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 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
530 u32 hw_access_flags = 0;
532 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
533 dest_rd_atomic = attr->max_dest_rd_atomic;
535 dest_rd_atomic = qp->resp_depth;
537 if (attr_mask & IB_QP_ACCESS_FLAGS)
538 access_flags = attr->qp_access_flags;
540 access_flags = qp->atomic_rd_en;
543 access_flags &= IB_ACCESS_REMOTE_WRITE;
545 if (access_flags & IB_ACCESS_REMOTE_READ)
546 hw_access_flags |= MTHCA_QP_BIT_RRE;
547 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
548 hw_access_flags |= MTHCA_QP_BIT_RAE;
549 if (access_flags & IB_ACCESS_REMOTE_WRITE)
550 hw_access_flags |= MTHCA_QP_BIT_RWE;
552 return cpu_to_be32(hw_access_flags);
555 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
557 struct mthca_dev *dev = to_mdev(ibqp->device);
558 struct mthca_qp *qp = to_mqp(ibqp);
559 enum ib_qp_state cur_state, new_state;
560 struct mthca_mailbox *mailbox;
561 struct mthca_qp_param *qp_param;
562 struct mthca_qp_context *qp_context;
563 u32 req_param, opt_param;
567 if (attr_mask & IB_QP_CUR_STATE) {
568 if (attr->cur_qp_state != IB_QPS_RTR &&
569 attr->cur_qp_state != IB_QPS_RTS &&
570 attr->cur_qp_state != IB_QPS_SQD &&
571 attr->cur_qp_state != IB_QPS_SQE)
574 cur_state = attr->cur_qp_state;
576 spin_lock_irq(&qp->sq.lock);
577 spin_lock(&qp->rq.lock);
578 cur_state = qp->state;
579 spin_unlock(&qp->rq.lock);
580 spin_unlock_irq(&qp->sq.lock);
583 if (attr_mask & IB_QP_STATE) {
584 if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
586 new_state = attr->qp_state;
588 new_state = cur_state;
590 if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
591 mthca_dbg(dev, "Illegal QP transition "
592 "%d->%d\n", cur_state, new_state);
596 req_param = state_table[cur_state][new_state].req_param[qp->transport];
597 opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
599 if ((req_param & attr_mask) != req_param) {
600 mthca_dbg(dev, "QP transition "
601 "%d->%d missing req attr 0x%08x\n",
602 cur_state, new_state,
603 req_param & ~attr_mask);
607 if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
608 mthca_dbg(dev, "QP transition (transport %d) "
609 "%d->%d has extra attr 0x%08x\n",
611 cur_state, new_state,
612 attr_mask & ~(req_param | opt_param |
617 if ((attr_mask & IB_QP_PKEY_INDEX) &&
618 attr->pkey_index >= dev->limits.pkey_table_len) {
619 mthca_dbg(dev, "PKey index (%u) too large. max is %d\n",
620 attr->pkey_index,dev->limits.pkey_table_len-1);
624 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
625 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
626 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
627 attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
631 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
632 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
633 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
634 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
638 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
640 return PTR_ERR(mailbox);
641 qp_param = mailbox->buf;
642 qp_context = &qp_param->context;
643 memset(qp_param, 0, sizeof *qp_param);
645 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
646 (to_mthca_st(qp->transport) << 16));
647 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
648 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
649 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
651 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
652 switch (attr->path_mig_state) {
653 case IB_MIG_MIGRATED:
654 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
657 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
660 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
665 /* leave tavor_sched_queue as 0 */
667 if (qp->transport == MLX || qp->transport == UD)
668 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
669 else if (attr_mask & IB_QP_PATH_MTU)
670 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
672 if (mthca_is_memfree(dev)) {
674 qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
675 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
678 qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
679 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
682 /* leave arbel_sched_queue as 0 */
684 if (qp->ibqp.uobject)
685 qp_context->usr_page =
686 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
688 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
689 qp_context->local_qpn = cpu_to_be32(qp->qpn);
690 if (attr_mask & IB_QP_DEST_QPN) {
691 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
694 if (qp->transport == MLX)
695 qp_context->pri_path.port_pkey |=
696 cpu_to_be32(to_msqp(qp)->port << 24);
698 if (attr_mask & IB_QP_PORT) {
699 qp_context->pri_path.port_pkey |=
700 cpu_to_be32(attr->port_num << 24);
701 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
705 if (attr_mask & IB_QP_PKEY_INDEX) {
706 qp_context->pri_path.port_pkey |=
707 cpu_to_be32(attr->pkey_index);
708 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
711 if (attr_mask & IB_QP_RNR_RETRY) {
712 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
713 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
716 if (attr_mask & IB_QP_AV) {
717 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
718 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
719 qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
720 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
721 qp_context->pri_path.g_mylmc |= 1 << 7;
722 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
723 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
724 qp_context->pri_path.sl_tclass_flowlabel =
725 cpu_to_be32((attr->ah_attr.sl << 28) |
726 (attr->ah_attr.grh.traffic_class << 20) |
727 (attr->ah_attr.grh.flow_label));
728 memcpy(qp_context->pri_path.rgid,
729 attr->ah_attr.grh.dgid.raw, 16);
731 qp_context->pri_path.sl_tclass_flowlabel =
732 cpu_to_be32(attr->ah_attr.sl << 28);
734 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
737 if (attr_mask & IB_QP_TIMEOUT) {
738 qp_context->pri_path.ackto = attr->timeout << 3;
739 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
745 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
746 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
747 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
748 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
749 (MTHCA_FLIGHT_LIMIT << 24) |
751 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
752 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
753 if (attr_mask & IB_QP_RETRY_CNT) {
754 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
755 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
758 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
759 if (attr->max_rd_atomic) {
760 qp_context->params1 |=
761 cpu_to_be32(MTHCA_QP_BIT_SRE |
763 qp_context->params1 |=
764 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
766 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
769 if (attr_mask & IB_QP_SQ_PSN)
770 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
771 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
773 if (mthca_is_memfree(dev)) {
774 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
775 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
778 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
779 if (attr->max_dest_rd_atomic)
780 qp_context->params2 |=
781 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
783 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
786 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
787 qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask);
788 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
789 MTHCA_QP_OPTPAR_RRE |
790 MTHCA_QP_OPTPAR_RAE);
793 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
796 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
798 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
799 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
800 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
802 if (attr_mask & IB_QP_RQ_PSN)
803 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
805 qp_context->ra_buff_indx =
806 cpu_to_be32(dev->qp_table.rdb_base +
807 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
808 dev->qp_table.rdb_shift));
810 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
812 if (mthca_is_memfree(dev))
813 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
815 if (attr_mask & IB_QP_QKEY) {
816 qp_context->qkey = cpu_to_be32(attr->qkey);
817 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
821 qp_context->srqn = cpu_to_be32(1 << 24 |
822 to_msrq(ibqp->srq)->srqn);
824 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
825 qp->qpn, 0, mailbox, 0, &status);
827 mthca_warn(dev, "modify QP %d returned status %02x.\n",
828 state_table[cur_state][new_state].trans, status);
833 qp->state = new_state;
834 if (attr_mask & IB_QP_ACCESS_FLAGS)
835 qp->atomic_rd_en = attr->qp_access_flags;
836 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
837 qp->resp_depth = attr->max_dest_rd_atomic;
840 mthca_free_mailbox(dev, mailbox);
843 store_attrs(to_msqp(qp), attr, attr_mask);
846 * If we moved QP0 to RTR, bring the IB link up; if we moved
847 * QP0 to RESET or ERROR, bring the link back down.
849 if (is_qp0(dev, qp)) {
850 if (cur_state != IB_QPS_RTR &&
851 new_state == IB_QPS_RTR)
852 init_port(dev, to_msqp(qp)->port);
854 if (cur_state != IB_QPS_RESET &&
855 cur_state != IB_QPS_ERR &&
856 (new_state == IB_QPS_RESET ||
857 new_state == IB_QPS_ERR))
858 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
862 * If we moved a kernel QP to RESET, clean up all old CQ
863 * entries and reinitialize the QP.
865 if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
866 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
867 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
868 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
869 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
870 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
872 mthca_wq_init(&qp->sq);
873 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
875 mthca_wq_init(&qp->rq);
876 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
878 if (mthca_is_memfree(dev)) {
887 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
894 * Calculate the maximum size of WQE s/g segments, excluding
895 * the next segment and other non-data segments.
897 max_data_size = min(dev->limits.max_desc_sz, 1 << qp->sq.wqe_shift) -
898 sizeof (struct mthca_next_seg);
900 switch (qp->transport) {
902 max_data_size -= 2 * sizeof (struct mthca_data_seg);
906 if (mthca_is_memfree(dev))
907 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
909 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
913 max_data_size -= sizeof (struct mthca_raddr_seg);
917 /* We don't support inline data for kernel QPs (yet). */
918 if (!pd->ibpd.uobject)
919 qp->max_inline_data = 0;
921 qp->max_inline_data = max_data_size - MTHCA_INLINE_HEADER_SIZE;
923 qp->sq.max_gs = min_t(int, dev->limits.max_sg,
924 max_data_size / sizeof (struct mthca_data_seg));
925 qp->rq.max_gs = min_t(int, dev->limits.max_sg,
926 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
927 sizeof (struct mthca_next_seg)) /
928 sizeof (struct mthca_data_seg));
932 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
933 * rq.max_gs and sq.max_gs must all be assigned.
934 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
935 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
938 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
945 size = sizeof (struct mthca_next_seg) +
946 qp->rq.max_gs * sizeof (struct mthca_data_seg);
948 if (size > dev->limits.max_desc_sz)
951 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
955 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
956 switch (qp->transport) {
958 size += 2 * sizeof (struct mthca_data_seg);
962 size += mthca_is_memfree(dev) ?
963 sizeof (struct mthca_arbel_ud_seg) :
964 sizeof (struct mthca_tavor_ud_seg);
968 size += sizeof (struct mthca_raddr_seg);
972 size += sizeof (struct mthca_raddr_seg);
974 * An atomic op will require an atomic segment, a
975 * remote address segment and one scatter entry.
977 size = max_t(int, size,
978 sizeof (struct mthca_atomic_seg) +
979 sizeof (struct mthca_raddr_seg) +
980 sizeof (struct mthca_data_seg));
987 /* Make sure that we have enough space for a bind request */
988 size = max_t(int, size, sizeof (struct mthca_bind_seg));
990 size += sizeof (struct mthca_next_seg);
992 if (size > dev->limits.max_desc_sz)
995 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
999 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1000 1 << qp->sq.wqe_shift);
1003 * If this is a userspace QP, we don't actually have to
1004 * allocate anything. All we need is to calculate the WQE
1005 * sizes and the send_wqe_offset, so we're done now.
1007 if (pd->ibpd.uobject)
1010 size = PAGE_ALIGN(qp->send_wqe_offset +
1011 (qp->sq.max << qp->sq.wqe_shift));
1013 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1018 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1019 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1030 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1031 struct mthca_qp *qp)
1033 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1034 (qp->sq.max << qp->sq.wqe_shift)),
1035 &qp->queue, qp->is_direct, &qp->mr);
1039 static int mthca_map_memfree(struct mthca_dev *dev,
1040 struct mthca_qp *qp)
1044 if (mthca_is_memfree(dev)) {
1045 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1049 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1053 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1054 qp->qpn << dev->qp_table.rdb_shift);
1063 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1066 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1071 static void mthca_unmap_memfree(struct mthca_dev *dev,
1072 struct mthca_qp *qp)
1074 mthca_table_put(dev, dev->qp_table.rdb_table,
1075 qp->qpn << dev->qp_table.rdb_shift);
1076 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1077 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1080 static int mthca_alloc_memfree(struct mthca_dev *dev,
1081 struct mthca_qp *qp)
1085 if (mthca_is_memfree(dev)) {
1086 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1087 qp->qpn, &qp->rq.db);
1088 if (qp->rq.db_index < 0)
1091 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1092 qp->qpn, &qp->sq.db);
1093 if (qp->sq.db_index < 0)
1094 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1100 static void mthca_free_memfree(struct mthca_dev *dev,
1101 struct mthca_qp *qp)
1103 if (mthca_is_memfree(dev)) {
1104 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1105 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1109 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1110 struct mthca_pd *pd,
1111 struct mthca_cq *send_cq,
1112 struct mthca_cq *recv_cq,
1113 enum ib_sig_type send_policy,
1114 struct mthca_qp *qp)
1119 atomic_set(&qp->refcount, 1);
1120 init_waitqueue_head(&qp->wait);
1121 qp->state = IB_QPS_RESET;
1122 qp->atomic_rd_en = 0;
1124 qp->sq_policy = send_policy;
1125 mthca_wq_init(&qp->sq);
1126 mthca_wq_init(&qp->rq);
1128 ret = mthca_map_memfree(dev, qp);
1132 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1134 mthca_unmap_memfree(dev, qp);
1138 mthca_adjust_qp_caps(dev, pd, qp);
1141 * If this is a userspace QP, we're done now. The doorbells
1142 * will be allocated and buffers will be initialized in
1145 if (pd->ibpd.uobject)
1148 ret = mthca_alloc_memfree(dev, qp);
1150 mthca_free_wqe_buf(dev, qp);
1151 mthca_unmap_memfree(dev, qp);
1155 if (mthca_is_memfree(dev)) {
1156 struct mthca_next_seg *next;
1157 struct mthca_data_seg *scatter;
1158 int size = (sizeof (struct mthca_next_seg) +
1159 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1161 for (i = 0; i < qp->rq.max; ++i) {
1162 next = get_recv_wqe(qp, i);
1163 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1165 next->ee_nds = cpu_to_be32(size);
1167 for (scatter = (void *) (next + 1);
1168 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1170 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1173 for (i = 0; i < qp->sq.max; ++i) {
1174 next = get_send_wqe(qp, i);
1175 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1177 qp->send_wqe_offset);
1181 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1182 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1187 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1188 struct mthca_qp *qp)
1190 /* Sanity check QP size before proceeding */
1191 if (cap->max_send_wr > dev->limits.max_wqes ||
1192 cap->max_recv_wr > dev->limits.max_wqes ||
1193 cap->max_send_sge > dev->limits.max_sg ||
1194 cap->max_recv_sge > dev->limits.max_sg)
1197 if (mthca_is_memfree(dev)) {
1198 qp->rq.max = cap->max_recv_wr ?
1199 roundup_pow_of_two(cap->max_recv_wr) : 0;
1200 qp->sq.max = cap->max_send_wr ?
1201 roundup_pow_of_two(cap->max_send_wr) : 0;
1203 qp->rq.max = cap->max_recv_wr;
1204 qp->sq.max = cap->max_send_wr;
1207 qp->rq.max_gs = cap->max_recv_sge;
1208 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1209 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1210 MTHCA_INLINE_CHUNK_SIZE) /
1211 sizeof (struct mthca_data_seg));
1214 * For MLX transport we need 2 extra S/G entries:
1215 * one for the header and one for the checksum at the end
1217 if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1218 qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1224 int mthca_alloc_qp(struct mthca_dev *dev,
1225 struct mthca_pd *pd,
1226 struct mthca_cq *send_cq,
1227 struct mthca_cq *recv_cq,
1228 enum ib_qp_type type,
1229 enum ib_sig_type send_policy,
1230 struct ib_qp_cap *cap,
1231 struct mthca_qp *qp)
1235 err = mthca_set_qp_size(dev, cap, qp);
1240 case IB_QPT_RC: qp->transport = RC; break;
1241 case IB_QPT_UC: qp->transport = UC; break;
1242 case IB_QPT_UD: qp->transport = UD; break;
1243 default: return -EINVAL;
1246 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1250 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1253 mthca_free(&dev->qp_table.alloc, qp->qpn);
1257 spin_lock_irq(&dev->qp_table.lock);
1258 mthca_array_set(&dev->qp_table.qp,
1259 qp->qpn & (dev->limits.num_qps - 1), qp);
1260 spin_unlock_irq(&dev->qp_table.lock);
1265 int mthca_alloc_sqp(struct mthca_dev *dev,
1266 struct mthca_pd *pd,
1267 struct mthca_cq *send_cq,
1268 struct mthca_cq *recv_cq,
1269 enum ib_sig_type send_policy,
1270 struct ib_qp_cap *cap,
1273 struct mthca_sqp *sqp)
1275 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1278 err = mthca_set_qp_size(dev, cap, &sqp->qp);
1282 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1283 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1284 &sqp->header_dma, GFP_KERNEL);
1285 if (!sqp->header_buf)
1288 spin_lock_irq(&dev->qp_table.lock);
1289 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1292 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1293 spin_unlock_irq(&dev->qp_table.lock);
1300 sqp->qp.transport = MLX;
1302 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1303 send_policy, &sqp->qp);
1307 atomic_inc(&pd->sqp_count);
1313 * Lock CQs here, so that CQ polling code can do QP lookup
1314 * without taking a lock.
1316 spin_lock_irq(&send_cq->lock);
1317 if (send_cq != recv_cq)
1318 spin_lock(&recv_cq->lock);
1320 spin_lock(&dev->qp_table.lock);
1321 mthca_array_clear(&dev->qp_table.qp, mqpn);
1322 spin_unlock(&dev->qp_table.lock);
1324 if (send_cq != recv_cq)
1325 spin_unlock(&recv_cq->lock);
1326 spin_unlock_irq(&send_cq->lock);
1329 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1330 sqp->header_buf, sqp->header_dma);
1335 void mthca_free_qp(struct mthca_dev *dev,
1336 struct mthca_qp *qp)
1339 struct mthca_cq *send_cq;
1340 struct mthca_cq *recv_cq;
1342 send_cq = to_mcq(qp->ibqp.send_cq);
1343 recv_cq = to_mcq(qp->ibqp.recv_cq);
1346 * Lock CQs here, so that CQ polling code can do QP lookup
1347 * without taking a lock.
1349 spin_lock_irq(&send_cq->lock);
1350 if (send_cq != recv_cq)
1351 spin_lock(&recv_cq->lock);
1353 spin_lock(&dev->qp_table.lock);
1354 mthca_array_clear(&dev->qp_table.qp,
1355 qp->qpn & (dev->limits.num_qps - 1));
1356 spin_unlock(&dev->qp_table.lock);
1358 if (send_cq != recv_cq)
1359 spin_unlock(&recv_cq->lock);
1360 spin_unlock_irq(&send_cq->lock);
1362 atomic_dec(&qp->refcount);
1363 wait_event(qp->wait, !atomic_read(&qp->refcount));
1365 if (qp->state != IB_QPS_RESET)
1366 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1369 * If this is a userspace QP, the buffers, MR, CQs and so on
1370 * will be cleaned up in userspace, so all we have to do is
1371 * unref the mem-free tables and free the QPN in our table.
1373 if (!qp->ibqp.uobject) {
1374 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1375 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1376 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1377 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1378 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1380 mthca_free_memfree(dev, qp);
1381 mthca_free_wqe_buf(dev, qp);
1384 mthca_unmap_memfree(dev, qp);
1386 if (is_sqp(dev, qp)) {
1387 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1388 dma_free_coherent(&dev->pdev->dev,
1389 to_msqp(qp)->header_buf_size,
1390 to_msqp(qp)->header_buf,
1391 to_msqp(qp)->header_dma);
1393 mthca_free(&dev->qp_table.alloc, qp->qpn);
1396 /* Create UD header for an MLX send and build a data segment for it */
1397 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1398 int ind, struct ib_send_wr *wr,
1399 struct mthca_mlx_seg *mlx,
1400 struct mthca_data_seg *data)
1406 ib_ud_header_init(256, /* assume a MAD */
1407 sqp->ud_header.grh_present,
1410 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1413 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1414 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1415 (sqp->ud_header.lrh.destination_lid ==
1416 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1417 (sqp->ud_header.lrh.service_level << 8));
1418 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1421 switch (wr->opcode) {
1423 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1424 sqp->ud_header.immediate_present = 0;
1426 case IB_WR_SEND_WITH_IMM:
1427 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1428 sqp->ud_header.immediate_present = 1;
1429 sqp->ud_header.immediate_data = wr->imm_data;
1435 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1436 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1437 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1438 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1439 if (!sqp->qp.ibqp.qp_num)
1440 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1441 sqp->pkey_index, &pkey);
1443 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1444 wr->wr.ud.pkey_index, &pkey);
1445 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1446 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1447 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1448 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1449 sqp->qkey : wr->wr.ud.remote_qkey);
1450 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1452 header_size = ib_ud_header_pack(&sqp->ud_header,
1454 ind * MTHCA_UD_HEADER_SIZE);
1456 data->byte_count = cpu_to_be32(header_size);
1457 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1458 data->addr = cpu_to_be64(sqp->header_dma +
1459 ind * MTHCA_UD_HEADER_SIZE);
1464 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1465 struct ib_cq *ib_cq)
1468 struct mthca_cq *cq;
1470 cur = wq->head - wq->tail;
1471 if (likely(cur + nreq < wq->max))
1475 spin_lock(&cq->lock);
1476 cur = wq->head - wq->tail;
1477 spin_unlock(&cq->lock);
1479 return cur + nreq >= wq->max;
1482 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1483 struct ib_send_wr **bad_wr)
1485 struct mthca_dev *dev = to_mdev(ibqp->device);
1486 struct mthca_qp *qp = to_mqp(ibqp);
1489 unsigned long flags;
1499 spin_lock_irqsave(&qp->sq.lock, flags);
1501 /* XXX check that state is OK to post send */
1503 ind = qp->sq.next_ind;
1505 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1506 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1507 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1508 " %d max, %d nreq)\n", qp->qpn,
1509 qp->sq.head, qp->sq.tail,
1516 wqe = get_send_wqe(qp, ind);
1517 prev_wqe = qp->sq.last;
1520 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1521 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1522 ((struct mthca_next_seg *) wqe)->flags =
1523 ((wr->send_flags & IB_SEND_SIGNALED) ?
1524 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1525 ((wr->send_flags & IB_SEND_SOLICITED) ?
1526 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1528 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1529 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1530 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1532 wqe += sizeof (struct mthca_next_seg);
1533 size = sizeof (struct mthca_next_seg) / 16;
1535 switch (qp->transport) {
1537 switch (wr->opcode) {
1538 case IB_WR_ATOMIC_CMP_AND_SWP:
1539 case IB_WR_ATOMIC_FETCH_AND_ADD:
1540 ((struct mthca_raddr_seg *) wqe)->raddr =
1541 cpu_to_be64(wr->wr.atomic.remote_addr);
1542 ((struct mthca_raddr_seg *) wqe)->rkey =
1543 cpu_to_be32(wr->wr.atomic.rkey);
1544 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1546 wqe += sizeof (struct mthca_raddr_seg);
1548 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1549 ((struct mthca_atomic_seg *) wqe)->swap_add =
1550 cpu_to_be64(wr->wr.atomic.swap);
1551 ((struct mthca_atomic_seg *) wqe)->compare =
1552 cpu_to_be64(wr->wr.atomic.compare_add);
1554 ((struct mthca_atomic_seg *) wqe)->swap_add =
1555 cpu_to_be64(wr->wr.atomic.compare_add);
1556 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1559 wqe += sizeof (struct mthca_atomic_seg);
1560 size += (sizeof (struct mthca_raddr_seg) +
1561 sizeof (struct mthca_atomic_seg)) / 16;
1564 case IB_WR_RDMA_WRITE:
1565 case IB_WR_RDMA_WRITE_WITH_IMM:
1566 case IB_WR_RDMA_READ:
1567 ((struct mthca_raddr_seg *) wqe)->raddr =
1568 cpu_to_be64(wr->wr.rdma.remote_addr);
1569 ((struct mthca_raddr_seg *) wqe)->rkey =
1570 cpu_to_be32(wr->wr.rdma.rkey);
1571 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1572 wqe += sizeof (struct mthca_raddr_seg);
1573 size += sizeof (struct mthca_raddr_seg) / 16;
1577 /* No extra segments required for sends */
1584 switch (wr->opcode) {
1585 case IB_WR_RDMA_WRITE:
1586 case IB_WR_RDMA_WRITE_WITH_IMM:
1587 ((struct mthca_raddr_seg *) wqe)->raddr =
1588 cpu_to_be64(wr->wr.rdma.remote_addr);
1589 ((struct mthca_raddr_seg *) wqe)->rkey =
1590 cpu_to_be32(wr->wr.rdma.rkey);
1591 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1592 wqe += sizeof (struct mthca_raddr_seg);
1593 size += sizeof (struct mthca_raddr_seg) / 16;
1597 /* No extra segments required for sends */
1604 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1605 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1606 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1607 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1608 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1609 cpu_to_be32(wr->wr.ud.remote_qpn);
1610 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1611 cpu_to_be32(wr->wr.ud.remote_qkey);
1613 wqe += sizeof (struct mthca_tavor_ud_seg);
1614 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1618 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1619 wqe - sizeof (struct mthca_next_seg),
1625 wqe += sizeof (struct mthca_data_seg);
1626 size += sizeof (struct mthca_data_seg) / 16;
1630 if (wr->num_sge > qp->sq.max_gs) {
1631 mthca_err(dev, "too many gathers\n");
1637 for (i = 0; i < wr->num_sge; ++i) {
1638 ((struct mthca_data_seg *) wqe)->byte_count =
1639 cpu_to_be32(wr->sg_list[i].length);
1640 ((struct mthca_data_seg *) wqe)->lkey =
1641 cpu_to_be32(wr->sg_list[i].lkey);
1642 ((struct mthca_data_seg *) wqe)->addr =
1643 cpu_to_be64(wr->sg_list[i].addr);
1644 wqe += sizeof (struct mthca_data_seg);
1645 size += sizeof (struct mthca_data_seg) / 16;
1648 /* Add one more inline data segment for ICRC */
1649 if (qp->transport == MLX) {
1650 ((struct mthca_data_seg *) wqe)->byte_count =
1651 cpu_to_be32((1 << 31) | 4);
1652 ((u32 *) wqe)[1] = 0;
1653 wqe += sizeof (struct mthca_data_seg);
1654 size += sizeof (struct mthca_data_seg) / 16;
1657 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1659 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1660 mthca_err(dev, "opcode invalid\n");
1666 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1667 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1668 qp->send_wqe_offset) |
1669 mthca_opcode[wr->opcode]);
1671 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1672 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1676 op0 = mthca_opcode[wr->opcode];
1680 if (unlikely(ind >= qp->sq.max))
1688 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1689 qp->send_wqe_offset) | f0 | op0);
1690 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1694 mthca_write64(doorbell,
1695 dev->kar + MTHCA_SEND_DOORBELL,
1696 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1699 qp->sq.next_ind = ind;
1700 qp->sq.head += nreq;
1702 spin_unlock_irqrestore(&qp->sq.lock, flags);
1706 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1707 struct ib_recv_wr **bad_wr)
1709 struct mthca_dev *dev = to_mdev(ibqp->device);
1710 struct mthca_qp *qp = to_mqp(ibqp);
1712 unsigned long flags;
1722 spin_lock_irqsave(&qp->rq.lock, flags);
1724 /* XXX check that state is OK to post receive */
1726 ind = qp->rq.next_ind;
1728 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1729 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1732 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1733 doorbell[1] = cpu_to_be32(qp->qpn << 8);
1737 mthca_write64(doorbell,
1738 dev->kar + MTHCA_RECEIVE_DOORBELL,
1739 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1741 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1745 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1746 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1747 " %d max, %d nreq)\n", qp->qpn,
1748 qp->rq.head, qp->rq.tail,
1755 wqe = get_recv_wqe(qp, ind);
1756 prev_wqe = qp->rq.last;
1759 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1760 ((struct mthca_next_seg *) wqe)->ee_nds =
1761 cpu_to_be32(MTHCA_NEXT_DBD);
1762 ((struct mthca_next_seg *) wqe)->flags = 0;
1764 wqe += sizeof (struct mthca_next_seg);
1765 size = sizeof (struct mthca_next_seg) / 16;
1767 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1773 for (i = 0; i < wr->num_sge; ++i) {
1774 ((struct mthca_data_seg *) wqe)->byte_count =
1775 cpu_to_be32(wr->sg_list[i].length);
1776 ((struct mthca_data_seg *) wqe)->lkey =
1777 cpu_to_be32(wr->sg_list[i].lkey);
1778 ((struct mthca_data_seg *) wqe)->addr =
1779 cpu_to_be64(wr->sg_list[i].addr);
1780 wqe += sizeof (struct mthca_data_seg);
1781 size += sizeof (struct mthca_data_seg) / 16;
1784 qp->wrid[ind] = wr->wr_id;
1786 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1787 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1789 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1790 cpu_to_be32(MTHCA_NEXT_DBD | size);
1796 if (unlikely(ind >= qp->rq.max))
1802 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1803 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1807 mthca_write64(doorbell,
1808 dev->kar + MTHCA_RECEIVE_DOORBELL,
1809 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1812 qp->rq.next_ind = ind;
1813 qp->rq.head += nreq;
1815 spin_unlock_irqrestore(&qp->rq.lock, flags);
1819 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1820 struct ib_send_wr **bad_wr)
1822 struct mthca_dev *dev = to_mdev(ibqp->device);
1823 struct mthca_qp *qp = to_mqp(ibqp);
1827 unsigned long flags;
1837 spin_lock_irqsave(&qp->sq.lock, flags);
1839 /* XXX check that state is OK to post send */
1841 ind = qp->sq.head & (qp->sq.max - 1);
1843 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1844 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1847 doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1848 ((qp->sq.head & 0xffff) << 8) |
1850 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1852 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1856 * Make sure that descriptors are written before
1860 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1863 * Make sure doorbell record is written before we
1864 * write MMIO send doorbell.
1867 mthca_write64(doorbell,
1868 dev->kar + MTHCA_SEND_DOORBELL,
1869 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1872 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1873 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1874 " %d max, %d nreq)\n", qp->qpn,
1875 qp->sq.head, qp->sq.tail,
1882 wqe = get_send_wqe(qp, ind);
1883 prev_wqe = qp->sq.last;
1886 ((struct mthca_next_seg *) wqe)->flags =
1887 ((wr->send_flags & IB_SEND_SIGNALED) ?
1888 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1889 ((wr->send_flags & IB_SEND_SOLICITED) ?
1890 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1892 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1893 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1894 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1896 wqe += sizeof (struct mthca_next_seg);
1897 size = sizeof (struct mthca_next_seg) / 16;
1899 switch (qp->transport) {
1901 switch (wr->opcode) {
1902 case IB_WR_ATOMIC_CMP_AND_SWP:
1903 case IB_WR_ATOMIC_FETCH_AND_ADD:
1904 ((struct mthca_raddr_seg *) wqe)->raddr =
1905 cpu_to_be64(wr->wr.atomic.remote_addr);
1906 ((struct mthca_raddr_seg *) wqe)->rkey =
1907 cpu_to_be32(wr->wr.atomic.rkey);
1908 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1910 wqe += sizeof (struct mthca_raddr_seg);
1912 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1913 ((struct mthca_atomic_seg *) wqe)->swap_add =
1914 cpu_to_be64(wr->wr.atomic.swap);
1915 ((struct mthca_atomic_seg *) wqe)->compare =
1916 cpu_to_be64(wr->wr.atomic.compare_add);
1918 ((struct mthca_atomic_seg *) wqe)->swap_add =
1919 cpu_to_be64(wr->wr.atomic.compare_add);
1920 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1923 wqe += sizeof (struct mthca_atomic_seg);
1924 size += (sizeof (struct mthca_raddr_seg) +
1925 sizeof (struct mthca_atomic_seg)) / 16;
1928 case IB_WR_RDMA_READ:
1929 case IB_WR_RDMA_WRITE:
1930 case IB_WR_RDMA_WRITE_WITH_IMM:
1931 ((struct mthca_raddr_seg *) wqe)->raddr =
1932 cpu_to_be64(wr->wr.rdma.remote_addr);
1933 ((struct mthca_raddr_seg *) wqe)->rkey =
1934 cpu_to_be32(wr->wr.rdma.rkey);
1935 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1936 wqe += sizeof (struct mthca_raddr_seg);
1937 size += sizeof (struct mthca_raddr_seg) / 16;
1941 /* No extra segments required for sends */
1948 switch (wr->opcode) {
1949 case IB_WR_RDMA_WRITE:
1950 case IB_WR_RDMA_WRITE_WITH_IMM:
1951 ((struct mthca_raddr_seg *) wqe)->raddr =
1952 cpu_to_be64(wr->wr.rdma.remote_addr);
1953 ((struct mthca_raddr_seg *) wqe)->rkey =
1954 cpu_to_be32(wr->wr.rdma.rkey);
1955 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1956 wqe += sizeof (struct mthca_raddr_seg);
1957 size += sizeof (struct mthca_raddr_seg) / 16;
1961 /* No extra segments required for sends */
1968 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1969 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1970 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1971 cpu_to_be32(wr->wr.ud.remote_qpn);
1972 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1973 cpu_to_be32(wr->wr.ud.remote_qkey);
1975 wqe += sizeof (struct mthca_arbel_ud_seg);
1976 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1980 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1981 wqe - sizeof (struct mthca_next_seg),
1987 wqe += sizeof (struct mthca_data_seg);
1988 size += sizeof (struct mthca_data_seg) / 16;
1992 if (wr->num_sge > qp->sq.max_gs) {
1993 mthca_err(dev, "too many gathers\n");
1999 for (i = 0; i < wr->num_sge; ++i) {
2000 ((struct mthca_data_seg *) wqe)->byte_count =
2001 cpu_to_be32(wr->sg_list[i].length);
2002 ((struct mthca_data_seg *) wqe)->lkey =
2003 cpu_to_be32(wr->sg_list[i].lkey);
2004 ((struct mthca_data_seg *) wqe)->addr =
2005 cpu_to_be64(wr->sg_list[i].addr);
2006 wqe += sizeof (struct mthca_data_seg);
2007 size += sizeof (struct mthca_data_seg) / 16;
2010 /* Add one more inline data segment for ICRC */
2011 if (qp->transport == MLX) {
2012 ((struct mthca_data_seg *) wqe)->byte_count =
2013 cpu_to_be32((1 << 31) | 4);
2014 ((u32 *) wqe)[1] = 0;
2015 wqe += sizeof (struct mthca_data_seg);
2016 size += sizeof (struct mthca_data_seg) / 16;
2019 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2021 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2022 mthca_err(dev, "opcode invalid\n");
2028 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2029 cpu_to_be32(((ind << qp->sq.wqe_shift) +
2030 qp->send_wqe_offset) |
2031 mthca_opcode[wr->opcode]);
2033 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2034 cpu_to_be32(MTHCA_NEXT_DBD | size);
2038 op0 = mthca_opcode[wr->opcode];
2042 if (unlikely(ind >= qp->sq.max))
2048 doorbell[0] = cpu_to_be32((nreq << 24) |
2049 ((qp->sq.head & 0xffff) << 8) |
2051 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2053 qp->sq.head += nreq;
2056 * Make sure that descriptors are written before
2060 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2063 * Make sure doorbell record is written before we
2064 * write MMIO send doorbell.
2067 mthca_write64(doorbell,
2068 dev->kar + MTHCA_SEND_DOORBELL,
2069 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2072 spin_unlock_irqrestore(&qp->sq.lock, flags);
2076 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2077 struct ib_recv_wr **bad_wr)
2079 struct mthca_dev *dev = to_mdev(ibqp->device);
2080 struct mthca_qp *qp = to_mqp(ibqp);
2081 unsigned long flags;
2088 spin_lock_irqsave(&qp->rq.lock, flags);
2090 /* XXX check that state is OK to post receive */
2092 ind = qp->rq.head & (qp->rq.max - 1);
2094 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2095 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2096 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2097 " %d max, %d nreq)\n", qp->qpn,
2098 qp->rq.head, qp->rq.tail,
2105 wqe = get_recv_wqe(qp, ind);
2107 ((struct mthca_next_seg *) wqe)->flags = 0;
2109 wqe += sizeof (struct mthca_next_seg);
2111 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2117 for (i = 0; i < wr->num_sge; ++i) {
2118 ((struct mthca_data_seg *) wqe)->byte_count =
2119 cpu_to_be32(wr->sg_list[i].length);
2120 ((struct mthca_data_seg *) wqe)->lkey =
2121 cpu_to_be32(wr->sg_list[i].lkey);
2122 ((struct mthca_data_seg *) wqe)->addr =
2123 cpu_to_be64(wr->sg_list[i].addr);
2124 wqe += sizeof (struct mthca_data_seg);
2127 if (i < qp->rq.max_gs) {
2128 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2129 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2130 ((struct mthca_data_seg *) wqe)->addr = 0;
2133 qp->wrid[ind] = wr->wr_id;
2136 if (unlikely(ind >= qp->rq.max))
2141 qp->rq.head += nreq;
2144 * Make sure that descriptors are written before
2148 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2151 spin_unlock_irqrestore(&qp->rq.lock, flags);
2155 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2156 int index, int *dbd, __be32 *new_wqe)
2158 struct mthca_next_seg *next;
2161 * For SRQs, all WQEs generate a CQE, so we're always at the
2162 * end of the doorbell chain.
2170 next = get_send_wqe(qp, index);
2172 next = get_recv_wqe(qp, index);
2174 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2175 if (next->ee_nds & cpu_to_be32(0x3f))
2176 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2177 (next->ee_nds & cpu_to_be32(0x3f));
2184 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2190 spin_lock_init(&dev->qp_table.lock);
2193 * We reserve 2 extra QPs per port for the special QPs. The
2194 * special QP for port 1 has to be even, so round up.
2196 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2197 err = mthca_alloc_init(&dev->qp_table.alloc,
2198 dev->limits.num_qps,
2200 dev->qp_table.sqp_start +
2201 MTHCA_MAX_PORTS * 2);
2205 err = mthca_array_init(&dev->qp_table.qp,
2206 dev->limits.num_qps);
2208 mthca_alloc_cleanup(&dev->qp_table.alloc);
2212 for (i = 0; i < 2; ++i) {
2213 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2214 dev->qp_table.sqp_start + i * 2,
2219 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2220 "status %02x, aborting.\n",
2229 for (i = 0; i < 2; ++i)
2230 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2232 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2233 mthca_alloc_cleanup(&dev->qp_table.alloc);
2238 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2243 for (i = 0; i < 2; ++i)
2244 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2246 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2247 mthca_alloc_cleanup(&dev->qp_table.alloc);