2 * Copyright (c) 2006 QLogic, Inc. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <rdma/ib_smi.h>
36 #include "ipath_verbs.h"
37 #include "ips_common.h"
40 * ipath_ud_loopback - handle send on loopback QPs
43 * @length: the length of the data to send
44 * @wr: the work request
45 * @wc: the work completion entry
47 * This is called from ipath_post_ud_send() to forward a WQE addressed
50 static void ipath_ud_loopback(struct ipath_qp *sqp,
51 struct ipath_sge_state *ss,
52 u32 length, struct ib_send_wr *wr,
55 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
57 struct ib_ah_attr *ah_attr;
60 struct ipath_srq *srq;
61 struct ipath_sge_state rsge;
62 struct ipath_sge *sge;
63 struct ipath_rwqe *wqe;
65 qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
70 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
71 * Qkeys with the high order bit set mean use the
72 * qkey from the QP context instead of the WR (see 10.2.5).
74 if (unlikely(qp->ibqp.qp_num &&
75 ((int) wr->wr.ud.remote_qkey < 0
76 ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
77 /* XXX OK to lose a count once in a while. */
78 dev->qkey_violations++;
84 * A GRH is expected to preceed the data even if not
85 * present on the wire.
87 wc->byte_len = length + sizeof(struct ib_grh);
89 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
90 wc->wc_flags = IB_WC_WITH_IMM;
91 wc->imm_data = wr->imm_data;
98 * Get the next work request entry to find where to put the data.
99 * Note that it is safe to drop the lock after changing rq->tail
100 * since ipath_post_receive() won't fill the empty slot.
103 srq = to_isrq(qp->ibqp.srq);
109 spin_lock_irqsave(&rq->lock, flags);
110 if (rq->tail == rq->head) {
111 spin_unlock_irqrestore(&rq->lock, flags);
115 /* Silently drop packets which are too big. */
116 wqe = get_rwqe_ptr(rq, rq->tail);
117 if (wc->byte_len > wqe->length) {
118 spin_unlock_irqrestore(&rq->lock, flags);
122 wc->wr_id = wqe->wr_id;
123 rsge.sge = wqe->sg_list[0];
124 rsge.sg_list = wqe->sg_list + 1;
125 rsge.num_sge = wqe->num_sge;
126 if (++rq->tail >= rq->size)
128 if (srq && srq->ibsrq.event_handler) {
131 if (rq->head < rq->tail)
132 n = rq->size + rq->head - rq->tail;
134 n = rq->head - rq->tail;
135 if (n < srq->limit) {
139 spin_unlock_irqrestore(&rq->lock, flags);
140 ev.device = qp->ibqp.device;
141 ev.element.srq = qp->ibqp.srq;
142 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
143 srq->ibsrq.event_handler(&ev,
144 srq->ibsrq.srq_context);
146 spin_unlock_irqrestore(&rq->lock, flags);
148 spin_unlock_irqrestore(&rq->lock, flags);
149 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
150 if (ah_attr->ah_flags & IB_AH_GRH) {
151 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
152 wc->wc_flags |= IB_WC_GRH;
154 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
157 u32 len = sge->length;
162 ipath_copy_sge(&rsge, sge->vaddr, len);
165 sge->sge_length -= len;
166 if (sge->sge_length == 0) {
168 *sge = *ss->sg_list++;
169 } else if (sge->length == 0 && sge->mr != NULL) {
170 if (++sge->n >= IPATH_SEGSZ) {
171 if (++sge->m >= sge->mr->mapsz)
176 sge->mr->map[sge->m]->segs[sge->n].vaddr;
178 sge->mr->map[sge->m]->segs[sge->n].length;
182 wc->status = IB_WC_SUCCESS;
183 wc->opcode = IB_WC_RECV;
185 wc->qp_num = qp->ibqp.qp_num;
186 wc->src_qp = sqp->ibqp.qp_num;
187 /* XXX do we know which pkey matched? Only needed for GSI. */
189 wc->slid = ipath_layer_get_lid(dev->dd) |
190 (ah_attr->src_path_bits &
191 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
192 wc->sl = ah_attr->sl;
194 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
195 /* Signal completion event if the solicited bit is set. */
196 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
197 wr->send_flags & IB_SEND_SOLICITED);
200 if (atomic_dec_and_test(&qp->refcount))
205 * ipath_post_ud_send - post a UD send on QP
207 * @wr: the work request
209 * Note that we actually send the data as it is posted instead of putting
210 * the request into a ring buffer. If we wanted to use a ring buffer,
211 * we would need to save a reference to the destination address in the SWQE.
213 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
215 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
216 struct ipath_other_headers *ohdr;
217 struct ib_ah_attr *ah_attr;
218 struct ipath_sge_state ss;
219 struct ipath_sge *sg_list;
231 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
236 /* IB spec says that num_sge == 0 is OK. */
237 if (wr->num_sge > qp->s_max_sge) {
242 if (wr->num_sge > 1) {
243 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
252 /* Check the buffer to send. */
253 ss.sg_list = sg_list;
257 ss.sge.sge_length = 0;
260 for (i = 0; i < wr->num_sge; i++) {
262 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
267 if (wr->sg_list[i].length == 0)
269 if (!ipath_lkey_ok(&dev->lk_table, ss.num_sge ?
270 sg_list + ss.num_sge - 1 : &ss.sge,
271 &wr->sg_list[i], 0)) {
275 len += wr->sg_list[i].length;
278 extra_bytes = (4 - len) & 3;
279 nwords = (len + extra_bytes) >> 2;
281 /* Construct the header. */
282 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
283 if (ah_attr->dlid == 0) {
287 if (ah_attr->dlid >= IPS_MULTICAST_LID_BASE) {
288 if (ah_attr->dlid != IPS_PERMISSIVE_LID)
289 dev->n_multicast_xmit++;
291 dev->n_unicast_xmit++;
293 dev->n_unicast_xmit++;
294 lid = ah_attr->dlid &
295 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
296 if (unlikely(lid == ipath_layer_get_lid(dev->dd))) {
298 * Pass in an uninitialized ib_wc to save stack
301 ipath_ud_loopback(qp, &ss, len, wr, &wc);
305 if (ah_attr->ah_flags & IB_AH_GRH) {
306 /* Header size in 32-bit words. */
309 ohdr = &qp->s_hdr.u.l.oth;
310 qp->s_hdr.u.l.grh.version_tclass_flow =
311 cpu_to_be32((6 << 28) |
312 (ah_attr->grh.traffic_class << 20) |
313 ah_attr->grh.flow_label);
314 qp->s_hdr.u.l.grh.paylen =
315 cpu_to_be16(((wr->opcode ==
316 IB_WR_SEND_WITH_IMM ? 6 : 5) +
317 nwords + SIZE_OF_CRC) << 2);
318 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
319 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
320 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
321 /* The SGID is 32-bit aligned. */
322 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
324 qp->s_hdr.u.l.grh.sgid.global.interface_id =
325 ipath_layer_get_guid(dev->dd);
326 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
328 * Don't worry about sending to locally attached multicast
329 * QPs. It is unspecified by the spec. what happens.
332 /* Header size in 32-bit words. */
335 ohdr = &qp->s_hdr.u.oth;
337 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
338 ohdr->u.ud.imm_data = wr->imm_data;
339 wc.imm_data = wr->imm_data;
341 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
342 } else if (wr->opcode == IB_WR_SEND) {
344 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
349 lrh0 |= ah_attr->sl << 4;
350 if (qp->ibqp.qp_type == IB_QPT_SMI)
351 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
352 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
353 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
354 qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
355 lid = ipath_layer_get_lid(dev->dd);
357 lid |= ah_attr->src_path_bits &
358 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
359 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
361 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
362 if (wr->send_flags & IB_SEND_SOLICITED)
364 bth0 |= extra_bytes << 20;
365 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPS_DEFAULT_P_KEY :
366 ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
367 ohdr->bth[0] = cpu_to_be32(bth0);
369 * Use the multicast QP if the destination LID is a multicast LID.
371 ohdr->bth[1] = ah_attr->dlid >= IPS_MULTICAST_LID_BASE &&
372 ah_attr->dlid != IPS_PERMISSIVE_LID ?
373 __constant_cpu_to_be32(IPS_MULTICAST_QPN) :
374 cpu_to_be32(wr->wr.ud.remote_qpn);
375 /* XXX Could lose a PSN count but not worth locking */
376 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPS_PSN_MASK);
378 * Qkeys with the high order bit set mean use the
379 * qkey from the QP context instead of the WR (see 10.2.5).
381 ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
382 qp->qkey : wr->wr.ud.remote_qkey);
383 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
384 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
389 /* Queue the completion status entry. */
390 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
391 (wr->send_flags & IB_SEND_SIGNALED)) {
392 wc.wr_id = wr->wr_id;
393 wc.status = IB_WC_SUCCESS;
395 wc.opcode = IB_WC_SEND;
397 wc.qp_num = qp->ibqp.qp_num;
400 /* XXX initialize other fields? */
401 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
412 * ipath_ud_rcv - receive an incoming UD packet
413 * @dev: the device the packet came in on
414 * @hdr: the packet header
415 * @has_grh: true if the packet has a GRH
416 * @data: the packet data
417 * @tlen: the packet length
418 * @qp: the QP the packet came on
420 * This is called from ipath_qp_rcv() to process an incoming UD packet
422 * Called at interrupt level.
424 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
425 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
427 struct ipath_other_headers *ohdr;
436 struct ipath_srq *srq;
437 struct ipath_rwqe *wqe;
444 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
445 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
446 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
449 ohdr = &hdr->u.l.oth;
450 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
452 * The header with GRH is 68 bytes and the core driver sets
453 * the eager header buffer size to 56 bytes so the last 12
454 * bytes of the IB header is in the data buffer.
457 ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
458 if (header_in_data) {
459 qkey = be32_to_cpu(((__be32 *) data)[1]);
460 src_qp = be32_to_cpu(((__be32 *) data)[2]);
463 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
464 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
467 src_qp &= IPS_QPN_MASK;
470 * Check that the permissive LID is only used on QP0
471 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
473 if (qp->ibqp.qp_num) {
474 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
475 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
479 if (unlikely(qkey != qp->qkey)) {
480 /* XXX OK to lose a count once in a while. */
481 dev->qkey_violations++;
485 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
486 hdr->lrh[3] == IB_LID_PERMISSIVE) {
487 struct ib_smp *smp = (struct ib_smp *) data;
489 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
495 /* Get the number of bytes the message was padded by. */
496 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
497 if (unlikely(tlen < (hdrsize + pad + 4))) {
498 /* Drop incomplete packets. */
502 tlen -= hdrsize + pad + 4;
504 /* Drop invalid MAD packets (see 13.5.3.1). */
505 if (unlikely((qp->ibqp.qp_num == 0 &&
507 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
508 (qp->ibqp.qp_num == 1 &&
510 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
516 * A GRH is expected to preceed the data even if not
517 * present on the wire.
519 wc.byte_len = tlen + sizeof(struct ib_grh);
522 * The opcode is in the low byte when its in network order
523 * (top byte when in host order).
525 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
526 if (qp->ibqp.qp_num > 1 &&
527 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
528 if (header_in_data) {
529 wc.imm_data = *(__be32 *) data;
530 data += sizeof(__be32);
532 wc.imm_data = ohdr->u.ud.imm_data;
533 wc.wc_flags = IB_WC_WITH_IMM;
534 hdrsize += sizeof(u32);
535 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
544 * Get the next work request entry to find where to put the data.
545 * Note that it is safe to drop the lock after changing rq->tail
546 * since ipath_post_receive() won't fill the empty slot.
549 srq = to_isrq(qp->ibqp.srq);
555 spin_lock_irqsave(&rq->lock, flags);
556 if (rq->tail == rq->head) {
557 spin_unlock_irqrestore(&rq->lock, flags);
561 /* Silently drop packets which are too big. */
562 wqe = get_rwqe_ptr(rq, rq->tail);
563 if (wc.byte_len > wqe->length) {
564 spin_unlock_irqrestore(&rq->lock, flags);
568 wc.wr_id = wqe->wr_id;
569 qp->r_sge.sge = wqe->sg_list[0];
570 qp->r_sge.sg_list = wqe->sg_list + 1;
571 qp->r_sge.num_sge = wqe->num_sge;
572 if (++rq->tail >= rq->size)
574 if (srq && srq->ibsrq.event_handler) {
577 if (rq->head < rq->tail)
578 n = rq->size + rq->head - rq->tail;
580 n = rq->head - rq->tail;
581 if (n < srq->limit) {
585 spin_unlock_irqrestore(&rq->lock, flags);
586 ev.device = qp->ibqp.device;
587 ev.element.srq = qp->ibqp.srq;
588 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
589 srq->ibsrq.event_handler(&ev,
590 srq->ibsrq.srq_context);
592 spin_unlock_irqrestore(&rq->lock, flags);
594 spin_unlock_irqrestore(&rq->lock, flags);
596 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
597 sizeof(struct ib_grh));
598 wc.wc_flags |= IB_WC_GRH;
600 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
601 ipath_copy_sge(&qp->r_sge, data,
602 wc.byte_len - sizeof(struct ib_grh));
603 wc.status = IB_WC_SUCCESS;
604 wc.opcode = IB_WC_RECV;
606 wc.qp_num = qp->ibqp.qp_num;
608 /* XXX do we know which pkey matched? Only needed for GSI. */
610 wc.slid = be16_to_cpu(hdr->lrh[3]);
611 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
612 dlid = be16_to_cpu(hdr->lrh[1]);
614 * Save the LMC lower bits if the destination LID is a unicast LID.
616 wc.dlid_path_bits = dlid >= IPS_MULTICAST_LID_BASE ? 0 :
617 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
618 /* Signal completion event if the solicited bit is set. */
619 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
621 __constant_cpu_to_be32(1 << 23)) != 0);