2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
38 #include <linux/init.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
46 #include "mthca_dev.h"
47 #include "mthca_cmd.h"
48 #include "mthca_memfree.h"
49 #include "mthca_wqe.h"
52 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
53 MTHCA_ACK_REQ_FREQ = 10,
54 MTHCA_FLIGHT_LIMIT = 9,
55 MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */
56 MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */
57 MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */
61 MTHCA_QP_STATE_RST = 0,
62 MTHCA_QP_STATE_INIT = 1,
63 MTHCA_QP_STATE_RTR = 2,
64 MTHCA_QP_STATE_RTS = 3,
65 MTHCA_QP_STATE_SQE = 4,
66 MTHCA_QP_STATE_SQD = 5,
67 MTHCA_QP_STATE_ERR = 6,
68 MTHCA_QP_STATE_DRAINING = 7
80 MTHCA_QP_PM_MIGRATED = 0x3,
81 MTHCA_QP_PM_ARMED = 0x0,
82 MTHCA_QP_PM_REARM = 0x1
86 /* qp_context flags */
87 MTHCA_QP_BIT_DE = 1 << 8,
89 MTHCA_QP_BIT_SRE = 1 << 15,
90 MTHCA_QP_BIT_SWE = 1 << 14,
91 MTHCA_QP_BIT_SAE = 1 << 13,
92 MTHCA_QP_BIT_SIC = 1 << 4,
93 MTHCA_QP_BIT_SSC = 1 << 3,
95 MTHCA_QP_BIT_RRE = 1 << 15,
96 MTHCA_QP_BIT_RWE = 1 << 14,
97 MTHCA_QP_BIT_RAE = 1 << 13,
98 MTHCA_QP_BIT_RIC = 1 << 4,
99 MTHCA_QP_BIT_RSC = 1 << 3
102 struct mthca_qp_path {
111 __be32 sl_tclass_flowlabel;
113 } __attribute__((packed));
115 struct mthca_qp_context {
117 __be32 tavor_sched_queue; /* Reserved on Arbel */
119 u8 rq_size_stride; /* Reserved on Tavor */
120 u8 sq_size_stride; /* Reserved on Tavor */
121 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
126 struct mthca_qp_path pri_path;
127 struct mthca_qp_path alt_path;
134 __be32 next_send_psn;
136 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
137 __be32 snd_db_index; /* (debugging only entries) */
138 __be32 last_acked_psn;
141 __be32 rnr_nextrecvpsn;
144 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
145 __be32 rcv_db_index; /* (debugging only entries) */
149 __be16 rq_wqe_counter; /* reserved on Tavor */
150 __be16 sq_wqe_counter; /* reserved on Tavor */
152 } __attribute__((packed));
154 struct mthca_qp_param {
155 __be32 opt_param_mask;
157 struct mthca_qp_context context;
159 } __attribute__((packed));
162 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
163 MTHCA_QP_OPTPAR_RRE = 1 << 1,
164 MTHCA_QP_OPTPAR_RAE = 1 << 2,
165 MTHCA_QP_OPTPAR_RWE = 1 << 3,
166 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
167 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
168 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
169 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
170 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
171 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
172 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
173 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
174 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
175 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
176 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
177 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
178 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
181 static const u8 mthca_opcode[] = {
182 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
183 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
184 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
185 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
186 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
187 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
188 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
191 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
193 return qp->qpn >= dev->qp_table.sqp_start &&
194 qp->qpn <= dev->qp_table.sqp_start + 3;
197 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
199 return qp->qpn >= dev->qp_table.sqp_start &&
200 qp->qpn <= dev->qp_table.sqp_start + 1;
203 static void *get_recv_wqe(struct mthca_qp *qp, int n)
206 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
208 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
209 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
212 static void *get_send_wqe(struct mthca_qp *qp, int n)
215 return qp->queue.direct.buf + qp->send_wqe_offset +
216 (n << qp->sq.wqe_shift);
218 return qp->queue.page_list[(qp->send_wqe_offset +
219 (n << qp->sq.wqe_shift)) >>
221 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
225 static void mthca_wq_init(struct mthca_wq *wq)
227 spin_lock_init(&wq->lock);
229 wq->last_comp = wq->max - 1;
234 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
235 enum ib_event_type event_type)
238 struct ib_event event;
240 spin_lock(&dev->qp_table.lock);
241 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
243 atomic_inc(&qp->refcount);
244 spin_unlock(&dev->qp_table.lock);
247 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
251 event.device = &dev->ib_dev;
252 event.event = event_type;
253 event.element.qp = &qp->ibqp;
254 if (qp->ibqp.event_handler)
255 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
257 if (atomic_dec_and_test(&qp->refcount))
261 static int to_mthca_state(enum ib_qp_state ib_state)
264 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
265 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
266 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
267 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
268 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
269 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
270 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
275 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
277 static int to_mthca_st(int transport)
280 case RC: return MTHCA_QP_ST_RC;
281 case UC: return MTHCA_QP_ST_UC;
282 case UD: return MTHCA_QP_ST_UD;
283 case RD: return MTHCA_QP_ST_RD;
284 case MLX: return MTHCA_QP_ST_MLX;
289 static const struct {
291 u32 req_param[NUM_TRANS];
292 u32 opt_param[NUM_TRANS];
293 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
295 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
296 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
298 .trans = MTHCA_TRANS_RST2INIT,
300 [UD] = (IB_QP_PKEY_INDEX |
303 [UC] = (IB_QP_PKEY_INDEX |
306 [RC] = (IB_QP_PKEY_INDEX |
309 [MLX] = (IB_QP_PKEY_INDEX |
312 /* bug-for-bug compatibility with VAPI: */
319 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
320 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
322 .trans = MTHCA_TRANS_INIT2INIT,
324 [UD] = (IB_QP_PKEY_INDEX |
327 [UC] = (IB_QP_PKEY_INDEX |
330 [RC] = (IB_QP_PKEY_INDEX |
333 [MLX] = (IB_QP_PKEY_INDEX |
338 .trans = MTHCA_TRANS_INIT2RTR,
348 IB_QP_MAX_DEST_RD_ATOMIC |
349 IB_QP_MIN_RNR_TIMER),
352 [UD] = (IB_QP_PKEY_INDEX |
354 [UC] = (IB_QP_ALT_PATH |
357 [RC] = (IB_QP_ALT_PATH |
360 [MLX] = (IB_QP_PKEY_INDEX |
366 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
367 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
369 .trans = MTHCA_TRANS_RTR2RTS,
373 [RC] = (IB_QP_TIMEOUT |
377 IB_QP_MAX_QP_RD_ATOMIC),
378 [MLX] = IB_QP_SQ_PSN,
381 [UD] = (IB_QP_CUR_STATE |
383 [UC] = (IB_QP_CUR_STATE |
387 IB_QP_PATH_MIG_STATE),
388 [RC] = (IB_QP_CUR_STATE |
392 IB_QP_MIN_RNR_TIMER |
393 IB_QP_PATH_MIG_STATE),
394 [MLX] = (IB_QP_CUR_STATE |
400 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
401 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
403 .trans = MTHCA_TRANS_RTS2RTS,
405 [UD] = (IB_QP_CUR_STATE |
407 [UC] = (IB_QP_ACCESS_FLAGS |
409 IB_QP_PATH_MIG_STATE),
410 [RC] = (IB_QP_ACCESS_FLAGS |
412 IB_QP_PATH_MIG_STATE |
413 IB_QP_MIN_RNR_TIMER),
414 [MLX] = (IB_QP_CUR_STATE |
419 .trans = MTHCA_TRANS_RTS2SQD,
423 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
424 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
426 .trans = MTHCA_TRANS_SQD2RTS,
428 [UD] = (IB_QP_CUR_STATE |
430 [UC] = (IB_QP_CUR_STATE |
433 IB_QP_PATH_MIG_STATE),
434 [RC] = (IB_QP_CUR_STATE |
437 IB_QP_MIN_RNR_TIMER |
438 IB_QP_PATH_MIG_STATE),
439 [MLX] = (IB_QP_CUR_STATE |
444 .trans = MTHCA_TRANS_SQD2SQD,
446 [UD] = (IB_QP_PKEY_INDEX |
453 IB_QP_PATH_MIG_STATE),
458 IB_QP_MAX_QP_RD_ATOMIC |
459 IB_QP_MAX_DEST_RD_ATOMIC |
464 IB_QP_MIN_RNR_TIMER |
465 IB_QP_PATH_MIG_STATE),
466 [MLX] = (IB_QP_PKEY_INDEX |
472 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
473 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
475 .trans = MTHCA_TRANS_SQERR2RTS,
477 [UD] = (IB_QP_CUR_STATE |
479 [UC] = IB_QP_CUR_STATE,
480 [RC] = (IB_QP_CUR_STATE |
481 IB_QP_MIN_RNR_TIMER),
482 [MLX] = (IB_QP_CUR_STATE |
488 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
489 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
493 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
496 if (attr_mask & IB_QP_PKEY_INDEX)
497 sqp->pkey_index = attr->pkey_index;
498 if (attr_mask & IB_QP_QKEY)
499 sqp->qkey = attr->qkey;
500 if (attr_mask & IB_QP_SQ_PSN)
501 sqp->send_psn = attr->sq_psn;
504 static void init_port(struct mthca_dev *dev, int port)
508 struct mthca_init_ib_param param;
510 memset(¶m, 0, sizeof param);
512 param.port_width = dev->limits.port_width_cap;
513 param.vl_cap = dev->limits.vl_cap;
514 param.mtu_cap = dev->limits.mtu_cap;
515 param.gid_cap = dev->limits.gid_table_len;
516 param.pkey_cap = dev->limits.pkey_table_len;
518 err = mthca_INIT_IB(dev, ¶m, port, &status);
520 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
522 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
525 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
527 struct mthca_dev *dev = to_mdev(ibqp->device);
528 struct mthca_qp *qp = to_mqp(ibqp);
529 enum ib_qp_state cur_state, new_state;
530 struct mthca_mailbox *mailbox;
531 struct mthca_qp_param *qp_param;
532 struct mthca_qp_context *qp_context;
533 u32 req_param, opt_param;
537 if (attr_mask & IB_QP_CUR_STATE) {
538 if (attr->cur_qp_state != IB_QPS_RTR &&
539 attr->cur_qp_state != IB_QPS_RTS &&
540 attr->cur_qp_state != IB_QPS_SQD &&
541 attr->cur_qp_state != IB_QPS_SQE)
544 cur_state = attr->cur_qp_state;
546 spin_lock_irq(&qp->sq.lock);
547 spin_lock(&qp->rq.lock);
548 cur_state = qp->state;
549 spin_unlock(&qp->rq.lock);
550 spin_unlock_irq(&qp->sq.lock);
553 if (attr_mask & IB_QP_STATE) {
554 if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
556 new_state = attr->qp_state;
558 new_state = cur_state;
560 if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
561 mthca_dbg(dev, "Illegal QP transition "
562 "%d->%d\n", cur_state, new_state);
566 req_param = state_table[cur_state][new_state].req_param[qp->transport];
567 opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
569 if ((req_param & attr_mask) != req_param) {
570 mthca_dbg(dev, "QP transition "
571 "%d->%d missing req attr 0x%08x\n",
572 cur_state, new_state,
573 req_param & ~attr_mask);
577 if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
578 mthca_dbg(dev, "QP transition (transport %d) "
579 "%d->%d has extra attr 0x%08x\n",
581 cur_state, new_state,
582 attr_mask & ~(req_param | opt_param |
587 if ((attr_mask & IB_QP_PKEY_INDEX) &&
588 attr->pkey_index >= dev->limits.pkey_table_len) {
589 mthca_dbg(dev, "PKey index (%u) too large. max is %d\n",
590 attr->pkey_index,dev->limits.pkey_table_len-1);
594 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
596 return PTR_ERR(mailbox);
597 qp_param = mailbox->buf;
598 qp_context = &qp_param->context;
599 memset(qp_param, 0, sizeof *qp_param);
601 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
602 (to_mthca_st(qp->transport) << 16));
603 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
604 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
605 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
607 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
608 switch (attr->path_mig_state) {
609 case IB_MIG_MIGRATED:
610 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
613 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
616 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
621 /* leave tavor_sched_queue as 0 */
623 if (qp->transport == MLX || qp->transport == UD)
624 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
625 else if (attr_mask & IB_QP_PATH_MTU)
626 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
628 if (mthca_is_memfree(dev)) {
630 qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
631 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
634 qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
635 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
638 /* leave arbel_sched_queue as 0 */
640 if (qp->ibqp.uobject)
641 qp_context->usr_page =
642 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
644 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
645 qp_context->local_qpn = cpu_to_be32(qp->qpn);
646 if (attr_mask & IB_QP_DEST_QPN) {
647 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
650 if (qp->transport == MLX)
651 qp_context->pri_path.port_pkey |=
652 cpu_to_be32(to_msqp(qp)->port << 24);
654 if (attr_mask & IB_QP_PORT) {
655 qp_context->pri_path.port_pkey |=
656 cpu_to_be32(attr->port_num << 24);
657 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
661 if (attr_mask & IB_QP_PKEY_INDEX) {
662 qp_context->pri_path.port_pkey |=
663 cpu_to_be32(attr->pkey_index);
664 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
667 if (attr_mask & IB_QP_RNR_RETRY) {
668 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
669 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
672 if (attr_mask & IB_QP_AV) {
673 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
674 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
675 qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
676 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
677 qp_context->pri_path.g_mylmc |= 1 << 7;
678 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
679 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
680 qp_context->pri_path.sl_tclass_flowlabel =
681 cpu_to_be32((attr->ah_attr.sl << 28) |
682 (attr->ah_attr.grh.traffic_class << 20) |
683 (attr->ah_attr.grh.flow_label));
684 memcpy(qp_context->pri_path.rgid,
685 attr->ah_attr.grh.dgid.raw, 16);
687 qp_context->pri_path.sl_tclass_flowlabel =
688 cpu_to_be32(attr->ah_attr.sl << 28);
690 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
693 if (attr_mask & IB_QP_TIMEOUT) {
694 qp_context->pri_path.ackto = attr->timeout << 3;
695 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
701 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
702 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
703 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
704 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
705 (MTHCA_FLIGHT_LIMIT << 24) |
709 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
710 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
711 if (attr_mask & IB_QP_RETRY_CNT) {
712 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
713 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
716 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
717 qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
718 ffs(attr->max_rd_atomic) - 1 : 0,
720 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
723 if (attr_mask & IB_QP_SQ_PSN)
724 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
725 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
727 if (mthca_is_memfree(dev)) {
728 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
729 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
732 if (attr_mask & IB_QP_ACCESS_FLAGS) {
733 qp_context->params2 |=
734 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
735 MTHCA_QP_BIT_RWE : 0);
738 * Only enable RDMA reads and atomics if we have
739 * responder resources set to a non-zero value.
741 if (qp->resp_depth) {
742 qp_context->params2 |=
743 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
744 MTHCA_QP_BIT_RRE : 0);
745 qp_context->params2 |=
746 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
747 MTHCA_QP_BIT_RAE : 0);
750 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
751 MTHCA_QP_OPTPAR_RRE |
752 MTHCA_QP_OPTPAR_RAE);
754 qp->atomic_rd_en = attr->qp_access_flags;
757 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
760 if (qp->resp_depth && !attr->max_dest_rd_atomic) {
762 * Lowering our responder resources to zero.
763 * Turn off reads RDMA and atomics as responder.
764 * (RRE/RAE in params2 already zero)
766 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
767 MTHCA_QP_OPTPAR_RAE);
770 if (!qp->resp_depth && attr->max_dest_rd_atomic) {
772 * Increasing our responder resources from
773 * zero. Turn on RDMA reads and atomics as
776 qp_context->params2 |=
777 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
778 MTHCA_QP_BIT_RRE : 0);
779 qp_context->params2 |=
780 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
781 MTHCA_QP_BIT_RAE : 0);
783 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
784 MTHCA_QP_OPTPAR_RAE);
788 1 << rra_max < attr->max_dest_rd_atomic &&
789 rra_max < dev->qp_table.rdb_shift;
793 qp_context->params2 |= cpu_to_be32(rra_max << 21);
794 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
796 qp->resp_depth = attr->max_dest_rd_atomic;
799 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
802 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
804 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
805 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
806 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
808 if (attr_mask & IB_QP_RQ_PSN)
809 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
811 qp_context->ra_buff_indx =
812 cpu_to_be32(dev->qp_table.rdb_base +
813 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
814 dev->qp_table.rdb_shift));
816 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
818 if (mthca_is_memfree(dev))
819 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
821 if (attr_mask & IB_QP_QKEY) {
822 qp_context->qkey = cpu_to_be32(attr->qkey);
823 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
827 qp_context->srqn = cpu_to_be32(1 << 24 |
828 to_msrq(ibqp->srq)->srqn);
830 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
831 qp->qpn, 0, mailbox, 0, &status);
833 mthca_warn(dev, "modify QP %d returned status %02x.\n",
834 state_table[cur_state][new_state].trans, status);
839 qp->state = new_state;
841 mthca_free_mailbox(dev, mailbox);
844 store_attrs(to_msqp(qp), attr, attr_mask);
847 * If we moved QP0 to RTR, bring the IB link up; if we moved
848 * QP0 to RESET or ERROR, bring the link back down.
850 if (is_qp0(dev, qp)) {
851 if (cur_state != IB_QPS_RTR &&
852 new_state == IB_QPS_RTR)
853 init_port(dev, to_msqp(qp)->port);
855 if (cur_state != IB_QPS_RESET &&
856 cur_state != IB_QPS_ERR &&
857 (new_state == IB_QPS_RESET ||
858 new_state == IB_QPS_ERR))
859 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
863 * If we moved a kernel QP to RESET, clean up all old CQ
864 * entries and reinitialize the QP.
866 if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
867 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
868 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
869 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
870 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
871 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
873 mthca_wq_init(&qp->sq);
874 mthca_wq_init(&qp->rq);
876 if (mthca_is_memfree(dev)) {
885 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
892 * Calculate the maximum size of WQE s/g segments, excluding
893 * the next segment and other non-data segments.
895 max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) -
896 sizeof (struct mthca_next_seg);
898 switch (qp->transport) {
900 max_data_size -= 2 * sizeof (struct mthca_data_seg);
904 if (mthca_is_memfree(dev))
905 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
907 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
911 max_data_size -= sizeof (struct mthca_raddr_seg);
915 /* We don't support inline data for kernel QPs (yet). */
916 if (!pd->ibpd.uobject)
917 qp->max_inline_data = 0;
919 qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE;
921 qp->sq.max_gs = max_data_size / sizeof (struct mthca_data_seg);
922 qp->rq.max_gs = (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
923 sizeof (struct mthca_next_seg)) /
924 sizeof (struct mthca_data_seg);
928 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
929 * rq.max_gs and sq.max_gs must all be assigned.
930 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
931 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
934 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
941 size = sizeof (struct mthca_next_seg) +
942 qp->rq.max_gs * sizeof (struct mthca_data_seg);
944 if (size > dev->limits.max_desc_sz)
947 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
951 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
952 switch (qp->transport) {
954 size += 2 * sizeof (struct mthca_data_seg);
958 size += mthca_is_memfree(dev) ?
959 sizeof (struct mthca_arbel_ud_seg) :
960 sizeof (struct mthca_tavor_ud_seg);
964 size += sizeof (struct mthca_raddr_seg);
968 size += sizeof (struct mthca_raddr_seg);
970 * An atomic op will require an atomic segment, a
971 * remote address segment and one scatter entry.
973 size = max_t(int, size,
974 sizeof (struct mthca_atomic_seg) +
975 sizeof (struct mthca_raddr_seg) +
976 sizeof (struct mthca_data_seg));
983 /* Make sure that we have enough space for a bind request */
984 size = max_t(int, size, sizeof (struct mthca_bind_seg));
986 size += sizeof (struct mthca_next_seg);
988 if (size > dev->limits.max_desc_sz)
991 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
995 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
996 1 << qp->sq.wqe_shift);
999 * If this is a userspace QP, we don't actually have to
1000 * allocate anything. All we need is to calculate the WQE
1001 * sizes and the send_wqe_offset, so we're done now.
1003 if (pd->ibpd.uobject)
1006 size = PAGE_ALIGN(qp->send_wqe_offset +
1007 (qp->sq.max << qp->sq.wqe_shift));
1009 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1014 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1015 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1026 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1027 struct mthca_qp *qp)
1029 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1030 (qp->sq.max << qp->sq.wqe_shift)),
1031 &qp->queue, qp->is_direct, &qp->mr);
1035 static int mthca_map_memfree(struct mthca_dev *dev,
1036 struct mthca_qp *qp)
1040 if (mthca_is_memfree(dev)) {
1041 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1045 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1049 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1050 qp->qpn << dev->qp_table.rdb_shift);
1059 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1062 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1067 static void mthca_unmap_memfree(struct mthca_dev *dev,
1068 struct mthca_qp *qp)
1070 mthca_table_put(dev, dev->qp_table.rdb_table,
1071 qp->qpn << dev->qp_table.rdb_shift);
1072 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1073 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1076 static int mthca_alloc_memfree(struct mthca_dev *dev,
1077 struct mthca_qp *qp)
1081 if (mthca_is_memfree(dev)) {
1082 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1083 qp->qpn, &qp->rq.db);
1084 if (qp->rq.db_index < 0)
1087 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1088 qp->qpn, &qp->sq.db);
1089 if (qp->sq.db_index < 0)
1090 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1096 static void mthca_free_memfree(struct mthca_dev *dev,
1097 struct mthca_qp *qp)
1099 if (mthca_is_memfree(dev)) {
1100 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1101 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1105 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1106 struct mthca_pd *pd,
1107 struct mthca_cq *send_cq,
1108 struct mthca_cq *recv_cq,
1109 enum ib_sig_type send_policy,
1110 struct mthca_qp *qp)
1115 atomic_set(&qp->refcount, 1);
1116 init_waitqueue_head(&qp->wait);
1117 qp->state = IB_QPS_RESET;
1118 qp->atomic_rd_en = 0;
1120 qp->sq_policy = send_policy;
1121 mthca_wq_init(&qp->sq);
1122 mthca_wq_init(&qp->rq);
1124 ret = mthca_map_memfree(dev, qp);
1128 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1130 mthca_unmap_memfree(dev, qp);
1134 mthca_adjust_qp_caps(dev, pd, qp);
1137 * If this is a userspace QP, we're done now. The doorbells
1138 * will be allocated and buffers will be initialized in
1141 if (pd->ibpd.uobject)
1144 ret = mthca_alloc_memfree(dev, qp);
1146 mthca_free_wqe_buf(dev, qp);
1147 mthca_unmap_memfree(dev, qp);
1151 if (mthca_is_memfree(dev)) {
1152 struct mthca_next_seg *next;
1153 struct mthca_data_seg *scatter;
1154 int size = (sizeof (struct mthca_next_seg) +
1155 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1157 for (i = 0; i < qp->rq.max; ++i) {
1158 next = get_recv_wqe(qp, i);
1159 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1161 next->ee_nds = cpu_to_be32(size);
1163 for (scatter = (void *) (next + 1);
1164 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1166 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1169 for (i = 0; i < qp->sq.max; ++i) {
1170 next = get_send_wqe(qp, i);
1171 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1173 qp->send_wqe_offset);
1177 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1178 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1183 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1184 struct mthca_qp *qp)
1186 /* Sanity check QP size before proceeding */
1187 if (cap->max_send_wr > dev->limits.max_wqes ||
1188 cap->max_recv_wr > dev->limits.max_wqes ||
1189 cap->max_send_sge > dev->limits.max_sg ||
1190 cap->max_recv_sge > dev->limits.max_sg)
1193 if (mthca_is_memfree(dev)) {
1194 qp->rq.max = cap->max_recv_wr ?
1195 roundup_pow_of_two(cap->max_recv_wr) : 0;
1196 qp->sq.max = cap->max_send_wr ?
1197 roundup_pow_of_two(cap->max_send_wr) : 0;
1199 qp->rq.max = cap->max_recv_wr;
1200 qp->sq.max = cap->max_send_wr;
1203 qp->rq.max_gs = cap->max_recv_sge;
1204 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1205 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1206 MTHCA_INLINE_CHUNK_SIZE) /
1207 sizeof (struct mthca_data_seg));
1210 * For MLX transport we need 2 extra S/G entries:
1211 * one for the header and one for the checksum at the end
1213 if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1214 qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1220 int mthca_alloc_qp(struct mthca_dev *dev,
1221 struct mthca_pd *pd,
1222 struct mthca_cq *send_cq,
1223 struct mthca_cq *recv_cq,
1224 enum ib_qp_type type,
1225 enum ib_sig_type send_policy,
1226 struct ib_qp_cap *cap,
1227 struct mthca_qp *qp)
1231 err = mthca_set_qp_size(dev, cap, qp);
1236 case IB_QPT_RC: qp->transport = RC; break;
1237 case IB_QPT_UC: qp->transport = UC; break;
1238 case IB_QPT_UD: qp->transport = UD; break;
1239 default: return -EINVAL;
1242 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1246 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1249 mthca_free(&dev->qp_table.alloc, qp->qpn);
1253 spin_lock_irq(&dev->qp_table.lock);
1254 mthca_array_set(&dev->qp_table.qp,
1255 qp->qpn & (dev->limits.num_qps - 1), qp);
1256 spin_unlock_irq(&dev->qp_table.lock);
1261 int mthca_alloc_sqp(struct mthca_dev *dev,
1262 struct mthca_pd *pd,
1263 struct mthca_cq *send_cq,
1264 struct mthca_cq *recv_cq,
1265 enum ib_sig_type send_policy,
1266 struct ib_qp_cap *cap,
1269 struct mthca_sqp *sqp)
1271 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1274 err = mthca_set_qp_size(dev, cap, &sqp->qp);
1278 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1279 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1280 &sqp->header_dma, GFP_KERNEL);
1281 if (!sqp->header_buf)
1284 spin_lock_irq(&dev->qp_table.lock);
1285 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1288 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1289 spin_unlock_irq(&dev->qp_table.lock);
1296 sqp->qp.transport = MLX;
1298 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1299 send_policy, &sqp->qp);
1303 atomic_inc(&pd->sqp_count);
1309 * Lock CQs here, so that CQ polling code can do QP lookup
1310 * without taking a lock.
1312 spin_lock_irq(&send_cq->lock);
1313 if (send_cq != recv_cq)
1314 spin_lock(&recv_cq->lock);
1316 spin_lock(&dev->qp_table.lock);
1317 mthca_array_clear(&dev->qp_table.qp, mqpn);
1318 spin_unlock(&dev->qp_table.lock);
1320 if (send_cq != recv_cq)
1321 spin_unlock(&recv_cq->lock);
1322 spin_unlock_irq(&send_cq->lock);
1325 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1326 sqp->header_buf, sqp->header_dma);
1331 void mthca_free_qp(struct mthca_dev *dev,
1332 struct mthca_qp *qp)
1335 struct mthca_cq *send_cq;
1336 struct mthca_cq *recv_cq;
1338 send_cq = to_mcq(qp->ibqp.send_cq);
1339 recv_cq = to_mcq(qp->ibqp.recv_cq);
1342 * Lock CQs here, so that CQ polling code can do QP lookup
1343 * without taking a lock.
1345 spin_lock_irq(&send_cq->lock);
1346 if (send_cq != recv_cq)
1347 spin_lock(&recv_cq->lock);
1349 spin_lock(&dev->qp_table.lock);
1350 mthca_array_clear(&dev->qp_table.qp,
1351 qp->qpn & (dev->limits.num_qps - 1));
1352 spin_unlock(&dev->qp_table.lock);
1354 if (send_cq != recv_cq)
1355 spin_unlock(&recv_cq->lock);
1356 spin_unlock_irq(&send_cq->lock);
1358 atomic_dec(&qp->refcount);
1359 wait_event(qp->wait, !atomic_read(&qp->refcount));
1361 if (qp->state != IB_QPS_RESET)
1362 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1365 * If this is a userspace QP, the buffers, MR, CQs and so on
1366 * will be cleaned up in userspace, so all we have to do is
1367 * unref the mem-free tables and free the QPN in our table.
1369 if (!qp->ibqp.uobject) {
1370 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1371 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1372 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1373 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1374 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1376 mthca_free_memfree(dev, qp);
1377 mthca_free_wqe_buf(dev, qp);
1380 mthca_unmap_memfree(dev, qp);
1382 if (is_sqp(dev, qp)) {
1383 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1384 dma_free_coherent(&dev->pdev->dev,
1385 to_msqp(qp)->header_buf_size,
1386 to_msqp(qp)->header_buf,
1387 to_msqp(qp)->header_dma);
1389 mthca_free(&dev->qp_table.alloc, qp->qpn);
1392 /* Create UD header for an MLX send and build a data segment for it */
1393 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1394 int ind, struct ib_send_wr *wr,
1395 struct mthca_mlx_seg *mlx,
1396 struct mthca_data_seg *data)
1402 ib_ud_header_init(256, /* assume a MAD */
1403 sqp->ud_header.grh_present,
1406 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1409 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1410 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1411 (sqp->ud_header.lrh.destination_lid ==
1412 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1413 (sqp->ud_header.lrh.service_level << 8));
1414 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1417 switch (wr->opcode) {
1419 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1420 sqp->ud_header.immediate_present = 0;
1422 case IB_WR_SEND_WITH_IMM:
1423 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1424 sqp->ud_header.immediate_present = 1;
1425 sqp->ud_header.immediate_data = wr->imm_data;
1431 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1432 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1433 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1434 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1435 if (!sqp->qp.ibqp.qp_num)
1436 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1437 sqp->pkey_index, &pkey);
1439 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1440 wr->wr.ud.pkey_index, &pkey);
1441 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1442 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1443 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1444 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1445 sqp->qkey : wr->wr.ud.remote_qkey);
1446 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1448 header_size = ib_ud_header_pack(&sqp->ud_header,
1450 ind * MTHCA_UD_HEADER_SIZE);
1452 data->byte_count = cpu_to_be32(header_size);
1453 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1454 data->addr = cpu_to_be64(sqp->header_dma +
1455 ind * MTHCA_UD_HEADER_SIZE);
1460 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1461 struct ib_cq *ib_cq)
1464 struct mthca_cq *cq;
1466 cur = wq->head - wq->tail;
1467 if (likely(cur + nreq < wq->max))
1471 spin_lock(&cq->lock);
1472 cur = wq->head - wq->tail;
1473 spin_unlock(&cq->lock);
1475 return cur + nreq >= wq->max;
1478 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1479 struct ib_send_wr **bad_wr)
1481 struct mthca_dev *dev = to_mdev(ibqp->device);
1482 struct mthca_qp *qp = to_mqp(ibqp);
1485 unsigned long flags;
1495 spin_lock_irqsave(&qp->sq.lock, flags);
1497 /* XXX check that state is OK to post send */
1499 ind = qp->sq.next_ind;
1501 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1502 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1503 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1504 " %d max, %d nreq)\n", qp->qpn,
1505 qp->sq.head, qp->sq.tail,
1512 wqe = get_send_wqe(qp, ind);
1513 prev_wqe = qp->sq.last;
1516 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1517 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1518 ((struct mthca_next_seg *) wqe)->flags =
1519 ((wr->send_flags & IB_SEND_SIGNALED) ?
1520 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1521 ((wr->send_flags & IB_SEND_SOLICITED) ?
1522 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1524 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1525 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1526 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1528 wqe += sizeof (struct mthca_next_seg);
1529 size = sizeof (struct mthca_next_seg) / 16;
1531 switch (qp->transport) {
1533 switch (wr->opcode) {
1534 case IB_WR_ATOMIC_CMP_AND_SWP:
1535 case IB_WR_ATOMIC_FETCH_AND_ADD:
1536 ((struct mthca_raddr_seg *) wqe)->raddr =
1537 cpu_to_be64(wr->wr.atomic.remote_addr);
1538 ((struct mthca_raddr_seg *) wqe)->rkey =
1539 cpu_to_be32(wr->wr.atomic.rkey);
1540 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1542 wqe += sizeof (struct mthca_raddr_seg);
1544 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1545 ((struct mthca_atomic_seg *) wqe)->swap_add =
1546 cpu_to_be64(wr->wr.atomic.swap);
1547 ((struct mthca_atomic_seg *) wqe)->compare =
1548 cpu_to_be64(wr->wr.atomic.compare_add);
1550 ((struct mthca_atomic_seg *) wqe)->swap_add =
1551 cpu_to_be64(wr->wr.atomic.compare_add);
1552 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1555 wqe += sizeof (struct mthca_atomic_seg);
1556 size += (sizeof (struct mthca_raddr_seg) +
1557 sizeof (struct mthca_atomic_seg)) / 16;
1560 case IB_WR_RDMA_WRITE:
1561 case IB_WR_RDMA_WRITE_WITH_IMM:
1562 case IB_WR_RDMA_READ:
1563 ((struct mthca_raddr_seg *) wqe)->raddr =
1564 cpu_to_be64(wr->wr.rdma.remote_addr);
1565 ((struct mthca_raddr_seg *) wqe)->rkey =
1566 cpu_to_be32(wr->wr.rdma.rkey);
1567 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1568 wqe += sizeof (struct mthca_raddr_seg);
1569 size += sizeof (struct mthca_raddr_seg) / 16;
1573 /* No extra segments required for sends */
1580 switch (wr->opcode) {
1581 case IB_WR_RDMA_WRITE:
1582 case IB_WR_RDMA_WRITE_WITH_IMM:
1583 ((struct mthca_raddr_seg *) wqe)->raddr =
1584 cpu_to_be64(wr->wr.rdma.remote_addr);
1585 ((struct mthca_raddr_seg *) wqe)->rkey =
1586 cpu_to_be32(wr->wr.rdma.rkey);
1587 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1588 wqe += sizeof (struct mthca_raddr_seg);
1589 size += sizeof (struct mthca_raddr_seg) / 16;
1593 /* No extra segments required for sends */
1600 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1601 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1602 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1603 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1604 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1605 cpu_to_be32(wr->wr.ud.remote_qpn);
1606 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1607 cpu_to_be32(wr->wr.ud.remote_qkey);
1609 wqe += sizeof (struct mthca_tavor_ud_seg);
1610 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1614 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1615 wqe - sizeof (struct mthca_next_seg),
1621 wqe += sizeof (struct mthca_data_seg);
1622 size += sizeof (struct mthca_data_seg) / 16;
1626 if (wr->num_sge > qp->sq.max_gs) {
1627 mthca_err(dev, "too many gathers\n");
1633 for (i = 0; i < wr->num_sge; ++i) {
1634 ((struct mthca_data_seg *) wqe)->byte_count =
1635 cpu_to_be32(wr->sg_list[i].length);
1636 ((struct mthca_data_seg *) wqe)->lkey =
1637 cpu_to_be32(wr->sg_list[i].lkey);
1638 ((struct mthca_data_seg *) wqe)->addr =
1639 cpu_to_be64(wr->sg_list[i].addr);
1640 wqe += sizeof (struct mthca_data_seg);
1641 size += sizeof (struct mthca_data_seg) / 16;
1644 /* Add one more inline data segment for ICRC */
1645 if (qp->transport == MLX) {
1646 ((struct mthca_data_seg *) wqe)->byte_count =
1647 cpu_to_be32((1 << 31) | 4);
1648 ((u32 *) wqe)[1] = 0;
1649 wqe += sizeof (struct mthca_data_seg);
1650 size += sizeof (struct mthca_data_seg) / 16;
1653 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1655 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1656 mthca_err(dev, "opcode invalid\n");
1662 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1663 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1664 qp->send_wqe_offset) |
1665 mthca_opcode[wr->opcode]);
1667 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1668 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1672 op0 = mthca_opcode[wr->opcode];
1676 if (unlikely(ind >= qp->sq.max))
1684 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1685 qp->send_wqe_offset) | f0 | op0);
1686 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1690 mthca_write64(doorbell,
1691 dev->kar + MTHCA_SEND_DOORBELL,
1692 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1695 qp->sq.next_ind = ind;
1696 qp->sq.head += nreq;
1698 spin_unlock_irqrestore(&qp->sq.lock, flags);
1702 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1703 struct ib_recv_wr **bad_wr)
1705 struct mthca_dev *dev = to_mdev(ibqp->device);
1706 struct mthca_qp *qp = to_mqp(ibqp);
1708 unsigned long flags;
1718 spin_lock_irqsave(&qp->rq.lock, flags);
1720 /* XXX check that state is OK to post receive */
1722 ind = qp->rq.next_ind;
1724 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1725 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1728 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1729 doorbell[1] = cpu_to_be32(qp->qpn << 8);
1733 mthca_write64(doorbell,
1734 dev->kar + MTHCA_RECEIVE_DOORBELL,
1735 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1737 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1741 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1742 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1743 " %d max, %d nreq)\n", qp->qpn,
1744 qp->rq.head, qp->rq.tail,
1751 wqe = get_recv_wqe(qp, ind);
1752 prev_wqe = qp->rq.last;
1755 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1756 ((struct mthca_next_seg *) wqe)->ee_nds =
1757 cpu_to_be32(MTHCA_NEXT_DBD);
1758 ((struct mthca_next_seg *) wqe)->flags = 0;
1760 wqe += sizeof (struct mthca_next_seg);
1761 size = sizeof (struct mthca_next_seg) / 16;
1763 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1769 for (i = 0; i < wr->num_sge; ++i) {
1770 ((struct mthca_data_seg *) wqe)->byte_count =
1771 cpu_to_be32(wr->sg_list[i].length);
1772 ((struct mthca_data_seg *) wqe)->lkey =
1773 cpu_to_be32(wr->sg_list[i].lkey);
1774 ((struct mthca_data_seg *) wqe)->addr =
1775 cpu_to_be64(wr->sg_list[i].addr);
1776 wqe += sizeof (struct mthca_data_seg);
1777 size += sizeof (struct mthca_data_seg) / 16;
1780 qp->wrid[ind] = wr->wr_id;
1782 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1783 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1785 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1786 cpu_to_be32(MTHCA_NEXT_DBD | size);
1792 if (unlikely(ind >= qp->rq.max))
1798 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1799 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1803 mthca_write64(doorbell,
1804 dev->kar + MTHCA_RECEIVE_DOORBELL,
1805 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1808 qp->rq.next_ind = ind;
1809 qp->rq.head += nreq;
1811 spin_unlock_irqrestore(&qp->rq.lock, flags);
1815 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1816 struct ib_send_wr **bad_wr)
1818 struct mthca_dev *dev = to_mdev(ibqp->device);
1819 struct mthca_qp *qp = to_mqp(ibqp);
1822 unsigned long flags;
1832 spin_lock_irqsave(&qp->sq.lock, flags);
1834 /* XXX check that state is OK to post send */
1836 ind = qp->sq.head & (qp->sq.max - 1);
1838 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1839 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1840 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1841 " %d max, %d nreq)\n", qp->qpn,
1842 qp->sq.head, qp->sq.tail,
1849 wqe = get_send_wqe(qp, ind);
1850 prev_wqe = qp->sq.last;
1853 ((struct mthca_next_seg *) wqe)->flags =
1854 ((wr->send_flags & IB_SEND_SIGNALED) ?
1855 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1856 ((wr->send_flags & IB_SEND_SOLICITED) ?
1857 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1859 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1860 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1861 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1863 wqe += sizeof (struct mthca_next_seg);
1864 size = sizeof (struct mthca_next_seg) / 16;
1866 switch (qp->transport) {
1868 switch (wr->opcode) {
1869 case IB_WR_ATOMIC_CMP_AND_SWP:
1870 case IB_WR_ATOMIC_FETCH_AND_ADD:
1871 ((struct mthca_raddr_seg *) wqe)->raddr =
1872 cpu_to_be64(wr->wr.atomic.remote_addr);
1873 ((struct mthca_raddr_seg *) wqe)->rkey =
1874 cpu_to_be32(wr->wr.atomic.rkey);
1875 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1877 wqe += sizeof (struct mthca_raddr_seg);
1879 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1880 ((struct mthca_atomic_seg *) wqe)->swap_add =
1881 cpu_to_be64(wr->wr.atomic.swap);
1882 ((struct mthca_atomic_seg *) wqe)->compare =
1883 cpu_to_be64(wr->wr.atomic.compare_add);
1885 ((struct mthca_atomic_seg *) wqe)->swap_add =
1886 cpu_to_be64(wr->wr.atomic.compare_add);
1887 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1890 wqe += sizeof (struct mthca_atomic_seg);
1891 size += (sizeof (struct mthca_raddr_seg) +
1892 sizeof (struct mthca_atomic_seg)) / 16;
1895 case IB_WR_RDMA_READ:
1896 case IB_WR_RDMA_WRITE:
1897 case IB_WR_RDMA_WRITE_WITH_IMM:
1898 ((struct mthca_raddr_seg *) wqe)->raddr =
1899 cpu_to_be64(wr->wr.rdma.remote_addr);
1900 ((struct mthca_raddr_seg *) wqe)->rkey =
1901 cpu_to_be32(wr->wr.rdma.rkey);
1902 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1903 wqe += sizeof (struct mthca_raddr_seg);
1904 size += sizeof (struct mthca_raddr_seg) / 16;
1908 /* No extra segments required for sends */
1915 switch (wr->opcode) {
1916 case IB_WR_RDMA_WRITE:
1917 case IB_WR_RDMA_WRITE_WITH_IMM:
1918 ((struct mthca_raddr_seg *) wqe)->raddr =
1919 cpu_to_be64(wr->wr.rdma.remote_addr);
1920 ((struct mthca_raddr_seg *) wqe)->rkey =
1921 cpu_to_be32(wr->wr.rdma.rkey);
1922 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1923 wqe += sizeof (struct mthca_raddr_seg);
1924 size += sizeof (struct mthca_raddr_seg) / 16;
1928 /* No extra segments required for sends */
1935 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1936 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1937 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1938 cpu_to_be32(wr->wr.ud.remote_qpn);
1939 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1940 cpu_to_be32(wr->wr.ud.remote_qkey);
1942 wqe += sizeof (struct mthca_arbel_ud_seg);
1943 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1947 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1948 wqe - sizeof (struct mthca_next_seg),
1954 wqe += sizeof (struct mthca_data_seg);
1955 size += sizeof (struct mthca_data_seg) / 16;
1959 if (wr->num_sge > qp->sq.max_gs) {
1960 mthca_err(dev, "too many gathers\n");
1966 for (i = 0; i < wr->num_sge; ++i) {
1967 ((struct mthca_data_seg *) wqe)->byte_count =
1968 cpu_to_be32(wr->sg_list[i].length);
1969 ((struct mthca_data_seg *) wqe)->lkey =
1970 cpu_to_be32(wr->sg_list[i].lkey);
1971 ((struct mthca_data_seg *) wqe)->addr =
1972 cpu_to_be64(wr->sg_list[i].addr);
1973 wqe += sizeof (struct mthca_data_seg);
1974 size += sizeof (struct mthca_data_seg) / 16;
1977 /* Add one more inline data segment for ICRC */
1978 if (qp->transport == MLX) {
1979 ((struct mthca_data_seg *) wqe)->byte_count =
1980 cpu_to_be32((1 << 31) | 4);
1981 ((u32 *) wqe)[1] = 0;
1982 wqe += sizeof (struct mthca_data_seg);
1983 size += sizeof (struct mthca_data_seg) / 16;
1986 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1988 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1989 mthca_err(dev, "opcode invalid\n");
1995 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1996 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1997 qp->send_wqe_offset) |
1998 mthca_opcode[wr->opcode]);
2000 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2001 cpu_to_be32(MTHCA_NEXT_DBD | size);
2005 op0 = mthca_opcode[wr->opcode];
2009 if (unlikely(ind >= qp->sq.max))
2017 doorbell[0] = cpu_to_be32((nreq << 24) |
2018 ((qp->sq.head & 0xffff) << 8) |
2020 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2022 qp->sq.head += nreq;
2025 * Make sure that descriptors are written before
2029 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2032 * Make sure doorbell record is written before we
2033 * write MMIO send doorbell.
2036 mthca_write64(doorbell,
2037 dev->kar + MTHCA_SEND_DOORBELL,
2038 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2041 spin_unlock_irqrestore(&qp->sq.lock, flags);
2045 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2046 struct ib_recv_wr **bad_wr)
2048 struct mthca_dev *dev = to_mdev(ibqp->device);
2049 struct mthca_qp *qp = to_mqp(ibqp);
2050 unsigned long flags;
2057 spin_lock_irqsave(&qp->rq.lock, flags);
2059 /* XXX check that state is OK to post receive */
2061 ind = qp->rq.head & (qp->rq.max - 1);
2063 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2064 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2065 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2066 " %d max, %d nreq)\n", qp->qpn,
2067 qp->rq.head, qp->rq.tail,
2074 wqe = get_recv_wqe(qp, ind);
2076 ((struct mthca_next_seg *) wqe)->flags = 0;
2078 wqe += sizeof (struct mthca_next_seg);
2080 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2086 for (i = 0; i < wr->num_sge; ++i) {
2087 ((struct mthca_data_seg *) wqe)->byte_count =
2088 cpu_to_be32(wr->sg_list[i].length);
2089 ((struct mthca_data_seg *) wqe)->lkey =
2090 cpu_to_be32(wr->sg_list[i].lkey);
2091 ((struct mthca_data_seg *) wqe)->addr =
2092 cpu_to_be64(wr->sg_list[i].addr);
2093 wqe += sizeof (struct mthca_data_seg);
2096 if (i < qp->rq.max_gs) {
2097 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2098 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2099 ((struct mthca_data_seg *) wqe)->addr = 0;
2102 qp->wrid[ind] = wr->wr_id;
2105 if (unlikely(ind >= qp->rq.max))
2110 qp->rq.head += nreq;
2113 * Make sure that descriptors are written before
2117 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2120 spin_unlock_irqrestore(&qp->rq.lock, flags);
2124 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2125 int index, int *dbd, __be32 *new_wqe)
2127 struct mthca_next_seg *next;
2130 * For SRQs, all WQEs generate a CQE, so we're always at the
2131 * end of the doorbell chain.
2139 next = get_send_wqe(qp, index);
2141 next = get_recv_wqe(qp, index);
2143 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2144 if (next->ee_nds & cpu_to_be32(0x3f))
2145 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2146 (next->ee_nds & cpu_to_be32(0x3f));
2153 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2159 spin_lock_init(&dev->qp_table.lock);
2162 * We reserve 2 extra QPs per port for the special QPs. The
2163 * special QP for port 1 has to be even, so round up.
2165 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2166 err = mthca_alloc_init(&dev->qp_table.alloc,
2167 dev->limits.num_qps,
2169 dev->qp_table.sqp_start +
2170 MTHCA_MAX_PORTS * 2);
2174 err = mthca_array_init(&dev->qp_table.qp,
2175 dev->limits.num_qps);
2177 mthca_alloc_cleanup(&dev->qp_table.alloc);
2181 for (i = 0; i < 2; ++i) {
2182 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2183 dev->qp_table.sqp_start + i * 2,
2188 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2189 "status %02x, aborting.\n",
2198 for (i = 0; i < 2; ++i)
2199 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2201 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2202 mthca_alloc_cleanup(&dev->qp_table.alloc);
2207 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2212 for (i = 0; i < 2; ++i)
2213 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2215 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2216 mthca_alloc_cleanup(&dev->qp_table.alloc);