2 * Copyright (c) 2006, 2007, 2008, 2009 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>
35 #include <rdma/ib_verbs.h>
41 * qib_ud_loopback - handle send on loopback QPs
42 * @sqp: the sending QP
43 * @swqe: the send work request
45 * This is called from qib_make_ud_req() to forward a WQE addressed
47 * Note that the receive interrupt handler may be calling qib_ud_rcv()
48 * while this is being called.
50 static void qib_ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
52 struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
53 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
54 struct qib_devdata *dd = ppd->dd;
55 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
57 struct ib_ah_attr *ah_attr;
59 struct rvt_sge_state ssge;
63 enum ib_qp_type sqptype, dqptype;
66 qp = rvt_lookup_qpn(rdi, &ibp->rvp, swqe->ud_wr.remote_qpn);
68 ibp->rvp.n_pkt_drops++;
73 sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
74 IB_QPT_UD : sqp->ibqp.qp_type;
75 dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
76 IB_QPT_UD : qp->ibqp.qp_type;
78 if (dqptype != sqptype ||
79 !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
80 ibp->rvp.n_pkt_drops++;
84 ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr;
85 ppd = ppd_from_ibp(ibp);
87 if (qp->ibqp.qp_num > 1) {
92 pkey1 = qib_get_pkey(ibp, sqp->s_pkey_index);
93 pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
94 if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
95 lid = ppd->lid | (ah_attr->src_path_bits &
96 ((1 << ppd->lmc) - 1));
97 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, pkey1,
99 sqp->ibqp.qp_num, qp->ibqp.qp_num,
101 cpu_to_be16(ah_attr->dlid));
107 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
108 * Qkeys with the high order bit set mean use the
109 * qkey from the QP context instead of the WR (see 10.2.5).
111 if (qp->ibqp.qp_num) {
114 qkey = (int)swqe->ud_wr.remote_qkey < 0 ?
115 sqp->qkey : swqe->ud_wr.remote_qkey;
116 if (unlikely(qkey != qp->qkey)) {
119 lid = ppd->lid | (ah_attr->src_path_bits &
120 ((1 << ppd->lmc) - 1));
121 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
123 sqp->ibqp.qp_num, qp->ibqp.qp_num,
125 cpu_to_be16(ah_attr->dlid));
131 * A GRH is expected to precede the data even if not
132 * present on the wire.
134 length = swqe->length;
135 memset(&wc, 0, sizeof(wc));
136 wc.byte_len = length + sizeof(struct ib_grh);
138 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
139 wc.wc_flags = IB_WC_WITH_IMM;
140 wc.ex.imm_data = swqe->wr.ex.imm_data;
143 spin_lock_irqsave(&qp->r_lock, flags);
146 * Get the next work request entry to find where to put the data.
148 if (qp->r_flags & RVT_R_REUSE_SGE)
149 qp->r_flags &= ~RVT_R_REUSE_SGE;
153 ret = qib_get_rwqe(qp, 0);
155 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
159 if (qp->ibqp.qp_num == 0)
160 ibp->rvp.n_vl15_dropped++;
164 /* Silently drop packets which are too big. */
165 if (unlikely(wc.byte_len > qp->r_len)) {
166 qp->r_flags |= RVT_R_REUSE_SGE;
167 ibp->rvp.n_pkt_drops++;
171 if (ah_attr->ah_flags & IB_AH_GRH) {
172 qib_copy_sge(&qp->r_sge, &ah_attr->grh,
173 sizeof(struct ib_grh), 1);
174 wc.wc_flags |= IB_WC_GRH;
176 qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
177 ssge.sg_list = swqe->sg_list + 1;
178 ssge.sge = *swqe->sg_list;
179 ssge.num_sge = swqe->wr.num_sge;
182 u32 len = sge->length;
186 if (len > sge->sge_length)
187 len = sge->sge_length;
189 qib_copy_sge(&qp->r_sge, sge->vaddr, len, 1);
192 sge->sge_length -= len;
193 if (sge->sge_length == 0) {
195 *sge = *ssge.sg_list++;
196 } else if (sge->length == 0 && sge->mr->lkey) {
197 if (++sge->n >= RVT_SEGSZ) {
198 if (++sge->m >= sge->mr->mapsz)
203 sge->mr->map[sge->m]->segs[sge->n].vaddr;
205 sge->mr->map[sge->m]->segs[sge->n].length;
209 rvt_put_ss(&qp->r_sge);
210 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
212 wc.wr_id = qp->r_wr_id;
213 wc.status = IB_WC_SUCCESS;
214 wc.opcode = IB_WC_RECV;
216 wc.src_qp = sqp->ibqp.qp_num;
217 wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
218 swqe->ud_wr.pkey_index : 0;
219 wc.slid = ppd->lid | (ah_attr->src_path_bits & ((1 << ppd->lmc) - 1));
221 wc.dlid_path_bits = ah_attr->dlid & ((1 << ppd->lmc) - 1);
222 wc.port_num = qp->port_num;
223 /* Signal completion event if the solicited bit is set. */
224 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
225 swqe->wr.send_flags & IB_SEND_SOLICITED);
226 ibp->rvp.n_loop_pkts++;
228 spin_unlock_irqrestore(&qp->r_lock, flags);
234 * qib_make_ud_req - construct a UD request packet
237 * Assumes the s_lock is held.
239 * Return 1 if constructed; otherwise, return 0.
241 int qib_make_ud_req(struct rvt_qp *qp, unsigned long *flags)
243 struct qib_qp_priv *priv = qp->priv;
244 struct qib_other_headers *ohdr;
245 struct ib_ah_attr *ah_attr;
246 struct qib_pportdata *ppd;
247 struct qib_ibport *ibp;
248 struct rvt_swqe *wqe;
257 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
258 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
260 /* We are in the error state, flush the work request. */
261 smp_read_barrier_depends(); /* see post_one_send */
262 if (qp->s_last == ACCESS_ONCE(qp->s_head))
264 /* If DMAs are in progress, we can't flush immediately. */
265 if (atomic_read(&priv->s_dma_busy)) {
266 qp->s_flags |= RVT_S_WAIT_DMA;
269 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
270 qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
274 /* see post_one_send() */
275 smp_read_barrier_depends();
276 if (qp->s_cur == ACCESS_ONCE(qp->s_head))
279 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
280 next_cur = qp->s_cur + 1;
281 if (next_cur >= qp->s_size)
284 /* Construct the header. */
285 ibp = to_iport(qp->ibqp.device, qp->port_num);
286 ppd = ppd_from_ibp(ibp);
287 ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
288 if (ah_attr->dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) {
289 if (ah_attr->dlid != be16_to_cpu(IB_LID_PERMISSIVE))
290 this_cpu_inc(ibp->pmastats->n_multicast_xmit);
292 this_cpu_inc(ibp->pmastats->n_unicast_xmit);
294 this_cpu_inc(ibp->pmastats->n_unicast_xmit);
295 lid = ah_attr->dlid & ~((1 << ppd->lmc) - 1);
296 if (unlikely(lid == ppd->lid)) {
297 unsigned long tflags = *flags;
299 * If DMAs are in progress, we can't generate
300 * a completion for the loopback packet since
301 * it would be out of order.
302 * XXX Instead of waiting, we could queue a
303 * zero length descriptor so we get a callback.
305 if (atomic_read(&priv->s_dma_busy)) {
306 qp->s_flags |= RVT_S_WAIT_DMA;
309 qp->s_cur = next_cur;
310 spin_unlock_irqrestore(&qp->s_lock, tflags);
311 qib_ud_loopback(qp, wqe);
312 spin_lock_irqsave(&qp->s_lock, tflags);
314 qib_send_complete(qp, wqe, IB_WC_SUCCESS);
319 qp->s_cur = next_cur;
320 extra_bytes = -wqe->length & 3;
321 nwords = (wqe->length + extra_bytes) >> 2;
323 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
325 qp->s_cur_size = wqe->length;
326 qp->s_cur_sge = &qp->s_sge;
327 qp->s_srate = ah_attr->static_rate;
329 qp->s_sge.sge = wqe->sg_list[0];
330 qp->s_sge.sg_list = wqe->sg_list + 1;
331 qp->s_sge.num_sge = wqe->wr.num_sge;
332 qp->s_sge.total_len = wqe->length;
334 if (ah_attr->ah_flags & IB_AH_GRH) {
335 /* Header size in 32-bit words. */
336 qp->s_hdrwords += qib_make_grh(ibp, &priv->s_hdr->u.l.grh,
338 qp->s_hdrwords, nwords);
340 ohdr = &priv->s_hdr->u.l.oth;
342 * Don't worry about sending to locally attached multicast
343 * QPs. It is unspecified by the spec. what happens.
346 /* Header size in 32-bit words. */
348 ohdr = &priv->s_hdr->u.oth;
350 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
352 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
353 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
355 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
356 lrh0 |= ah_attr->sl << 4;
357 if (qp->ibqp.qp_type == IB_QPT_SMI)
358 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
360 lrh0 |= ibp->sl_to_vl[ah_attr->sl] << 12;
361 priv->s_hdr->lrh[0] = cpu_to_be16(lrh0);
362 priv->s_hdr->lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
363 priv->s_hdr->lrh[2] =
364 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
367 lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
368 priv->s_hdr->lrh[3] = cpu_to_be16(lid);
370 priv->s_hdr->lrh[3] = IB_LID_PERMISSIVE;
371 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
372 bth0 |= IB_BTH_SOLICITED;
373 bth0 |= extra_bytes << 20;
374 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? QIB_DEFAULT_P_KEY :
375 qib_get_pkey(ibp, qp->ibqp.qp_type == IB_QPT_GSI ?
376 wqe->ud_wr.pkey_index : qp->s_pkey_index);
377 ohdr->bth[0] = cpu_to_be32(bth0);
379 * Use the multicast QP if the destination LID is a multicast LID.
381 ohdr->bth[1] = ah_attr->dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) &&
382 ah_attr->dlid != be16_to_cpu(IB_LID_PERMISSIVE) ?
383 cpu_to_be32(QIB_MULTICAST_QPN) :
384 cpu_to_be32(wqe->ud_wr.remote_qpn);
385 ohdr->bth[2] = cpu_to_be32(wqe->psn & QIB_PSN_MASK);
387 * Qkeys with the high order bit set mean use the
388 * qkey from the QP context instead of the WR (see 10.2.5).
390 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
391 qp->qkey : wqe->ud_wr.remote_qkey);
392 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
397 qp->s_flags &= ~RVT_S_BUSY;
401 static unsigned qib_lookup_pkey(struct qib_ibport *ibp, u16 pkey)
403 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
404 struct qib_devdata *dd = ppd->dd;
405 unsigned ctxt = ppd->hw_pidx;
408 pkey &= 0x7fff; /* remove limited/full membership bit */
410 for (i = 0; i < ARRAY_SIZE(dd->rcd[ctxt]->pkeys); ++i)
411 if ((dd->rcd[ctxt]->pkeys[i] & 0x7fff) == pkey)
415 * Should not get here, this means hardware failed to validate pkeys.
416 * Punt and return index 0.
422 * qib_ud_rcv - receive an incoming UD packet
423 * @ibp: the port the packet came in on
424 * @hdr: the packet header
425 * @has_grh: true if the packet has a GRH
426 * @data: the packet data
427 * @tlen: the packet length
428 * @qp: the QP the packet came on
430 * This is called from qib_qp_rcv() to process an incoming UD packet
432 * Called at interrupt level.
434 void qib_ud_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
435 int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
437 struct qib_other_headers *ohdr;
449 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
451 ohdr = &hdr->u.l.oth;
452 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
454 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
455 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK;
458 * Get the number of bytes the message was padded by
459 * and drop incomplete packets.
461 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
462 if (unlikely(tlen < (hdrsize + pad + 4)))
465 tlen -= hdrsize + pad + 4;
468 * Check that the permissive LID is only used on QP0
469 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
471 if (qp->ibqp.qp_num) {
472 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
473 hdr->lrh[3] == IB_LID_PERMISSIVE))
475 if (qp->ibqp.qp_num > 1) {
478 pkey1 = be32_to_cpu(ohdr->bth[0]);
479 pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
480 if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
481 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
483 (be16_to_cpu(hdr->lrh[0]) >> 4) &
485 src_qp, qp->ibqp.qp_num,
486 hdr->lrh[3], hdr->lrh[1]);
490 if (unlikely(qkey != qp->qkey)) {
491 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
492 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
493 src_qp, qp->ibqp.qp_num,
494 hdr->lrh[3], hdr->lrh[1]);
497 /* Drop invalid MAD packets (see 13.5.3.1). */
498 if (unlikely(qp->ibqp.qp_num == 1 &&
500 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))
505 /* Drop invalid MAD packets (see 13.5.3.1). */
506 if (tlen != 256 || (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)
508 smp = (struct ib_smp *) data;
509 if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
510 hdr->lrh[3] == IB_LID_PERMISSIVE) &&
511 smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
516 * The opcode is in the low byte when its in network order
517 * (top byte when in host order).
519 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
520 if (qp->ibqp.qp_num > 1 &&
521 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
522 wc.ex.imm_data = ohdr->u.ud.imm_data;
523 wc.wc_flags = IB_WC_WITH_IMM;
525 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
532 * A GRH is expected to precede the data even if not
533 * present on the wire.
535 wc.byte_len = tlen + sizeof(struct ib_grh);
538 * Get the next work request entry to find where to put the data.
540 if (qp->r_flags & RVT_R_REUSE_SGE)
541 qp->r_flags &= ~RVT_R_REUSE_SGE;
545 ret = qib_get_rwqe(qp, 0);
547 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
551 if (qp->ibqp.qp_num == 0)
552 ibp->rvp.n_vl15_dropped++;
556 /* Silently drop packets which are too big. */
557 if (unlikely(wc.byte_len > qp->r_len)) {
558 qp->r_flags |= RVT_R_REUSE_SGE;
562 qib_copy_sge(&qp->r_sge, &hdr->u.l.grh,
563 sizeof(struct ib_grh), 1);
564 wc.wc_flags |= IB_WC_GRH;
566 qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
567 qib_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 1);
568 rvt_put_ss(&qp->r_sge);
569 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
571 wc.wr_id = qp->r_wr_id;
572 wc.status = IB_WC_SUCCESS;
573 wc.opcode = IB_WC_RECV;
577 wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
578 qib_lookup_pkey(ibp, be32_to_cpu(ohdr->bth[0])) : 0;
579 wc.slid = be16_to_cpu(hdr->lrh[3]);
580 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
581 dlid = be16_to_cpu(hdr->lrh[1]);
583 * Save the LMC lower bits if the destination LID is a unicast LID.
585 wc.dlid_path_bits = dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) ? 0 :
586 dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
587 wc.port_num = qp->port_num;
588 /* Signal completion event if the solicited bit is set. */
589 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
591 cpu_to_be32(IB_BTH_SOLICITED)) != 0);
595 ibp->rvp.n_pkt_drops++;