2 * Copyright (c) 2006, 2007, 2008 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"
40 * ipath_ud_loopback - handle send on loopback QPs
41 * @sqp: the sending QP
42 * @swqe: the send work request
44 * This is called from ipath_make_ud_req() to forward a WQE addressed
46 * Note that the receive interrupt handler may be calling ipath_ud_rcv()
47 * while this is being called.
49 static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
51 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
53 struct ib_ah_attr *ah_attr;
56 struct ipath_srq *srq;
57 struct ipath_sge_state rsge;
58 struct ipath_sge *sge;
60 struct ipath_rwqe *wqe;
61 void (*handler)(struct ib_event *, void *);
67 qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
68 if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
74 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
75 * Qkeys with the high order bit set mean use the
76 * qkey from the QP context instead of the WR (see 10.2.5).
78 if (unlikely(qp->ibqp.qp_num &&
79 ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
80 sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
81 /* XXX OK to lose a count once in a while. */
82 dev->qkey_violations++;
88 * A GRH is expected to preceed the data even if not
89 * present on the wire.
91 length = swqe->length;
92 memset(&wc, 0, sizeof wc);
93 wc.byte_len = length + sizeof(struct ib_grh);
95 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
96 wc.wc_flags = IB_WC_WITH_IMM;
97 wc.ex.imm_data = swqe->wr.ex.imm_data;
101 * This would be a lot simpler if we could call ipath_get_rwqe()
102 * but that uses state that the receive interrupt handler uses
103 * so we would need to lock out receive interrupts while doing
107 srq = to_isrq(qp->ibqp.srq);
108 handler = srq->ibsrq.event_handler;
117 * Get the next work request entry to find where to put the data.
118 * Note that it is safe to drop the lock after changing rq->tail
119 * since ipath_post_receive() won't fill the empty slot.
121 spin_lock_irqsave(&rq->lock, flags);
124 /* Validate tail before using it since it is user writable. */
125 if (tail >= rq->size)
127 if (unlikely(tail == wq->head)) {
128 spin_unlock_irqrestore(&rq->lock, flags);
132 wqe = get_rwqe_ptr(rq, tail);
133 rsge.sg_list = qp->r_ud_sg_list;
134 if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
135 spin_unlock_irqrestore(&rq->lock, flags);
139 /* Silently drop packets which are too big. */
140 if (wc.byte_len > rlen) {
141 spin_unlock_irqrestore(&rq->lock, flags);
145 if (++tail >= rq->size)
148 wc.wr_id = wqe->wr_id;
153 * validate head pointer value and compute
154 * the number of remaining WQEs.
160 n += rq->size - tail;
163 if (n < srq->limit) {
167 spin_unlock_irqrestore(&rq->lock, flags);
168 ev.device = qp->ibqp.device;
169 ev.element.srq = qp->ibqp.srq;
170 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
171 handler(&ev, srq->ibsrq.srq_context);
173 spin_unlock_irqrestore(&rq->lock, flags);
175 spin_unlock_irqrestore(&rq->lock, flags);
177 ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
178 if (ah_attr->ah_flags & IB_AH_GRH) {
179 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
180 wc.wc_flags |= IB_WC_GRH;
182 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
185 u32 len = sge->length;
189 if (len > sge->sge_length)
190 len = sge->sge_length;
192 ipath_copy_sge(&rsge, sge->vaddr, len);
195 sge->sge_length -= len;
196 if (sge->sge_length == 0) {
197 if (--swqe->wr.num_sge)
199 } else if (sge->length == 0 && sge->mr != NULL) {
200 if (++sge->n >= IPATH_SEGSZ) {
201 if (++sge->m >= sge->mr->mapsz)
206 sge->mr->map[sge->m]->segs[sge->n].vaddr;
208 sge->mr->map[sge->m]->segs[sge->n].length;
212 wc.status = IB_WC_SUCCESS;
213 wc.opcode = IB_WC_RECV;
215 wc.src_qp = sqp->ibqp.qp_num;
216 /* XXX do we know which pkey matched? Only needed for GSI. */
218 wc.slid = dev->dd->ipath_lid |
219 (ah_attr->src_path_bits &
220 ((1 << dev->dd->ipath_lmc) - 1));
223 ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
225 /* Signal completion event if the solicited bit is set. */
226 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
227 swqe->wr.send_flags & IB_SEND_SOLICITED);
229 if (atomic_dec_and_test(&qp->refcount))
235 * ipath_make_ud_req - construct a UD request packet
238 * Return 1 if constructed; otherwise, return 0.
240 int ipath_make_ud_req(struct ipath_qp *qp)
242 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
243 struct ipath_other_headers *ohdr;
244 struct ib_ah_attr *ah_attr;
245 struct ipath_swqe *wqe;
255 spin_lock_irqsave(&qp->s_lock, flags);
257 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
258 if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
260 /* We are in the error state, flush the work request. */
261 if (qp->s_last == qp->s_head)
263 /* If DMAs are in progress, we can't flush immediately. */
264 if (atomic_read(&qp->s_dma_busy)) {
265 qp->s_flags |= IPATH_S_WAIT_DMA;
268 wqe = get_swqe_ptr(qp, qp->s_last);
269 ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
273 if (qp->s_cur == qp->s_head)
276 wqe = get_swqe_ptr(qp, qp->s_cur);
277 next_cur = qp->s_cur + 1;
278 if (next_cur >= qp->s_size)
281 /* Construct the header. */
282 ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
283 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
284 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
285 dev->n_multicast_xmit++;
287 dev->n_unicast_xmit++;
289 dev->n_unicast_xmit++;
290 lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
291 if (unlikely(lid == dev->dd->ipath_lid)) {
293 * If DMAs are in progress, we can't generate
294 * a completion for the loopback packet since
295 * it would be out of order.
296 * XXX Instead of waiting, we could queue a
297 * zero length descriptor so we get a callback.
299 if (atomic_read(&qp->s_dma_busy)) {
300 qp->s_flags |= IPATH_S_WAIT_DMA;
303 qp->s_cur = next_cur;
304 spin_unlock_irqrestore(&qp->s_lock, flags);
305 ipath_ud_loopback(qp, wqe);
306 spin_lock_irqsave(&qp->s_lock, flags);
307 ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
312 qp->s_cur = next_cur;
313 extra_bytes = -wqe->length & 3;
314 nwords = (wqe->length + extra_bytes) >> 2;
316 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
318 qp->s_cur_size = wqe->length;
319 qp->s_cur_sge = &qp->s_sge;
320 qp->s_dmult = ah_attr->static_rate;
322 qp->s_sge.sge = wqe->sg_list[0];
323 qp->s_sge.sg_list = wqe->sg_list + 1;
324 qp->s_sge.num_sge = wqe->wr.num_sge;
326 if (ah_attr->ah_flags & IB_AH_GRH) {
327 /* Header size in 32-bit words. */
328 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
330 qp->s_hdrwords, nwords);
331 lrh0 = IPATH_LRH_GRH;
332 ohdr = &qp->s_hdr.u.l.oth;
334 * Don't worry about sending to locally attached multicast
335 * QPs. It is unspecified by the spec. what happens.
338 /* Header size in 32-bit words. */
339 lrh0 = IPATH_LRH_BTH;
340 ohdr = &qp->s_hdr.u.oth;
342 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
344 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
345 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
347 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(qp->s_hdrwords + nwords +
355 lid = dev->dd->ipath_lid;
357 lid |= ah_attr->src_path_bits &
358 ((1 << dev->dd->ipath_lmc) - 1);
359 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
361 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
362 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
364 bth0 |= extra_bytes << 20;
365 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
366 ipath_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 >= IPATH_MULTICAST_LID_BASE &&
372 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
373 cpu_to_be32(IPATH_MULTICAST_QPN) :
374 cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
375 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_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)wqe->wr.wr.ud.remote_qkey < 0 ?
381 qp->qkey : wqe->wr.wr.ud.remote_qkey);
382 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
389 qp->s_flags &= ~IPATH_S_BUSY;
391 spin_unlock_irqrestore(&qp->s_lock, flags);
396 * ipath_ud_rcv - receive an incoming UD packet
397 * @dev: the device the packet came in on
398 * @hdr: the packet header
399 * @has_grh: true if the packet has a GRH
400 * @data: the packet data
401 * @tlen: the packet length
402 * @qp: the QP the packet came on
404 * This is called from ipath_qp_rcv() to process an incoming UD packet
406 * Called at interrupt level.
408 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
409 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
411 struct ipath_other_headers *ohdr;
424 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
425 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
426 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
429 ohdr = &hdr->u.l.oth;
430 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
432 * The header with GRH is 68 bytes and the core driver sets
433 * the eager header buffer size to 56 bytes so the last 12
434 * bytes of the IB header is in the data buffer.
436 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
437 if (header_in_data) {
438 qkey = be32_to_cpu(((__be32 *) data)[1]);
439 src_qp = be32_to_cpu(((__be32 *) data)[2]);
442 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
443 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
446 src_qp &= IPATH_QPN_MASK;
449 * Check that the permissive LID is only used on QP0
450 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
452 if (qp->ibqp.qp_num) {
453 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
454 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
458 if (unlikely(qkey != qp->qkey)) {
459 /* XXX OK to lose a count once in a while. */
460 dev->qkey_violations++;
464 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
465 hdr->lrh[3] == IB_LID_PERMISSIVE) {
466 struct ib_smp *smp = (struct ib_smp *) data;
468 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
475 * The opcode is in the low byte when its in network order
476 * (top byte when in host order).
478 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
479 if (qp->ibqp.qp_num > 1 &&
480 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
481 if (header_in_data) {
482 wc.ex.imm_data = *(__be32 *) data;
483 data += sizeof(__be32);
485 wc.ex.imm_data = ohdr->u.ud.imm_data;
486 wc.wc_flags = IB_WC_WITH_IMM;
487 hdrsize += sizeof(u32);
488 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
496 /* Get the number of bytes the message was padded by. */
497 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
498 if (unlikely(tlen < (hdrsize + pad + 4))) {
499 /* Drop incomplete packets. */
503 tlen -= hdrsize + pad + 4;
505 /* Drop invalid MAD packets (see 13.5.3.1). */
506 if (unlikely((qp->ibqp.qp_num == 0 &&
508 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
509 (qp->ibqp.qp_num == 1 &&
511 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
517 * A GRH is expected to preceed the data even if not
518 * present on the wire.
520 wc.byte_len = tlen + sizeof(struct ib_grh);
523 * Get the next work request entry to find where to put the data.
525 if (qp->r_flags & IPATH_R_REUSE_SGE)
526 qp->r_flags &= ~IPATH_R_REUSE_SGE;
527 else if (!ipath_get_rwqe(qp, 0)) {
529 * Count VL15 packets dropped due to no receive buffer.
530 * Otherwise, count them as buffer overruns since usually,
531 * the HW will be able to receive packets even if there are
532 * no QPs with posted receive buffers.
534 if (qp->ibqp.qp_num == 0)
535 dev->n_vl15_dropped++;
540 /* Silently drop packets which are too big. */
541 if (wc.byte_len > qp->r_len) {
542 qp->r_flags |= IPATH_R_REUSE_SGE;
547 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
548 sizeof(struct ib_grh));
549 wc.wc_flags |= IB_WC_GRH;
551 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
552 ipath_copy_sge(&qp->r_sge, data,
553 wc.byte_len - sizeof(struct ib_grh));
554 if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
556 wc.wr_id = qp->r_wr_id;
557 wc.status = IB_WC_SUCCESS;
558 wc.opcode = IB_WC_RECV;
562 /* XXX do we know which pkey matched? Only needed for GSI. */
564 wc.slid = be16_to_cpu(hdr->lrh[3]);
565 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
566 dlid = be16_to_cpu(hdr->lrh[1]);
568 * Save the LMC lower bits if the destination LID is a unicast LID.
570 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
571 dlid & ((1 << dev->dd->ipath_lmc) - 1);
573 /* Signal completion event if the solicited bit is set. */
574 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
576 cpu_to_be32(1 << 23)) != 0);