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
103 MTHCA_SEND_DOORBELL_FENCE = 1 << 5
106 struct mthca_qp_path {
115 __be32 sl_tclass_flowlabel;
117 } __attribute__((packed));
119 struct mthca_qp_context {
121 __be32 tavor_sched_queue; /* Reserved on Arbel */
123 u8 rq_size_stride; /* Reserved on Tavor */
124 u8 sq_size_stride; /* Reserved on Tavor */
125 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
130 struct mthca_qp_path pri_path;
131 struct mthca_qp_path alt_path;
138 __be32 next_send_psn;
140 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
141 __be32 snd_db_index; /* (debugging only entries) */
142 __be32 last_acked_psn;
145 __be32 rnr_nextrecvpsn;
148 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
149 __be32 rcv_db_index; /* (debugging only entries) */
153 __be16 rq_wqe_counter; /* reserved on Tavor */
154 __be16 sq_wqe_counter; /* reserved on Tavor */
156 } __attribute__((packed));
158 struct mthca_qp_param {
159 __be32 opt_param_mask;
161 struct mthca_qp_context context;
163 } __attribute__((packed));
166 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
167 MTHCA_QP_OPTPAR_RRE = 1 << 1,
168 MTHCA_QP_OPTPAR_RAE = 1 << 2,
169 MTHCA_QP_OPTPAR_RWE = 1 << 3,
170 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
171 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
172 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
173 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
174 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
175 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
176 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
177 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
178 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
179 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
180 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
181 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
182 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
185 static const u8 mthca_opcode[] = {
186 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
187 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
188 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
189 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
190 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
191 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
192 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
195 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
197 return qp->qpn >= dev->qp_table.sqp_start &&
198 qp->qpn <= dev->qp_table.sqp_start + 3;
201 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
203 return qp->qpn >= dev->qp_table.sqp_start &&
204 qp->qpn <= dev->qp_table.sqp_start + 1;
207 static void *get_recv_wqe(struct mthca_qp *qp, int n)
210 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
212 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
213 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
216 static void *get_send_wqe(struct mthca_qp *qp, int n)
219 return qp->queue.direct.buf + qp->send_wqe_offset +
220 (n << qp->sq.wqe_shift);
222 return qp->queue.page_list[(qp->send_wqe_offset +
223 (n << qp->sq.wqe_shift)) >>
225 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
229 static void mthca_wq_reset(struct mthca_wq *wq)
232 wq->last_comp = wq->max - 1;
237 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
238 enum ib_event_type event_type)
241 struct ib_event event;
243 spin_lock(&dev->qp_table.lock);
244 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
247 spin_unlock(&dev->qp_table.lock);
250 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
254 if (event_type == IB_EVENT_PATH_MIG)
255 qp->port = qp->alt_port;
257 event.device = &dev->ib_dev;
258 event.event = event_type;
259 event.element.qp = &qp->ibqp;
260 if (qp->ibqp.event_handler)
261 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
263 spin_lock(&dev->qp_table.lock);
266 spin_unlock(&dev->qp_table.lock);
269 static int to_mthca_state(enum ib_qp_state ib_state)
272 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
273 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
274 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
275 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
276 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
277 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
278 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
283 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
285 static int to_mthca_st(int transport)
288 case RC: return MTHCA_QP_ST_RC;
289 case UC: return MTHCA_QP_ST_UC;
290 case UD: return MTHCA_QP_ST_UD;
291 case RD: return MTHCA_QP_ST_RD;
292 case MLX: return MTHCA_QP_ST_MLX;
297 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
300 if (attr_mask & IB_QP_PKEY_INDEX)
301 sqp->pkey_index = attr->pkey_index;
302 if (attr_mask & IB_QP_QKEY)
303 sqp->qkey = attr->qkey;
304 if (attr_mask & IB_QP_SQ_PSN)
305 sqp->send_psn = attr->sq_psn;
308 static void init_port(struct mthca_dev *dev, int port)
312 struct mthca_init_ib_param param;
314 memset(¶m, 0, sizeof param);
316 param.port_width = dev->limits.port_width_cap;
317 param.vl_cap = dev->limits.vl_cap;
318 param.mtu_cap = dev->limits.mtu_cap;
319 param.gid_cap = dev->limits.gid_table_len;
320 param.pkey_cap = dev->limits.pkey_table_len;
322 err = mthca_INIT_IB(dev, ¶m, port, &status);
324 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
326 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
329 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
334 u32 hw_access_flags = 0;
336 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
337 dest_rd_atomic = attr->max_dest_rd_atomic;
339 dest_rd_atomic = qp->resp_depth;
341 if (attr_mask & IB_QP_ACCESS_FLAGS)
342 access_flags = attr->qp_access_flags;
344 access_flags = qp->atomic_rd_en;
347 access_flags &= IB_ACCESS_REMOTE_WRITE;
349 if (access_flags & IB_ACCESS_REMOTE_READ)
350 hw_access_flags |= MTHCA_QP_BIT_RRE;
351 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
352 hw_access_flags |= MTHCA_QP_BIT_RAE;
353 if (access_flags & IB_ACCESS_REMOTE_WRITE)
354 hw_access_flags |= MTHCA_QP_BIT_RWE;
356 return cpu_to_be32(hw_access_flags);
359 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
361 switch (mthca_state) {
362 case MTHCA_QP_STATE_RST: return IB_QPS_RESET;
363 case MTHCA_QP_STATE_INIT: return IB_QPS_INIT;
364 case MTHCA_QP_STATE_RTR: return IB_QPS_RTR;
365 case MTHCA_QP_STATE_RTS: return IB_QPS_RTS;
366 case MTHCA_QP_STATE_DRAINING:
367 case MTHCA_QP_STATE_SQD: return IB_QPS_SQD;
368 case MTHCA_QP_STATE_SQE: return IB_QPS_SQE;
369 case MTHCA_QP_STATE_ERR: return IB_QPS_ERR;
374 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
376 switch (mthca_mig_state) {
377 case 0: return IB_MIG_ARMED;
378 case 1: return IB_MIG_REARM;
379 case 3: return IB_MIG_MIGRATED;
384 static int to_ib_qp_access_flags(int mthca_flags)
388 if (mthca_flags & MTHCA_QP_BIT_RRE)
389 ib_flags |= IB_ACCESS_REMOTE_READ;
390 if (mthca_flags & MTHCA_QP_BIT_RWE)
391 ib_flags |= IB_ACCESS_REMOTE_WRITE;
392 if (mthca_flags & MTHCA_QP_BIT_RAE)
393 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
398 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
399 struct mthca_qp_path *path)
401 memset(ib_ah_attr, 0, sizeof *path);
402 ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
404 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
407 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
408 ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
409 ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
410 ib_ah_attr->static_rate = mthca_rate_to_ib(dev,
411 path->static_rate & 0x7,
412 ib_ah_attr->port_num);
413 ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
414 if (ib_ah_attr->ah_flags) {
415 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
416 ib_ah_attr->grh.hop_limit = path->hop_limit;
417 ib_ah_attr->grh.traffic_class =
418 (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
419 ib_ah_attr->grh.flow_label =
420 be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
421 memcpy(ib_ah_attr->grh.dgid.raw,
422 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
426 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
427 struct ib_qp_init_attr *qp_init_attr)
429 struct mthca_dev *dev = to_mdev(ibqp->device);
430 struct mthca_qp *qp = to_mqp(ibqp);
432 struct mthca_mailbox *mailbox;
433 struct mthca_qp_param *qp_param;
434 struct mthca_qp_context *context;
438 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
440 return PTR_ERR(mailbox);
442 err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
446 mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
451 qp_param = mailbox->buf;
452 context = &qp_param->context;
453 mthca_state = be32_to_cpu(context->flags) >> 28;
455 qp_attr->qp_state = to_ib_qp_state(mthca_state);
456 qp_attr->cur_qp_state = qp_attr->qp_state;
457 qp_attr->path_mtu = context->mtu_msgmax >> 5;
458 qp_attr->path_mig_state =
459 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
460 qp_attr->qkey = be32_to_cpu(context->qkey);
461 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
462 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
463 qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff;
464 qp_attr->qp_access_flags =
465 to_ib_qp_access_flags(be32_to_cpu(context->params2));
466 qp_attr->cap.max_send_wr = qp->sq.max;
467 qp_attr->cap.max_recv_wr = qp->rq.max;
468 qp_attr->cap.max_send_sge = qp->sq.max_gs;
469 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
470 qp_attr->cap.max_inline_data = qp->max_inline_data;
472 if (qp->transport == RC || qp->transport == UC) {
473 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
474 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
477 qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
478 qp_attr->alt_pkey_index = be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
480 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
481 qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
483 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
485 qp_attr->max_dest_rd_atomic =
486 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
487 qp_attr->min_rnr_timer =
488 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
489 qp_attr->port_num = qp_attr->ah_attr.port_num;
490 qp_attr->timeout = context->pri_path.ackto >> 3;
491 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
492 qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5;
493 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
494 qp_attr->alt_timeout = context->alt_path.ackto >> 3;
495 qp_init_attr->cap = qp_attr->cap;
498 mthca_free_mailbox(dev, mailbox);
502 static int mthca_path_set(struct mthca_dev *dev, struct ib_ah_attr *ah,
503 struct mthca_qp_path *path, u8 port)
505 path->g_mylmc = ah->src_path_bits & 0x7f;
506 path->rlid = cpu_to_be16(ah->dlid);
507 path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
509 if (ah->ah_flags & IB_AH_GRH) {
510 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
511 mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
512 ah->grh.sgid_index, dev->limits.gid_table_len-1);
516 path->g_mylmc |= 1 << 7;
517 path->mgid_index = ah->grh.sgid_index;
518 path->hop_limit = ah->grh.hop_limit;
519 path->sl_tclass_flowlabel =
520 cpu_to_be32((ah->sl << 28) |
521 (ah->grh.traffic_class << 20) |
522 (ah->grh.flow_label));
523 memcpy(path->rgid, ah->grh.dgid.raw, 16);
525 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
530 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
532 struct mthca_dev *dev = to_mdev(ibqp->device);
533 struct mthca_qp *qp = to_mqp(ibqp);
534 enum ib_qp_state cur_state, new_state;
535 struct mthca_mailbox *mailbox;
536 struct mthca_qp_param *qp_param;
537 struct mthca_qp_context *qp_context;
542 mutex_lock(&qp->mutex);
544 if (attr_mask & IB_QP_CUR_STATE) {
545 cur_state = attr->cur_qp_state;
547 spin_lock_irq(&qp->sq.lock);
548 spin_lock(&qp->rq.lock);
549 cur_state = qp->state;
550 spin_unlock(&qp->rq.lock);
551 spin_unlock_irq(&qp->sq.lock);
554 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
556 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
557 mthca_dbg(dev, "Bad QP transition (transport %d) "
558 "%d->%d with attr 0x%08x\n",
559 qp->transport, cur_state, new_state,
564 if ((attr_mask & IB_QP_PKEY_INDEX) &&
565 attr->pkey_index >= dev->limits.pkey_table_len) {
566 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
567 attr->pkey_index, dev->limits.pkey_table_len-1);
571 if ((attr_mask & IB_QP_PORT) &&
572 (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
573 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
577 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
578 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
579 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
580 attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
584 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
585 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
586 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
587 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
591 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
592 if (IS_ERR(mailbox)) {
593 err = PTR_ERR(mailbox);
596 qp_param = mailbox->buf;
597 qp_context = &qp_param->context;
598 memset(qp_param, 0, sizeof *qp_param);
600 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
601 (to_mthca_st(qp->transport) << 16));
602 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
603 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
604 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
606 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
607 switch (attr->path_mig_state) {
608 case IB_MIG_MIGRATED:
609 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
612 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
615 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
620 /* leave tavor_sched_queue as 0 */
622 if (qp->transport == MLX || qp->transport == UD)
623 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
624 else if (attr_mask & IB_QP_PATH_MTU) {
625 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
626 mthca_dbg(dev, "path MTU (%u) is invalid\n",
630 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
633 if (mthca_is_memfree(dev)) {
635 qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
636 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
639 qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
640 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
643 /* leave arbel_sched_queue as 0 */
645 if (qp->ibqp.uobject)
646 qp_context->usr_page =
647 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
649 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
650 qp_context->local_qpn = cpu_to_be32(qp->qpn);
651 if (attr_mask & IB_QP_DEST_QPN) {
652 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
655 if (qp->transport == MLX)
656 qp_context->pri_path.port_pkey |=
657 cpu_to_be32(qp->port << 24);
659 if (attr_mask & IB_QP_PORT) {
660 qp_context->pri_path.port_pkey |=
661 cpu_to_be32(attr->port_num << 24);
662 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
666 if (attr_mask & IB_QP_PKEY_INDEX) {
667 qp_context->pri_path.port_pkey |=
668 cpu_to_be32(attr->pkey_index);
669 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
672 if (attr_mask & IB_QP_RNR_RETRY) {
673 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
674 attr->rnr_retry << 5;
675 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
676 MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
679 if (attr_mask & IB_QP_AV) {
680 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
681 attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
684 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
687 if (attr_mask & IB_QP_TIMEOUT) {
688 qp_context->pri_path.ackto = attr->timeout << 3;
689 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
692 if (attr_mask & IB_QP_ALT_PATH) {
693 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
694 mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
695 attr->alt_pkey_index, dev->limits.pkey_table_len-1);
699 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
700 mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
705 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
706 attr->alt_ah_attr.port_num))
709 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
710 attr->alt_port_num << 24);
711 qp_context->alt_path.ackto = attr->alt_timeout << 3;
712 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
716 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
717 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
718 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
719 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
720 (MTHCA_FLIGHT_LIMIT << 24) |
722 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
723 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
724 if (attr_mask & IB_QP_RETRY_CNT) {
725 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
726 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
729 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
730 if (attr->max_rd_atomic) {
731 qp_context->params1 |=
732 cpu_to_be32(MTHCA_QP_BIT_SRE |
734 qp_context->params1 |=
735 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
737 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
740 if (attr_mask & IB_QP_SQ_PSN)
741 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
742 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
744 if (mthca_is_memfree(dev)) {
745 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
746 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
749 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
750 if (attr->max_dest_rd_atomic)
751 qp_context->params2 |=
752 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
754 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
757 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
758 qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask);
759 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
760 MTHCA_QP_OPTPAR_RRE |
761 MTHCA_QP_OPTPAR_RAE);
764 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
767 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
769 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
770 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
771 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
773 if (attr_mask & IB_QP_RQ_PSN)
774 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
776 qp_context->ra_buff_indx =
777 cpu_to_be32(dev->qp_table.rdb_base +
778 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
779 dev->qp_table.rdb_shift));
781 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
783 if (mthca_is_memfree(dev))
784 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
786 if (attr_mask & IB_QP_QKEY) {
787 qp_context->qkey = cpu_to_be32(attr->qkey);
788 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
792 qp_context->srqn = cpu_to_be32(1 << 24 |
793 to_msrq(ibqp->srq)->srqn);
795 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
796 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY &&
797 attr->en_sqd_async_notify)
800 err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
801 mailbox, sqd_event, &status);
805 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
806 cur_state, new_state, status);
811 qp->state = new_state;
812 if (attr_mask & IB_QP_ACCESS_FLAGS)
813 qp->atomic_rd_en = attr->qp_access_flags;
814 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
815 qp->resp_depth = attr->max_dest_rd_atomic;
816 if (attr_mask & IB_QP_PORT)
817 qp->port = attr->port_num;
818 if (attr_mask & IB_QP_ALT_PATH)
819 qp->alt_port = attr->alt_port_num;
822 store_attrs(to_msqp(qp), attr, attr_mask);
825 * If we moved QP0 to RTR, bring the IB link up; if we moved
826 * QP0 to RESET or ERROR, bring the link back down.
828 if (is_qp0(dev, qp)) {
829 if (cur_state != IB_QPS_RTR &&
830 new_state == IB_QPS_RTR)
831 init_port(dev, qp->port);
833 if (cur_state != IB_QPS_RESET &&
834 cur_state != IB_QPS_ERR &&
835 (new_state == IB_QPS_RESET ||
836 new_state == IB_QPS_ERR))
837 mthca_CLOSE_IB(dev, qp->port, &status);
841 * If we moved a kernel QP to RESET, clean up all old CQ
842 * entries and reinitialize the QP.
844 if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
845 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
846 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
847 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
848 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
849 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
851 mthca_wq_reset(&qp->sq);
852 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
854 mthca_wq_reset(&qp->rq);
855 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
857 if (mthca_is_memfree(dev)) {
864 mthca_free_mailbox(dev, mailbox);
867 mutex_unlock(&qp->mutex);
871 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
874 * Calculate the maximum size of WQE s/g segments, excluding
875 * the next segment and other non-data segments.
877 int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
879 switch (qp->transport) {
881 max_data_size -= 2 * sizeof (struct mthca_data_seg);
885 if (mthca_is_memfree(dev))
886 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
888 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
892 max_data_size -= sizeof (struct mthca_raddr_seg);
896 return max_data_size;
899 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
901 /* We don't support inline data for kernel QPs (yet). */
902 return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
905 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
909 int max_data_size = mthca_max_data_size(dev, qp,
910 min(dev->limits.max_desc_sz,
911 1 << qp->sq.wqe_shift));
913 qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
915 qp->sq.max_gs = min_t(int, dev->limits.max_sg,
916 max_data_size / sizeof (struct mthca_data_seg));
917 qp->rq.max_gs = min_t(int, dev->limits.max_sg,
918 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
919 sizeof (struct mthca_next_seg)) /
920 sizeof (struct mthca_data_seg));
924 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
925 * rq.max_gs and sq.max_gs must all be assigned.
926 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
927 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
930 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
937 size = sizeof (struct mthca_next_seg) +
938 qp->rq.max_gs * sizeof (struct mthca_data_seg);
940 if (size > dev->limits.max_desc_sz)
943 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
947 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
948 switch (qp->transport) {
950 size += 2 * sizeof (struct mthca_data_seg);
954 size += mthca_is_memfree(dev) ?
955 sizeof (struct mthca_arbel_ud_seg) :
956 sizeof (struct mthca_tavor_ud_seg);
960 size += sizeof (struct mthca_raddr_seg);
964 size += sizeof (struct mthca_raddr_seg);
966 * An atomic op will require an atomic segment, a
967 * remote address segment and one scatter entry.
969 size = max_t(int, size,
970 sizeof (struct mthca_atomic_seg) +
971 sizeof (struct mthca_raddr_seg) +
972 sizeof (struct mthca_data_seg));
979 /* Make sure that we have enough space for a bind request */
980 size = max_t(int, size, sizeof (struct mthca_bind_seg));
982 size += sizeof (struct mthca_next_seg);
984 if (size > dev->limits.max_desc_sz)
987 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
991 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
992 1 << qp->sq.wqe_shift);
995 * If this is a userspace QP, we don't actually have to
996 * allocate anything. All we need is to calculate the WQE
997 * sizes and the send_wqe_offset, so we're done now.
999 if (pd->ibpd.uobject)
1002 size = PAGE_ALIGN(qp->send_wqe_offset +
1003 (qp->sq.max << qp->sq.wqe_shift));
1005 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1010 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1011 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1022 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1023 struct mthca_qp *qp)
1025 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1026 (qp->sq.max << qp->sq.wqe_shift)),
1027 &qp->queue, qp->is_direct, &qp->mr);
1031 static int mthca_map_memfree(struct mthca_dev *dev,
1032 struct mthca_qp *qp)
1036 if (mthca_is_memfree(dev)) {
1037 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1041 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1045 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1046 qp->qpn << dev->qp_table.rdb_shift);
1055 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1058 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1063 static void mthca_unmap_memfree(struct mthca_dev *dev,
1064 struct mthca_qp *qp)
1066 mthca_table_put(dev, dev->qp_table.rdb_table,
1067 qp->qpn << dev->qp_table.rdb_shift);
1068 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1069 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1072 static int mthca_alloc_memfree(struct mthca_dev *dev,
1073 struct mthca_qp *qp)
1077 if (mthca_is_memfree(dev)) {
1078 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1079 qp->qpn, &qp->rq.db);
1080 if (qp->rq.db_index < 0)
1083 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1084 qp->qpn, &qp->sq.db);
1085 if (qp->sq.db_index < 0)
1086 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1092 static void mthca_free_memfree(struct mthca_dev *dev,
1093 struct mthca_qp *qp)
1095 if (mthca_is_memfree(dev)) {
1096 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1097 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1101 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1102 struct mthca_pd *pd,
1103 struct mthca_cq *send_cq,
1104 struct mthca_cq *recv_cq,
1105 enum ib_sig_type send_policy,
1106 struct mthca_qp *qp)
1112 init_waitqueue_head(&qp->wait);
1113 mutex_init(&qp->mutex);
1114 qp->state = IB_QPS_RESET;
1115 qp->atomic_rd_en = 0;
1117 qp->sq_policy = send_policy;
1118 mthca_wq_reset(&qp->sq);
1119 mthca_wq_reset(&qp->rq);
1121 spin_lock_init(&qp->sq.lock);
1122 spin_lock_init(&qp->rq.lock);
1124 ret = mthca_map_memfree(dev, qp);
1128 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1130 mthca_unmap_memfree(dev, qp);
1134 mthca_adjust_qp_caps(dev, pd, qp);
1137 * If this is a userspace QP, we're done now. The doorbells
1138 * will be allocated and buffers will be initialized in
1141 if (pd->ibpd.uobject)
1144 ret = mthca_alloc_memfree(dev, qp);
1146 mthca_free_wqe_buf(dev, qp);
1147 mthca_unmap_memfree(dev, qp);
1151 if (mthca_is_memfree(dev)) {
1152 struct mthca_next_seg *next;
1153 struct mthca_data_seg *scatter;
1154 int size = (sizeof (struct mthca_next_seg) +
1155 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1157 for (i = 0; i < qp->rq.max; ++i) {
1158 next = get_recv_wqe(qp, i);
1159 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1161 next->ee_nds = cpu_to_be32(size);
1163 for (scatter = (void *) (next + 1);
1164 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1166 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1169 for (i = 0; i < qp->sq.max; ++i) {
1170 next = get_send_wqe(qp, i);
1171 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1173 qp->send_wqe_offset);
1177 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1178 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1183 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1184 struct mthca_pd *pd, struct mthca_qp *qp)
1186 int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1188 /* Sanity check QP size before proceeding */
1189 if (cap->max_send_wr > dev->limits.max_wqes ||
1190 cap->max_recv_wr > dev->limits.max_wqes ||
1191 cap->max_send_sge > dev->limits.max_sg ||
1192 cap->max_recv_sge > dev->limits.max_sg ||
1193 cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1197 * For MLX transport we need 2 extra S/G entries:
1198 * one for the header and one for the checksum at the end
1200 if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1203 if (mthca_is_memfree(dev)) {
1204 qp->rq.max = cap->max_recv_wr ?
1205 roundup_pow_of_two(cap->max_recv_wr) : 0;
1206 qp->sq.max = cap->max_send_wr ?
1207 roundup_pow_of_two(cap->max_send_wr) : 0;
1209 qp->rq.max = cap->max_recv_wr;
1210 qp->sq.max = cap->max_send_wr;
1213 qp->rq.max_gs = cap->max_recv_sge;
1214 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1215 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1216 MTHCA_INLINE_CHUNK_SIZE) /
1217 sizeof (struct mthca_data_seg));
1222 int mthca_alloc_qp(struct mthca_dev *dev,
1223 struct mthca_pd *pd,
1224 struct mthca_cq *send_cq,
1225 struct mthca_cq *recv_cq,
1226 enum ib_qp_type type,
1227 enum ib_sig_type send_policy,
1228 struct ib_qp_cap *cap,
1229 struct mthca_qp *qp)
1234 case IB_QPT_RC: qp->transport = RC; break;
1235 case IB_QPT_UC: qp->transport = UC; break;
1236 case IB_QPT_UD: qp->transport = UD; break;
1237 default: return -EINVAL;
1240 err = mthca_set_qp_size(dev, cap, pd, qp);
1244 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1248 /* initialize port to zero for error-catching. */
1251 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1254 mthca_free(&dev->qp_table.alloc, qp->qpn);
1258 spin_lock_irq(&dev->qp_table.lock);
1259 mthca_array_set(&dev->qp_table.qp,
1260 qp->qpn & (dev->limits.num_qps - 1), qp);
1261 spin_unlock_irq(&dev->qp_table.lock);
1266 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1268 if (send_cq == recv_cq)
1269 spin_lock_irq(&send_cq->lock);
1270 else if (send_cq->cqn < recv_cq->cqn) {
1271 spin_lock_irq(&send_cq->lock);
1272 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1274 spin_lock_irq(&recv_cq->lock);
1275 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1279 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1281 if (send_cq == recv_cq)
1282 spin_unlock_irq(&send_cq->lock);
1283 else if (send_cq->cqn < recv_cq->cqn) {
1284 spin_unlock(&recv_cq->lock);
1285 spin_unlock_irq(&send_cq->lock);
1287 spin_unlock(&send_cq->lock);
1288 spin_unlock_irq(&recv_cq->lock);
1292 int mthca_alloc_sqp(struct mthca_dev *dev,
1293 struct mthca_pd *pd,
1294 struct mthca_cq *send_cq,
1295 struct mthca_cq *recv_cq,
1296 enum ib_sig_type send_policy,
1297 struct ib_qp_cap *cap,
1300 struct mthca_sqp *sqp)
1302 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1305 sqp->qp.transport = MLX;
1306 err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1310 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1311 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1312 &sqp->header_dma, GFP_KERNEL);
1313 if (!sqp->header_buf)
1316 spin_lock_irq(&dev->qp_table.lock);
1317 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1320 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1321 spin_unlock_irq(&dev->qp_table.lock);
1326 sqp->qp.port = port;
1328 sqp->qp.transport = MLX;
1330 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1331 send_policy, &sqp->qp);
1335 atomic_inc(&pd->sqp_count);
1341 * Lock CQs here, so that CQ polling code can do QP lookup
1342 * without taking a lock.
1344 mthca_lock_cqs(send_cq, recv_cq);
1346 spin_lock(&dev->qp_table.lock);
1347 mthca_array_clear(&dev->qp_table.qp, mqpn);
1348 spin_unlock(&dev->qp_table.lock);
1350 mthca_unlock_cqs(send_cq, recv_cq);
1353 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1354 sqp->header_buf, sqp->header_dma);
1359 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1363 spin_lock_irq(&dev->qp_table.lock);
1365 spin_unlock_irq(&dev->qp_table.lock);
1370 void mthca_free_qp(struct mthca_dev *dev,
1371 struct mthca_qp *qp)
1374 struct mthca_cq *send_cq;
1375 struct mthca_cq *recv_cq;
1377 send_cq = to_mcq(qp->ibqp.send_cq);
1378 recv_cq = to_mcq(qp->ibqp.recv_cq);
1381 * Lock CQs here, so that CQ polling code can do QP lookup
1382 * without taking a lock.
1384 mthca_lock_cqs(send_cq, recv_cq);
1386 spin_lock(&dev->qp_table.lock);
1387 mthca_array_clear(&dev->qp_table.qp,
1388 qp->qpn & (dev->limits.num_qps - 1));
1390 spin_unlock(&dev->qp_table.lock);
1392 mthca_unlock_cqs(send_cq, recv_cq);
1394 wait_event(qp->wait, !get_qp_refcount(dev, qp));
1396 if (qp->state != IB_QPS_RESET)
1397 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1401 * If this is a userspace QP, the buffers, MR, CQs and so on
1402 * will be cleaned up in userspace, so all we have to do is
1403 * unref the mem-free tables and free the QPN in our table.
1405 if (!qp->ibqp.uobject) {
1406 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn,
1407 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1408 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1409 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
1410 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1412 mthca_free_memfree(dev, qp);
1413 mthca_free_wqe_buf(dev, qp);
1416 mthca_unmap_memfree(dev, qp);
1418 if (is_sqp(dev, qp)) {
1419 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1420 dma_free_coherent(&dev->pdev->dev,
1421 to_msqp(qp)->header_buf_size,
1422 to_msqp(qp)->header_buf,
1423 to_msqp(qp)->header_dma);
1425 mthca_free(&dev->qp_table.alloc, qp->qpn);
1428 /* Create UD header for an MLX send and build a data segment for it */
1429 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1430 int ind, struct ib_send_wr *wr,
1431 struct mthca_mlx_seg *mlx,
1432 struct mthca_data_seg *data)
1438 ib_ud_header_init(256, /* assume a MAD */
1439 mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1442 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1445 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1446 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1447 (sqp->ud_header.lrh.destination_lid ==
1448 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1449 (sqp->ud_header.lrh.service_level << 8));
1450 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1453 switch (wr->opcode) {
1455 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1456 sqp->ud_header.immediate_present = 0;
1458 case IB_WR_SEND_WITH_IMM:
1459 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1460 sqp->ud_header.immediate_present = 1;
1461 sqp->ud_header.immediate_data = wr->imm_data;
1467 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1468 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1469 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1470 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1471 if (!sqp->qp.ibqp.qp_num)
1472 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1473 sqp->pkey_index, &pkey);
1475 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1476 wr->wr.ud.pkey_index, &pkey);
1477 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1478 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1479 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1480 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1481 sqp->qkey : wr->wr.ud.remote_qkey);
1482 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1484 header_size = ib_ud_header_pack(&sqp->ud_header,
1486 ind * MTHCA_UD_HEADER_SIZE);
1488 data->byte_count = cpu_to_be32(header_size);
1489 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1490 data->addr = cpu_to_be64(sqp->header_dma +
1491 ind * MTHCA_UD_HEADER_SIZE);
1496 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1497 struct ib_cq *ib_cq)
1500 struct mthca_cq *cq;
1502 cur = wq->head - wq->tail;
1503 if (likely(cur + nreq < wq->max))
1507 spin_lock(&cq->lock);
1508 cur = wq->head - wq->tail;
1509 spin_unlock(&cq->lock);
1511 return cur + nreq >= wq->max;
1514 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1515 struct ib_send_wr **bad_wr)
1517 struct mthca_dev *dev = to_mdev(ibqp->device);
1518 struct mthca_qp *qp = to_mqp(ibqp);
1521 unsigned long flags;
1531 spin_lock_irqsave(&qp->sq.lock, flags);
1533 /* XXX check that state is OK to post send */
1535 ind = qp->sq.next_ind;
1537 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1538 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1539 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1540 " %d max, %d nreq)\n", qp->qpn,
1541 qp->sq.head, qp->sq.tail,
1548 wqe = get_send_wqe(qp, ind);
1549 prev_wqe = qp->sq.last;
1552 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1553 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1554 ((struct mthca_next_seg *) wqe)->flags =
1555 ((wr->send_flags & IB_SEND_SIGNALED) ?
1556 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1557 ((wr->send_flags & IB_SEND_SOLICITED) ?
1558 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1560 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1561 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1562 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1564 wqe += sizeof (struct mthca_next_seg);
1565 size = sizeof (struct mthca_next_seg) / 16;
1567 switch (qp->transport) {
1569 switch (wr->opcode) {
1570 case IB_WR_ATOMIC_CMP_AND_SWP:
1571 case IB_WR_ATOMIC_FETCH_AND_ADD:
1572 ((struct mthca_raddr_seg *) wqe)->raddr =
1573 cpu_to_be64(wr->wr.atomic.remote_addr);
1574 ((struct mthca_raddr_seg *) wqe)->rkey =
1575 cpu_to_be32(wr->wr.atomic.rkey);
1576 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1578 wqe += sizeof (struct mthca_raddr_seg);
1580 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1581 ((struct mthca_atomic_seg *) wqe)->swap_add =
1582 cpu_to_be64(wr->wr.atomic.swap);
1583 ((struct mthca_atomic_seg *) wqe)->compare =
1584 cpu_to_be64(wr->wr.atomic.compare_add);
1586 ((struct mthca_atomic_seg *) wqe)->swap_add =
1587 cpu_to_be64(wr->wr.atomic.compare_add);
1588 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1591 wqe += sizeof (struct mthca_atomic_seg);
1592 size += (sizeof (struct mthca_raddr_seg) +
1593 sizeof (struct mthca_atomic_seg)) / 16;
1596 case IB_WR_RDMA_WRITE:
1597 case IB_WR_RDMA_WRITE_WITH_IMM:
1598 case IB_WR_RDMA_READ:
1599 ((struct mthca_raddr_seg *) wqe)->raddr =
1600 cpu_to_be64(wr->wr.rdma.remote_addr);
1601 ((struct mthca_raddr_seg *) wqe)->rkey =
1602 cpu_to_be32(wr->wr.rdma.rkey);
1603 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1604 wqe += sizeof (struct mthca_raddr_seg);
1605 size += sizeof (struct mthca_raddr_seg) / 16;
1609 /* No extra segments required for sends */
1616 switch (wr->opcode) {
1617 case IB_WR_RDMA_WRITE:
1618 case IB_WR_RDMA_WRITE_WITH_IMM:
1619 ((struct mthca_raddr_seg *) wqe)->raddr =
1620 cpu_to_be64(wr->wr.rdma.remote_addr);
1621 ((struct mthca_raddr_seg *) wqe)->rkey =
1622 cpu_to_be32(wr->wr.rdma.rkey);
1623 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1624 wqe += sizeof (struct mthca_raddr_seg);
1625 size += sizeof (struct mthca_raddr_seg) / 16;
1629 /* No extra segments required for sends */
1636 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1637 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1638 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1639 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1640 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1641 cpu_to_be32(wr->wr.ud.remote_qpn);
1642 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1643 cpu_to_be32(wr->wr.ud.remote_qkey);
1645 wqe += sizeof (struct mthca_tavor_ud_seg);
1646 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1650 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1651 wqe - sizeof (struct mthca_next_seg),
1657 wqe += sizeof (struct mthca_data_seg);
1658 size += sizeof (struct mthca_data_seg) / 16;
1662 if (wr->num_sge > qp->sq.max_gs) {
1663 mthca_err(dev, "too many gathers\n");
1669 for (i = 0; i < wr->num_sge; ++i) {
1670 ((struct mthca_data_seg *) wqe)->byte_count =
1671 cpu_to_be32(wr->sg_list[i].length);
1672 ((struct mthca_data_seg *) wqe)->lkey =
1673 cpu_to_be32(wr->sg_list[i].lkey);
1674 ((struct mthca_data_seg *) wqe)->addr =
1675 cpu_to_be64(wr->sg_list[i].addr);
1676 wqe += sizeof (struct mthca_data_seg);
1677 size += sizeof (struct mthca_data_seg) / 16;
1680 /* Add one more inline data segment for ICRC */
1681 if (qp->transport == MLX) {
1682 ((struct mthca_data_seg *) wqe)->byte_count =
1683 cpu_to_be32((1 << 31) | 4);
1684 ((u32 *) wqe)[1] = 0;
1685 wqe += sizeof (struct mthca_data_seg);
1686 size += sizeof (struct mthca_data_seg) / 16;
1689 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1691 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1692 mthca_err(dev, "opcode invalid\n");
1698 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1699 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1700 qp->send_wqe_offset) |
1701 mthca_opcode[wr->opcode]);
1703 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1704 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size |
1705 ((wr->send_flags & IB_SEND_FENCE) ?
1706 MTHCA_NEXT_FENCE : 0));
1710 op0 = mthca_opcode[wr->opcode];
1711 f0 = wr->send_flags & IB_SEND_FENCE ?
1712 MTHCA_SEND_DOORBELL_FENCE : 0;
1716 if (unlikely(ind >= qp->sq.max))
1724 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1725 qp->send_wqe_offset) | f0 | op0);
1726 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1730 mthca_write64(doorbell,
1731 dev->kar + MTHCA_SEND_DOORBELL,
1732 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1735 qp->sq.next_ind = ind;
1736 qp->sq.head += nreq;
1738 spin_unlock_irqrestore(&qp->sq.lock, flags);
1742 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1743 struct ib_recv_wr **bad_wr)
1745 struct mthca_dev *dev = to_mdev(ibqp->device);
1746 struct mthca_qp *qp = to_mqp(ibqp);
1748 unsigned long flags;
1758 spin_lock_irqsave(&qp->rq.lock, flags);
1760 /* XXX check that state is OK to post receive */
1762 ind = qp->rq.next_ind;
1764 for (nreq = 0; wr; wr = wr->next) {
1765 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1766 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1767 " %d max, %d nreq)\n", qp->qpn,
1768 qp->rq.head, qp->rq.tail,
1775 wqe = get_recv_wqe(qp, ind);
1776 prev_wqe = qp->rq.last;
1779 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1780 ((struct mthca_next_seg *) wqe)->ee_nds =
1781 cpu_to_be32(MTHCA_NEXT_DBD);
1782 ((struct mthca_next_seg *) wqe)->flags = 0;
1784 wqe += sizeof (struct mthca_next_seg);
1785 size = sizeof (struct mthca_next_seg) / 16;
1787 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1793 for (i = 0; i < wr->num_sge; ++i) {
1794 ((struct mthca_data_seg *) wqe)->byte_count =
1795 cpu_to_be32(wr->sg_list[i].length);
1796 ((struct mthca_data_seg *) wqe)->lkey =
1797 cpu_to_be32(wr->sg_list[i].lkey);
1798 ((struct mthca_data_seg *) wqe)->addr =
1799 cpu_to_be64(wr->sg_list[i].addr);
1800 wqe += sizeof (struct mthca_data_seg);
1801 size += sizeof (struct mthca_data_seg) / 16;
1804 qp->wrid[ind] = wr->wr_id;
1806 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1807 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1809 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1810 cpu_to_be32(MTHCA_NEXT_DBD | size);
1816 if (unlikely(ind >= qp->rq.max))
1820 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1823 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1824 doorbell[1] = cpu_to_be32(qp->qpn << 8);
1828 mthca_write64(doorbell,
1829 dev->kar + MTHCA_RECEIVE_DOORBELL,
1830 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1832 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1839 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1840 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1844 mthca_write64(doorbell,
1845 dev->kar + MTHCA_RECEIVE_DOORBELL,
1846 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1849 qp->rq.next_ind = ind;
1850 qp->rq.head += nreq;
1852 spin_unlock_irqrestore(&qp->rq.lock, flags);
1856 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1857 struct ib_send_wr **bad_wr)
1859 struct mthca_dev *dev = to_mdev(ibqp->device);
1860 struct mthca_qp *qp = to_mqp(ibqp);
1864 unsigned long flags;
1874 spin_lock_irqsave(&qp->sq.lock, flags);
1876 /* XXX check that state is OK to post send */
1878 ind = qp->sq.head & (qp->sq.max - 1);
1880 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1881 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1884 doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1885 ((qp->sq.head & 0xffff) << 8) |
1887 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1889 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1893 * Make sure that descriptors are written before
1897 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1900 * Make sure doorbell record is written before we
1901 * write MMIO send doorbell.
1904 mthca_write64(doorbell,
1905 dev->kar + MTHCA_SEND_DOORBELL,
1906 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1909 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1910 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1911 " %d max, %d nreq)\n", qp->qpn,
1912 qp->sq.head, qp->sq.tail,
1919 wqe = get_send_wqe(qp, ind);
1920 prev_wqe = qp->sq.last;
1923 ((struct mthca_next_seg *) wqe)->flags =
1924 ((wr->send_flags & IB_SEND_SIGNALED) ?
1925 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1926 ((wr->send_flags & IB_SEND_SOLICITED) ?
1927 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1929 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1930 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1931 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1933 wqe += sizeof (struct mthca_next_seg);
1934 size = sizeof (struct mthca_next_seg) / 16;
1936 switch (qp->transport) {
1938 switch (wr->opcode) {
1939 case IB_WR_ATOMIC_CMP_AND_SWP:
1940 case IB_WR_ATOMIC_FETCH_AND_ADD:
1941 ((struct mthca_raddr_seg *) wqe)->raddr =
1942 cpu_to_be64(wr->wr.atomic.remote_addr);
1943 ((struct mthca_raddr_seg *) wqe)->rkey =
1944 cpu_to_be32(wr->wr.atomic.rkey);
1945 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1947 wqe += sizeof (struct mthca_raddr_seg);
1949 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1950 ((struct mthca_atomic_seg *) wqe)->swap_add =
1951 cpu_to_be64(wr->wr.atomic.swap);
1952 ((struct mthca_atomic_seg *) wqe)->compare =
1953 cpu_to_be64(wr->wr.atomic.compare_add);
1955 ((struct mthca_atomic_seg *) wqe)->swap_add =
1956 cpu_to_be64(wr->wr.atomic.compare_add);
1957 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1960 wqe += sizeof (struct mthca_atomic_seg);
1961 size += (sizeof (struct mthca_raddr_seg) +
1962 sizeof (struct mthca_atomic_seg)) / 16;
1965 case IB_WR_RDMA_READ:
1966 case IB_WR_RDMA_WRITE:
1967 case IB_WR_RDMA_WRITE_WITH_IMM:
1968 ((struct mthca_raddr_seg *) wqe)->raddr =
1969 cpu_to_be64(wr->wr.rdma.remote_addr);
1970 ((struct mthca_raddr_seg *) wqe)->rkey =
1971 cpu_to_be32(wr->wr.rdma.rkey);
1972 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1973 wqe += sizeof (struct mthca_raddr_seg);
1974 size += sizeof (struct mthca_raddr_seg) / 16;
1978 /* No extra segments required for sends */
1985 switch (wr->opcode) {
1986 case IB_WR_RDMA_WRITE:
1987 case IB_WR_RDMA_WRITE_WITH_IMM:
1988 ((struct mthca_raddr_seg *) wqe)->raddr =
1989 cpu_to_be64(wr->wr.rdma.remote_addr);
1990 ((struct mthca_raddr_seg *) wqe)->rkey =
1991 cpu_to_be32(wr->wr.rdma.rkey);
1992 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1993 wqe += sizeof (struct mthca_raddr_seg);
1994 size += sizeof (struct mthca_raddr_seg) / 16;
1998 /* No extra segments required for sends */
2005 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
2006 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
2007 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
2008 cpu_to_be32(wr->wr.ud.remote_qpn);
2009 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
2010 cpu_to_be32(wr->wr.ud.remote_qkey);
2012 wqe += sizeof (struct mthca_arbel_ud_seg);
2013 size += sizeof (struct mthca_arbel_ud_seg) / 16;
2017 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2018 wqe - sizeof (struct mthca_next_seg),
2024 wqe += sizeof (struct mthca_data_seg);
2025 size += sizeof (struct mthca_data_seg) / 16;
2029 if (wr->num_sge > qp->sq.max_gs) {
2030 mthca_err(dev, "too many gathers\n");
2036 for (i = 0; i < wr->num_sge; ++i) {
2037 ((struct mthca_data_seg *) wqe)->byte_count =
2038 cpu_to_be32(wr->sg_list[i].length);
2039 ((struct mthca_data_seg *) wqe)->lkey =
2040 cpu_to_be32(wr->sg_list[i].lkey);
2041 ((struct mthca_data_seg *) wqe)->addr =
2042 cpu_to_be64(wr->sg_list[i].addr);
2043 wqe += sizeof (struct mthca_data_seg);
2044 size += sizeof (struct mthca_data_seg) / 16;
2047 /* Add one more inline data segment for ICRC */
2048 if (qp->transport == MLX) {
2049 ((struct mthca_data_seg *) wqe)->byte_count =
2050 cpu_to_be32((1 << 31) | 4);
2051 ((u32 *) wqe)[1] = 0;
2052 wqe += sizeof (struct mthca_data_seg);
2053 size += sizeof (struct mthca_data_seg) / 16;
2056 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2058 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2059 mthca_err(dev, "opcode invalid\n");
2065 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2066 cpu_to_be32(((ind << qp->sq.wqe_shift) +
2067 qp->send_wqe_offset) |
2068 mthca_opcode[wr->opcode]);
2070 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2071 cpu_to_be32(MTHCA_NEXT_DBD | size |
2072 ((wr->send_flags & IB_SEND_FENCE) ?
2073 MTHCA_NEXT_FENCE : 0));
2077 op0 = mthca_opcode[wr->opcode];
2078 f0 = wr->send_flags & IB_SEND_FENCE ?
2079 MTHCA_SEND_DOORBELL_FENCE : 0;
2083 if (unlikely(ind >= qp->sq.max))
2089 doorbell[0] = cpu_to_be32((nreq << 24) |
2090 ((qp->sq.head & 0xffff) << 8) |
2092 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2094 qp->sq.head += nreq;
2097 * Make sure that descriptors are written before
2101 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2104 * Make sure doorbell record is written before we
2105 * write MMIO send doorbell.
2108 mthca_write64(doorbell,
2109 dev->kar + MTHCA_SEND_DOORBELL,
2110 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2113 spin_unlock_irqrestore(&qp->sq.lock, flags);
2117 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2118 struct ib_recv_wr **bad_wr)
2120 struct mthca_dev *dev = to_mdev(ibqp->device);
2121 struct mthca_qp *qp = to_mqp(ibqp);
2122 unsigned long flags;
2129 spin_lock_irqsave(&qp->rq.lock, flags);
2131 /* XXX check that state is OK to post receive */
2133 ind = qp->rq.head & (qp->rq.max - 1);
2135 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2136 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2137 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2138 " %d max, %d nreq)\n", qp->qpn,
2139 qp->rq.head, qp->rq.tail,
2146 wqe = get_recv_wqe(qp, ind);
2148 ((struct mthca_next_seg *) wqe)->flags = 0;
2150 wqe += sizeof (struct mthca_next_seg);
2152 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2158 for (i = 0; i < wr->num_sge; ++i) {
2159 ((struct mthca_data_seg *) wqe)->byte_count =
2160 cpu_to_be32(wr->sg_list[i].length);
2161 ((struct mthca_data_seg *) wqe)->lkey =
2162 cpu_to_be32(wr->sg_list[i].lkey);
2163 ((struct mthca_data_seg *) wqe)->addr =
2164 cpu_to_be64(wr->sg_list[i].addr);
2165 wqe += sizeof (struct mthca_data_seg);
2168 if (i < qp->rq.max_gs) {
2169 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2170 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2171 ((struct mthca_data_seg *) wqe)->addr = 0;
2174 qp->wrid[ind] = wr->wr_id;
2177 if (unlikely(ind >= qp->rq.max))
2182 qp->rq.head += nreq;
2185 * Make sure that descriptors are written before
2189 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2192 spin_unlock_irqrestore(&qp->rq.lock, flags);
2196 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2197 int index, int *dbd, __be32 *new_wqe)
2199 struct mthca_next_seg *next;
2202 * For SRQs, all WQEs generate a CQE, so we're always at the
2203 * end of the doorbell chain.
2211 next = get_send_wqe(qp, index);
2213 next = get_recv_wqe(qp, index);
2215 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2216 if (next->ee_nds & cpu_to_be32(0x3f))
2217 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2218 (next->ee_nds & cpu_to_be32(0x3f));
2223 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2229 spin_lock_init(&dev->qp_table.lock);
2232 * We reserve 2 extra QPs per port for the special QPs. The
2233 * special QP for port 1 has to be even, so round up.
2235 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2236 err = mthca_alloc_init(&dev->qp_table.alloc,
2237 dev->limits.num_qps,
2239 dev->qp_table.sqp_start +
2240 MTHCA_MAX_PORTS * 2);
2244 err = mthca_array_init(&dev->qp_table.qp,
2245 dev->limits.num_qps);
2247 mthca_alloc_cleanup(&dev->qp_table.alloc);
2251 for (i = 0; i < 2; ++i) {
2252 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2253 dev->qp_table.sqp_start + i * 2,
2258 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2259 "status %02x, aborting.\n",
2268 for (i = 0; i < 2; ++i)
2269 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2271 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2272 mthca_alloc_cleanup(&dev->qp_table.alloc);
2277 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2282 for (i = 0; i < 2; ++i)
2283 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2285 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2286 mthca_alloc_cleanup(&dev->qp_table.alloc);