1 /*
2 * Copyright(c) 2015 - 2019 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 "trace_ibhdrs.h"
55 #include "qp.h"
56
57 /* We support only two types - 9B and 16B for now */
58 static const hfi1_make_req hfi1_make_ud_req_tbl[2] = {
59 [HFI1_PKT_TYPE_9B] = &hfi1_make_ud_req_9B,
60 [HFI1_PKT_TYPE_16B] = &hfi1_make_ud_req_16B
61 };
62
63 /**
64 * ud_loopback - handle send on loopback QPs
65 * @sqp: the sending QP
66 * @swqe: the send work request
67 *
68 * This is called from hfi1_make_ud_req() to forward a WQE addressed
69 * to the same HFI.
70 * Note that the receive interrupt handler may be calling hfi1_ud_rcv()
71 * while this is being called.
72 */
ud_loopback(struct rvt_qp * sqp,struct rvt_swqe * swqe)73 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
74 {
75 struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
76 struct hfi1_pportdata *ppd;
77 struct hfi1_qp_priv *priv = sqp->priv;
78 struct rvt_qp *qp;
79 struct rdma_ah_attr *ah_attr;
80 unsigned long flags;
81 struct rvt_sge_state ssge;
82 struct rvt_sge *sge;
83 struct ib_wc wc;
84 u32 length;
85 enum ib_qp_type sqptype, dqptype;
86
87 rcu_read_lock();
88
89 qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
90 rvt_get_swqe_remote_qpn(swqe));
91 if (!qp) {
92 ibp->rvp.n_pkt_drops++;
93 rcu_read_unlock();
94 return;
95 }
96
97 sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
98 IB_QPT_UD : sqp->ibqp.qp_type;
99 dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
100 IB_QPT_UD : qp->ibqp.qp_type;
101
102 if (dqptype != sqptype ||
103 !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
104 ibp->rvp.n_pkt_drops++;
105 goto drop;
106 }
107
108 ah_attr = rvt_get_swqe_ah_attr(swqe);
109 ppd = ppd_from_ibp(ibp);
110
111 if (qp->ibqp.qp_num > 1) {
112 u16 pkey;
113 u32 slid;
114 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
115
116 pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
117 slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
118 ((1 << ppd->lmc) - 1));
119 if (unlikely(ingress_pkey_check(ppd, pkey, sc5,
120 qp->s_pkey_index,
121 slid, false))) {
122 hfi1_bad_pkey(ibp, pkey,
123 rdma_ah_get_sl(ah_attr),
124 sqp->ibqp.qp_num, qp->ibqp.qp_num,
125 slid, rdma_ah_get_dlid(ah_attr));
126 goto drop;
127 }
128 }
129
130 /*
131 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
132 * Qkeys with the high order bit set mean use the
133 * qkey from the QP context instead of the WR (see 10.2.5).
134 */
135 if (qp->ibqp.qp_num) {
136 u32 qkey;
137
138 qkey = (int)rvt_get_swqe_remote_qkey(swqe) < 0 ?
139 sqp->qkey : rvt_get_swqe_remote_qkey(swqe);
140 if (unlikely(qkey != qp->qkey))
141 goto drop; /* silently drop per IBTA spec */
142 }
143
144 /*
145 * A GRH is expected to precede the data even if not
146 * present on the wire.
147 */
148 length = swqe->length;
149 memset(&wc, 0, sizeof(wc));
150 wc.byte_len = length + sizeof(struct ib_grh);
151
152 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
153 wc.wc_flags = IB_WC_WITH_IMM;
154 wc.ex.imm_data = swqe->wr.ex.imm_data;
155 }
156
157 spin_lock_irqsave(&qp->r_lock, flags);
158
159 /*
160 * Get the next work request entry to find where to put the data.
161 */
162 if (qp->r_flags & RVT_R_REUSE_SGE) {
163 qp->r_flags &= ~RVT_R_REUSE_SGE;
164 } else {
165 int ret;
166
167 ret = rvt_get_rwqe(qp, false);
168 if (ret < 0) {
169 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
170 goto bail_unlock;
171 }
172 if (!ret) {
173 if (qp->ibqp.qp_num == 0)
174 ibp->rvp.n_vl15_dropped++;
175 goto bail_unlock;
176 }
177 }
178 /* Silently drop packets which are too big. */
179 if (unlikely(wc.byte_len > qp->r_len)) {
180 qp->r_flags |= RVT_R_REUSE_SGE;
181 ibp->rvp.n_pkt_drops++;
182 goto bail_unlock;
183 }
184
185 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
186 struct ib_grh grh;
187 struct ib_global_route grd = *(rdma_ah_read_grh(ah_attr));
188
189 /*
190 * For loopback packets with extended LIDs, the
191 * sgid_index in the GRH is 0 and the dgid is
192 * OPA GID of the sender. While creating a response
193 * to the loopback packet, IB core creates the new
194 * sgid_index from the DGID and that will be the
195 * OPA_GID_INDEX. The new dgid is from the sgid
196 * index and that will be in the IB GID format.
197 *
198 * We now have a case where the sent packet had a
199 * different sgid_index and dgid compared to the
200 * one that was received in response.
201 *
202 * Fix this inconsistency.
203 */
204 if (priv->hdr_type == HFI1_PKT_TYPE_16B) {
205 if (grd.sgid_index == 0)
206 grd.sgid_index = OPA_GID_INDEX;
207
208 if (ib_is_opa_gid(&grd.dgid))
209 grd.dgid.global.interface_id =
210 cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]);
211 }
212
213 hfi1_make_grh(ibp, &grh, &grd, 0, 0);
214 rvt_copy_sge(qp, &qp->r_sge, &grh,
215 sizeof(grh), true, false);
216 wc.wc_flags |= IB_WC_GRH;
217 } else {
218 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
219 }
220 ssge.sg_list = swqe->sg_list + 1;
221 ssge.sge = *swqe->sg_list;
222 ssge.num_sge = swqe->wr.num_sge;
223 sge = &ssge.sge;
224 while (length) {
225 u32 len = rvt_get_sge_length(sge, length);
226
227 WARN_ON_ONCE(len == 0);
228 rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, true, false);
229 rvt_update_sge(&ssge, len, false);
230 length -= len;
231 }
232 rvt_put_ss(&qp->r_sge);
233 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
234 goto bail_unlock;
235 wc.wr_id = qp->r_wr_id;
236 wc.status = IB_WC_SUCCESS;
237 wc.opcode = IB_WC_RECV;
238 wc.qp = &qp->ibqp;
239 wc.src_qp = sqp->ibqp.qp_num;
240 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) {
241 if (sqp->ibqp.qp_type == IB_QPT_GSI ||
242 sqp->ibqp.qp_type == IB_QPT_SMI)
243 wc.pkey_index = rvt_get_swqe_pkey_index(swqe);
244 else
245 wc.pkey_index = sqp->s_pkey_index;
246 } else {
247 wc.pkey_index = 0;
248 }
249 wc.slid = (ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
250 ((1 << ppd->lmc) - 1))) & U16_MAX;
251 /* Check for loopback when the port lid is not set */
252 if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
253 wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
254 wc.sl = rdma_ah_get_sl(ah_attr);
255 wc.dlid_path_bits = rdma_ah_get_dlid(ah_attr) & ((1 << ppd->lmc) - 1);
256 wc.port_num = qp->port_num;
257 /* Signal completion event if the solicited bit is set. */
258 rvt_recv_cq(qp, &wc, swqe->wr.send_flags & IB_SEND_SOLICITED);
259 ibp->rvp.n_loop_pkts++;
260 bail_unlock:
261 spin_unlock_irqrestore(&qp->r_lock, flags);
262 drop:
263 rcu_read_unlock();
264 }
265
hfi1_make_bth_deth(struct rvt_qp * qp,struct rvt_swqe * wqe,struct ib_other_headers * ohdr,u16 * pkey,u32 extra_bytes,bool bypass)266 static void hfi1_make_bth_deth(struct rvt_qp *qp, struct rvt_swqe *wqe,
267 struct ib_other_headers *ohdr,
268 u16 *pkey, u32 extra_bytes, bool bypass)
269 {
270 u32 bth0;
271 struct hfi1_ibport *ibp;
272
273 ibp = to_iport(qp->ibqp.device, qp->port_num);
274 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
275 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
276 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
277 } else {
278 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
279 }
280
281 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
282 bth0 |= IB_BTH_SOLICITED;
283 bth0 |= extra_bytes << 20;
284 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
285 *pkey = hfi1_get_pkey(ibp, rvt_get_swqe_pkey_index(wqe));
286 else
287 *pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
288 if (!bypass)
289 bth0 |= *pkey;
290 ohdr->bth[0] = cpu_to_be32(bth0);
291 ohdr->bth[1] = cpu_to_be32(rvt_get_swqe_remote_qpn(wqe));
292 ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
293 /*
294 * Qkeys with the high order bit set mean use the
295 * qkey from the QP context instead of the WR (see 10.2.5).
296 */
297 ohdr->u.ud.deth[0] =
298 cpu_to_be32((int)rvt_get_swqe_remote_qkey(wqe) < 0 ? qp->qkey :
299 rvt_get_swqe_remote_qkey(wqe));
300 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
301 }
302
hfi1_make_ud_req_9B(struct rvt_qp * qp,struct hfi1_pkt_state * ps,struct rvt_swqe * wqe)303 void hfi1_make_ud_req_9B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
304 struct rvt_swqe *wqe)
305 {
306 u32 nwords, extra_bytes;
307 u16 len, slid, dlid, pkey;
308 u16 lrh0 = 0;
309 u8 sc5;
310 struct hfi1_qp_priv *priv = qp->priv;
311 struct ib_other_headers *ohdr;
312 struct rdma_ah_attr *ah_attr;
313 struct hfi1_pportdata *ppd;
314 struct hfi1_ibport *ibp;
315 struct ib_grh *grh;
316
317 ibp = to_iport(qp->ibqp.device, qp->port_num);
318 ppd = ppd_from_ibp(ibp);
319 ah_attr = rvt_get_swqe_ah_attr(wqe);
320
321 extra_bytes = -wqe->length & 3;
322 nwords = ((wqe->length + extra_bytes) >> 2) + SIZE_OF_CRC;
323 /* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
324 ps->s_txreq->hdr_dwords = 7;
325 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
326 ps->s_txreq->hdr_dwords++;
327
328 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
329 grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
330 ps->s_txreq->hdr_dwords +=
331 hfi1_make_grh(ibp, grh, rdma_ah_read_grh(ah_attr),
332 ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
333 nwords);
334 lrh0 = HFI1_LRH_GRH;
335 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
336 } else {
337 lrh0 = HFI1_LRH_BTH;
338 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
339 }
340
341 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
342 lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
343 if (qp->ibqp.qp_type == IB_QPT_SMI) {
344 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
345 priv->s_sc = 0xf;
346 } else {
347 lrh0 |= (sc5 & 0xf) << 12;
348 priv->s_sc = sc5;
349 }
350
351 dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
352 if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
353 slid = be16_to_cpu(IB_LID_PERMISSIVE);
354 } else {
355 u16 lid = (u16)ppd->lid;
356
357 if (lid) {
358 lid |= rdma_ah_get_path_bits(ah_attr) &
359 ((1 << ppd->lmc) - 1);
360 slid = lid;
361 } else {
362 slid = be16_to_cpu(IB_LID_PERMISSIVE);
363 }
364 }
365 hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, false);
366 len = ps->s_txreq->hdr_dwords + nwords;
367
368 /* Setup the packet */
369 ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_9B;
370 hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
371 lrh0, len, dlid, slid);
372 }
373
hfi1_make_ud_req_16B(struct rvt_qp * qp,struct hfi1_pkt_state * ps,struct rvt_swqe * wqe)374 void hfi1_make_ud_req_16B(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
375 struct rvt_swqe *wqe)
376 {
377 struct hfi1_qp_priv *priv = qp->priv;
378 struct ib_other_headers *ohdr;
379 struct rdma_ah_attr *ah_attr;
380 struct hfi1_pportdata *ppd;
381 struct hfi1_ibport *ibp;
382 u32 dlid, slid, nwords, extra_bytes;
383 u32 dest_qp = rvt_get_swqe_remote_qpn(wqe);
384 u32 src_qp = qp->ibqp.qp_num;
385 u16 len, pkey;
386 u8 l4, sc5;
387 bool is_mgmt = false;
388
389 ibp = to_iport(qp->ibqp.device, qp->port_num);
390 ppd = ppd_from_ibp(ibp);
391 ah_attr = rvt_get_swqe_ah_attr(wqe);
392
393 /*
394 * Build 16B Management Packet if either the destination
395 * or source queue pair number is 0 or 1.
396 */
397 if (dest_qp == 0 || src_qp == 0 || dest_qp == 1 || src_qp == 1) {
398 /* header size in dwords 16B LRH+L4_FM = (16+8)/4. */
399 ps->s_txreq->hdr_dwords = 6;
400 is_mgmt = true;
401 } else {
402 /* header size in dwords 16B LRH+BTH+DETH = (16+12+8)/4. */
403 ps->s_txreq->hdr_dwords = 9;
404 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
405 ps->s_txreq->hdr_dwords++;
406 }
407
408 /* SW provides space for CRC and LT for bypass packets. */
409 extra_bytes = hfi1_get_16b_padding((ps->s_txreq->hdr_dwords << 2),
410 wqe->length);
411 nwords = ((wqe->length + extra_bytes + SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
412
413 if ((rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) &&
414 hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) {
415 struct ib_grh *grh;
416 struct ib_global_route *grd = rdma_ah_retrieve_grh(ah_attr);
417 /*
418 * Ensure OPA GIDs are transformed to IB gids
419 * before creating the GRH.
420 */
421 if (grd->sgid_index == OPA_GID_INDEX) {
422 dd_dev_warn(ppd->dd, "Bad sgid_index. sgid_index: %d\n",
423 grd->sgid_index);
424 grd->sgid_index = 0;
425 }
426 grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
427 ps->s_txreq->hdr_dwords += hfi1_make_grh(
428 ibp, grh, grd,
429 ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
430 nwords);
431 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
432 l4 = OPA_16B_L4_IB_GLOBAL;
433 } else {
434 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
435 l4 = OPA_16B_L4_IB_LOCAL;
436 }
437
438 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
439 if (qp->ibqp.qp_type == IB_QPT_SMI)
440 priv->s_sc = 0xf;
441 else
442 priv->s_sc = sc5;
443
444 dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 16B);
445 if (!ppd->lid)
446 slid = be32_to_cpu(OPA_LID_PERMISSIVE);
447 else
448 slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) &
449 ((1 << ppd->lmc) - 1));
450
451 if (is_mgmt) {
452 l4 = OPA_16B_L4_FM;
453 pkey = hfi1_get_pkey(ibp, rvt_get_swqe_pkey_index(wqe));
454 hfi1_16B_set_qpn(&ps->s_txreq->phdr.hdr.opah.u.mgmt,
455 dest_qp, src_qp);
456 } else {
457 hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, true);
458 }
459 /* Convert dwords to flits */
460 len = (ps->s_txreq->hdr_dwords + nwords) >> 1;
461
462 /* Setup the packet */
463 ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_16B;
464 hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
465 slid, dlid, len, pkey, 0, 0, l4, priv->s_sc);
466 }
467
468 /**
469 * hfi1_make_ud_req - construct a UD request packet
470 * @qp: the QP
471 *
472 * Assume s_lock is held.
473 *
474 * Return 1 if constructed; otherwise, return 0.
475 */
hfi1_make_ud_req(struct rvt_qp * qp,struct hfi1_pkt_state * ps)476 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
477 {
478 struct hfi1_qp_priv *priv = qp->priv;
479 struct rdma_ah_attr *ah_attr;
480 struct hfi1_pportdata *ppd;
481 struct hfi1_ibport *ibp;
482 struct rvt_swqe *wqe;
483 int next_cur;
484 u32 lid;
485
486 ps->s_txreq = get_txreq(ps->dev, qp);
487 if (!ps->s_txreq)
488 goto bail_no_tx;
489
490 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
491 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
492 goto bail;
493 /* We are in the error state, flush the work request. */
494 if (qp->s_last == READ_ONCE(qp->s_head))
495 goto bail;
496 /* If DMAs are in progress, we can't flush immediately. */
497 if (iowait_sdma_pending(&priv->s_iowait)) {
498 qp->s_flags |= RVT_S_WAIT_DMA;
499 goto bail;
500 }
501 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
502 rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
503 goto done_free_tx;
504 }
505
506 /* see post_one_send() */
507 if (qp->s_cur == READ_ONCE(qp->s_head))
508 goto bail;
509
510 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
511 next_cur = qp->s_cur + 1;
512 if (next_cur >= qp->s_size)
513 next_cur = 0;
514
515 /* Construct the header. */
516 ibp = to_iport(qp->ibqp.device, qp->port_num);
517 ppd = ppd_from_ibp(ibp);
518 ah_attr = rvt_get_swqe_ah_attr(wqe);
519 priv->hdr_type = hfi1_get_hdr_type(ppd->lid, ah_attr);
520 if ((!hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) ||
521 (rdma_ah_get_dlid(ah_attr) == be32_to_cpu(OPA_LID_PERMISSIVE))) {
522 lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1);
523 if (unlikely(!loopback &&
524 ((lid == ppd->lid) ||
525 ((lid == be32_to_cpu(OPA_LID_PERMISSIVE)) &&
526 (qp->ibqp.qp_type == IB_QPT_GSI))))) {
527 unsigned long tflags = ps->flags;
528 /*
529 * If DMAs are in progress, we can't generate
530 * a completion for the loopback packet since
531 * it would be out of order.
532 * Instead of waiting, we could queue a
533 * zero length descriptor so we get a callback.
534 */
535 if (iowait_sdma_pending(&priv->s_iowait)) {
536 qp->s_flags |= RVT_S_WAIT_DMA;
537 goto bail;
538 }
539 qp->s_cur = next_cur;
540 spin_unlock_irqrestore(&qp->s_lock, tflags);
541 ud_loopback(qp, wqe);
542 spin_lock_irqsave(&qp->s_lock, tflags);
543 ps->flags = tflags;
544 rvt_send_complete(qp, wqe, IB_WC_SUCCESS);
545 goto done_free_tx;
546 }
547 }
548
549 qp->s_cur = next_cur;
550 ps->s_txreq->s_cur_size = wqe->length;
551 ps->s_txreq->ss = &qp->s_sge;
552 qp->s_srate = rdma_ah_get_static_rate(ah_attr);
553 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
554 qp->s_wqe = wqe;
555 qp->s_sge.sge = wqe->sg_list[0];
556 qp->s_sge.sg_list = wqe->sg_list + 1;
557 qp->s_sge.num_sge = wqe->wr.num_sge;
558 qp->s_sge.total_len = wqe->length;
559
560 /* Make the appropriate header */
561 hfi1_make_ud_req_tbl[priv->hdr_type](qp, ps, qp->s_wqe);
562 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
563 ps->s_txreq->sde = priv->s_sde;
564 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
565 ps->s_txreq->psc = priv->s_sendcontext;
566 /* disarm any ahg */
567 priv->s_ahg->ahgcount = 0;
568 priv->s_ahg->ahgidx = 0;
569 priv->s_ahg->tx_flags = 0;
570
571 return 1;
572
573 done_free_tx:
574 hfi1_put_txreq(ps->s_txreq);
575 ps->s_txreq = NULL;
576 return 1;
577
578 bail:
579 hfi1_put_txreq(ps->s_txreq);
580
581 bail_no_tx:
582 ps->s_txreq = NULL;
583 qp->s_flags &= ~RVT_S_BUSY;
584 return 0;
585 }
586
587 /*
588 * Hardware can't check this so we do it here.
589 *
590 * This is a slightly different algorithm than the standard pkey check. It
591 * special cases the management keys and allows for 0x7fff and 0xffff to be in
592 * the table at the same time.
593 *
594 * @returns the index found or -1 if not found
595 */
hfi1_lookup_pkey_idx(struct hfi1_ibport * ibp,u16 pkey)596 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
597 {
598 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
599 unsigned i;
600
601 if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) {
602 unsigned lim_idx = -1;
603
604 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) {
605 /* here we look for an exact match */
606 if (ppd->pkeys[i] == pkey)
607 return i;
608 if (ppd->pkeys[i] == LIM_MGMT_P_KEY)
609 lim_idx = i;
610 }
611
612 /* did not find 0xffff return 0x7fff idx if found */
613 if (pkey == FULL_MGMT_P_KEY)
614 return lim_idx;
615
616 /* no match... */
617 return -1;
618 }
619
620 pkey &= 0x7fff; /* remove limited/full membership bit */
621
622 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i)
623 if ((ppd->pkeys[i] & 0x7fff) == pkey)
624 return i;
625
626 /*
627 * Should not get here, this means hardware failed to validate pkeys.
628 */
629 return -1;
630 }
631
return_cnp_16B(struct hfi1_ibport * ibp,struct rvt_qp * qp,u32 remote_qpn,u16 pkey,u32 slid,u32 dlid,u8 sc5,const struct ib_grh * old_grh)632 void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp,
633 u32 remote_qpn, u16 pkey, u32 slid, u32 dlid,
634 u8 sc5, const struct ib_grh *old_grh)
635 {
636 u64 pbc, pbc_flags = 0;
637 u32 bth0, plen, vl, hwords = 7;
638 u16 len;
639 u8 l4;
640 struct hfi1_opa_header hdr;
641 struct ib_other_headers *ohdr;
642 struct pio_buf *pbuf;
643 struct send_context *ctxt = qp_to_send_context(qp, sc5);
644 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
645 u32 nwords;
646
647 hdr.hdr_type = HFI1_PKT_TYPE_16B;
648 /* Populate length */
649 nwords = ((hfi1_get_16b_padding(hwords << 2, 0) +
650 SIZE_OF_LT) >> 2) + SIZE_OF_CRC;
651 if (old_grh) {
652 struct ib_grh *grh = &hdr.opah.u.l.grh;
653
654 grh->version_tclass_flow = old_grh->version_tclass_flow;
655 grh->paylen = cpu_to_be16(
656 (hwords - LRH_16B_DWORDS + nwords) << 2);
657 grh->hop_limit = 0xff;
658 grh->sgid = old_grh->dgid;
659 grh->dgid = old_grh->sgid;
660 ohdr = &hdr.opah.u.l.oth;
661 l4 = OPA_16B_L4_IB_GLOBAL;
662 hwords += sizeof(struct ib_grh) / sizeof(u32);
663 } else {
664 ohdr = &hdr.opah.u.oth;
665 l4 = OPA_16B_L4_IB_LOCAL;
666 }
667
668 /* BIT 16 to 19 is TVER. Bit 20 to 22 is pad cnt */
669 bth0 = (IB_OPCODE_CNP << 24) | (1 << 16) |
670 (hfi1_get_16b_padding(hwords << 2, 0) << 20);
671 ohdr->bth[0] = cpu_to_be32(bth0);
672
673 ohdr->bth[1] = cpu_to_be32(remote_qpn);
674 ohdr->bth[2] = 0; /* PSN 0 */
675
676 /* Convert dwords to flits */
677 len = (hwords + nwords) >> 1;
678 hfi1_make_16b_hdr(&hdr.opah, slid, dlid, len, pkey, 1, 0, l4, sc5);
679
680 plen = 2 /* PBC */ + hwords + nwords;
681 pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
682 vl = sc_to_vlt(ppd->dd, sc5);
683 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
684 if (ctxt) {
685 pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
686 if (!IS_ERR_OR_NULL(pbuf)) {
687 trace_pio_output_ibhdr(ppd->dd, &hdr, sc5);
688 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
689 &hdr, hwords);
690 }
691 }
692 }
693
return_cnp(struct hfi1_ibport * ibp,struct rvt_qp * qp,u32 remote_qpn,u16 pkey,u32 slid,u32 dlid,u8 sc5,const struct ib_grh * old_grh)694 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
695 u16 pkey, u32 slid, u32 dlid, u8 sc5,
696 const struct ib_grh *old_grh)
697 {
698 u64 pbc, pbc_flags = 0;
699 u32 bth0, plen, vl, hwords = 5;
700 u16 lrh0;
701 u8 sl = ibp->sc_to_sl[sc5];
702 struct hfi1_opa_header hdr;
703 struct ib_other_headers *ohdr;
704 struct pio_buf *pbuf;
705 struct send_context *ctxt = qp_to_send_context(qp, sc5);
706 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
707
708 hdr.hdr_type = HFI1_PKT_TYPE_9B;
709 if (old_grh) {
710 struct ib_grh *grh = &hdr.ibh.u.l.grh;
711
712 grh->version_tclass_flow = old_grh->version_tclass_flow;
713 grh->paylen = cpu_to_be16(
714 (hwords - LRH_9B_DWORDS + SIZE_OF_CRC) << 2);
715 grh->hop_limit = 0xff;
716 grh->sgid = old_grh->dgid;
717 grh->dgid = old_grh->sgid;
718 ohdr = &hdr.ibh.u.l.oth;
719 lrh0 = HFI1_LRH_GRH;
720 hwords += sizeof(struct ib_grh) / sizeof(u32);
721 } else {
722 ohdr = &hdr.ibh.u.oth;
723 lrh0 = HFI1_LRH_BTH;
724 }
725
726 lrh0 |= (sc5 & 0xf) << 12 | sl << 4;
727
728 bth0 = pkey | (IB_OPCODE_CNP << 24);
729 ohdr->bth[0] = cpu_to_be32(bth0);
730
731 ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT));
732 ohdr->bth[2] = 0; /* PSN 0 */
733
734 hfi1_make_ib_hdr(&hdr.ibh, lrh0, hwords + SIZE_OF_CRC, dlid, slid);
735 plen = 2 /* PBC */ + hwords;
736 pbc_flags |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
737 vl = sc_to_vlt(ppd->dd, sc5);
738 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
739 if (ctxt) {
740 pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
741 if (!IS_ERR_OR_NULL(pbuf)) {
742 trace_pio_output_ibhdr(ppd->dd, &hdr, sc5);
743 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
744 &hdr, hwords);
745 }
746 }
747 }
748
749 /*
750 * opa_smp_check() - Do the regular pkey checking, and the additional
751 * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section
752 * 9.10.25 ("SMA Packet Checks").
753 *
754 * Note that:
755 * - Checks are done using the pkey directly from the packet's BTH,
756 * and specifically _not_ the pkey that we attach to the completion,
757 * which may be different.
758 * - These checks are specifically for "non-local" SMPs (i.e., SMPs
759 * which originated on another node). SMPs which are sent from, and
760 * destined to this node are checked in opa_local_smp_check().
761 *
762 * At the point where opa_smp_check() is called, we know:
763 * - destination QP is QP0
764 *
765 * opa_smp_check() returns 0 if all checks succeed, 1 otherwise.
766 */
opa_smp_check(struct hfi1_ibport * ibp,u16 pkey,u8 sc5,struct rvt_qp * qp,u16 slid,struct opa_smp * smp)767 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5,
768 struct rvt_qp *qp, u16 slid, struct opa_smp *smp)
769 {
770 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
771
772 /*
773 * I don't think it's possible for us to get here with sc != 0xf,
774 * but check it to be certain.
775 */
776 if (sc5 != 0xf)
777 return 1;
778
779 if (rcv_pkey_check(ppd, pkey, sc5, slid))
780 return 1;
781
782 /*
783 * At this point we know (and so don't need to check again) that
784 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY
785 * (see ingress_pkey_check).
786 */
787 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE &&
788 smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) {
789 ingress_pkey_table_fail(ppd, pkey, slid);
790 return 1;
791 }
792
793 /*
794 * SMPs fall into one of four (disjoint) categories:
795 * SMA request, SMA response, SMA trap, or SMA trap repress.
796 * Our response depends, in part, on which type of SMP we're
797 * processing.
798 *
799 * If this is an SMA response, skip the check here.
800 *
801 * If this is an SMA request or SMA trap repress:
802 * - pkey != FULL_MGMT_P_KEY =>
803 * increment port recv constraint errors, drop MAD
804 *
805 * Otherwise:
806 * - accept if the port is running an SM
807 * - drop MAD if it's an SMA trap
808 * - pkey == FULL_MGMT_P_KEY =>
809 * reply with unsupported method
810 * - pkey != FULL_MGMT_P_KEY =>
811 * increment port recv constraint errors, drop MAD
812 */
813 switch (smp->method) {
814 case IB_MGMT_METHOD_GET_RESP:
815 case IB_MGMT_METHOD_REPORT_RESP:
816 break;
817 case IB_MGMT_METHOD_GET:
818 case IB_MGMT_METHOD_SET:
819 case IB_MGMT_METHOD_REPORT:
820 case IB_MGMT_METHOD_TRAP_REPRESS:
821 if (pkey != FULL_MGMT_P_KEY) {
822 ingress_pkey_table_fail(ppd, pkey, slid);
823 return 1;
824 }
825 break;
826 default:
827 if (ibp->rvp.port_cap_flags & IB_PORT_SM)
828 return 0;
829 if (smp->method == IB_MGMT_METHOD_TRAP)
830 return 1;
831 if (pkey == FULL_MGMT_P_KEY) {
832 smp->status |= IB_SMP_UNSUP_METHOD;
833 return 0;
834 }
835 ingress_pkey_table_fail(ppd, pkey, slid);
836 return 1;
837 }
838 return 0;
839 }
840
841 /**
842 * hfi1_ud_rcv - receive an incoming UD packet
843 * @ibp: the port the packet came in on
844 * @hdr: the packet header
845 * @rcv_flags: flags relevant to rcv processing
846 * @data: the packet data
847 * @tlen: the packet length
848 * @qp: the QP the packet came on
849 *
850 * This is called from qp_rcv() to process an incoming UD packet
851 * for the given QP.
852 * Called at interrupt level.
853 */
hfi1_ud_rcv(struct hfi1_packet * packet)854 void hfi1_ud_rcv(struct hfi1_packet *packet)
855 {
856 u32 hdrsize = packet->hlen;
857 struct ib_wc wc;
858 u32 src_qp;
859 u16 pkey;
860 int mgmt_pkey_idx = -1;
861 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
862 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
863 void *data = packet->payload;
864 u32 tlen = packet->tlen;
865 struct rvt_qp *qp = packet->qp;
866 u8 sc5 = packet->sc;
867 u8 sl_from_sc;
868 u8 opcode = packet->opcode;
869 u8 sl = packet->sl;
870 u32 dlid = packet->dlid;
871 u32 slid = packet->slid;
872 u8 extra_bytes;
873 u8 l4 = 0;
874 bool dlid_is_permissive;
875 bool slid_is_permissive;
876 bool solicited = false;
877
878 extra_bytes = packet->pad + packet->extra_byte + (SIZE_OF_CRC << 2);
879
880 if (packet->etype == RHF_RCV_TYPE_BYPASS) {
881 u32 permissive_lid =
882 opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B);
883
884 l4 = hfi1_16B_get_l4(packet->hdr);
885 pkey = hfi1_16B_get_pkey(packet->hdr);
886 dlid_is_permissive = (dlid == permissive_lid);
887 slid_is_permissive = (slid == permissive_lid);
888 } else {
889 pkey = ib_bth_get_pkey(packet->ohdr);
890 dlid_is_permissive = (dlid == be16_to_cpu(IB_LID_PERMISSIVE));
891 slid_is_permissive = (slid == be16_to_cpu(IB_LID_PERMISSIVE));
892 }
893 sl_from_sc = ibp->sc_to_sl[sc5];
894
895 if (likely(l4 != OPA_16B_L4_FM)) {
896 src_qp = ib_get_sqpn(packet->ohdr);
897 solicited = ib_bth_is_solicited(packet->ohdr);
898 } else {
899 src_qp = hfi1_16B_get_src_qpn(packet->mgmt);
900 }
901
902 process_ecn(qp, packet);
903 /*
904 * Get the number of bytes the message was padded by
905 * and drop incomplete packets.
906 */
907 if (unlikely(tlen < (hdrsize + extra_bytes)))
908 goto drop;
909
910 tlen -= hdrsize + extra_bytes;
911
912 /*
913 * Check that the permissive LID is only used on QP0
914 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
915 */
916 if (qp->ibqp.qp_num) {
917 if (unlikely(dlid_is_permissive || slid_is_permissive))
918 goto drop;
919 if (qp->ibqp.qp_num > 1) {
920 if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
921 /*
922 * Traps will not be sent for packets dropped
923 * by the HW. This is fine, as sending trap
924 * for invalid pkeys is optional according to
925 * IB spec (release 1.3, section 10.9.4)
926 */
927 hfi1_bad_pkey(ibp,
928 pkey, sl,
929 src_qp, qp->ibqp.qp_num,
930 slid, dlid);
931 return;
932 }
933 } else {
934 /* GSI packet */
935 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
936 if (mgmt_pkey_idx < 0)
937 goto drop;
938 }
939 if (unlikely(l4 != OPA_16B_L4_FM &&
940 ib_get_qkey(packet->ohdr) != qp->qkey))
941 return; /* Silent drop */
942
943 /* Drop invalid MAD packets (see 13.5.3.1). */
944 if (unlikely(qp->ibqp.qp_num == 1 &&
945 (tlen > 2048 || (sc5 == 0xF))))
946 goto drop;
947 } else {
948 /* Received on QP0, and so by definition, this is an SMP */
949 struct opa_smp *smp = (struct opa_smp *)data;
950
951 if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp))
952 goto drop;
953
954 if (tlen > 2048)
955 goto drop;
956 if ((dlid_is_permissive || slid_is_permissive) &&
957 smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
958 goto drop;
959
960 /* look up SMI pkey */
961 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
962 if (mgmt_pkey_idx < 0)
963 goto drop;
964 }
965
966 if (qp->ibqp.qp_num > 1 &&
967 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
968 wc.ex.imm_data = packet->ohdr->u.ud.imm_data;
969 wc.wc_flags = IB_WC_WITH_IMM;
970 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
971 wc.ex.imm_data = 0;
972 wc.wc_flags = 0;
973 } else {
974 goto drop;
975 }
976
977 /*
978 * A GRH is expected to precede the data even if not
979 * present on the wire.
980 */
981 wc.byte_len = tlen + sizeof(struct ib_grh);
982
983 /*
984 * Get the next work request entry to find where to put the data.
985 */
986 if (qp->r_flags & RVT_R_REUSE_SGE) {
987 qp->r_flags &= ~RVT_R_REUSE_SGE;
988 } else {
989 int ret;
990
991 ret = rvt_get_rwqe(qp, false);
992 if (ret < 0) {
993 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
994 return;
995 }
996 if (!ret) {
997 if (qp->ibqp.qp_num == 0)
998 ibp->rvp.n_vl15_dropped++;
999 return;
1000 }
1001 }
1002 /* Silently drop packets which are too big. */
1003 if (unlikely(wc.byte_len > qp->r_len)) {
1004 qp->r_flags |= RVT_R_REUSE_SGE;
1005 goto drop;
1006 }
1007 if (packet->grh) {
1008 rvt_copy_sge(qp, &qp->r_sge, packet->grh,
1009 sizeof(struct ib_grh), true, false);
1010 wc.wc_flags |= IB_WC_GRH;
1011 } else if (packet->etype == RHF_RCV_TYPE_BYPASS) {
1012 struct ib_grh grh;
1013 /*
1014 * Assuming we only created 16B on the send side
1015 * if we want to use large LIDs, since GRH was stripped
1016 * out when creating 16B, add back the GRH here.
1017 */
1018 hfi1_make_ext_grh(packet, &grh, slid, dlid);
1019 rvt_copy_sge(qp, &qp->r_sge, &grh,
1020 sizeof(struct ib_grh), true, false);
1021 wc.wc_flags |= IB_WC_GRH;
1022 } else {
1023 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
1024 }
1025 rvt_copy_sge(qp, &qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
1026 true, false);
1027 rvt_put_ss(&qp->r_sge);
1028 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1029 return;
1030 wc.wr_id = qp->r_wr_id;
1031 wc.status = IB_WC_SUCCESS;
1032 wc.opcode = IB_WC_RECV;
1033 wc.vendor_err = 0;
1034 wc.qp = &qp->ibqp;
1035 wc.src_qp = src_qp;
1036
1037 if (qp->ibqp.qp_type == IB_QPT_GSI ||
1038 qp->ibqp.qp_type == IB_QPT_SMI) {
1039 if (mgmt_pkey_idx < 0) {
1040 if (net_ratelimit()) {
1041 struct hfi1_devdata *dd = ppd->dd;
1042
1043 dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n",
1044 qp->ibqp.qp_type);
1045 mgmt_pkey_idx = 0;
1046 }
1047 }
1048 wc.pkey_index = (unsigned)mgmt_pkey_idx;
1049 } else {
1050 wc.pkey_index = 0;
1051 }
1052 if (slid_is_permissive)
1053 slid = be32_to_cpu(OPA_LID_PERMISSIVE);
1054 wc.slid = slid & U16_MAX;
1055 wc.sl = sl_from_sc;
1056
1057 /*
1058 * Save the LMC lower bits if the destination LID is a unicast LID.
1059 */
1060 wc.dlid_path_bits = hfi1_check_mcast(dlid) ? 0 :
1061 dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
1062 wc.port_num = qp->port_num;
1063 /* Signal completion event if the solicited bit is set. */
1064 rvt_recv_cq(qp, &wc, solicited);
1065 return;
1066
1067 drop:
1068 ibp->rvp.n_pkt_drops++;
1069 }
1070