2 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <rdma/ib_smi.h>
35 #include "ipath_verbs.h"
36 #include "ips_common.h"
39 * ipath_ud_loopback - handle send on loopback QPs
42 * @length: the length of the data to send
43 * @wr: the work request
44 * @wc: the work completion entry
46 * This is called from ipath_post_ud_send() to forward a WQE addressed
49 static void ipath_ud_loopback(struct ipath_qp *sqp,
50 struct ipath_sge_state *ss,
51 u32 length, struct ib_send_wr *wr,
54 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
56 struct ib_ah_attr *ah_attr;
59 struct ipath_srq *srq;
60 struct ipath_sge_state rsge;
61 struct ipath_sge *sge;
62 struct ipath_rwqe *wqe;
64 qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn);
69 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
70 * Qkeys with the high order bit set mean use the
71 * qkey from the QP context instead of the WR (see 10.2.5).
73 if (unlikely(qp->ibqp.qp_num &&
74 ((int) wr->wr.ud.remote_qkey < 0
75 ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) {
76 /* XXX OK to lose a count once in a while. */
77 dev->qkey_violations++;
83 * A GRH is expected to preceed the data even if not
84 * present on the wire.
86 wc->byte_len = length + sizeof(struct ib_grh);
88 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
89 wc->wc_flags = IB_WC_WITH_IMM;
90 wc->imm_data = wr->imm_data;
97 * Get the next work request entry to find where to put the data.
98 * Note that it is safe to drop the lock after changing rq->tail
99 * since ipath_post_receive() won't fill the empty slot.
102 srq = to_isrq(qp->ibqp.srq);
108 spin_lock_irqsave(&rq->lock, flags);
109 if (rq->tail == rq->head) {
110 spin_unlock_irqrestore(&rq->lock, flags);
114 /* Silently drop packets which are too big. */
115 wqe = get_rwqe_ptr(rq, rq->tail);
116 if (wc->byte_len > wqe->length) {
117 spin_unlock_irqrestore(&rq->lock, flags);
121 wc->wr_id = wqe->wr_id;
122 rsge.sge = wqe->sg_list[0];
123 rsge.sg_list = wqe->sg_list + 1;
124 rsge.num_sge = wqe->num_sge;
125 if (++rq->tail >= rq->size)
127 if (srq && srq->ibsrq.event_handler) {
130 if (rq->head < rq->tail)
131 n = rq->size + rq->head - rq->tail;
133 n = rq->head - rq->tail;
134 if (n < srq->limit) {
138 spin_unlock_irqrestore(&rq->lock, flags);
139 ev.device = qp->ibqp.device;
140 ev.element.srq = qp->ibqp.srq;
141 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
142 srq->ibsrq.event_handler(&ev,
143 srq->ibsrq.srq_context);
145 spin_unlock_irqrestore(&rq->lock, flags);
147 spin_unlock_irqrestore(&rq->lock, flags);
148 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
149 if (ah_attr->ah_flags & IB_AH_GRH) {
150 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
151 wc->wc_flags |= IB_WC_GRH;
153 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
156 u32 len = sge->length;
161 ipath_copy_sge(&rsge, sge->vaddr, len);
164 sge->sge_length -= len;
165 if (sge->sge_length == 0) {
167 *sge = *ss->sg_list++;
168 } else if (sge->length == 0 && sge->mr != NULL) {
169 if (++sge->n >= IPATH_SEGSZ) {
170 if (++sge->m >= sge->mr->mapsz)
175 sge->mr->map[sge->m]->segs[sge->n].vaddr;
177 sge->mr->map[sge->m]->segs[sge->n].length;
181 wc->status = IB_WC_SUCCESS;
182 wc->opcode = IB_WC_RECV;
184 wc->qp_num = qp->ibqp.qp_num;
185 wc->src_qp = sqp->ibqp.qp_num;
186 /* XXX do we know which pkey matched? Only needed for GSI. */
188 wc->slid = ipath_layer_get_lid(dev->dd) |
189 (ah_attr->src_path_bits &
190 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
191 wc->sl = ah_attr->sl;
193 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
194 /* Signal completion event if the solicited bit is set. */
195 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
196 wr->send_flags & IB_SEND_SOLICITED);
199 if (atomic_dec_and_test(&qp->refcount))
204 * ipath_post_ud_send - post a UD send on QP
206 * @wr: the work request
208 * Note that we actually send the data as it is posted instead of putting
209 * the request into a ring buffer. If we wanted to use a ring buffer,
210 * we would need to save a reference to the destination address in the SWQE.
212 int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
214 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
215 struct ipath_other_headers *ohdr;
216 struct ib_ah_attr *ah_attr;
217 struct ipath_sge_state ss;
218 struct ipath_sge *sg_list;
230 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
235 /* IB spec says that num_sge == 0 is OK. */
236 if (wr->num_sge > qp->s_max_sge) {
241 if (wr->num_sge > 1) {
242 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
251 /* Check the buffer to send. */
252 ss.sg_list = sg_list;
256 ss.sge.sge_length = 0;
259 for (i = 0; i < wr->num_sge; i++) {
261 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
266 if (wr->sg_list[i].length == 0)
268 if (!ipath_lkey_ok(&dev->lk_table, ss.num_sge ?
269 sg_list + ss.num_sge - 1 : &ss.sge,
270 &wr->sg_list[i], 0)) {
274 len += wr->sg_list[i].length;
277 extra_bytes = (4 - len) & 3;
278 nwords = (len + extra_bytes) >> 2;
280 /* Construct the header. */
281 ah_attr = &to_iah(wr->wr.ud.ah)->attr;
282 if (ah_attr->dlid == 0) {
286 if (ah_attr->dlid >= IPS_MULTICAST_LID_BASE) {
287 if (ah_attr->dlid != IPS_PERMISSIVE_LID)
288 dev->n_multicast_xmit++;
290 dev->n_unicast_xmit++;
292 dev->n_unicast_xmit++;
293 lid = ah_attr->dlid &
294 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
295 if (unlikely(lid == ipath_layer_get_lid(dev->dd))) {
297 * Pass in an uninitialized ib_wc to save stack
300 ipath_ud_loopback(qp, &ss, len, wr, &wc);
304 if (ah_attr->ah_flags & IB_AH_GRH) {
305 /* Header size in 32-bit words. */
308 ohdr = &qp->s_hdr.u.l.oth;
309 qp->s_hdr.u.l.grh.version_tclass_flow =
310 cpu_to_be32((6 << 28) |
311 (ah_attr->grh.traffic_class << 20) |
312 ah_attr->grh.flow_label);
313 qp->s_hdr.u.l.grh.paylen =
314 cpu_to_be16(((wr->opcode ==
315 IB_WR_SEND_WITH_IMM ? 6 : 5) +
316 nwords + SIZE_OF_CRC) << 2);
317 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
318 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
319 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
320 /* The SGID is 32-bit aligned. */
321 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
323 qp->s_hdr.u.l.grh.sgid.global.interface_id =
324 ipath_layer_get_guid(dev->dd);
325 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
327 * Don't worry about sending to locally attached multicast
328 * QPs. It is unspecified by the spec. what happens.
331 /* Header size in 32-bit words. */
334 ohdr = &qp->s_hdr.u.oth;
336 if (wr->opcode == IB_WR_SEND_WITH_IMM) {
337 ohdr->u.ud.imm_data = wr->imm_data;
338 wc.imm_data = wr->imm_data;
340 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
341 } else if (wr->opcode == IB_WR_SEND) {
343 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
348 lrh0 |= ah_attr->sl << 4;
349 if (qp->ibqp.qp_type == IB_QPT_SMI)
350 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
351 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
352 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
353 qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
354 lid = ipath_layer_get_lid(dev->dd);
356 lid |= ah_attr->src_path_bits &
357 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
358 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
360 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
361 if (wr->send_flags & IB_SEND_SOLICITED)
363 bth0 |= extra_bytes << 20;
364 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPS_DEFAULT_P_KEY :
365 ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
366 ohdr->bth[0] = cpu_to_be32(bth0);
368 * Use the multicast QP if the destination LID is a multicast LID.
370 ohdr->bth[1] = ah_attr->dlid >= IPS_MULTICAST_LID_BASE &&
371 ah_attr->dlid != IPS_PERMISSIVE_LID ?
372 __constant_cpu_to_be32(IPS_MULTICAST_QPN) :
373 cpu_to_be32(wr->wr.ud.remote_qpn);
374 /* XXX Could lose a PSN count but not worth locking */
375 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPS_PSN_MASK);
377 * Qkeys with the high order bit set mean use the
378 * qkey from the QP context instead of the WR (see 10.2.5).
380 ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ?
381 qp->qkey : wr->wr.ud.remote_qkey);
382 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
383 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
388 /* Queue the completion status entry. */
389 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
390 (wr->send_flags & IB_SEND_SIGNALED)) {
391 wc.wr_id = wr->wr_id;
392 wc.status = IB_WC_SUCCESS;
394 wc.opcode = IB_WC_SEND;
396 wc.qp_num = qp->ibqp.qp_num;
399 /* XXX initialize other fields? */
400 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
411 * ipath_ud_rcv - receive an incoming UD packet
412 * @dev: the device the packet came in on
413 * @hdr: the packet header
414 * @has_grh: true if the packet has a GRH
415 * @data: the packet data
416 * @tlen: the packet length
417 * @qp: the QP the packet came on
419 * This is called from ipath_qp_rcv() to process an incoming UD packet
421 * Called at interrupt level.
423 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
424 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
426 struct ipath_other_headers *ohdr;
435 struct ipath_srq *srq;
436 struct ipath_rwqe *wqe;
443 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
444 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
445 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
448 ohdr = &hdr->u.l.oth;
449 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
451 * The header with GRH is 68 bytes and the core driver sets
452 * the eager header buffer size to 56 bytes so the last 12
453 * bytes of the IB header is in the data buffer.
456 ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
457 if (header_in_data) {
458 qkey = be32_to_cpu(((__be32 *) data)[1]);
459 src_qp = be32_to_cpu(((__be32 *) data)[2]);
462 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
463 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
466 src_qp &= IPS_QPN_MASK;
469 * Check that the permissive LID is only used on QP0
470 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
472 if (qp->ibqp.qp_num) {
473 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
474 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
478 if (unlikely(qkey != qp->qkey)) {
479 /* XXX OK to lose a count once in a while. */
480 dev->qkey_violations++;
484 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
485 hdr->lrh[3] == IB_LID_PERMISSIVE) {
486 struct ib_smp *smp = (struct ib_smp *) data;
488 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
494 /* Get the number of bytes the message was padded by. */
495 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
496 if (unlikely(tlen < (hdrsize + pad + 4))) {
497 /* Drop incomplete packets. */
501 tlen -= hdrsize + pad + 4;
503 /* Drop invalid MAD packets (see 13.5.3.1). */
504 if (unlikely((qp->ibqp.qp_num == 0 &&
506 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
507 (qp->ibqp.qp_num == 1 &&
509 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
515 * A GRH is expected to preceed the data even if not
516 * present on the wire.
518 wc.byte_len = tlen + sizeof(struct ib_grh);
521 * The opcode is in the low byte when its in network order
522 * (top byte when in host order).
524 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
525 if (qp->ibqp.qp_num > 1 &&
526 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
527 if (header_in_data) {
528 wc.imm_data = *(__be32 *) data;
529 data += sizeof(__be32);
531 wc.imm_data = ohdr->u.ud.imm_data;
532 wc.wc_flags = IB_WC_WITH_IMM;
533 hdrsize += sizeof(u32);
534 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
543 * Get the next work request entry to find where to put the data.
544 * Note that it is safe to drop the lock after changing rq->tail
545 * since ipath_post_receive() won't fill the empty slot.
548 srq = to_isrq(qp->ibqp.srq);
554 spin_lock_irqsave(&rq->lock, flags);
555 if (rq->tail == rq->head) {
556 spin_unlock_irqrestore(&rq->lock, flags);
560 /* Silently drop packets which are too big. */
561 wqe = get_rwqe_ptr(rq, rq->tail);
562 if (wc.byte_len > wqe->length) {
563 spin_unlock_irqrestore(&rq->lock, flags);
567 wc.wr_id = wqe->wr_id;
568 qp->r_sge.sge = wqe->sg_list[0];
569 qp->r_sge.sg_list = wqe->sg_list + 1;
570 qp->r_sge.num_sge = wqe->num_sge;
571 if (++rq->tail >= rq->size)
573 if (srq && srq->ibsrq.event_handler) {
576 if (rq->head < rq->tail)
577 n = rq->size + rq->head - rq->tail;
579 n = rq->head - rq->tail;
580 if (n < srq->limit) {
584 spin_unlock_irqrestore(&rq->lock, flags);
585 ev.device = qp->ibqp.device;
586 ev.element.srq = qp->ibqp.srq;
587 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
588 srq->ibsrq.event_handler(&ev,
589 srq->ibsrq.srq_context);
591 spin_unlock_irqrestore(&rq->lock, flags);
593 spin_unlock_irqrestore(&rq->lock, flags);
595 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
596 sizeof(struct ib_grh));
597 wc.wc_flags |= IB_WC_GRH;
599 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
600 ipath_copy_sge(&qp->r_sge, data,
601 wc.byte_len - sizeof(struct ib_grh));
602 wc.status = IB_WC_SUCCESS;
603 wc.opcode = IB_WC_RECV;
605 wc.qp_num = qp->ibqp.qp_num;
607 /* XXX do we know which pkey matched? Only needed for GSI. */
609 wc.slid = be16_to_cpu(hdr->lrh[3]);
610 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
611 dlid = be16_to_cpu(hdr->lrh[1]);
613 * Save the LMC lower bits if the destination LID is a unicast LID.
615 wc.dlid_path_bits = dlid >= IPS_MULTICAST_LID_BASE ? 0 :
616 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
617 /* Signal completion event if the solicited bit is set. */
618 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
620 __constant_cpu_to_be32(1 << 23)) != 0);