibp, &grh, &grd, 0, 0);
172 		rvt_copy_sge(qp, &qp->r_sge, &grh,
173 			     sizeof(grh), true, false);
174 		wc.wc_flags |= IB_WC_GRH;
175 	} else {
176 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
177 	}
178 	ssge.sg_list = swqe->sg_list + 1;
179 	ssge.sge = *swqe->sg_list;
180 	ssge.num_sge = swqe->wr.num_sge;
181 	sge = &ssge.sge;
182 	while (length) {
183 		u32 len = rvt_get_sge_length(sge, length);
184 
185 		WARN_ON_ONCE(len == 0);
186 		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, true, false);
187 		rvt_update_sge(&ssge, len, false);
188 		length -= len;
189 	}
190 	rvt_put_ss(&qp->r_sge);
191 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
192 		goto bail_unlock;
193 	wc.wr_id = qp->r_wr_id;
194 	wc.status = IB_WC_SUCCESS;
195 	wc.opcode = IB_WC_RECV;
196 	wc.qp = &qp->ibqp;
197 	wc.src_qp = sqp->ibqp.qp_num;
198 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) {
199 		if (sqp->ibqp.qp_type == IB_QPT_GSI ||
200 		    sqp->ibqp.qp_type == IB_QPT_SMI)
201 			wc.pkey_index = rvt_get_swqe_pkey_index(swqe);
202 		else
203 			wc.pkey_index = sqp->s_pkey_index;
204 	} else {
205 		wc.pkey_index = 0;
206 	}
207 	wc.slid = (ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
208 				   ((1 << ppd->lmc) - 1))) & U16_MAX;
209 	/* Check for loopback when the port lid is not set */
210 	if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
211 		wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
212 	wc.sl = rdma_ah_get_sl(ah_attr);
213 	wc.dlid_path_bits = rdma_ah_get_dlid(ah_attr) & ((1 << ppd->lmc) - 1);
214 	wc.port_num = qp->port_num;
215 	/* Signal completion event if the solicited bit is set. */
216 	rvt_recv_cq(qp, &wc, swqe->wr.send_flags & IB_SEND_SOLICITED);
217 	ibp->rvp.n_loop_pkts++;
218 bail_unlock:
219 	spin_unlock_irqrestore(&qp->r_lock, flags);
220 drop:
221 	rcu_read_unlock();
222 }
223 
224 static void hfi1_make_bth_deth(struct rvt_qp *qp, struct rvt_swqe *wqe,
225 			       struct ib_other_headers *ohdr,
226 			       u16 *pkey, u32 extra_bytes, bool bypass)
227 {
228 	u32 bth0;
229 	struct hfi1_ibport *ibp;
230 
231 	ibp = to_iport(qp->ibqp.device, qp->port_num);
232 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
233 		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
234 		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
235 	} else {
236 		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
237 	}
238 
239 	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
240 		bth0 |= IB_BTH_SOLICITED;
241 	bth0 |= extra_bytes << 20;
242 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
243 		*pkey = hfi1_get_pkey(ibp, rvt_get_swqe_pkey_index(wqe));
244 	else
245 		*pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
246 	if (!bypass)
247 		bth0 |= *pkey;
248 	ohdr->bth[0] = cpu_to_be32(bth0);
249 	ohdr->bth[1] = cpu_to_be32(rvt_get_swqe_remote_qpn(wqe));
250 	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
251 	/*
252 	 * Qkeys with the high order bit set mean use the
253 	 * qkey from the QP context instead of the WR (see 10.2.5).
254 	 */
255 	ohdr->u.ud.deth[0] =
256 		cpu_to_be32((int)rvt_get_swqe_remote_qkey(wqe) < 0 ? qp->qkey :
257 			    rvt_get_swqe_remote_qkey(wqe));
258 	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
259 }
260 
261 void hfi1_make_ud_req_9B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
262 			 struct rvt_swqe *wqe)
263 {
264 	u32 nwords, extra_bytes;
265 	u16 len, slid, dlid, pkey;
266 	u16 lrh0 = 0;
267 	u8 sc5;
268 	struct hfi1_qp_priv *priv = qp->priv;
269 	struct ib_other_headers *ohdr;
270 	struct rdma_ah_attr *ah_attr;
271 	struct hfi1_pportdata *ppd;
272 	struct hfi1_ibport *ibp;
273 	struct ib_grh *grh;
274 
275 	ibp = to_iport(qp->ibqp.device, qp->port_num);
276 	ppd = ppd_from_ibp(ibp);
277 	ah_attr = rvt_get_swqe_ah_attr(wqe);
278 
279 	extra_bytes = -wqe->length & 3;
280 	nwords = ((wqe->length + extra_bytes) >> 2) + SIZE_OF_CRC;
281 	/* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
282 	ps->s_txreq->hdr_dwords = 7;
283 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
284 		ps->s_txreq->hdr_dwords++;
285 
286 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
287 		grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
288 		ps->s_txreq->hdr_dwords +=
289 			hfi1_make_grh(ibp, grh, rdma_ah_read_grh(ah_attr),
290 				      ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
291 				      nwords);
292 		lrh0 = HFI1_LRH_GRH;
293 		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
294 	} else {
295 		lrh0 = HFI1_LRH_BTH;
296 		ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
297 	}
298 
299 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
300 	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
301 	if (qp->ibqp.qp_type == IB_QPT_SMI) {
302 		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
303 		priv->s_sc = 0xf;
304 	} else {
305 		lrh0 |= (sc5 & 0xf) << 12;
306 		priv->s_sc = sc5;
307 	}
308 
309 	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
310 	if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
311 		slid = be16_to_cpu(IB_LID_PERMISSIVE);
312 	} else {
313 		u16 lid = (u16)ppd->lid;
314 
315 		if (lid) {
316 			lid |= rdma_ah_get_path_bits(ah_attr) &
317 				((1 << ppd->lmc) - 1);
318 			slid = lid;
319 		} else {
320 			slid = be16_to_cpu(IB_LID_PERMISSIVE);
321 		}
322 	}
323 	hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, false);
324 	len = ps->s_txreq->hdr_dwords + nwords;
325 
326 	/* Setup the packet */
327 	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_9B;
328 	hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
329 			 lrh0, len, dlid, slid);
330 }
331 
332 void hfi1_make_ud_req_16B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
333 			  struct rvt_swqe *wqe)
334 {
335 	struct hfi1_qp_priv *priv = qp->priv;
336 	struct ib_other_headers *ohdr;
337 	struct rdma_ah_attr *ah_attr;
338 	struct hfi1_pportdata *ppd;
339 	struct hfi1_ibport *ibp;
340 	u32 dlid, slid, nwords, extra_bytes;
341 	u32 dest_qp = rvt_get_swqe_remote_qpn(wqe);
342 	u32 src_qp = qp->ibqp.qp_num;
343 	u16 len, pkey;
344 	u8 l4, sc5;
345 	bool is_mgmt = false;
346 
347 	ibp = to_iport(qp->ibqp.device, qp->port_num);
348 	ppd = ppd_from_ibp(ibp);
349 	ah_attr = rvt_get_swqe_ah_attr(wqe);
350 
351 	/*
352 	 * Build 16B Management Packet if either the destination
353 	 * or source queue pair number is 0 or 1.
354 	 */
355 	if (dest_qp == 0 || src_qp == 0 || dest_qp == 1 || src_qp == 1) {
356 		/* header size in dwords 16B LRH+L4_FM = (16+8)/4. */
357 		ps->s_txreq->hdr_dwords = 6;
358 		is_mgmt = true;
359 	} else {
360 		/* header size in dwords 16B LRH+BTH+DETH = (16+12+8)/4. */
361 		ps->s_txreq->hdr_dwords = 9;
362 		if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
363 			ps->s_txreq->hdr_dwords++;
364 	}
365 
366 	/* SW provides space for CRC and LT for bypass packets. */
367 	extra_bytes = hfi1_get_16b_padding((ps->s_txreq->hdr_dwords << 2),
368 					   wqe->length);
369 	nwords = ((wqe->length + extra_bytes + SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
370 
371 	if ((rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) &&
372 	    hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) {
373 		struct ib_grh *grh;
374 		struct ib_global_route *grd = rdma_ah_retrieve_grh(ah_attr);
375 		/*
376 		 * Ensure OPA GIDs are transformed to IB gids
377 		 * before creating the GRH.
378 		 */
379 		if (grd->sgid_index == OPA_GID_INDEX) {
380 			dd_dev_warn(ppd->dd, "Bad sgid_index. sgid_index: %d\n",
381 				    grd->sgid_index);
382 			grd->sgid_index = 0;
383 		}
384 		grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
385 		ps->s_txreq->hdr_dwords += hfi1_make_grh(
386 			ibp, grh, grd,
387 			ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
388 			nwords);
389 		ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
390 		l4 = OPA_16B_L4_IB_GLOBAL;
391 	} else {
392 		ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
393 		l4 = OPA_16B_L4_IB_LOCAL;
394 	}
395 
396 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
397 	if (qp->ibqp.qp_type == IB_QPT_SMI)
398 		priv->s_sc = 0xf;
399 	else
400 		priv->s_sc = sc5;
401 
402 	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 16B);
403 	if (!ppd->lid)
404 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
405 	else
406 		slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
407 			   ((1 << ppd->lmc) - 1));
408 
409 	if (is_mgmt) {
410 		l4 = OPA_16B_L4_FM;
411 		pkey = hfi1_get_pkey(ibp, rvt_get_swqe_pkey_index(wqe));
412 		hfi1_16B_set_qpn(&ps->s_txreq->phdr.hdr.opah.u.mgmt,
413 				 dest_qp, src_qp);
414 	} else {
415 		hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, true);
416 	}
417 	/* Convert dwords to flits */
418 	len = (ps->s_txreq->hdr_dwords + nwords) >> 1;
419 
420 	/* Setup the packet */
421 	ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_16B;
422 	hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
423 			  slid, dlid, len, pkey, 0, 0, l4, priv->s_sc);
424 }
425 
426 /**
427  * hfi1_make_ud_req - construct a UD request packet
428  * @qp: the QP
429  * @ps: the current packet state
430  *
431  * Assume s_lock is held.
432  *
433  * Return 1 if constructed; otherwise, return 0.
434  */
435 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
436 {
437 	struct hfi1_qp_priv *priv = qp->priv;
438 	struct rdma_ah_attr *ah_attr;
439 	struct hfi1_pportdata *ppd;
440 	struct hfi1_ibport *ibp;
441 	struct rvt_swqe *wqe;
442 	int next_cur;
443 	u32 lid;
444 
445 	ps->s_txreq = get_txreq(ps->dev, qp);
446 	if (!ps->s_txreq)
447 		goto bail_no_tx;
448 
449 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
450 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
451 			goto bail;
452 		/* We are in the error state, flush the work request. */
453 		if (qp->s_last == READ_ONCE(qp->s_head))
454 			goto bail;
455 		/* If DMAs are in progress, we can't flush immediately. */
456 		if (iowait_sdma_pending(&priv->s_iowait)) {
457 			qp->s_flags |= RVT_S_WAIT_DMA;
458 			goto bail;
459 		}
460 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
461 		rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
462 		goto done_free_tx;
463 	}
464 
465 	/* see post_one_send() */
466 	if (qp->s_cur == READ_ONCE(qp->s_head))
467 		goto bail;
468 
469 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
470 	next_cur = qp->s_cur + 1;
471 	if (next_cur >= qp->s_size)
472 		next_cur = 0;
473 
474 	/* Construct the header. */
475 	ibp = to_iport(qp->ibqp.device, qp->port_num);
476 	ppd = ppd_from_ibp(ibp);
477 	ah_attr = rvt_get_swqe_ah_attr(wqe);
478 	priv->hdr_type = hfi1_get_hdr_type(ppd->lid, ah_attr);
479 	if ((!hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) ||
480 	    (rdma_ah_get_dlid(ah_attr) == be32_to_cpu(OPA_LID_PERMISSIVE))) {
481 		lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1);
482 		if (unlikely(!loopback &&
483 			     ((lid == ppd->lid) ||
484 			      ((lid == be32_to_cpu(OPA_LID_PERMISSIVE)) &&
485 			       (qp->ibqp.qp_type == IB_QPT_GSI))))) {
486 			unsigned long tflags = ps->flags;
487 			/*
488 			 * If DMAs are in progress, we can't generate
489 			 * a completion for the loopback packet since
490 			 * it would be out of order.
491 			 * Instead of waiting, we could queue a
492 			 * zero length descriptor so we get a callback.
493 			 */
494 			if (iowait_sdma_pending(&priv->s_iowait)) {
495 				qp->s_flags |= RVT_S_WAIT_DMA;
496 				goto bail;
497 			}
498 			qp->s_cur = next_cur;
499 			spin_unlock_irqrestore(&qp->s_lock, tflags);
500 			ud_loopback(qp, wqe);
501 			spin_lock_irqsave(&qp->s_lock, tflags);
502 			ps->flags = tflags;
503 			rvt_send_complete(qp, wqe, IB_WC_SUCCESS);
504 			goto done_free_tx;
505 		}
506 	}
507 
508 	qp->s_cur = next_cur;
509 	ps->s_txreq->s_cur_size = wqe->length;
510 	ps->s_txreq->ss = &qp->s_sge;
511 	qp->s_srate = rdma_ah_get_static_rate(ah_attr);
512 	qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
513 	qp->s_wqe = wqe;
514 	qp->s_sge.sge = wqe->sg_list[0];
515 	qp->s_sge.sg_list = wqe->sg_list + 1;
516 	qp->s_sge.num_sge = wqe->wr.num_sge;
517 	qp->s_sge.total_len = wqe->length;
518 
519 	/* Make the appropriate header */
520 	hfi1_make_ud_req_tbl[priv->hdr_type](qp, ps, qp->s_wqe);
521 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
522 	ps->s_txreq->sde = priv->s_sde;
523 	priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
524 	ps->s_txreq->psc = priv->s_sendcontext;
525 	/* disarm any ahg */
526 	priv->s_ahg->ahgcount = 0;
527 	priv->s_ahg->ahgidx = 0;
528 	priv->s_ahg->tx_flags = 0;
529 
530 	return 1;
531 
532 done_free_tx:
533 	hfi1_put_txreq(ps->s_txreq);
534 	ps->s_txreq = NULL;
535 	return 1;
536 
537 bail:
538 	hfi1_put_txreq(ps->s_txreq);
539 
540 bail_no_tx:
541 	ps->s_txreq = NULL;
542 	qp->s_flags &= ~RVT_S_BUSY;
543 	return 0;
544 }
545 
546 /*
547  * Hardware can't check this so we do it here.
548  *
549  * This is a slightly different algorithm than the standard pkey check.  It
550  * special cases the management keys and allows for 0x7fff and 0xffff to be in
551  * the table at the same time.
552  *
553  * @returns the index found or -1 if not found
554  */
555 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
556 {
557 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
558 	unsigned i;
559 
560 	if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) {
561 		unsigned lim_idx = -1;
562 
563 		for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) {
564 			/* here we look for an exact match */
565 			if (ppd->pkeys[i] == pkey)
566 				return i;
567 			if (ppd->pkeys[i] == LIM_MGMT_P_KEY)
568 				lim_idx = i;
569 		}
570 
571 		/* did not find 0xffff return 0x7fff idx if found */
572 		if (pkey == FULL_MGMT_P_KEY)
573 			return lim_idx;
574 
575 		/* no match...  */
576 		return -1;
577 	}
578 
579 	pkey &= 0x7fff; /* remove limited/full membership bit */
580 
581 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i)
582 		if ((ppd->pkeys[i] & 0x7fff) == pkey)
583 			return i;
584 
585 	/*
586 	 * Should not get here, this means hardware failed to validate pkeys.
587 	 */
588 	return -1;
589 }
590 
591 void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp,
592 		    u32 remote_qpn, u16 pkey, u32 slid, u32 dlid,
593 		    u8 sc5, const struct ib_grh *old_grh)
594 {
595 	u64 pbc, pbc_flags = 0;
596 	u32 bth0, plen, vl, hwords = 7;
597 	u16 len;
598 	u8 l4;
599 	struct hfi1_opa_header hdr;
600 	struct ib_other_headers *ohdr;
601 	struct pio_buf *pbuf;
602 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
603 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
604 	u32 nwords;
605 
606 	hdr.hdr_type = HFI1_PKT_TYPE_16B;
607 	/* Populate length */
608 	nwords = ((hfi1_get_16b_padding(hwords << 2, 0) +
609 		   SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
610 	if (old_grh) {
611 		struct ib_grh *grh = &hdr.opah.u.l.grh;
612 
613 		grh->version_tclass_flow = old_grh->version_tclass_flow;
614 		grh->paylen = cpu_to_be16(
615 			(hwords - LRH_16B_DWORDS + nwords) << 2);
616 		grh->hop_limit = 0xff;
617 		grh->sgid = old_grh->dgid;
618 		grh->dgid = old_grh->sgid;
619 		ohdr = &hdr.opah.u.l.oth;
620 		l4 = OPA_16B_L4_IB_GLOBAL;
621 		hwords += sizeof(struct ib_grh) / sizeof(u32);
622 	} else {
623 		ohdr = &hdr.opah.u.oth;
624 		l4 = OPA_16B_L4_IB_LOCAL;
625 	}
626 
627 	/* BIT 16 to 19 is TVER. Bit 20 to 22 is pad cnt */
628 	bth0 = (IB_OPCODE_CNP << 24) | (1 << 16) |
629 	       (hfi1_get_16b_padding(hwords << 2, 0) << 20);
630 	ohdr->bth[0] = cpu_to_be32(bth0);
631 
632 	ohdr->bth[1] = cpu_to_be32(remote_qpn);
633 	ohdr->bth[2] = 0; /* PSN 0 */
634 
635 	/* Convert dwords to flits */
636 	len = (hwords + nwords) >> 1;
637 	hfi1_make_16b_hdr(&hdr.opah, slid, dlid, len, pkey, 1, 0, l4, sc5);
638 
639 	plen = 2 /* PBC */ + hwords + nwords;
640 	pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
641 	vl = sc_to_vlt(ppd->dd, sc5);
642 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
643 	if (ctxt) {
644 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
645 		if (!IS_ERR_OR_NULL(pbuf)) {
646 			trace_pio_output_ibhdr(ppd->dd, &hdr, sc5);
647 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
648 						 &hdr, hwords);
649 		}
650 	}
651 }
652 
653 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
654 		u16 pkey, u32 slid, u32 dlid, u8 sc5,
655 		const struct ib_grh *old_grh)
656 {
657 	u64 pbc, pbc_flags = 0;
658 	u32 bth0, plen, vl, hwords = 5;
659 	u16 lrh0;
660 	u8 sl = ibp->sc_to_sl[sc5];
661 	struct hfi1_opa_header hdr;
662 	struct ib_other_headers *ohdr;
663 	struct pio_buf *pbuf;
664 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
665 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
666 
667 	hdr.hdr_type = HFI1_PKT_TYPE_9B;
668 	if (old_grh) {
669 		struct ib_grh *grh = &hdr.ibh.u.l.grh;
670 
671 		grh->version_tclass_flow = old_grh->version_tclass_flow;
672 		grh->paylen = cpu_to_be16(
673 			(hwords - LRH_9B_DWORDS + SIZE_OF_CRC) << 2);
674 		grh->hop_limit = 0xff;
675 		grh->sgid = old_grh->dgid;
676 		grh->dgid = old_grh->sgid;
677 		ohdr = &hdr.ibh.u.l.oth;
678 		lrh0 = HFI1_LRH_GRH;
679 		hwords += sizeof(struct ib_grh) / sizeof(u32);
680 	} else {
681 		ohdr = &hdr.ibh.u.oth;
682 		lrh0 = HFI1_LRH_BTH;
683 	}
684 
685 	lrh0 |= (sc5 & 0xf) << 12 | sl << 4;
686 
687 	bth0 = pkey | (IB_OPCODE_CNP << 24);
688 	ohdr->bth[0] = cpu_to_be32(bth0);
689 
690 	ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT));
691 	ohdr->bth[2] = 0; /* PSN 0 */
692 
693 	hfi1_make_ib_hdr(&hdr.ibh, lrh0, hwords + SIZE_OF_CRC, dlid, slid);
694 	plen = 2 /* PBC */ + hwords;
695 	pbc_flags |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
696 	vl = sc_to_vlt(ppd->dd, sc5);
697 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
698 	if (ctxt) {
699 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
700 		if (!IS_ERR_OR_NULL(pbuf)) {
701 			trace_pio_output_ibhdr(ppd->dd, &hdr, sc5);
702 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
703 						 &hdr, hwords);
704 		}
705 	}
706 }
707 
708 /*
709  * opa_smp_check() - Do the regular pkey checking, and the additional
710  * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section
711  * 9.10.25 ("SMA Packet Checks").
712  *
713  * Note that:
714  *   - Checks are done using the pkey directly from the packet's BTH,
715  *     and specifically _not_ the pkey that we attach to the completion,
716  *     which may be different.
717  *   - These checks are specifically for "non-local" SMPs (i.e., SMPs
718  *     which originated on another node). SMPs which are sent from, and
719  *     destined to this node are checked in opa_local_smp_check().
720  *
721  * At the point where opa_smp_check() is called, we know:
722  *   - destination QP is QP0
723  *
724  * opa_smp_check() returns 0 if all checks succeed, 1 otherwise.
725  */
726 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5,
727 			 struct rvt_qp *qp, u16 slid, struct opa_smp *smp)
728 {
729 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
730 
731 	/*
732 	 * I don't think it's possible for us to get here with sc != 0xf,
733 	 * but check it to be certain.
734 	 */
735 	if (sc5 != 0xf)
736 		return 1;
737 
738 	if (rcv_pkey_check(ppd, pkey, sc5, slid))
739 		return 1;
740 
741 	/*
742 	 * At this point we know (and so don't need to check again) that
743 	 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY
744 	 * (see ingress_pkey_check).
745 	 */
746 	if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE &&
747 	    smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) {
748 		ingress_pkey_table_fail(ppd, pkey, slid);
749 		return 1;
750 	}
751 
752 	/*
753 	 * SMPs fall into one of four (disjoint) categories:
754 	 * SMA request, SMA response, SMA trap, or SMA trap repress.
755 	 * Our response depends, in part, on which type of SMP we're
756 	 * processing.
757 	 *
758 	 * If this is an SMA response, skip the check here.
759 	 *
760 	 * If this is an SMA request or SMA trap repress:
761 	 *   - pkey != FULL_MGMT_P_KEY =>
762 	 *       increment port recv constraint errors, drop MAD
763 	 *
764 	 * Otherwise:
765 	 *    - accept if the port is running an SM
766 	 *    - drop MAD if it's an SMA trap
767 	 *    - pkey == FULL_MGMT_P_KEY =>
768 	 *        reply with unsupported method
769 	 *    - pkey != FULL_MGMT_P_KEY =>
770 	 *	  increment port recv constraint errors, drop MAD
771 	 */
772 	switch (smp->method) {
773 	case IB_MGMT_METHOD_GET_RESP:
774 	case IB_MGMT_METHOD_REPORT_RESP:
775 		break;
776 	case IB_MGMT_METHOD_GET:
777 	case IB_MGMT_METHOD_SET:
778 	case IB_MGMT_METHOD_REPORT:
779 	case IB_MGMT_METHOD_TRAP_REPRESS:
780 		if (pkey != FULL_MGMT_P_KEY) {
781 			ingress_pkey_table_fail(ppd, pkey, slid);
782 			return 1;
783 		}
784 		break;
785 	default:
786 		if (ibp->rvp.port_cap_flags & IB_PORT_SM)
787 			return 0;
788 		if (smp->method == IB_MGMT_METHOD_TRAP)
789 			return 1;
790 		if (pkey == FULL_MGMT_P_KEY) {
791 			smp->status |= IB_SMP_UNSUP_METHOD;
792 			return 0;
793 		}
794 		ingress_pkey_table_fail(ppd, pkey, slid);
795 		return 1;
796 	}
797 	return 0;
798 }
799 
800 /**
801  * hfi1_ud_rcv - receive an incoming UD packet
802  * @packet: the packet structure
803  *
804  * This is called from qp_rcv() to process an incoming UD packet
805  * for the given QP.
806  * Called at interrupt level.
807  */
808 void hfi1_ud_rcv(struct hfi1_packet *packet)
809 {
810 	u32 hdrsize = packet->hlen;
811 	struct ib_wc wc;
812 	u32 src_qp;
813 	u16 pkey;
814 	int mgmt_pkey_idx = -1;
815 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
816 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
817 	void *data = packet->payload;
818 	u32 tlen = packet->tlen;
819 	struct rvt_qp *qp = packet->qp;
820 	u8 sc5 = packet->sc;
821 	u8 sl_from_sc;
822 	u8 opcode = packet->opcode;
823 	u8 sl = packet->sl;
824 	u32 dlid = packet->dlid;
825 	u32 slid = packet->slid;
826 	u8 extra_bytes;
827 	u8 l4 = 0;
828 	bool dlid_is_permissive;
829 	bool slid_is_permissive;
830 	bool solicited = false;
831 
832 	extra_bytes = packet->pad + packet->extra_byte + (SIZE_OF_CRC << 2);
833 
834 	if (packet->etype == RHF_RCV_TYPE_BYPASS) {
835 		u32 permissive_lid =
836 			opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B);
837 
838 		l4 = hfi1_16B_get_l4(packet->hdr);
839 		pkey = hfi1_16B_get_pkey(packet->hdr);
840 		dlid_is_permissive = (dlid == permissive_lid);
841 		slid_is_permissive = (slid == permissive_lid);
842 	} else {
843 		pkey = ib_bth_get_pkey(packet->ohdr);
844 		dlid_is_permissive = (dlid == be16_to_cpu(IB_LID_PERMISSIVE));
845 		slid_is_permissive = (slid == be16_to_cpu(IB_LID_PERMISSIVE));
846 	}
847 	sl_from_sc = ibp->sc_to_sl[sc5];
848 
849 	if (likely(l4 != OPA_16B_L4_FM)) {
850 		src_qp = ib_get_sqpn(packet->ohdr);
851 		solicited = ib_bth_is_solicited(packet->ohdr);
852 	} else {
853 		src_qp = hfi1_16B_get_src_qpn(packet->mgmt);
854 	}
855 
856 	process_ecn(qp, packet);
857 	/*
858 	 * Get the number of bytes the message was padded by
859 	 * and drop incomplete packets.
860 	 */
861 	if (unlikely(tlen < (hdrsize + extra_bytes)))
862 		goto drop;
863 
864 	tlen -= hdrsize + extra_bytes;
865 
866 	/*
867 	 * Check that the permissive LID is only used on QP0
868 	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
869 	 */
870 	if (qp->ibqp.qp_num) {
871 		if (unlikely(dlid_is_permissive || slid_is_permissive))
872 			goto drop;
873 		if (qp->ibqp.qp_num > 1) {
874 			if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
875 				/*
876 				 * Traps will not be sent for packets dropped
877 				 * by the HW. This is fine, as sending trap
878 				 * for invalid pkeys is optional according to
879 				 * IB spec (release 1.3, section 10.9.4)
880 				 */
881 				hfi1_bad_pkey(ibp,
882 					      pkey, sl,
883 					      src_qp, qp->ibqp.qp_num,
884 					      slid, dlid);
885 				return;
886 			}
887 		} else {
888 			/* GSI packet */
889 			mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
890 			if (mgmt_pkey_idx < 0)
891 				goto drop;
892 		}
893 		if (unlikely(l4 != OPA_16B_L4_FM &&
894 			     ib_get_qkey(packet->ohdr) != qp->qkey))
895 			return; /* Silent drop */
896 
897 		/* Drop invalid MAD packets (see 13.5.3.1). */
898 		if (unlikely(qp->ibqp.qp_num == 1 &&
899 			     (tlen > 2048 || (sc5 == 0xF))))
900 			goto drop;
901 	} else {
902 		/* Received on QP0, and so by definition, this is an SMP */
903 		struct opa_smp *smp = (struct opa_smp *)data;
904 
905 		if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp))
906 			goto drop;
907 
908 		if (tlen > 2048)
909 			goto drop;
910 		if ((dlid_is_permissive || slid_is_permissive) &&
911 		    smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
912 			goto drop;
913 
914 		/* look up SMI pkey */
915 		mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
916 		if (mgmt_pkey_idx < 0)
917 			goto drop;
918 	}
919 
920 	if (qp->ibqp.qp_num > 1 &&
921 	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
922 		wc.ex.imm_data = packet->ohdr->u.ud.imm_data;
923 		wc.wc_flags = IB_WC_WITH_IMM;
924 	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
925 		wc.ex.imm_data = 0;
926 		wc.wc_flags = 0;
927 	} else {
928 		goto drop;
929 	}
930 
931 	/*
932 	 * A GRH is expected to precede the data even if not
933 	 * present on the wire.
934 	 */
935 	wc.byte_len = tlen + sizeof(struct ib_grh);
936 
937 	/*
938 	 * Get the next work request entry to find where to put the data.
939 	 */
940 	if (qp->r_flags & RVT_R_REUSE_SGE) {
941 		qp->r_flags &= ~RVT_R_REUSE_SGE;
942 	} else {
943 		int ret;
944 
945 		ret = rvt_get_rwqe(qp, false);
946 		if (ret < 0) {
947 			rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
948 			return;
949 		}
950 		if (!ret) {
951 			if (qp->ibqp.qp_num == 0)
952 				ibp->rvp.n_vl15_dropped++;
953 			return;
954 		}
955 	}
956 	/* Silently drop packets which are too big. */
957 	if (unlikely(wc.byte_len > qp->r_len)) {
958 		qp->r_flags |= RVT_R_REUSE_SGE;
959 		goto drop;
960 	}
961 	if (packet->grh) {
962 		rvt_copy_sge(qp, &qp->r_sge, packet->grh,
963 			     sizeof(struct ib_grh), true, false);
964 		wc.wc_flags |= IB_WC_GRH;
965 	} else if (packet->etype == RHF_RCV_TYPE_BYPASS) {
966 		struct ib_grh grh;
967 		/*
968 		 * Assuming we only created 16B on the send side
969 		 * if we want to use large LIDs, since GRH was stripped
970 		 * out when creating 16B, add back the GRH here.
971 		 */
972 		hfi1_make_ext_grh(packet, &grh, slid, dlid);
973 		rvt_copy_sge(qp, &qp->r_sge, &grh,
974 			     sizeof(struct ib_grh), true, false);
975 		wc.wc_flags |= IB_WC_GRH;
976 	} else {
977 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
978 	}
979 	rvt_copy_sge(qp, &qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
980 		     true, false);
981 	rvt_put_ss(&qp->r_sge);
982 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
983 		return;
984 	wc.wr_id = qp->r_wr_id;
985 	wc.status = IB_WC_SUCCESS;
986 	wc.opcode = IB_WC_RECV;
987 	wc.vendor_err = 0;
988 	wc.qp = &qp->ibqp;
989 	wc.src_qp = src_qp;
990 
991 	if (qp->ibqp.qp_type == IB_QPT_GSI ||
992 	    qp->ibqp.qp_type == IB_QPT_SMI) {
993 		if (mgmt_pkey_idx < 0) {
994 			if (net_ratelimit()) {
995 				struct hfi1_devdata *dd = ppd->dd;
996 
997 				dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n",
998 					   qp->ibqp.qp_type);
999 				mgmt_pkey_idx = 0;
1000 			}
1001 		}
1002 		wc.pkey_index = (unsigned)mgmt_pkey_idx;
1003 	} else {
1004 		wc.pkey_index = 0;
1005 	}
1006 	if (slid_is_permissive)
1007 		slid = be32_to_cpu(OPA_LID_PERMISSIVE);
1008 	wc.slid = slid & U16_MAX;
1009 	wc.sl = sl_from_sc;
1010 
1011 	/*
1012 	 * Save the LMC lower bits if the destination LID is a unicast LID.
1013 	 */
1014 	wc.dlid_path_bits = hfi1_check_mcast(dlid) ? 0 :
1015 		dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
1016 	wc.port_num = qp->port_num;
1017 	/* Signal completion event if the solicited bit is set. */
1018 	rvt_recv_cq(qp, &wc, solicited);
1019 	return;
1020 
1021 drop:
1022 	ibp->rvp.n_pkt_drops++;
1023 }
1024