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)) {
76 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
77 * Qkeys with the high order bit set mean use the
78 * qkey from the QP context instead of the WR (see 10.2.5).
80 if (unlikely(qp->ibqp.qp_num &&
81 ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
82 sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
83 /* XXX OK to lose a count once in a while. */
84 dev->qkey_violations++;
90 * A GRH is expected to preceed the data even if not
91 * present on the wire.
93 length = swqe->length;
94 memset(&wc, 0, sizeof wc);
95 wc.byte_len = length + sizeof(struct ib_grh);
97 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
98 wc.wc_flags = IB_WC_WITH_IMM;
99 wc.ex.imm_data = swqe->wr.ex.imm_data;
103 * This would be a lot simpler if we could call ipath_get_rwqe()
104 * but that uses state that the receive interrupt handler uses
105 * so we would need to lock out receive interrupts while doing
109 srq = to_isrq(qp->ibqp.srq);
110 handler = srq->ibsrq.event_handler;
118 if (rq->max_sge > 1) {
120 * XXX We could use GFP_KERNEL if ipath_do_send()
121 * was always called from the tasklet instead of
122 * from ipath_post_send().
124 rsge.sg_list = kmalloc((rq->max_sge - 1) *
125 sizeof(struct ipath_sge),
134 * Get the next work request entry to find where to put the data.
135 * Note that it is safe to drop the lock after changing rq->tail
136 * since ipath_post_receive() won't fill the empty slot.
138 spin_lock_irqsave(&rq->lock, flags);
141 /* Validate tail before using it since it is user writable. */
142 if (tail >= rq->size)
144 if (unlikely(tail == wq->head)) {
145 spin_unlock_irqrestore(&rq->lock, flags);
149 wqe = get_rwqe_ptr(rq, tail);
150 if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
151 spin_unlock_irqrestore(&rq->lock, flags);
155 /* Silently drop packets which are too big. */
156 if (wc.byte_len > rlen) {
157 spin_unlock_irqrestore(&rq->lock, flags);
161 if (++tail >= rq->size)
164 wc.wr_id = wqe->wr_id;
169 * validate head pointer value and compute
170 * the number of remaining WQEs.
176 n += rq->size - tail;
179 if (n < srq->limit) {
183 spin_unlock_irqrestore(&rq->lock, flags);
184 ev.device = qp->ibqp.device;
185 ev.element.srq = qp->ibqp.srq;
186 ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
187 handler(&ev, srq->ibsrq.srq_context);
189 spin_unlock_irqrestore(&rq->lock, flags);
191 spin_unlock_irqrestore(&rq->lock, flags);
193 ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
194 if (ah_attr->ah_flags & IB_AH_GRH) {
195 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
196 wc.wc_flags |= IB_WC_GRH;
198 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
201 u32 len = sge->length;
205 if (len > sge->sge_length)
206 len = sge->sge_length;
208 ipath_copy_sge(&rsge, sge->vaddr, len);
211 sge->sge_length -= len;
212 if (sge->sge_length == 0) {
213 if (--swqe->wr.num_sge)
215 } else if (sge->length == 0 && sge->mr != NULL) {
216 if (++sge->n >= IPATH_SEGSZ) {
217 if (++sge->m >= sge->mr->mapsz)
222 sge->mr->map[sge->m]->segs[sge->n].vaddr;
224 sge->mr->map[sge->m]->segs[sge->n].length;
228 wc.status = IB_WC_SUCCESS;
229 wc.opcode = IB_WC_RECV;
231 wc.src_qp = sqp->ibqp.qp_num;
232 /* XXX do we know which pkey matched? Only needed for GSI. */
234 wc.slid = dev->dd->ipath_lid |
235 (ah_attr->src_path_bits &
236 ((1 << dev->dd->ipath_lmc) - 1));
239 ah_attr->dlid & ((1 << dev->dd->ipath_lmc) - 1);
241 /* Signal completion event if the solicited bit is set. */
242 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
243 swqe->wr.send_flags & IB_SEND_SOLICITED);
246 if (atomic_dec_and_test(&qp->refcount))
252 * ipath_make_ud_req - construct a UD request packet
255 * Return 1 if constructed; otherwise, return 0.
257 int ipath_make_ud_req(struct ipath_qp *qp)
259 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
260 struct ipath_other_headers *ohdr;
261 struct ib_ah_attr *ah_attr;
262 struct ipath_swqe *wqe;
271 spin_lock_irqsave(&qp->s_lock, flags);
273 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
274 if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
276 /* We are in the error state, flush the work request. */
277 if (qp->s_last == qp->s_head)
279 /* If DMAs are in progress, we can't flush immediately. */
280 if (atomic_read(&qp->s_dma_busy)) {
281 qp->s_flags |= IPATH_S_WAIT_DMA;
284 wqe = get_swqe_ptr(qp, qp->s_last);
285 ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
289 if (qp->s_cur == qp->s_head)
292 wqe = get_swqe_ptr(qp, qp->s_cur);
293 if (++qp->s_cur >= qp->s_size)
296 /* Construct the header. */
297 ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
298 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
299 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
300 dev->n_multicast_xmit++;
302 dev->n_unicast_xmit++;
304 dev->n_unicast_xmit++;
305 lid = ah_attr->dlid & ~((1 << dev->dd->ipath_lmc) - 1);
306 if (unlikely(lid == dev->dd->ipath_lid)) {
308 * If DMAs are in progress, we can't generate
309 * a completion for the loopback packet since
310 * it would be out of order.
311 * XXX Instead of waiting, we could queue a
312 * zero length descriptor so we get a callback.
314 if (atomic_read(&qp->s_dma_busy)) {
315 qp->s_flags |= IPATH_S_WAIT_DMA;
318 spin_unlock_irqrestore(&qp->s_lock, flags);
319 ipath_ud_loopback(qp, wqe);
320 spin_lock_irqsave(&qp->s_lock, flags);
321 ipath_send_complete(qp, wqe, IB_WC_SUCCESS);
326 extra_bytes = -wqe->length & 3;
327 nwords = (wqe->length + extra_bytes) >> 2;
329 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
331 qp->s_cur_size = wqe->length;
332 qp->s_cur_sge = &qp->s_sge;
333 qp->s_dmult = ah_attr->static_rate;
335 qp->s_sge.sge = wqe->sg_list[0];
336 qp->s_sge.sg_list = wqe->sg_list + 1;
337 qp->s_sge.num_sge = wqe->wr.num_sge;
339 if (ah_attr->ah_flags & IB_AH_GRH) {
340 /* Header size in 32-bit words. */
341 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
343 qp->s_hdrwords, nwords);
344 lrh0 = IPATH_LRH_GRH;
345 ohdr = &qp->s_hdr.u.l.oth;
347 * Don't worry about sending to locally attached multicast
348 * QPs. It is unspecified by the spec. what happens.
351 /* Header size in 32-bit words. */
352 lrh0 = IPATH_LRH_BTH;
353 ohdr = &qp->s_hdr.u.oth;
355 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
357 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
358 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
360 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
361 lrh0 |= ah_attr->sl << 4;
362 if (qp->ibqp.qp_type == IB_QPT_SMI)
363 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
364 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
365 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
366 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
368 lid = dev->dd->ipath_lid;
370 lid |= ah_attr->src_path_bits &
371 ((1 << dev->dd->ipath_lmc) - 1);
372 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
374 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
375 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
377 bth0 |= extra_bytes << 20;
378 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
379 ipath_get_pkey(dev->dd, qp->s_pkey_index);
380 ohdr->bth[0] = cpu_to_be32(bth0);
382 * Use the multicast QP if the destination LID is a multicast LID.
384 ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
385 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
386 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
387 cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
388 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
390 * Qkeys with the high order bit set mean use the
391 * qkey from the QP context instead of the WR (see 10.2.5).
393 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
394 qp->qkey : wqe->wr.wr.ud.remote_qkey);
395 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
402 qp->s_flags &= ~IPATH_S_BUSY;
404 spin_unlock_irqrestore(&qp->s_lock, flags);
409 * ipath_ud_rcv - receive an incoming UD packet
410 * @dev: the device the packet came in on
411 * @hdr: the packet header
412 * @has_grh: true if the packet has a GRH
413 * @data: the packet data
414 * @tlen: the packet length
415 * @qp: the QP the packet came on
417 * This is called from ipath_qp_rcv() to process an incoming UD packet
419 * Called at interrupt level.
421 void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
422 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
424 struct ipath_other_headers *ohdr;
437 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
438 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
439 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
442 ohdr = &hdr->u.l.oth;
443 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
445 * The header with GRH is 68 bytes and the core driver sets
446 * the eager header buffer size to 56 bytes so the last 12
447 * bytes of the IB header is in the data buffer.
449 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
450 if (header_in_data) {
451 qkey = be32_to_cpu(((__be32 *) data)[1]);
452 src_qp = be32_to_cpu(((__be32 *) data)[2]);
455 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
456 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]);
459 src_qp &= IPATH_QPN_MASK;
462 * Check that the permissive LID is only used on QP0
463 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
465 if (qp->ibqp.qp_num) {
466 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
467 hdr->lrh[3] == IB_LID_PERMISSIVE)) {
471 if (unlikely(qkey != qp->qkey)) {
472 /* XXX OK to lose a count once in a while. */
473 dev->qkey_violations++;
477 } else if (hdr->lrh[1] == IB_LID_PERMISSIVE ||
478 hdr->lrh[3] == IB_LID_PERMISSIVE) {
479 struct ib_smp *smp = (struct ib_smp *) data;
481 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
488 * The opcode is in the low byte when its in network order
489 * (top byte when in host order).
491 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
492 if (qp->ibqp.qp_num > 1 &&
493 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
494 if (header_in_data) {
495 wc.ex.imm_data = *(__be32 *) data;
496 data += sizeof(__be32);
498 wc.ex.imm_data = ohdr->u.ud.imm_data;
499 wc.wc_flags = IB_WC_WITH_IMM;
500 hdrsize += sizeof(u32);
501 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
509 /* Get the number of bytes the message was padded by. */
510 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
511 if (unlikely(tlen < (hdrsize + pad + 4))) {
512 /* Drop incomplete packets. */
516 tlen -= hdrsize + pad + 4;
518 /* Drop invalid MAD packets (see 13.5.3.1). */
519 if (unlikely((qp->ibqp.qp_num == 0 &&
521 (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)) ||
522 (qp->ibqp.qp_num == 1 &&
524 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))) {
530 * A GRH is expected to preceed the data even if not
531 * present on the wire.
533 wc.byte_len = tlen + sizeof(struct ib_grh);
536 * Get the next work request entry to find where to put the data.
538 if (qp->r_flags & IPATH_R_REUSE_SGE)
539 qp->r_flags &= ~IPATH_R_REUSE_SGE;
540 else if (!ipath_get_rwqe(qp, 0)) {
542 * Count VL15 packets dropped due to no receive buffer.
543 * Otherwise, count them as buffer overruns since usually,
544 * the HW will be able to receive packets even if there are
545 * no QPs with posted receive buffers.
547 if (qp->ibqp.qp_num == 0)
548 dev->n_vl15_dropped++;
553 /* Silently drop packets which are too big. */
554 if (wc.byte_len > qp->r_len) {
555 qp->r_flags |= IPATH_R_REUSE_SGE;
560 ipath_copy_sge(&qp->r_sge, &hdr->u.l.grh,
561 sizeof(struct ib_grh));
562 wc.wc_flags |= IB_WC_GRH;
564 ipath_skip_sge(&qp->r_sge, sizeof(struct ib_grh));
565 ipath_copy_sge(&qp->r_sge, data,
566 wc.byte_len - sizeof(struct ib_grh));
567 if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
569 wc.wr_id = qp->r_wr_id;
570 wc.status = IB_WC_SUCCESS;
571 wc.opcode = IB_WC_RECV;
575 /* XXX do we know which pkey matched? Only needed for GSI. */
577 wc.slid = be16_to_cpu(hdr->lrh[3]);
578 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
579 dlid = be16_to_cpu(hdr->lrh[1]);
581 * Save the LMC lower bits if the destination LID is a unicast LID.
583 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
584 dlid & ((1 << dev->dd->ipath_lmc) - 1);
586 /* Signal completion event if the solicited bit is set. */
587 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
589 __constant_cpu_to_be32(1 << 23)) != 0);