2 * Copyright (c) 2006, 2007 QLogic Corporation. 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 "ipath_kernel.h"
39 static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
40 u32 *lengthp, struct ipath_sge_state *ss)
42 int user = to_ipd(qp->ibqp.pd)->user;
47 for (i = j = 0; i < wqe->num_sge; i++) {
48 if (wqe->sg_list[i].length == 0)
51 if ((user && wqe->sg_list[i].lkey == 0) ||
52 !ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
53 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
55 *lengthp += wqe->sg_list[i].length;
63 wc.wr_id = wqe->wr_id;
64 wc.status = IB_WC_LOC_PROT_ERR;
65 wc.opcode = IB_WC_RECV;
75 wc.dlid_path_bits = 0;
77 /* Signal solicited completion event. */
78 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
85 * ipath_ud_loopback - handle send on loopback QPs
88 * @length: the length of the data to send
89 * @wr: the work request
90 * @wc: the work completion entry
92 * This is called from ipath_post_ud_send() to forward a WQE addressed
94 * Note that the receive interrupt handler may be calling ipath_ud_rcv()
95 * while this is being called.
97 static void ipath_ud_loopback(struct ipath_qp *sqp,
98 struct ipath_sge_state *ss,
99 u32 length, struct ib_send_wr *wr,
102 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
104 struct ib_ah_attr *ah_attr;
107 struct ipath_srq *srq;
108 struct ipath_sge_state rsge;
109 struct ipath_sge *sge;
110 struct ipath_rwq *wq;
111 struct ipath_rwqe *wqe;
112 void (*handler)(struct ib_event *, void *);
116 qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
121 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
122 * Qkeys with the high order bit set mean use the
123 * qkey from the QP context instead of the WR (see 10.2.5).
125 if (unlikely(qp->ibqp.qp_num &&
126 ((int) wr->wr.ud.remote_qkey < 0
127 ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
128 /* XXX OK to lose a count once in a while. */
129 dev->qkey_violations++;
135 * A GRH is expected to preceed the data even if not
136 * present on the wire.
138 wc->byte_len = length + sizeof(struct ib_grh);
140 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
141 wc->wc_flags = IB_WC_WITH_IMM;
142 wc->imm_data = wr->imm_data;
148 if (wr->num_sge > 1) {
149 rsge.sg_list = kmalloc((wr->num_sge - 1) *
150 sizeof(struct ipath_sge),
156 * Get the next work request entry to find where to put the data.
157 * Note that it is safe to drop the lock after changing rq->tail
158 * since ipath_post_receive() won't fill the empty slot.
161 srq = to_isrq(qp->ibqp.srq);
162 handler = srq->ibsrq.event_handler;
170 spin_lock_irqsave(&rq->lock, flags);
174 if (unlikely(tail == wq->head)) {
175 spin_unlock_irqrestore(&rq->lock, flags);
179 /* Make sure entry is read after head index is read. */
181 wqe = get_rwqe_ptr(rq, tail);
182 if (++tail >= rq->size)
184 if (init_sge(qp, wqe, &rlen, &rsge))
188 /* Silently drop packets which are too big. */
189 if (wc->byte_len > rlen) {
190 spin_unlock_irqrestore(&rq->lock, flags);
195 wc->wr_id = wqe->wr_id;
200 * validate head pointer value and compute
201 * the number of remaining WQEs.
207 n += rq->size - tail;
210 if (n < srq->limit) {
214 spin_unlock_irqrestore(&rq->lock, flags);
215 ev.device = qp->ibqp.device;
216 ev.element.srq = qp->ibqp.srq;
217 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
218 handler(&ev, srq->ibsrq.srq_context);
220 spin_unlock_irqrestore(&rq->lock, flags);
222 spin_unlock_irqrestore(&rq->lock, flags);
224 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
225 if (ah_attr->ah_flags & IB_AH_GRH) {
226 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
227 wc->wc_flags |= IB_WC_GRH;
229 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
232 u32 len = sge->length;
236 if (len > sge->sge_length)
237 len = sge->sge_length;
239 ipath_copy_sge(&rsge, sge->vaddr, len);
242 sge->sge_length -= len;
243 if (sge->sge_length == 0) {
245 *sge = *ss->sg_list++;
246 } else if (sge->length == 0 && sge->mr != NULL) {
247 if (++sge->n >= IPATH_SEGSZ) {
248 if (++sge->m >= sge->mr->mapsz)
253 sge->mr->map[sge->m]->segs[sge->n].vaddr;
255 sge->mr->map[sge->m]->segs[sge->n].length;
259 wc->status = IB_WC_SUCCESS;
260 wc->opcode = IB_WC_RECV;
263 wc->src_qp = sqp->ibqp.qp_num;
264 /* XXX do we know which pkey matched? Only needed for GSI. */
266 wc->slid = dev->dd->ipath_lid |
267 (ah_attr->src_path_bits &
268 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
269 wc->sl = ah_attr->sl;
271 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
272 /* Signal completion event if the solicited bit is set. */
273 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
274 wr->send_flags & IB_SEND_SOLICITED);
279 if (atomic_dec_and_test(&qp->refcount))
284 * ipath_post_ud_send - post a UD send on QP
286 * @wr: the work request
288 * Note that we actually send the data as it is posted instead of putting
289 * the request into a ring buffer. If we wanted to use a ring buffer,
290 * we would need to save a reference to the destination address in the SWQE.
292 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
294 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
295 struct ipath_other_headers *ohdr;
296 struct ib_ah_attr *ah_attr;
297 struct ipath_sge_state ss;
298 struct ipath_sge *sg_list;
310 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
315 if (wr->wr.ud.ah->pd != qp->ibqp.pd) {
320 /* IB spec says that num_sge == 0 is OK. */
321 if (wr->num_sge > qp->s_max_sge) {
326 if (wr->num_sge > 1) {
327 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
336 /* Check the buffer to send. */
337 ss.sg_list = sg_list;
341 ss.sge.sge_length = 0;
344 for (i = 0; i < wr->num_sge; i++) {
346 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
351 if (wr->sg_list[i].length == 0)
353 if (!ipath_lkey_ok(qp, ss.num_sge ?
354 sg_list + ss.num_sge - 1 : &ss.sge,
355 &wr->sg_list[i], 0)) {
359 len += wr->sg_list[i].length;
362 /* Check for invalid packet size. */
363 if (len > dev->dd->ipath_ibmtu) {
367 extra_bytes = (4 - len) & 3;
368 nwords = (len + extra_bytes) >> 2;
370 /* Construct the header. */
371 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
372 if (ah_attr->dlid == 0) {
376 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
377 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
378 dev->n_multicast_xmit++;
380 dev->n_unicast_xmit++;
382 dev->n_unicast_xmit++;
383 lid = ah_attr->dlid &
384 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
385 if (unlikely(lid == dev->dd->ipath_lid)) {
387 * Pass in an uninitialized ib_wc to save stack
390 ipath_ud_loopback(qp, &ss, len, wr, &wc);
394 if (ah_attr->ah_flags & IB_AH_GRH) {
395 /* Header size in 32-bit words. */
397 lrh0 = IPATH_LRH_GRH;
398 ohdr = &qp->s_hdr.u.l.oth;
399 qp->s_hdr.u.l.grh.version_tclass_flow =
400 cpu_to_be32((6 << 28) |
401 (ah_attr->grh.traffic_class << 20) |
402 ah_attr->grh.flow_label);
403 qp->s_hdr.u.l.grh.paylen =
404 cpu_to_be16(((wr->opcode ==
405 IB_WR_SEND_WITH_IMM ? 6 : 5) +
406 nwords + SIZE_OF_CRC) << 2);
407 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
408 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
409 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
410 /* The SGID is 32-bit aligned. */
411 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
413 qp->s_hdr.u.l.grh.sgid.global.interface_id =
415 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
417 * Don't worry about sending to locally attached multicast
418 * QPs. It is unspecified by the spec. what happens.
421 /* Header size in 32-bit words. */
423 lrh0 = IPATH_LRH_BTH;
424 ohdr = &qp->s_hdr.u.oth;
426 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
427 ohdr->u.ud.imm_data = wr->imm_data;
428 wc.imm_data = wr->imm_data;
430 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
431 } else if (wr->opcode == IB_WR_SEND) {
433 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
438 lrh0 |= ah_attr->sl << 4;
439 if (qp->ibqp.qp_type == IB_QPT_SMI)
440 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
441 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
442 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
443 qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
444 lid = dev->dd->ipath_lid;
446 lid |= ah_attr->src_path_bits &
447 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
448 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
450 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
451 if (wr->send_flags & IB_SEND_SOLICITED)
453 bth0 |= extra_bytes << 20;
454 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
455 ipath_get_pkey(dev->dd, qp->s_pkey_index);
456 ohdr->bth[0] = cpu_to_be32(bth0);
458 * Use the multicast QP if the destination LID is a multicast LID.
460 ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
461 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
462 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
463 cpu_to_be32(wr->wr.ud.remote_qpn);
464 /* XXX Could lose a PSN count but not worth locking */
465 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
467 * Qkeys with the high order bit set mean use the
468 * qkey from the QP context instead of the WR (see 10.2.5).
470 ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
471 qp->qkey : wr->wr.ud.remote_qkey);
472 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
473 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
478 /* Queue the completion status entry. */
479 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
480 (wr->send_flags & IB_SEND_SIGNALED)) {
481 wc.wr_id = wr->wr_id;
482 wc.status = IB_WC_SUCCESS;
484 wc.opcode = IB_WC_SEND;
489 /* XXX initialize other fields? */
490 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
501 * ipath_ud_rcv - receive an incoming UD packet
502 * @dev: the device the packet came in on
503 * @hdr: the packet header
504 * @has_grh: true if the packet has a GRH
505 * @data: the packet data
506 * @tlen: the packet length
507 * @qp: the QP the packet came on
509 * This is called from ipath_qp_rcv() to process an incoming UD packet
511 * Called at interrupt level.
513 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
514 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
516 struct ipath_other_headers *ohdr;
529 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
530 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
531 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
534 ohdr = &hdr->u.l.oth;
535 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
537 * The header with GRH is 68 bytes and the core driver sets
538 * the eager header buffer size to 56 bytes so the last 12
539 * bytes of the IB header is in the data buffer.
541 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
542 if (header_in_data) {
543 qkey = be32_to_cpu(((__be32 *) data)[1]);
544 src_qp = be32_to_cpu(((__be32 *) data)[2]);
547 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
548 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
551 src_qp &= IPATH_QPN_MASK;
554 * Check that the permissive LID is only used on QP0
555 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
557 if (qp->ibqp.qp_num) {
558 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
559 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
563 if (unlikely(qkey != qp->qkey)) {
564 /* XXX OK to lose a count once in a while. */
565 dev->qkey_violations++;
569 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
570 hdr->lrh[3] == IB_LID_PERMISSIVE) {
571 struct ib_smp *smp = (struct ib_smp *) data;
573 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
579 /* Get the number of bytes the message was padded by. */
580 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
581 if (unlikely(tlen < (hdrsize + pad + 4))) {
582 /* Drop incomplete packets. */
586 tlen -= hdrsize + pad + 4;
588 /* Drop invalid MAD packets (see 13.5.3.1). */
589 if (unlikely((qp->ibqp.qp_num == 0 &&
591 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
592 (qp->ibqp.qp_num == 1 &&
594 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
600 * A GRH is expected to preceed the data even if not
601 * present on the wire.
603 wc.byte_len = tlen + sizeof(struct ib_grh);
606 * The opcode is in the low byte when its in network order
607 * (top byte when in host order).
609 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
610 if (qp->ibqp.qp_num > 1 &&
611 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
612 if (header_in_data) {
613 wc.imm_data = *(__be32 *) data;
614 data += sizeof(__be32);
616 wc.imm_data = ohdr->u.ud.imm_data;
617 wc.wc_flags = IB_WC_WITH_IMM;
618 hdrsize += sizeof(u32);
619 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
628 * Get the next work request entry to find where to put the data.
632 else if (!ipath_get_rwqe(qp, 0)) {
634 * Count VL15 packets dropped due to no receive buffer.
635 * Otherwise, count them as buffer overruns since usually,
636 * the HW will be able to receive packets even if there are
637 * no QPs with posted receive buffers.
639 if (qp->ibqp.qp_num == 0)
640 dev->n_vl15_dropped++;
645 /* Silently drop packets which are too big. */
646 if (wc.byte_len > qp->r_len) {
652 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
653 sizeof(struct ib_grh));
654 wc.wc_flags |= IB_WC_GRH;
656 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
657 ipath_copy_sge(&qp->r_sge, data,
658 wc.byte_len - sizeof(struct ib_grh));
659 qp->r_wrid_valid = 0;
660 wc.wr_id = qp->r_wr_id;
661 wc.status = IB_WC_SUCCESS;
662 wc.opcode = IB_WC_RECV;
666 /* XXX do we know which pkey matched? Only needed for GSI. */
668 wc.slid = be16_to_cpu(hdr->lrh[3]);
669 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
670 dlid = be16_to_cpu(hdr->lrh[1]);
672 * Save the LMC lower bits if the destination LID is a unicast LID.
674 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
675 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
676 /* Signal completion event if the solicited bit is set. */
677 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
679 __constant_cpu_to_be32(1 << 23)) != 0);