Linux 3.11-rc3
[cris-mirror.git] / drivers / infiniband / hw / qib / qib_ud.c
blobd6c7fe7f88d5ecd8b133e892e2f31e7f87bc74fb
1 /*
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
13 * conditions are met:
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
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
31 * SOFTWARE.
34 #include <rdma/ib_smi.h>
36 #include "qib.h"
37 #include "qib_mad.h"
39 /**
40 * qib_ud_loopback - handle send on loopback QPs
41 * @sqp: the sending QP
42 * @swqe: the send work request
44 * This is called from qib_make_ud_req() to forward a WQE addressed
45 * to the same HCA.
46 * Note that the receive interrupt handler may be calling qib_ud_rcv()
47 * while this is being called.
49 static void qib_ud_loopback(struct qib_qp *sqp, struct qib_swqe *swqe)
51 struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
52 struct qib_pportdata *ppd;
53 struct qib_qp *qp;
54 struct ib_ah_attr *ah_attr;
55 unsigned long flags;
56 struct qib_sge_state ssge;
57 struct qib_sge *sge;
58 struct ib_wc wc;
59 u32 length;
61 qp = qib_lookup_qpn(ibp, swqe->wr.wr.ud.remote_qpn);
62 if (!qp) {
63 ibp->n_pkt_drops++;
64 return;
66 if (qp->ibqp.qp_type != sqp->ibqp.qp_type ||
67 !(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) {
68 ibp->n_pkt_drops++;
69 goto drop;
72 ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
73 ppd = ppd_from_ibp(ibp);
75 if (qp->ibqp.qp_num > 1) {
76 u16 pkey1;
77 u16 pkey2;
78 u16 lid;
80 pkey1 = qib_get_pkey(ibp, sqp->s_pkey_index);
81 pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
82 if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
83 lid = ppd->lid | (ah_attr->src_path_bits &
84 ((1 << ppd->lmc) - 1));
85 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, pkey1,
86 ah_attr->sl,
87 sqp->ibqp.qp_num, qp->ibqp.qp_num,
88 cpu_to_be16(lid),
89 cpu_to_be16(ah_attr->dlid));
90 goto drop;
95 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
96 * Qkeys with the high order bit set mean use the
97 * qkey from the QP context instead of the WR (see 10.2.5).
99 if (qp->ibqp.qp_num) {
100 u32 qkey;
102 qkey = (int)swqe->wr.wr.ud.remote_qkey < 0 ?
103 sqp->qkey : swqe->wr.wr.ud.remote_qkey;
104 if (unlikely(qkey != qp->qkey)) {
105 u16 lid;
107 lid = ppd->lid | (ah_attr->src_path_bits &
108 ((1 << ppd->lmc) - 1));
109 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
110 ah_attr->sl,
111 sqp->ibqp.qp_num, qp->ibqp.qp_num,
112 cpu_to_be16(lid),
113 cpu_to_be16(ah_attr->dlid));
114 goto drop;
119 * A GRH is expected to precede the data even if not
120 * present on the wire.
122 length = swqe->length;
123 memset(&wc, 0, sizeof wc);
124 wc.byte_len = length + sizeof(struct ib_grh);
126 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
127 wc.wc_flags = IB_WC_WITH_IMM;
128 wc.ex.imm_data = swqe->wr.ex.imm_data;
131 spin_lock_irqsave(&qp->r_lock, flags);
134 * Get the next work request entry to find where to put the data.
136 if (qp->r_flags & QIB_R_REUSE_SGE)
137 qp->r_flags &= ~QIB_R_REUSE_SGE;
138 else {
139 int ret;
141 ret = qib_get_rwqe(qp, 0);
142 if (ret < 0) {
143 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
144 goto bail_unlock;
146 if (!ret) {
147 if (qp->ibqp.qp_num == 0)
148 ibp->n_vl15_dropped++;
149 goto bail_unlock;
152 /* Silently drop packets which are too big. */
153 if (unlikely(wc.byte_len > qp->r_len)) {
154 qp->r_flags |= QIB_R_REUSE_SGE;
155 ibp->n_pkt_drops++;
156 goto bail_unlock;
159 if (ah_attr->ah_flags & IB_AH_GRH) {
160 qib_copy_sge(&qp->r_sge, &ah_attr->grh,
161 sizeof(struct ib_grh), 1);
162 wc.wc_flags |= IB_WC_GRH;
163 } else
164 qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
165 ssge.sg_list = swqe->sg_list + 1;
166 ssge.sge = *swqe->sg_list;
167 ssge.num_sge = swqe->wr.num_sge;
168 sge = &ssge.sge;
169 while (length) {
170 u32 len = sge->length;
172 if (len > length)
173 len = length;
174 if (len > sge->sge_length)
175 len = sge->sge_length;
176 BUG_ON(len == 0);
177 qib_copy_sge(&qp->r_sge, sge->vaddr, len, 1);
178 sge->vaddr += len;
179 sge->length -= len;
180 sge->sge_length -= len;
181 if (sge->sge_length == 0) {
182 if (--ssge.num_sge)
183 *sge = *ssge.sg_list++;
184 } else if (sge->length == 0 && sge->mr->lkey) {
185 if (++sge->n >= QIB_SEGSZ) {
186 if (++sge->m >= sge->mr->mapsz)
187 break;
188 sge->n = 0;
190 sge->vaddr =
191 sge->mr->map[sge->m]->segs[sge->n].vaddr;
192 sge->length =
193 sge->mr->map[sge->m]->segs[sge->n].length;
195 length -= len;
197 qib_put_ss(&qp->r_sge);
198 if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
199 goto bail_unlock;
200 wc.wr_id = qp->r_wr_id;
201 wc.status = IB_WC_SUCCESS;
202 wc.opcode = IB_WC_RECV;
203 wc.qp = &qp->ibqp;
204 wc.src_qp = sqp->ibqp.qp_num;
205 wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
206 swqe->wr.wr.ud.pkey_index : 0;
207 wc.slid = ppd->lid | (ah_attr->src_path_bits & ((1 << ppd->lmc) - 1));
208 wc.sl = ah_attr->sl;
209 wc.dlid_path_bits = ah_attr->dlid & ((1 << ppd->lmc) - 1);
210 wc.port_num = qp->port_num;
211 /* Signal completion event if the solicited bit is set. */
212 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
213 swqe->wr.send_flags & IB_SEND_SOLICITED);
214 ibp->n_loop_pkts++;
215 bail_unlock:
216 spin_unlock_irqrestore(&qp->r_lock, flags);
217 drop:
218 if (atomic_dec_and_test(&qp->refcount))
219 wake_up(&qp->wait);
223 * qib_make_ud_req - construct a UD request packet
224 * @qp: the QP
226 * Return 1 if constructed; otherwise, return 0.
228 int qib_make_ud_req(struct qib_qp *qp)
230 struct qib_other_headers *ohdr;
231 struct ib_ah_attr *ah_attr;
232 struct qib_pportdata *ppd;
233 struct qib_ibport *ibp;
234 struct qib_swqe *wqe;
235 unsigned long flags;
236 u32 nwords;
237 u32 extra_bytes;
238 u32 bth0;
239 u16 lrh0;
240 u16 lid;
241 int ret = 0;
242 int next_cur;
244 spin_lock_irqsave(&qp->s_lock, flags);
246 if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_NEXT_SEND_OK)) {
247 if (!(ib_qib_state_ops[qp->state] & QIB_FLUSH_SEND))
248 goto bail;
249 /* We are in the error state, flush the work request. */
250 if (qp->s_last == qp->s_head)
251 goto bail;
252 /* If DMAs are in progress, we can't flush immediately. */
253 if (atomic_read(&qp->s_dma_busy)) {
254 qp->s_flags |= QIB_S_WAIT_DMA;
255 goto bail;
257 wqe = get_swqe_ptr(qp, qp->s_last);
258 qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
259 goto done;
262 if (qp->s_cur == qp->s_head)
263 goto bail;
265 wqe = get_swqe_ptr(qp, qp->s_cur);
266 next_cur = qp->s_cur + 1;
267 if (next_cur >= qp->s_size)
268 next_cur = 0;
270 /* Construct the header. */
271 ibp = to_iport(qp->ibqp.device, qp->port_num);
272 ppd = ppd_from_ibp(ibp);
273 ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
274 if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE) {
275 if (ah_attr->dlid != QIB_PERMISSIVE_LID)
276 ibp->n_multicast_xmit++;
277 else
278 ibp->n_unicast_xmit++;
279 } else {
280 ibp->n_unicast_xmit++;
281 lid = ah_attr->dlid & ~((1 << ppd->lmc) - 1);
282 if (unlikely(lid == ppd->lid)) {
284 * If DMAs are in progress, we can't generate
285 * a completion for the loopback packet since
286 * it would be out of order.
287 * XXX Instead of waiting, we could queue a
288 * zero length descriptor so we get a callback.
290 if (atomic_read(&qp->s_dma_busy)) {
291 qp->s_flags |= QIB_S_WAIT_DMA;
292 goto bail;
294 qp->s_cur = next_cur;
295 spin_unlock_irqrestore(&qp->s_lock, flags);
296 qib_ud_loopback(qp, wqe);
297 spin_lock_irqsave(&qp->s_lock, flags);
298 qib_send_complete(qp, wqe, IB_WC_SUCCESS);
299 goto done;
303 qp->s_cur = next_cur;
304 extra_bytes = -wqe->length & 3;
305 nwords = (wqe->length + extra_bytes) >> 2;
307 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
308 qp->s_hdrwords = 7;
309 qp->s_cur_size = wqe->length;
310 qp->s_cur_sge = &qp->s_sge;
311 qp->s_srate = ah_attr->static_rate;
312 qp->s_wqe = wqe;
313 qp->s_sge.sge = wqe->sg_list[0];
314 qp->s_sge.sg_list = wqe->sg_list + 1;
315 qp->s_sge.num_sge = wqe->wr.num_sge;
316 qp->s_sge.total_len = wqe->length;
318 if (ah_attr->ah_flags & IB_AH_GRH) {
319 /* Header size in 32-bit words. */
320 qp->s_hdrwords += qib_make_grh(ibp, &qp->s_hdr->u.l.grh,
321 &ah_attr->grh,
322 qp->s_hdrwords, nwords);
323 lrh0 = QIB_LRH_GRH;
324 ohdr = &qp->s_hdr->u.l.oth;
326 * Don't worry about sending to locally attached multicast
327 * QPs. It is unspecified by the spec. what happens.
329 } else {
330 /* Header size in 32-bit words. */
331 lrh0 = QIB_LRH_BTH;
332 ohdr = &qp->s_hdr->u.oth;
334 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
335 qp->s_hdrwords++;
336 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
337 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
338 } else
339 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
340 lrh0 |= ah_attr->sl << 4;
341 if (qp->ibqp.qp_type == IB_QPT_SMI)
342 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
343 else
344 lrh0 |= ibp->sl_to_vl[ah_attr->sl] << 12;
345 qp->s_hdr->lrh[0] = cpu_to_be16(lrh0);
346 qp->s_hdr->lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
347 qp->s_hdr->lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
348 lid = ppd->lid;
349 if (lid) {
350 lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
351 qp->s_hdr->lrh[3] = cpu_to_be16(lid);
352 } else
353 qp->s_hdr->lrh[3] = IB_LID_PERMISSIVE;
354 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
355 bth0 |= IB_BTH_SOLICITED;
356 bth0 |= extra_bytes << 20;
357 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? QIB_DEFAULT_P_KEY :
358 qib_get_pkey(ibp, qp->ibqp.qp_type == IB_QPT_GSI ?
359 wqe->wr.wr.ud.pkey_index : qp->s_pkey_index);
360 ohdr->bth[0] = cpu_to_be32(bth0);
362 * Use the multicast QP if the destination LID is a multicast LID.
364 ohdr->bth[1] = ah_attr->dlid >= QIB_MULTICAST_LID_BASE &&
365 ah_attr->dlid != QIB_PERMISSIVE_LID ?
366 cpu_to_be32(QIB_MULTICAST_QPN) :
367 cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
368 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & QIB_PSN_MASK);
370 * Qkeys with the high order bit set mean use the
371 * qkey from the QP context instead of the WR (see 10.2.5).
373 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
374 qp->qkey : wqe->wr.wr.ud.remote_qkey);
375 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
377 done:
378 ret = 1;
379 goto unlock;
381 bail:
382 qp->s_flags &= ~QIB_S_BUSY;
383 unlock:
384 spin_unlock_irqrestore(&qp->s_lock, flags);
385 return ret;
388 static unsigned qib_lookup_pkey(struct qib_ibport *ibp, u16 pkey)
390 struct qib_pportdata *ppd = ppd_from_ibp(ibp);
391 struct qib_devdata *dd = ppd->dd;
392 unsigned ctxt = ppd->hw_pidx;
393 unsigned i;
395 pkey &= 0x7fff; /* remove limited/full membership bit */
397 for (i = 0; i < ARRAY_SIZE(dd->rcd[ctxt]->pkeys); ++i)
398 if ((dd->rcd[ctxt]->pkeys[i] & 0x7fff) == pkey)
399 return i;
402 * Should not get here, this means hardware failed to validate pkeys.
403 * Punt and return index 0.
405 return 0;
409 * qib_ud_rcv - receive an incoming UD packet
410 * @ibp: the port 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 qib_qp_rcv() to process an incoming UD packet
418 * for the given QP.
419 * Called at interrupt level.
421 void qib_ud_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
422 int has_grh, void *data, u32 tlen, struct qib_qp *qp)
424 struct qib_other_headers *ohdr;
425 int opcode;
426 u32 hdrsize;
427 u32 pad;
428 struct ib_wc wc;
429 u32 qkey;
430 u32 src_qp;
431 u16 dlid;
433 /* Check for GRH */
434 if (!has_grh) {
435 ohdr = &hdr->u.oth;
436 hdrsize = 8 + 12 + 8; /* LRH + BTH + DETH */
437 } else {
438 ohdr = &hdr->u.l.oth;
439 hdrsize = 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
441 qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
442 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & QIB_QPN_MASK;
445 * Get the number of bytes the message was padded by
446 * and drop incomplete packets.
448 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
449 if (unlikely(tlen < (hdrsize + pad + 4)))
450 goto drop;
452 tlen -= hdrsize + pad + 4;
455 * Check that the permissive LID is only used on QP0
456 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
458 if (qp->ibqp.qp_num) {
459 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
460 hdr->lrh[3] == IB_LID_PERMISSIVE))
461 goto drop;
462 if (qp->ibqp.qp_num > 1) {
463 u16 pkey1, pkey2;
465 pkey1 = be32_to_cpu(ohdr->bth[0]);
466 pkey2 = qib_get_pkey(ibp, qp->s_pkey_index);
467 if (unlikely(!qib_pkey_ok(pkey1, pkey2))) {
468 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
469 pkey1,
470 (be16_to_cpu(hdr->lrh[0]) >> 4) &
471 0xF,
472 src_qp, qp->ibqp.qp_num,
473 hdr->lrh[3], hdr->lrh[1]);
474 return;
477 if (unlikely(qkey != qp->qkey)) {
478 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_QKEY, qkey,
479 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
480 src_qp, qp->ibqp.qp_num,
481 hdr->lrh[3], hdr->lrh[1]);
482 return;
484 /* Drop invalid MAD packets (see 13.5.3.1). */
485 if (unlikely(qp->ibqp.qp_num == 1 &&
486 (tlen != 256 ||
487 (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))
488 goto drop;
489 } else {
490 struct ib_smp *smp;
492 /* Drop invalid MAD packets (see 13.5.3.1). */
493 if (tlen != 256 || (be16_to_cpu(hdr->lrh[0]) >> 12) != 15)
494 goto drop;
495 smp = (struct ib_smp *) data;
496 if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
497 hdr->lrh[3] == IB_LID_PERMISSIVE) &&
498 smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
499 goto drop;
503 * The opcode is in the low byte when its in network order
504 * (top byte when in host order).
506 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
507 if (qp->ibqp.qp_num > 1 &&
508 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
509 wc.ex.imm_data = ohdr->u.ud.imm_data;
510 wc.wc_flags = IB_WC_WITH_IMM;
511 tlen -= sizeof(u32);
512 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
513 wc.ex.imm_data = 0;
514 wc.wc_flags = 0;
515 } else
516 goto drop;
519 * A GRH is expected to precede the data even if not
520 * present on the wire.
522 wc.byte_len = tlen + sizeof(struct ib_grh);
525 * Get the next work request entry to find where to put the data.
527 if (qp->r_flags & QIB_R_REUSE_SGE)
528 qp->r_flags &= ~QIB_R_REUSE_SGE;
529 else {
530 int ret;
532 ret = qib_get_rwqe(qp, 0);
533 if (ret < 0) {
534 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
535 return;
537 if (!ret) {
538 if (qp->ibqp.qp_num == 0)
539 ibp->n_vl15_dropped++;
540 return;
543 /* Silently drop packets which are too big. */
544 if (unlikely(wc.byte_len > qp->r_len)) {
545 qp->r_flags |= QIB_R_REUSE_SGE;
546 goto drop;
548 if (has_grh) {
549 qib_copy_sge(&qp->r_sge, &hdr->u.l.grh,
550 sizeof(struct ib_grh), 1);
551 wc.wc_flags |= IB_WC_GRH;
552 } else
553 qib_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
554 qib_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 1);
555 qib_put_ss(&qp->r_sge);
556 if (!test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags))
557 return;
558 wc.wr_id = qp->r_wr_id;
559 wc.status = IB_WC_SUCCESS;
560 wc.opcode = IB_WC_RECV;
561 wc.vendor_err = 0;
562 wc.qp = &qp->ibqp;
563 wc.src_qp = src_qp;
564 wc.pkey_index = qp->ibqp.qp_type == IB_QPT_GSI ?
565 qib_lookup_pkey(ibp, be32_to_cpu(ohdr->bth[0])) : 0;
566 wc.slid = be16_to_cpu(hdr->lrh[3]);
567 wc.sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF;
568 dlid = be16_to_cpu(hdr->lrh[1]);
570 * Save the LMC lower bits if the destination LID is a unicast LID.
572 wc.dlid_path_bits = dlid >= QIB_MULTICAST_LID_BASE ? 0 :
573 dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
574 wc.port_num = qp->port_num;
575 /* Signal completion event if the solicited bit is set. */
576 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
577 (ohdr->bth[0] &
578 cpu_to_be32(IB_BTH_SOLICITED)) != 0);
579 return;
581 drop:
582 ibp->n_pkt_drops++;