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 "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 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
43 int user = to_ipd(qp->ibqp.pd)->user;
48 for (i = j = 0; i < wqe->num_sge; i++) {
49 if (wqe->sg_list[i].length == 0)
52 if ((user && wqe->sg_list[i].lkey == 0) ||
53 !ipath_lkey_ok(&dev->lk_table,
54 j ? &ss->sg_list[j - 1] : &ss->sge,
55 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
57 *lengthp += wqe->sg_list[i].length;
65 wc.wr_id = wqe->wr_id;
66 wc.status = IB_WC_LOC_PROT_ERR;
67 wc.opcode = IB_WC_RECV;
71 wc.qp_num = qp->ibqp.qp_num;
77 wc.dlid_path_bits = 0;
79 /* Signal solicited completion event. */
80 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
87 * ipath_ud_loopback - handle send on loopback QPs
90 * @length: the length of the data to send
91 * @wr: the work request
92 * @wc: the work completion entry
94 * This is called from ipath_post_ud_send() to forward a WQE addressed
96 * Note that the receive interrupt handler may be calling ipath_ud_rcv()
97 * while this is being called.
99 static void ipath_ud_loopback(struct ipath_qp *sqp,
100 struct ipath_sge_state *ss,
101 u32 length, struct ib_send_wr *wr,
104 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
106 struct ib_ah_attr *ah_attr;
109 struct ipath_srq *srq;
110 struct ipath_sge_state rsge;
111 struct ipath_sge *sge;
112 struct ipath_rwq *wq;
113 struct ipath_rwqe *wqe;
114 void (*handler)(struct ib_event *, void *);
118 qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
123 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
124 * Qkeys with the high order bit set mean use the
125 * qkey from the QP context instead of the WR (see 10.2.5).
127 if (unlikely(qp->ibqp.qp_num &&
128 ((int) wr->wr.ud.remote_qkey < 0
129 ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
130 /* XXX OK to lose a count once in a while. */
131 dev->qkey_violations++;
137 * A GRH is expected to preceed the data even if not
138 * present on the wire.
140 wc->byte_len = length + sizeof(struct ib_grh);
142 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
143 wc->wc_flags = IB_WC_WITH_IMM;
144 wc->imm_data = wr->imm_data;
150 if (wr->num_sge > 1) {
151 rsge.sg_list = kmalloc((wr->num_sge - 1) *
152 sizeof(struct ipath_sge),
158 * Get the next work request entry to find where to put the data.
159 * Note that it is safe to drop the lock after changing rq->tail
160 * since ipath_post_receive() won't fill the empty slot.
163 srq = to_isrq(qp->ibqp.srq);
164 handler = srq->ibsrq.event_handler;
172 spin_lock_irqsave(&rq->lock, flags);
176 if (unlikely(tail == wq->head)) {
177 spin_unlock_irqrestore(&rq->lock, flags);
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;
237 ipath_copy_sge(&rsge, sge->vaddr, len);
240 sge->sge_length -= len;
241 if (sge->sge_length == 0) {
243 *sge = *ss->sg_list++;
244 } else if (sge->length == 0 && sge->mr != NULL) {
245 if (++sge->n >= IPATH_SEGSZ) {
246 if (++sge->m >= sge->mr->mapsz)
251 sge->mr->map[sge->m]->segs[sge->n].vaddr;
253 sge->mr->map[sge->m]->segs[sge->n].length;
257 wc->status = IB_WC_SUCCESS;
258 wc->opcode = IB_WC_RECV;
260 wc->qp_num = qp->ibqp.qp_num;
261 wc->src_qp = sqp->ibqp.qp_num;
262 /* XXX do we know which pkey matched? Only needed for GSI. */
264 wc->slid = dev->dd->ipath_lid |
265 (ah_attr->src_path_bits &
266 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
267 wc->sl = ah_attr->sl;
269 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
270 /* Signal completion event if the solicited bit is set. */
271 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
272 wr->send_flags & IB_SEND_SOLICITED);
277 if (atomic_dec_and_test(&qp->refcount))
282 * ipath_post_ud_send - post a UD send on QP
284 * @wr: the work request
286 * Note that we actually send the data as it is posted instead of putting
287 * the request into a ring buffer. If we wanted to use a ring buffer,
288 * we would need to save a reference to the destination address in the SWQE.
290 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
292 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
293 struct ipath_other_headers *ohdr;
294 struct ib_ah_attr *ah_attr;
295 struct ipath_sge_state ss;
296 struct ipath_sge *sg_list;
308 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
313 /* IB spec says that num_sge == 0 is OK. */
314 if (wr->num_sge > qp->s_max_sge) {
319 if (wr->num_sge > 1) {
320 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
329 /* Check the buffer to send. */
330 ss.sg_list = sg_list;
334 ss.sge.sge_length = 0;
337 for (i = 0; i < wr->num_sge; i++) {
339 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
344 if (wr->sg_list[i].length == 0)
346 if (!ipath_lkey_ok(&dev->lk_table, ss.num_sge ?
347 sg_list + ss.num_sge - 1 : &ss.sge,
348 &wr->sg_list[i], 0)) {
352 len += wr->sg_list[i].length;
355 /* Check for invalid packet size. */
356 if (len > dev->dd->ipath_ibmtu) {
360 extra_bytes = (4 - len) & 3;
361 nwords = (len + extra_bytes) >> 2;
363 /* Construct the header. */
364 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
365 if (ah_attr->dlid == 0) {
369 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
370 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
371 dev->n_multicast_xmit++;
373 dev->n_unicast_xmit++;
375 dev->n_unicast_xmit++;
376 lid = ah_attr->dlid &
377 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
378 if (unlikely(lid == dev->dd->ipath_lid)) {
380 * Pass in an uninitialized ib_wc to save stack
383 ipath_ud_loopback(qp, &ss, len, wr, &wc);
387 if (ah_attr->ah_flags & IB_AH_GRH) {
388 /* Header size in 32-bit words. */
390 lrh0 = IPATH_LRH_GRH;
391 ohdr = &qp->s_hdr.u.l.oth;
392 qp->s_hdr.u.l.grh.version_tclass_flow =
393 cpu_to_be32((6 << 28) |
394 (ah_attr->grh.traffic_class << 20) |
395 ah_attr->grh.flow_label);
396 qp->s_hdr.u.l.grh.paylen =
397 cpu_to_be16(((wr->opcode ==
398 IB_WR_SEND_WITH_IMM ? 6 : 5) +
399 nwords + SIZE_OF_CRC) << 2);
400 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
401 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
402 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
403 /* The SGID is 32-bit aligned. */
404 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
406 qp->s_hdr.u.l.grh.sgid.global.interface_id =
408 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
410 * Don't worry about sending to locally attached multicast
411 * QPs. It is unspecified by the spec. what happens.
414 /* Header size in 32-bit words. */
416 lrh0 = IPATH_LRH_BTH;
417 ohdr = &qp->s_hdr.u.oth;
419 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
420 ohdr->u.ud.imm_data = wr->imm_data;
421 wc.imm_data = wr->imm_data;
423 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
424 } else if (wr->opcode == IB_WR_SEND) {
426 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
431 lrh0 |= ah_attr->sl << 4;
432 if (qp->ibqp.qp_type == IB_QPT_SMI)
433 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
434 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
435 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
436 qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
437 lid = dev->dd->ipath_lid;
439 lid |= ah_attr->src_path_bits &
440 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
441 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
443 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
444 if (wr->send_flags & IB_SEND_SOLICITED)
446 bth0 |= extra_bytes << 20;
447 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
448 ipath_get_pkey(dev->dd, qp->s_pkey_index);
449 ohdr->bth[0] = cpu_to_be32(bth0);
451 * Use the multicast QP if the destination LID is a multicast LID.
453 ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
454 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
455 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
456 cpu_to_be32(wr->wr.ud.remote_qpn);
457 /* XXX Could lose a PSN count but not worth locking */
458 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
460 * Qkeys with the high order bit set mean use the
461 * qkey from the QP context instead of the WR (see 10.2.5).
463 ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
464 qp->qkey : wr->wr.ud.remote_qkey);
465 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
466 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
471 /* Queue the completion status entry. */
472 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
473 (wr->send_flags & IB_SEND_SIGNALED)) {
474 wc.wr_id = wr->wr_id;
475 wc.status = IB_WC_SUCCESS;
477 wc.opcode = IB_WC_SEND;
479 wc.qp_num = qp->ibqp.qp_num;
482 /* XXX initialize other fields? */
483 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
494 * ipath_ud_rcv - receive an incoming UD packet
495 * @dev: the device the packet came in on
496 * @hdr: the packet header
497 * @has_grh: true if the packet has a GRH
498 * @data: the packet data
499 * @tlen: the packet length
500 * @qp: the QP the packet came on
502 * This is called from ipath_qp_rcv() to process an incoming UD packet
504 * Called at interrupt level.
506 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
507 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
509 struct ipath_other_headers *ohdr;
522 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
523 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
524 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
527 ohdr = &hdr->u.l.oth;
528 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
530 * The header with GRH is 68 bytes and the core driver sets
531 * the eager header buffer size to 56 bytes so the last 12
532 * bytes of the IB header is in the data buffer.
534 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
535 if (header_in_data) {
536 qkey = be32_to_cpu(((__be32 *) data)[1]);
537 src_qp = be32_to_cpu(((__be32 *) data)[2]);
540 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
541 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
544 src_qp &= IPATH_QPN_MASK;
547 * Check that the permissive LID is only used on QP0
548 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
550 if (qp->ibqp.qp_num) {
551 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
552 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
556 if (unlikely(qkey != qp->qkey)) {
557 /* XXX OK to lose a count once in a while. */
558 dev->qkey_violations++;
562 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
563 hdr->lrh[3] == IB_LID_PERMISSIVE) {
564 struct ib_smp *smp = (struct ib_smp *) data;
566 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
572 /* Get the number of bytes the message was padded by. */
573 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
574 if (unlikely(tlen < (hdrsize + pad + 4))) {
575 /* Drop incomplete packets. */
579 tlen -= hdrsize + pad + 4;
581 /* Drop invalid MAD packets (see 13.5.3.1). */
582 if (unlikely((qp->ibqp.qp_num == 0 &&
584 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
585 (qp->ibqp.qp_num == 1 &&
587 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
593 * A GRH is expected to preceed the data even if not
594 * present on the wire.
596 wc.byte_len = tlen + sizeof(struct ib_grh);
599 * The opcode is in the low byte when its in network order
600 * (top byte when in host order).
602 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
603 if (qp->ibqp.qp_num > 1 &&
604 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
605 if (header_in_data) {
606 wc.imm_data = *(__be32 *) data;
607 data += sizeof(__be32);
609 wc.imm_data = ohdr->u.ud.imm_data;
610 wc.wc_flags = IB_WC_WITH_IMM;
611 hdrsize += sizeof(u32);
612 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
621 * Get the next work request entry to find where to put the data.
625 else if (!ipath_get_rwqe(qp, 0)) {
627 * Count VL15 packets dropped due to no receive buffer.
628 * Otherwise, count them as buffer overruns since usually,
629 * the HW will be able to receive packets even if there are
630 * no QPs with posted receive buffers.
632 if (qp->ibqp.qp_num == 0)
633 dev->n_vl15_dropped++;
638 /* Silently drop packets which are too big. */
639 if (wc.byte_len > qp->r_len) {
645 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
646 sizeof(struct ib_grh));
647 wc.wc_flags |= IB_WC_GRH;
649 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
650 ipath_copy_sge(&qp->r_sge, data,
651 wc.byte_len - sizeof(struct ib_grh));
652 wc.wr_id = qp->r_wr_id;
653 wc.status = IB_WC_SUCCESS;
654 wc.opcode = IB_WC_RECV;
656 wc.qp_num = qp->ibqp.qp_num;
658 /* XXX do we know which pkey matched? Only needed for GSI. */
660 wc.slid = be16_to_cpu(hdr->lrh[3]);
661 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
662 dlid = be16_to_cpu(hdr->lrh[1]);
664 * Save the LMC lower bits if the destination LID is a unicast LID.
666 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
667 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
668 /* Signal completion event if the solicited bit is set. */
669 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
671 __constant_cpu_to_be32(1 << 23)) != 0);