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 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
595 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
596 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
597 attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
601 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
602 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
603 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
604 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
608 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
610 return PTR_ERR(mailbox);
611 qp_param = mailbox->buf;
612 qp_context = &qp_param->context;
613 memset(qp_param, 0, sizeof *qp_param);
615 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
616 (to_mthca_st(qp->transport) << 16));
617 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
618 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
619 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
621 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
622 switch (attr->path_mig_state) {
623 case IB_MIG_MIGRATED:
624 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
627 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
630 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
635 /* leave tavor_sched_queue as 0 */
637 if (qp->transport == MLX || qp->transport == UD)
638 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
639 else if (attr_mask & IB_QP_PATH_MTU)
640 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
642 if (mthca_is_memfree(dev)) {
644 qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
645 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
648 qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
649 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
652 /* leave arbel_sched_queue as 0 */
654 if (qp->ibqp.uobject)
655 qp_context->usr_page =
656 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
658 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
659 qp_context->local_qpn = cpu_to_be32(qp->qpn);
660 if (attr_mask & IB_QP_DEST_QPN) {
661 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
664 if (qp->transport == MLX)
665 qp_context->pri_path.port_pkey |=
666 cpu_to_be32(to_msqp(qp)->port << 24);
668 if (attr_mask & IB_QP_PORT) {
669 qp_context->pri_path.port_pkey |=
670 cpu_to_be32(attr->port_num << 24);
671 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
675 if (attr_mask & IB_QP_PKEY_INDEX) {
676 qp_context->pri_path.port_pkey |=
677 cpu_to_be32(attr->pkey_index);
678 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
681 if (attr_mask & IB_QP_RNR_RETRY) {
682 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
683 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
686 if (attr_mask & IB_QP_AV) {
687 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
688 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
689 qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
690 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
691 qp_context->pri_path.g_mylmc |= 1 << 7;
692 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
693 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
694 qp_context->pri_path.sl_tclass_flowlabel =
695 cpu_to_be32((attr->ah_attr.sl << 28) |
696 (attr->ah_attr.grh.traffic_class << 20) |
697 (attr->ah_attr.grh.flow_label));
698 memcpy(qp_context->pri_path.rgid,
699 attr->ah_attr.grh.dgid.raw, 16);
701 qp_context->pri_path.sl_tclass_flowlabel =
702 cpu_to_be32(attr->ah_attr.sl << 28);
704 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
707 if (attr_mask & IB_QP_TIMEOUT) {
708 qp_context->pri_path.ackto = attr->timeout << 3;
709 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
715 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
716 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
717 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
718 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
719 (MTHCA_FLIGHT_LIMIT << 24) |
723 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
724 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
725 if (attr_mask & IB_QP_RETRY_CNT) {
726 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
727 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
730 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
731 qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
732 ffs(attr->max_rd_atomic) - 1 : 0,
734 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
737 if (attr_mask & IB_QP_SQ_PSN)
738 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
739 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
741 if (mthca_is_memfree(dev)) {
742 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
743 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
746 if (attr_mask & IB_QP_ACCESS_FLAGS) {
747 qp_context->params2 |=
748 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
749 MTHCA_QP_BIT_RWE : 0);
752 * Only enable RDMA reads and atomics if we have
753 * responder resources set to a non-zero value.
755 if (qp->resp_depth) {
756 qp_context->params2 |=
757 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
758 MTHCA_QP_BIT_RRE : 0);
759 qp_context->params2 |=
760 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
761 MTHCA_QP_BIT_RAE : 0);
764 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
765 MTHCA_QP_OPTPAR_RRE |
766 MTHCA_QP_OPTPAR_RAE);
768 qp->atomic_rd_en = attr->qp_access_flags;
771 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
774 if (qp->resp_depth && !attr->max_dest_rd_atomic) {
776 * Lowering our responder resources to zero.
777 * Turn off reads RDMA and atomics as responder.
778 * (RRE/RAE in params2 already zero)
780 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
781 MTHCA_QP_OPTPAR_RAE);
784 if (!qp->resp_depth && attr->max_dest_rd_atomic) {
786 * Increasing our responder resources from
787 * zero. Turn on RDMA reads and atomics as
790 qp_context->params2 |=
791 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
792 MTHCA_QP_BIT_RRE : 0);
793 qp_context->params2 |=
794 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
795 MTHCA_QP_BIT_RAE : 0);
797 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRE |
798 MTHCA_QP_OPTPAR_RAE);
802 1 << rra_max < attr->max_dest_rd_atomic &&
803 rra_max < dev->qp_table.rdb_shift;
807 qp_context->params2 |= cpu_to_be32(rra_max << 21);
808 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
810 qp->resp_depth = attr->max_dest_rd_atomic;
813 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
816 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
818 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
819 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
820 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
822 if (attr_mask & IB_QP_RQ_PSN)
823 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
825 qp_context->ra_buff_indx =
826 cpu_to_be32(dev->qp_table.rdb_base +
827 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
828 dev->qp_table.rdb_shift));
830 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
832 if (mthca_is_memfree(dev))
833 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
835 if (attr_mask & IB_QP_QKEY) {
836 qp_context->qkey = cpu_to_be32(attr->qkey);
837 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
841 qp_context->srqn = cpu_to_be32(1 << 24 |
842 to_msrq(ibqp->srq)->srqn);
844 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
845 qp->qpn, 0, mailbox, 0, &status);
847 mthca_warn(dev, "modify QP %d returned status %02x.\n",
848 state_table[cur_state][new_state].trans, status);
853 qp->state = new_state;
855 mthca_free_mailbox(dev, mailbox);
858 store_attrs(to_msqp(qp), attr, attr_mask);
861 * If we moved QP0 to RTR, bring the IB link up; if we moved
862 * QP0 to RESET or ERROR, bring the link back down.
864 if (is_qp0(dev, qp)) {
865 if (cur_state != IB_QPS_RTR &&
866 new_state == IB_QPS_RTR)
867 init_port(dev, to_msqp(qp)->port);
869 if (cur_state != IB_QPS_RESET &&
870 cur_state != IB_QPS_ERR &&
871 (new_state == IB_QPS_RESET ||
872 new_state == IB_QPS_ERR))
873 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
877 * If we moved a kernel QP to RESET, clean up all old CQ
878 * entries and reinitialize the QP.
880 if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
881 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
882 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
883 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
884 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
885 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
887 mthca_wq_init(&qp->sq);
888 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
890 mthca_wq_init(&qp->rq);
891 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
893 if (mthca_is_memfree(dev)) {
902 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
909 * Calculate the maximum size of WQE s/g segments, excluding
910 * the next segment and other non-data segments.
912 max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) -
913 sizeof (struct mthca_next_seg);
915 switch (qp->transport) {
917 max_data_size -= 2 * sizeof (struct mthca_data_seg);
921 if (mthca_is_memfree(dev))
922 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
924 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
928 max_data_size -= sizeof (struct mthca_raddr_seg);
932 /* We don't support inline data for kernel QPs (yet). */
933 if (!pd->ibpd.uobject)
934 qp->max_inline_data = 0;
936 qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE;
938 qp->sq.max_gs = min_t(int, dev->limits.max_sg,
939 max_data_size / sizeof (struct mthca_data_seg));
940 qp->rq.max_gs = min_t(int, dev->limits.max_sg,
941 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
942 sizeof (struct mthca_next_seg)) /
943 sizeof (struct mthca_data_seg));
947 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
948 * rq.max_gs and sq.max_gs must all be assigned.
949 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
950 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
953 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
960 size = sizeof (struct mthca_next_seg) +
961 qp->rq.max_gs * sizeof (struct mthca_data_seg);
963 if (size > dev->limits.max_desc_sz)
966 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
970 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
971 switch (qp->transport) {
973 size += 2 * sizeof (struct mthca_data_seg);
977 size += mthca_is_memfree(dev) ?
978 sizeof (struct mthca_arbel_ud_seg) :
979 sizeof (struct mthca_tavor_ud_seg);
983 size += sizeof (struct mthca_raddr_seg);
987 size += sizeof (struct mthca_raddr_seg);
989 * An atomic op will require an atomic segment, a
990 * remote address segment and one scatter entry.
992 size = max_t(int, size,
993 sizeof (struct mthca_atomic_seg) +
994 sizeof (struct mthca_raddr_seg) +
995 sizeof (struct mthca_data_seg));
1002 /* Make sure that we have enough space for a bind request */
1003 size = max_t(int, size, sizeof (struct mthca_bind_seg));
1005 size += sizeof (struct mthca_next_seg);
1007 if (size > dev->limits.max_desc_sz)
1010 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1014 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1015 1 << qp->sq.wqe_shift);
1018 * If this is a userspace QP, we don't actually have to
1019 * allocate anything. All we need is to calculate the WQE
1020 * sizes and the send_wqe_offset, so we're done now.
1022 if (pd->ibpd.uobject)
1025 size = PAGE_ALIGN(qp->send_wqe_offset +
1026 (qp->sq.max << qp->sq.wqe_shift));
1028 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1033 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1034 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1045 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1046 struct mthca_qp *qp)
1048 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1049 (qp->sq.max << qp->sq.wqe_shift)),
1050 &qp->queue, qp->is_direct, &qp->mr);
1054 static int mthca_map_memfree(struct mthca_dev *dev,
1055 struct mthca_qp *qp)
1059 if (mthca_is_memfree(dev)) {
1060 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1064 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1068 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1069 qp->qpn << dev->qp_table.rdb_shift);
1078 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1081 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1086 static void mthca_unmap_memfree(struct mthca_dev *dev,
1087 struct mthca_qp *qp)
1089 mthca_table_put(dev, dev->qp_table.rdb_table,
1090 qp->qpn << dev->qp_table.rdb_shift);
1091 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1092 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1095 static int mthca_alloc_memfree(struct mthca_dev *dev,
1096 struct mthca_qp *qp)
1100 if (mthca_is_memfree(dev)) {
1101 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1102 qp->qpn, &qp->rq.db);
1103 if (qp->rq.db_index < 0)
1106 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1107 qp->qpn, &qp->sq.db);
1108 if (qp->sq.db_index < 0)
1109 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1115 static void mthca_free_memfree(struct mthca_dev *dev,
1116 struct mthca_qp *qp)
1118 if (mthca_is_memfree(dev)) {
1119 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1120 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1124 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1125 struct mthca_pd *pd,
1126 struct mthca_cq *send_cq,
1127 struct mthca_cq *recv_cq,
1128 enum ib_sig_type send_policy,
1129 struct mthca_qp *qp)
1134 atomic_set(&qp->refcount, 1);
1135 init_waitqueue_head(&qp->wait);
1136 qp->state = IB_QPS_RESET;
1137 qp->atomic_rd_en = 0;
1139 qp->sq_policy = send_policy;
1140 mthca_wq_init(&qp->sq);
1141 mthca_wq_init(&qp->rq);
1143 ret = mthca_map_memfree(dev, qp);
1147 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1149 mthca_unmap_memfree(dev, qp);
1153 mthca_adjust_qp_caps(dev, pd, qp);
1156 * If this is a userspace QP, we're done now. The doorbells
1157 * will be allocated and buffers will be initialized in
1160 if (pd->ibpd.uobject)
1163 ret = mthca_alloc_memfree(dev, qp);
1165 mthca_free_wqe_buf(dev, qp);
1166 mthca_unmap_memfree(dev, qp);
1170 if (mthca_is_memfree(dev)) {
1171 struct mthca_next_seg *next;
1172 struct mthca_data_seg *scatter;
1173 int size = (sizeof (struct mthca_next_seg) +
1174 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1176 for (i = 0; i < qp->rq.max; ++i) {
1177 next = get_recv_wqe(qp, i);
1178 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1180 next->ee_nds = cpu_to_be32(size);
1182 for (scatter = (void *) (next + 1);
1183 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1185 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1188 for (i = 0; i < qp->sq.max; ++i) {
1189 next = get_send_wqe(qp, i);
1190 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1192 qp->send_wqe_offset);
1196 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1197 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1202 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1203 struct mthca_qp *qp)
1205 /* Sanity check QP size before proceeding */
1206 if (cap->max_send_wr > dev->limits.max_wqes ||
1207 cap->max_recv_wr > dev->limits.max_wqes ||
1208 cap->max_send_sge > dev->limits.max_sg ||
1209 cap->max_recv_sge > dev->limits.max_sg)
1212 if (mthca_is_memfree(dev)) {
1213 qp->rq.max = cap->max_recv_wr ?
1214 roundup_pow_of_two(cap->max_recv_wr) : 0;
1215 qp->sq.max = cap->max_send_wr ?
1216 roundup_pow_of_two(cap->max_send_wr) : 0;
1218 qp->rq.max = cap->max_recv_wr;
1219 qp->sq.max = cap->max_send_wr;
1222 qp->rq.max_gs = cap->max_recv_sge;
1223 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1224 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1225 MTHCA_INLINE_CHUNK_SIZE) /
1226 sizeof (struct mthca_data_seg));
1229 * For MLX transport we need 2 extra S/G entries:
1230 * one for the header and one for the checksum at the end
1232 if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1233 qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1239 int mthca_alloc_qp(struct mthca_dev *dev,
1240 struct mthca_pd *pd,
1241 struct mthca_cq *send_cq,
1242 struct mthca_cq *recv_cq,
1243 enum ib_qp_type type,
1244 enum ib_sig_type send_policy,
1245 struct ib_qp_cap *cap,
1246 struct mthca_qp *qp)
1250 err = mthca_set_qp_size(dev, cap, qp);
1255 case IB_QPT_RC: qp->transport = RC; break;
1256 case IB_QPT_UC: qp->transport = UC; break;
1257 case IB_QPT_UD: qp->transport = UD; break;
1258 default: return -EINVAL;
1261 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1265 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1268 mthca_free(&dev->qp_table.alloc, qp->qpn);
1272 spin_lock_irq(&dev->qp_table.lock);
1273 mthca_array_set(&dev->qp_table.qp,
1274 qp->qpn & (dev->limits.num_qps - 1), qp);
1275 spin_unlock_irq(&dev->qp_table.lock);
1280 int mthca_alloc_sqp(struct mthca_dev *dev,
1281 struct mthca_pd *pd,
1282 struct mthca_cq *send_cq,
1283 struct mthca_cq *recv_cq,
1284 enum ib_sig_type send_policy,
1285 struct ib_qp_cap *cap,
1288 struct mthca_sqp *sqp)
1290 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1293 err = mthca_set_qp_size(dev, cap, &sqp->qp);
1297 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1298 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1299 &sqp->header_dma, GFP_KERNEL);
1300 if (!sqp->header_buf)
1303 spin_lock_irq(&dev->qp_table.lock);
1304 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1307 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1308 spin_unlock_irq(&dev->qp_table.lock);
1315 sqp->qp.transport = MLX;
1317 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1318 send_policy, &sqp->qp);
1322 atomic_inc(&pd->sqp_count);
1328 * Lock CQs here, so that CQ polling code can do QP lookup
1329 * without taking a lock.
1331 spin_lock_irq(&send_cq->lock);
1332 if (send_cq != recv_cq)
1333 spin_lock(&recv_cq->lock);
1335 spin_lock(&dev->qp_table.lock);
1336 mthca_array_clear(&dev->qp_table.qp, mqpn);
1337 spin_unlock(&dev->qp_table.lock);
1339 if (send_cq != recv_cq)
1340 spin_unlock(&recv_cq->lock);
1341 spin_unlock_irq(&send_cq->lock);
1344 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1345 sqp->header_buf, sqp->header_dma);
1350 void mthca_free_qp(struct mthca_dev *dev,
1351 struct mthca_qp *qp)
1354 struct mthca_cq *send_cq;
1355 struct mthca_cq *recv_cq;
1357 send_cq = to_mcq(qp->ibqp.send_cq);
1358 recv_cq = to_mcq(qp->ibqp.recv_cq);
1361 * Lock CQs here, so that CQ polling code can do QP lookup
1362 * without taking a lock.
1364 spin_lock_irq(&send_cq->lock);
1365 if (send_cq != recv_cq)
1366 spin_lock(&recv_cq->lock);
1368 spin_lock(&dev->qp_table.lock);
1369 mthca_array_clear(&dev->qp_table.qp,
1370 qp->qpn & (dev->limits.num_qps - 1));
1371 spin_unlock(&dev->qp_table.lock);
1373 if (send_cq != recv_cq)
1374 spin_unlock(&recv_cq->lock);
1375 spin_unlock_irq(&send_cq->lock);
1377 atomic_dec(&qp->refcount);
1378 wait_event(qp->wait, !atomic_read(&qp->refcount));
1380 if (qp->state != IB_QPS_RESET)
1381 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1384 * If this is a userspace QP, the buffers, MR, CQs and so on
1385 * will be cleaned up in userspace, so all we have to do is
1386 * unref the mem-free tables and free the QPN in our table.
1388 if (!qp->ibqp.uobject) {
1389 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1390 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1391 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1392 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1393 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1395 mthca_free_memfree(dev, qp);
1396 mthca_free_wqe_buf(dev, qp);
1399 mthca_unmap_memfree(dev, qp);
1401 if (is_sqp(dev, qp)) {
1402 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1403 dma_free_coherent(&dev->pdev->dev,
1404 to_msqp(qp)->header_buf_size,
1405 to_msqp(qp)->header_buf,
1406 to_msqp(qp)->header_dma);
1408 mthca_free(&dev->qp_table.alloc, qp->qpn);
1411 /* Create UD header for an MLX send and build a data segment for it */
1412 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1413 int ind, struct ib_send_wr *wr,
1414 struct mthca_mlx_seg *mlx,
1415 struct mthca_data_seg *data)
1421 ib_ud_header_init(256, /* assume a MAD */
1422 sqp->ud_header.grh_present,
1425 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1428 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1429 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1430 (sqp->ud_header.lrh.destination_lid ==
1431 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1432 (sqp->ud_header.lrh.service_level << 8));
1433 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1436 switch (wr->opcode) {
1438 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1439 sqp->ud_header.immediate_present = 0;
1441 case IB_WR_SEND_WITH_IMM:
1442 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1443 sqp->ud_header.immediate_present = 1;
1444 sqp->ud_header.immediate_data = wr->imm_data;
1450 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1451 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1452 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1453 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1454 if (!sqp->qp.ibqp.qp_num)
1455 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1456 sqp->pkey_index, &pkey);
1458 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1459 wr->wr.ud.pkey_index, &pkey);
1460 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1461 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1462 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1463 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1464 sqp->qkey : wr->wr.ud.remote_qkey);
1465 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1467 header_size = ib_ud_header_pack(&sqp->ud_header,
1469 ind * MTHCA_UD_HEADER_SIZE);
1471 data->byte_count = cpu_to_be32(header_size);
1472 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1473 data->addr = cpu_to_be64(sqp->header_dma +
1474 ind * MTHCA_UD_HEADER_SIZE);
1479 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1480 struct ib_cq *ib_cq)
1483 struct mthca_cq *cq;
1485 cur = wq->head - wq->tail;
1486 if (likely(cur + nreq < wq->max))
1490 spin_lock(&cq->lock);
1491 cur = wq->head - wq->tail;
1492 spin_unlock(&cq->lock);
1494 return cur + nreq >= wq->max;
1497 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1498 struct ib_send_wr **bad_wr)
1500 struct mthca_dev *dev = to_mdev(ibqp->device);
1501 struct mthca_qp *qp = to_mqp(ibqp);
1504 unsigned long flags;
1514 spin_lock_irqsave(&qp->sq.lock, flags);
1516 /* XXX check that state is OK to post send */
1518 ind = qp->sq.next_ind;
1520 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1521 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1522 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1523 " %d max, %d nreq)\n", qp->qpn,
1524 qp->sq.head, qp->sq.tail,
1531 wqe = get_send_wqe(qp, ind);
1532 prev_wqe = qp->sq.last;
1535 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1536 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1537 ((struct mthca_next_seg *) wqe)->flags =
1538 ((wr->send_flags & IB_SEND_SIGNALED) ?
1539 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1540 ((wr->send_flags & IB_SEND_SOLICITED) ?
1541 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1543 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1544 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1545 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1547 wqe += sizeof (struct mthca_next_seg);
1548 size = sizeof (struct mthca_next_seg) / 16;
1550 switch (qp->transport) {
1552 switch (wr->opcode) {
1553 case IB_WR_ATOMIC_CMP_AND_SWP:
1554 case IB_WR_ATOMIC_FETCH_AND_ADD:
1555 ((struct mthca_raddr_seg *) wqe)->raddr =
1556 cpu_to_be64(wr->wr.atomic.remote_addr);
1557 ((struct mthca_raddr_seg *) wqe)->rkey =
1558 cpu_to_be32(wr->wr.atomic.rkey);
1559 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1561 wqe += sizeof (struct mthca_raddr_seg);
1563 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1564 ((struct mthca_atomic_seg *) wqe)->swap_add =
1565 cpu_to_be64(wr->wr.atomic.swap);
1566 ((struct mthca_atomic_seg *) wqe)->compare =
1567 cpu_to_be64(wr->wr.atomic.compare_add);
1569 ((struct mthca_atomic_seg *) wqe)->swap_add =
1570 cpu_to_be64(wr->wr.atomic.compare_add);
1571 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1574 wqe += sizeof (struct mthca_atomic_seg);
1575 size += (sizeof (struct mthca_raddr_seg) +
1576 sizeof (struct mthca_atomic_seg)) / 16;
1579 case IB_WR_RDMA_WRITE:
1580 case IB_WR_RDMA_WRITE_WITH_IMM:
1581 case IB_WR_RDMA_READ:
1582 ((struct mthca_raddr_seg *) wqe)->raddr =
1583 cpu_to_be64(wr->wr.rdma.remote_addr);
1584 ((struct mthca_raddr_seg *) wqe)->rkey =
1585 cpu_to_be32(wr->wr.rdma.rkey);
1586 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1587 wqe += sizeof (struct mthca_raddr_seg);
1588 size += sizeof (struct mthca_raddr_seg) / 16;
1592 /* No extra segments required for sends */
1599 switch (wr->opcode) {
1600 case IB_WR_RDMA_WRITE:
1601 case IB_WR_RDMA_WRITE_WITH_IMM:
1602 ((struct mthca_raddr_seg *) wqe)->raddr =
1603 cpu_to_be64(wr->wr.rdma.remote_addr);
1604 ((struct mthca_raddr_seg *) wqe)->rkey =
1605 cpu_to_be32(wr->wr.rdma.rkey);
1606 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1607 wqe += sizeof (struct mthca_raddr_seg);
1608 size += sizeof (struct mthca_raddr_seg) / 16;
1612 /* No extra segments required for sends */
1619 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1620 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1621 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1622 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1623 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1624 cpu_to_be32(wr->wr.ud.remote_qpn);
1625 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1626 cpu_to_be32(wr->wr.ud.remote_qkey);
1628 wqe += sizeof (struct mthca_tavor_ud_seg);
1629 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1633 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1634 wqe - sizeof (struct mthca_next_seg),
1640 wqe += sizeof (struct mthca_data_seg);
1641 size += sizeof (struct mthca_data_seg) / 16;
1645 if (wr->num_sge > qp->sq.max_gs) {
1646 mthca_err(dev, "too many gathers\n");
1652 for (i = 0; i < wr->num_sge; ++i) {
1653 ((struct mthca_data_seg *) wqe)->byte_count =
1654 cpu_to_be32(wr->sg_list[i].length);
1655 ((struct mthca_data_seg *) wqe)->lkey =
1656 cpu_to_be32(wr->sg_list[i].lkey);
1657 ((struct mthca_data_seg *) wqe)->addr =
1658 cpu_to_be64(wr->sg_list[i].addr);
1659 wqe += sizeof (struct mthca_data_seg);
1660 size += sizeof (struct mthca_data_seg) / 16;
1663 /* Add one more inline data segment for ICRC */
1664 if (qp->transport == MLX) {
1665 ((struct mthca_data_seg *) wqe)->byte_count =
1666 cpu_to_be32((1 << 31) | 4);
1667 ((u32 *) wqe)[1] = 0;
1668 wqe += sizeof (struct mthca_data_seg);
1669 size += sizeof (struct mthca_data_seg) / 16;
1672 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1674 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1675 mthca_err(dev, "opcode invalid\n");
1681 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1682 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1683 qp->send_wqe_offset) |
1684 mthca_opcode[wr->opcode]);
1686 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1687 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1691 op0 = mthca_opcode[wr->opcode];
1695 if (unlikely(ind >= qp->sq.max))
1703 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1704 qp->send_wqe_offset) | f0 | op0);
1705 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1709 mthca_write64(doorbell,
1710 dev->kar + MTHCA_SEND_DOORBELL,
1711 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1714 qp->sq.next_ind = ind;
1715 qp->sq.head += nreq;
1717 spin_unlock_irqrestore(&qp->sq.lock, flags);
1721 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1722 struct ib_recv_wr **bad_wr)
1724 struct mthca_dev *dev = to_mdev(ibqp->device);
1725 struct mthca_qp *qp = to_mqp(ibqp);
1727 unsigned long flags;
1737 spin_lock_irqsave(&qp->rq.lock, flags);
1739 /* XXX check that state is OK to post receive */
1741 ind = qp->rq.next_ind;
1743 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1744 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1747 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1748 doorbell[1] = cpu_to_be32(qp->qpn << 8);
1752 mthca_write64(doorbell,
1753 dev->kar + MTHCA_RECEIVE_DOORBELL,
1754 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1756 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1760 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1761 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1762 " %d max, %d nreq)\n", qp->qpn,
1763 qp->rq.head, qp->rq.tail,
1770 wqe = get_recv_wqe(qp, ind);
1771 prev_wqe = qp->rq.last;
1774 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1775 ((struct mthca_next_seg *) wqe)->ee_nds =
1776 cpu_to_be32(MTHCA_NEXT_DBD);
1777 ((struct mthca_next_seg *) wqe)->flags = 0;
1779 wqe += sizeof (struct mthca_next_seg);
1780 size = sizeof (struct mthca_next_seg) / 16;
1782 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1788 for (i = 0; i < wr->num_sge; ++i) {
1789 ((struct mthca_data_seg *) wqe)->byte_count =
1790 cpu_to_be32(wr->sg_list[i].length);
1791 ((struct mthca_data_seg *) wqe)->lkey =
1792 cpu_to_be32(wr->sg_list[i].lkey);
1793 ((struct mthca_data_seg *) wqe)->addr =
1794 cpu_to_be64(wr->sg_list[i].addr);
1795 wqe += sizeof (struct mthca_data_seg);
1796 size += sizeof (struct mthca_data_seg) / 16;
1799 qp->wrid[ind] = wr->wr_id;
1801 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1802 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1804 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1805 cpu_to_be32(MTHCA_NEXT_DBD | size);
1811 if (unlikely(ind >= qp->rq.max))
1817 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1818 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1822 mthca_write64(doorbell,
1823 dev->kar + MTHCA_RECEIVE_DOORBELL,
1824 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1827 qp->rq.next_ind = ind;
1828 qp->rq.head += nreq;
1830 spin_unlock_irqrestore(&qp->rq.lock, flags);
1834 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1835 struct ib_send_wr **bad_wr)
1837 struct mthca_dev *dev = to_mdev(ibqp->device);
1838 struct mthca_qp *qp = to_mqp(ibqp);
1842 unsigned long flags;
1852 spin_lock_irqsave(&qp->sq.lock, flags);
1854 /* XXX check that state is OK to post send */
1856 ind = qp->sq.head & (qp->sq.max - 1);
1858 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1859 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1862 doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1863 ((qp->sq.head & 0xffff) << 8) |
1865 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1867 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1871 * Make sure that descriptors are written before
1875 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1878 * Make sure doorbell record is written before we
1879 * write MMIO send doorbell.
1882 mthca_write64(doorbell,
1883 dev->kar + MTHCA_SEND_DOORBELL,
1884 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1887 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1888 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1889 " %d max, %d nreq)\n", qp->qpn,
1890 qp->sq.head, qp->sq.tail,
1897 wqe = get_send_wqe(qp, ind);
1898 prev_wqe = qp->sq.last;
1901 ((struct mthca_next_seg *) wqe)->flags =
1902 ((wr->send_flags & IB_SEND_SIGNALED) ?
1903 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1904 ((wr->send_flags & IB_SEND_SOLICITED) ?
1905 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1907 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1908 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1909 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1911 wqe += sizeof (struct mthca_next_seg);
1912 size = sizeof (struct mthca_next_seg) / 16;
1914 switch (qp->transport) {
1916 switch (wr->opcode) {
1917 case IB_WR_ATOMIC_CMP_AND_SWP:
1918 case IB_WR_ATOMIC_FETCH_AND_ADD:
1919 ((struct mthca_raddr_seg *) wqe)->raddr =
1920 cpu_to_be64(wr->wr.atomic.remote_addr);
1921 ((struct mthca_raddr_seg *) wqe)->rkey =
1922 cpu_to_be32(wr->wr.atomic.rkey);
1923 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1925 wqe += sizeof (struct mthca_raddr_seg);
1927 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1928 ((struct mthca_atomic_seg *) wqe)->swap_add =
1929 cpu_to_be64(wr->wr.atomic.swap);
1930 ((struct mthca_atomic_seg *) wqe)->compare =
1931 cpu_to_be64(wr->wr.atomic.compare_add);
1933 ((struct mthca_atomic_seg *) wqe)->swap_add =
1934 cpu_to_be64(wr->wr.atomic.compare_add);
1935 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1938 wqe += sizeof (struct mthca_atomic_seg);
1939 size += (sizeof (struct mthca_raddr_seg) +
1940 sizeof (struct mthca_atomic_seg)) / 16;
1943 case IB_WR_RDMA_READ:
1944 case IB_WR_RDMA_WRITE:
1945 case IB_WR_RDMA_WRITE_WITH_IMM:
1946 ((struct mthca_raddr_seg *) wqe)->raddr =
1947 cpu_to_be64(wr->wr.rdma.remote_addr);
1948 ((struct mthca_raddr_seg *) wqe)->rkey =
1949 cpu_to_be32(wr->wr.rdma.rkey);
1950 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1951 wqe += sizeof (struct mthca_raddr_seg);
1952 size += sizeof (struct mthca_raddr_seg) / 16;
1956 /* No extra segments required for sends */
1963 switch (wr->opcode) {
1964 case IB_WR_RDMA_WRITE:
1965 case IB_WR_RDMA_WRITE_WITH_IMM:
1966 ((struct mthca_raddr_seg *) wqe)->raddr =
1967 cpu_to_be64(wr->wr.rdma.remote_addr);
1968 ((struct mthca_raddr_seg *) wqe)->rkey =
1969 cpu_to_be32(wr->wr.rdma.rkey);
1970 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1971 wqe += sizeof (struct mthca_raddr_seg);
1972 size += sizeof (struct mthca_raddr_seg) / 16;
1976 /* No extra segments required for sends */
1983 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1984 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1985 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1986 cpu_to_be32(wr->wr.ud.remote_qpn);
1987 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1988 cpu_to_be32(wr->wr.ud.remote_qkey);
1990 wqe += sizeof (struct mthca_arbel_ud_seg);
1991 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1995 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1996 wqe - sizeof (struct mthca_next_seg),
2002 wqe += sizeof (struct mthca_data_seg);
2003 size += sizeof (struct mthca_data_seg) / 16;
2007 if (wr->num_sge > qp->sq.max_gs) {
2008 mthca_err(dev, "too many gathers\n");
2014 for (i = 0; i < wr->num_sge; ++i) {
2015 ((struct mthca_data_seg *) wqe)->byte_count =
2016 cpu_to_be32(wr->sg_list[i].length);
2017 ((struct mthca_data_seg *) wqe)->lkey =
2018 cpu_to_be32(wr->sg_list[i].lkey);
2019 ((struct mthca_data_seg *) wqe)->addr =
2020 cpu_to_be64(wr->sg_list[i].addr);
2021 wqe += sizeof (struct mthca_data_seg);
2022 size += sizeof (struct mthca_data_seg) / 16;
2025 /* Add one more inline data segment for ICRC */
2026 if (qp->transport == MLX) {
2027 ((struct mthca_data_seg *) wqe)->byte_count =
2028 cpu_to_be32((1 << 31) | 4);
2029 ((u32 *) wqe)[1] = 0;
2030 wqe += sizeof (struct mthca_data_seg);
2031 size += sizeof (struct mthca_data_seg) / 16;
2034 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2036 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2037 mthca_err(dev, "opcode invalid\n");
2043 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2044 cpu_to_be32(((ind << qp->sq.wqe_shift) +
2045 qp->send_wqe_offset) |
2046 mthca_opcode[wr->opcode]);
2048 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2049 cpu_to_be32(MTHCA_NEXT_DBD | size);
2053 op0 = mthca_opcode[wr->opcode];
2057 if (unlikely(ind >= qp->sq.max))
2063 doorbell[0] = cpu_to_be32((nreq << 24) |
2064 ((qp->sq.head & 0xffff) << 8) |
2066 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2068 qp->sq.head += nreq;
2071 * Make sure that descriptors are written before
2075 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2078 * Make sure doorbell record is written before we
2079 * write MMIO send doorbell.
2082 mthca_write64(doorbell,
2083 dev->kar + MTHCA_SEND_DOORBELL,
2084 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2087 spin_unlock_irqrestore(&qp->sq.lock, flags);
2091 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2092 struct ib_recv_wr **bad_wr)
2094 struct mthca_dev *dev = to_mdev(ibqp->device);
2095 struct mthca_qp *qp = to_mqp(ibqp);
2096 unsigned long flags;
2103 spin_lock_irqsave(&qp->rq.lock, flags);
2105 /* XXX check that state is OK to post receive */
2107 ind = qp->rq.head & (qp->rq.max - 1);
2109 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2110 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2111 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2112 " %d max, %d nreq)\n", qp->qpn,
2113 qp->rq.head, qp->rq.tail,
2120 wqe = get_recv_wqe(qp, ind);
2122 ((struct mthca_next_seg *) wqe)->flags = 0;
2124 wqe += sizeof (struct mthca_next_seg);
2126 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2132 for (i = 0; i < wr->num_sge; ++i) {
2133 ((struct mthca_data_seg *) wqe)->byte_count =
2134 cpu_to_be32(wr->sg_list[i].length);
2135 ((struct mthca_data_seg *) wqe)->lkey =
2136 cpu_to_be32(wr->sg_list[i].lkey);
2137 ((struct mthca_data_seg *) wqe)->addr =
2138 cpu_to_be64(wr->sg_list[i].addr);
2139 wqe += sizeof (struct mthca_data_seg);
2142 if (i < qp->rq.max_gs) {
2143 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2144 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2145 ((struct mthca_data_seg *) wqe)->addr = 0;
2148 qp->wrid[ind] = wr->wr_id;
2151 if (unlikely(ind >= qp->rq.max))
2156 qp->rq.head += nreq;
2159 * Make sure that descriptors are written before
2163 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2166 spin_unlock_irqrestore(&qp->rq.lock, flags);
2170 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2171 int index, int *dbd, __be32 *new_wqe)
2173 struct mthca_next_seg *next;
2176 * For SRQs, all WQEs generate a CQE, so we're always at the
2177 * end of the doorbell chain.
2185 next = get_send_wqe(qp, index);
2187 next = get_recv_wqe(qp, index);
2189 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2190 if (next->ee_nds & cpu_to_be32(0x3f))
2191 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2192 (next->ee_nds & cpu_to_be32(0x3f));
2199 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2205 spin_lock_init(&dev->qp_table.lock);
2208 * We reserve 2 extra QPs per port for the special QPs. The
2209 * special QP for port 1 has to be even, so round up.
2211 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2212 err = mthca_alloc_init(&dev->qp_table.alloc,
2213 dev->limits.num_qps,
2215 dev->qp_table.sqp_start +
2216 MTHCA_MAX_PORTS * 2);
2220 err = mthca_array_init(&dev->qp_table.qp,
2221 dev->limits.num_qps);
2223 mthca_alloc_cleanup(&dev->qp_table.alloc);
2227 for (i = 0; i < 2; ++i) {
2228 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2229 dev->qp_table.sqp_start + i * 2,
2234 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2235 "status %02x, aborting.\n",
2244 for (i = 0; i < 2; ++i)
2245 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2247 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2248 mthca_alloc_cleanup(&dev->qp_table.alloc);
2253 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2258 for (i = 0; i < 2; ++i)
2259 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2261 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2262 mthca_alloc_cleanup(&dev->qp_table.alloc);