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/string.h>
39 #include <linux/slab.h>
40 #include <linux/sched.h>
44 #include <rdma/ib_verbs.h>
45 #include <rdma/ib_cache.h>
46 #include <rdma/ib_pack.h>
48 #include "mthca_dev.h"
49 #include "mthca_cmd.h"
50 #include "mthca_memfree.h"
51 #include "mthca_wqe.h"
54 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
55 MTHCA_ACK_REQ_FREQ = 10,
56 MTHCA_FLIGHT_LIMIT = 9,
57 MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */
58 MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */
59 MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */
63 MTHCA_QP_STATE_RST = 0,
64 MTHCA_QP_STATE_INIT = 1,
65 MTHCA_QP_STATE_RTR = 2,
66 MTHCA_QP_STATE_RTS = 3,
67 MTHCA_QP_STATE_SQE = 4,
68 MTHCA_QP_STATE_SQD = 5,
69 MTHCA_QP_STATE_ERR = 6,
70 MTHCA_QP_STATE_DRAINING = 7
82 MTHCA_QP_PM_MIGRATED = 0x3,
83 MTHCA_QP_PM_ARMED = 0x0,
84 MTHCA_QP_PM_REARM = 0x1
88 /* qp_context flags */
89 MTHCA_QP_BIT_DE = 1 << 8,
91 MTHCA_QP_BIT_SRE = 1 << 15,
92 MTHCA_QP_BIT_SWE = 1 << 14,
93 MTHCA_QP_BIT_SAE = 1 << 13,
94 MTHCA_QP_BIT_SIC = 1 << 4,
95 MTHCA_QP_BIT_SSC = 1 << 3,
97 MTHCA_QP_BIT_RRE = 1 << 15,
98 MTHCA_QP_BIT_RWE = 1 << 14,
99 MTHCA_QP_BIT_RAE = 1 << 13,
100 MTHCA_QP_BIT_RIC = 1 << 4,
101 MTHCA_QP_BIT_RSC = 1 << 3
105 MTHCA_SEND_DOORBELL_FENCE = 1 << 5
108 struct mthca_qp_path {
117 __be32 sl_tclass_flowlabel;
119 } __attribute__((packed));
121 struct mthca_qp_context {
123 __be32 tavor_sched_queue; /* Reserved on Arbel */
125 u8 rq_size_stride; /* Reserved on Tavor */
126 u8 sq_size_stride; /* Reserved on Tavor */
127 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
132 struct mthca_qp_path pri_path;
133 struct mthca_qp_path alt_path;
140 __be32 next_send_psn;
142 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
143 __be32 snd_db_index; /* (debugging only entries) */
144 __be32 last_acked_psn;
147 __be32 rnr_nextrecvpsn;
150 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
151 __be32 rcv_db_index; /* (debugging only entries) */
155 __be16 rq_wqe_counter; /* reserved on Tavor */
156 __be16 sq_wqe_counter; /* reserved on Tavor */
158 } __attribute__((packed));
160 struct mthca_qp_param {
161 __be32 opt_param_mask;
163 struct mthca_qp_context context;
165 } __attribute__((packed));
168 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
169 MTHCA_QP_OPTPAR_RRE = 1 << 1,
170 MTHCA_QP_OPTPAR_RAE = 1 << 2,
171 MTHCA_QP_OPTPAR_RWE = 1 << 3,
172 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
173 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
174 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
175 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
176 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
177 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
178 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
179 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
180 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
181 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
182 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
183 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
184 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
187 static const u8 mthca_opcode[] = {
188 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
189 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
190 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
191 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
192 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
193 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
194 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
197 static int is_sqp(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 + 3;
203 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
205 return qp->qpn >= dev->qp_table.sqp_start &&
206 qp->qpn <= dev->qp_table.sqp_start + 1;
209 static void *get_recv_wqe(struct mthca_qp *qp, int n)
212 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
214 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
215 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
218 static void *get_send_wqe(struct mthca_qp *qp, int n)
221 return qp->queue.direct.buf + qp->send_wqe_offset +
222 (n << qp->sq.wqe_shift);
224 return qp->queue.page_list[(qp->send_wqe_offset +
225 (n << qp->sq.wqe_shift)) >>
227 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
231 static void mthca_wq_reset(struct mthca_wq *wq)
234 wq->last_comp = wq->max - 1;
239 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
240 enum ib_event_type event_type)
243 struct ib_event event;
245 spin_lock(&dev->qp_table.lock);
246 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
249 spin_unlock(&dev->qp_table.lock);
252 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
256 if (event_type == IB_EVENT_PATH_MIG)
257 qp->port = qp->alt_port;
259 event.device = &dev->ib_dev;
260 event.event = event_type;
261 event.element.qp = &qp->ibqp;
262 if (qp->ibqp.event_handler)
263 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
265 spin_lock(&dev->qp_table.lock);
268 spin_unlock(&dev->qp_table.lock);
271 static int to_mthca_state(enum ib_qp_state ib_state)
274 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
275 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
276 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
277 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
278 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
279 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
280 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
285 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
287 static int to_mthca_st(int transport)
290 case RC: return MTHCA_QP_ST_RC;
291 case UC: return MTHCA_QP_ST_UC;
292 case UD: return MTHCA_QP_ST_UD;
293 case RD: return MTHCA_QP_ST_RD;
294 case MLX: return MTHCA_QP_ST_MLX;
299 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
302 if (attr_mask & IB_QP_PKEY_INDEX)
303 sqp->pkey_index = attr->pkey_index;
304 if (attr_mask & IB_QP_QKEY)
305 sqp->qkey = attr->qkey;
306 if (attr_mask & IB_QP_SQ_PSN)
307 sqp->send_psn = attr->sq_psn;
310 static void init_port(struct mthca_dev *dev, int port)
314 struct mthca_init_ib_param param;
316 memset(¶m, 0, sizeof param);
318 param.port_width = dev->limits.port_width_cap;
319 param.vl_cap = dev->limits.vl_cap;
320 param.mtu_cap = dev->limits.mtu_cap;
321 param.gid_cap = dev->limits.gid_table_len;
322 param.pkey_cap = dev->limits.pkey_table_len;
324 err = mthca_INIT_IB(dev, ¶m, port, &status);
326 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
328 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
331 static __be32 get_hw_access_flags(struct mthca_qp *qp, struct ib_qp_attr *attr,
336 u32 hw_access_flags = 0;
338 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
339 dest_rd_atomic = attr->max_dest_rd_atomic;
341 dest_rd_atomic = qp->resp_depth;
343 if (attr_mask & IB_QP_ACCESS_FLAGS)
344 access_flags = attr->qp_access_flags;
346 access_flags = qp->atomic_rd_en;
349 access_flags &= IB_ACCESS_REMOTE_WRITE;
351 if (access_flags & IB_ACCESS_REMOTE_READ)
352 hw_access_flags |= MTHCA_QP_BIT_RRE;
353 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
354 hw_access_flags |= MTHCA_QP_BIT_RAE;
355 if (access_flags & IB_ACCESS_REMOTE_WRITE)
356 hw_access_flags |= MTHCA_QP_BIT_RWE;
358 return cpu_to_be32(hw_access_flags);
361 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
363 switch (mthca_state) {
364 case MTHCA_QP_STATE_RST: return IB_QPS_RESET;
365 case MTHCA_QP_STATE_INIT: return IB_QPS_INIT;
366 case MTHCA_QP_STATE_RTR: return IB_QPS_RTR;
367 case MTHCA_QP_STATE_RTS: return IB_QPS_RTS;
368 case MTHCA_QP_STATE_DRAINING:
369 case MTHCA_QP_STATE_SQD: return IB_QPS_SQD;
370 case MTHCA_QP_STATE_SQE: return IB_QPS_SQE;
371 case MTHCA_QP_STATE_ERR: return IB_QPS_ERR;
376 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
378 switch (mthca_mig_state) {
379 case 0: return IB_MIG_ARMED;
380 case 1: return IB_MIG_REARM;
381 case 3: return IB_MIG_MIGRATED;
386 static int to_ib_qp_access_flags(int mthca_flags)
390 if (mthca_flags & MTHCA_QP_BIT_RRE)
391 ib_flags |= IB_ACCESS_REMOTE_READ;
392 if (mthca_flags & MTHCA_QP_BIT_RWE)
393 ib_flags |= IB_ACCESS_REMOTE_WRITE;
394 if (mthca_flags & MTHCA_QP_BIT_RAE)
395 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
400 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
401 struct mthca_qp_path *path)
403 memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
404 ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
406 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
409 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
410 ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
411 ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
412 ib_ah_attr->static_rate = mthca_rate_to_ib(dev,
413 path->static_rate & 0xf,
414 ib_ah_attr->port_num);
415 ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
416 if (ib_ah_attr->ah_flags) {
417 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
418 ib_ah_attr->grh.hop_limit = path->hop_limit;
419 ib_ah_attr->grh.traffic_class =
420 (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
421 ib_ah_attr->grh.flow_label =
422 be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
423 memcpy(ib_ah_attr->grh.dgid.raw,
424 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
428 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
429 struct ib_qp_init_attr *qp_init_attr)
431 struct mthca_dev *dev = to_mdev(ibqp->device);
432 struct mthca_qp *qp = to_mqp(ibqp);
434 struct mthca_mailbox *mailbox = NULL;
435 struct mthca_qp_param *qp_param;
436 struct mthca_qp_context *context;
440 if (qp->state == IB_QPS_RESET) {
441 qp_attr->qp_state = IB_QPS_RESET;
445 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
447 return PTR_ERR(mailbox);
449 err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status);
453 mthca_warn(dev, "QUERY_QP returned status %02x\n", status);
458 qp_param = mailbox->buf;
459 context = &qp_param->context;
460 mthca_state = be32_to_cpu(context->flags) >> 28;
462 qp_attr->qp_state = to_ib_qp_state(mthca_state);
463 qp_attr->path_mtu = context->mtu_msgmax >> 5;
464 qp_attr->path_mig_state =
465 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
466 qp_attr->qkey = be32_to_cpu(context->qkey);
467 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
468 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
469 qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff;
470 qp_attr->qp_access_flags =
471 to_ib_qp_access_flags(be32_to_cpu(context->params2));
473 if (qp->transport == RC || qp->transport == UC) {
474 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
475 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
476 qp_attr->alt_pkey_index =
477 be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
478 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
481 qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
483 (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
485 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
486 qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
488 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
490 qp_attr->max_dest_rd_atomic =
491 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
492 qp_attr->min_rnr_timer =
493 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
494 qp_attr->timeout = context->pri_path.ackto >> 3;
495 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
496 qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5;
497 qp_attr->alt_timeout = context->alt_path.ackto >> 3;
500 qp_attr->cur_qp_state = qp_attr->qp_state;
501 qp_attr->cap.max_send_wr = qp->sq.max;
502 qp_attr->cap.max_recv_wr = qp->rq.max;
503 qp_attr->cap.max_send_sge = qp->sq.max_gs;
504 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
505 qp_attr->cap.max_inline_data = qp->max_inline_data;
507 qp_init_attr->cap = qp_attr->cap;
510 mthca_free_mailbox(dev, mailbox);
514 static int mthca_path_set(struct mthca_dev *dev, struct ib_ah_attr *ah,
515 struct mthca_qp_path *path, u8 port)
517 path->g_mylmc = ah->src_path_bits & 0x7f;
518 path->rlid = cpu_to_be16(ah->dlid);
519 path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
521 if (ah->ah_flags & IB_AH_GRH) {
522 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
523 mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
524 ah->grh.sgid_index, dev->limits.gid_table_len-1);
528 path->g_mylmc |= 1 << 7;
529 path->mgid_index = ah->grh.sgid_index;
530 path->hop_limit = ah->grh.hop_limit;
531 path->sl_tclass_flowlabel =
532 cpu_to_be32((ah->sl << 28) |
533 (ah->grh.traffic_class << 20) |
534 (ah->grh.flow_label));
535 memcpy(path->rgid, ah->grh.dgid.raw, 16);
537 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
542 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
543 struct ib_udata *udata)
545 struct mthca_dev *dev = to_mdev(ibqp->device);
546 struct mthca_qp *qp = to_mqp(ibqp);
547 enum ib_qp_state cur_state, new_state;
548 struct mthca_mailbox *mailbox;
549 struct mthca_qp_param *qp_param;
550 struct mthca_qp_context *qp_context;
555 mutex_lock(&qp->mutex);
557 if (attr_mask & IB_QP_CUR_STATE) {
558 cur_state = attr->cur_qp_state;
560 spin_lock_irq(&qp->sq.lock);
561 spin_lock(&qp->rq.lock);
562 cur_state = qp->state;
563 spin_unlock(&qp->rq.lock);
564 spin_unlock_irq(&qp->sq.lock);
567 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
569 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
570 mthca_dbg(dev, "Bad QP transition (transport %d) "
571 "%d->%d with attr 0x%08x\n",
572 qp->transport, cur_state, new_state,
577 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
582 if ((attr_mask & IB_QP_PKEY_INDEX) &&
583 attr->pkey_index >= dev->limits.pkey_table_len) {
584 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
585 attr->pkey_index, dev->limits.pkey_table_len-1);
589 if ((attr_mask & IB_QP_PORT) &&
590 (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
591 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
595 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
596 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
597 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
598 attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
602 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
603 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
604 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
605 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
609 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
610 if (IS_ERR(mailbox)) {
611 err = PTR_ERR(mailbox);
614 qp_param = mailbox->buf;
615 qp_context = &qp_param->context;
616 memset(qp_param, 0, sizeof *qp_param);
618 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
619 (to_mthca_st(qp->transport) << 16));
620 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
621 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
622 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
624 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
625 switch (attr->path_mig_state) {
626 case IB_MIG_MIGRATED:
627 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
630 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
633 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
638 /* leave tavor_sched_queue as 0 */
640 if (qp->transport == MLX || qp->transport == UD)
641 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
642 else if (attr_mask & IB_QP_PATH_MTU) {
643 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
644 mthca_dbg(dev, "path MTU (%u) is invalid\n",
648 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
651 if (mthca_is_memfree(dev)) {
653 qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
654 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
657 qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
658 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
661 /* leave arbel_sched_queue as 0 */
663 if (qp->ibqp.uobject)
664 qp_context->usr_page =
665 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
667 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
668 qp_context->local_qpn = cpu_to_be32(qp->qpn);
669 if (attr_mask & IB_QP_DEST_QPN) {
670 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
673 if (qp->transport == MLX)
674 qp_context->pri_path.port_pkey |=
675 cpu_to_be32(qp->port << 24);
677 if (attr_mask & IB_QP_PORT) {
678 qp_context->pri_path.port_pkey |=
679 cpu_to_be32(attr->port_num << 24);
680 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
684 if (attr_mask & IB_QP_PKEY_INDEX) {
685 qp_context->pri_path.port_pkey |=
686 cpu_to_be32(attr->pkey_index);
687 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
690 if (attr_mask & IB_QP_RNR_RETRY) {
691 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
692 attr->rnr_retry << 5;
693 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
694 MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
697 if (attr_mask & IB_QP_AV) {
698 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
699 attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
702 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
705 if (ibqp->qp_type == IB_QPT_RC &&
706 cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
707 u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
709 if (mthca_is_memfree(dev))
710 qp_context->rlkey_arbel_sched_queue |= sched_queue;
712 qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
714 qp_param->opt_param_mask |=
715 cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
718 if (attr_mask & IB_QP_TIMEOUT) {
719 qp_context->pri_path.ackto = attr->timeout << 3;
720 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
723 if (attr_mask & IB_QP_ALT_PATH) {
724 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
725 mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
726 attr->alt_pkey_index, dev->limits.pkey_table_len-1);
730 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
731 mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
736 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
737 attr->alt_ah_attr.port_num))
740 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
741 attr->alt_port_num << 24);
742 qp_context->alt_path.ackto = attr->alt_timeout << 3;
743 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
747 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
748 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
749 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
750 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
751 (MTHCA_FLIGHT_LIMIT << 24) |
753 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
754 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
755 if (attr_mask & IB_QP_RETRY_CNT) {
756 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
757 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
760 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
761 if (attr->max_rd_atomic) {
762 qp_context->params1 |=
763 cpu_to_be32(MTHCA_QP_BIT_SRE |
765 qp_context->params1 |=
766 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
768 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
771 if (attr_mask & IB_QP_SQ_PSN)
772 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
773 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
775 if (mthca_is_memfree(dev)) {
776 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
777 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
780 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
781 if (attr->max_dest_rd_atomic)
782 qp_context->params2 |=
783 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
785 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
788 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
789 qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask);
790 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
791 MTHCA_QP_OPTPAR_RRE |
792 MTHCA_QP_OPTPAR_RAE);
795 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
798 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
800 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
801 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
802 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
804 if (attr_mask & IB_QP_RQ_PSN)
805 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
807 qp_context->ra_buff_indx =
808 cpu_to_be32(dev->qp_table.rdb_base +
809 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
810 dev->qp_table.rdb_shift));
812 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
814 if (mthca_is_memfree(dev))
815 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
817 if (attr_mask & IB_QP_QKEY) {
818 qp_context->qkey = cpu_to_be32(attr->qkey);
819 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
823 qp_context->srqn = cpu_to_be32(1 << 24 |
824 to_msrq(ibqp->srq)->srqn);
826 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
827 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY &&
828 attr->en_sqd_async_notify)
831 err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
832 mailbox, sqd_event, &status);
836 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n",
837 cur_state, new_state, status);
842 qp->state = new_state;
843 if (attr_mask & IB_QP_ACCESS_FLAGS)
844 qp->atomic_rd_en = attr->qp_access_flags;
845 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
846 qp->resp_depth = attr->max_dest_rd_atomic;
847 if (attr_mask & IB_QP_PORT)
848 qp->port = attr->port_num;
849 if (attr_mask & IB_QP_ALT_PATH)
850 qp->alt_port = attr->alt_port_num;
853 store_attrs(to_msqp(qp), attr, attr_mask);
856 * If we moved QP0 to RTR, bring the IB link up; if we moved
857 * QP0 to RESET or ERROR, bring the link back down.
859 if (is_qp0(dev, qp)) {
860 if (cur_state != IB_QPS_RTR &&
861 new_state == IB_QPS_RTR)
862 init_port(dev, qp->port);
864 if (cur_state != IB_QPS_RESET &&
865 cur_state != IB_QPS_ERR &&
866 (new_state == IB_QPS_RESET ||
867 new_state == IB_QPS_ERR))
868 mthca_CLOSE_IB(dev, qp->port, &status);
872 * If we moved a kernel QP to RESET, clean up all old CQ
873 * entries and reinitialize the QP.
875 if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
876 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
877 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
878 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
879 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
881 mthca_wq_reset(&qp->sq);
882 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
884 mthca_wq_reset(&qp->rq);
885 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
887 if (mthca_is_memfree(dev)) {
894 mthca_free_mailbox(dev, mailbox);
897 mutex_unlock(&qp->mutex);
901 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
904 * Calculate the maximum size of WQE s/g segments, excluding
905 * the next segment and other non-data segments.
907 int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
909 switch (qp->transport) {
911 max_data_size -= 2 * sizeof (struct mthca_data_seg);
915 if (mthca_is_memfree(dev))
916 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
918 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
922 max_data_size -= sizeof (struct mthca_raddr_seg);
926 return max_data_size;
929 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
931 /* We don't support inline data for kernel QPs (yet). */
932 return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
935 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
939 int max_data_size = mthca_max_data_size(dev, qp,
940 min(dev->limits.max_desc_sz,
941 1 << qp->sq.wqe_shift));
943 qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
945 qp->sq.max_gs = min_t(int, dev->limits.max_sg,
946 max_data_size / sizeof (struct mthca_data_seg));
947 qp->rq.max_gs = min_t(int, dev->limits.max_sg,
948 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
949 sizeof (struct mthca_next_seg)) /
950 sizeof (struct mthca_data_seg));
954 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
955 * rq.max_gs and sq.max_gs must all be assigned.
956 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
957 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
960 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
967 size = sizeof (struct mthca_next_seg) +
968 qp->rq.max_gs * sizeof (struct mthca_data_seg);
970 if (size > dev->limits.max_desc_sz)
973 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
977 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
978 switch (qp->transport) {
980 size += 2 * sizeof (struct mthca_data_seg);
984 size += mthca_is_memfree(dev) ?
985 sizeof (struct mthca_arbel_ud_seg) :
986 sizeof (struct mthca_tavor_ud_seg);
990 size += sizeof (struct mthca_raddr_seg);
994 size += sizeof (struct mthca_raddr_seg);
996 * An atomic op will require an atomic segment, a
997 * remote address segment and one scatter entry.
999 size = max_t(int, size,
1000 sizeof (struct mthca_atomic_seg) +
1001 sizeof (struct mthca_raddr_seg) +
1002 sizeof (struct mthca_data_seg));
1009 /* Make sure that we have enough space for a bind request */
1010 size = max_t(int, size, sizeof (struct mthca_bind_seg));
1012 size += sizeof (struct mthca_next_seg);
1014 if (size > dev->limits.max_desc_sz)
1017 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1021 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1022 1 << qp->sq.wqe_shift);
1025 * If this is a userspace QP, we don't actually have to
1026 * allocate anything. All we need is to calculate the WQE
1027 * sizes and the send_wqe_offset, so we're done now.
1029 if (pd->ibpd.uobject)
1032 size = PAGE_ALIGN(qp->send_wqe_offset +
1033 (qp->sq.max << qp->sq.wqe_shift));
1035 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1040 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1041 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1052 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1053 struct mthca_qp *qp)
1055 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1056 (qp->sq.max << qp->sq.wqe_shift)),
1057 &qp->queue, qp->is_direct, &qp->mr);
1061 static int mthca_map_memfree(struct mthca_dev *dev,
1062 struct mthca_qp *qp)
1066 if (mthca_is_memfree(dev)) {
1067 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1071 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1075 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1076 qp->qpn << dev->qp_table.rdb_shift);
1085 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1088 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1093 static void mthca_unmap_memfree(struct mthca_dev *dev,
1094 struct mthca_qp *qp)
1096 mthca_table_put(dev, dev->qp_table.rdb_table,
1097 qp->qpn << dev->qp_table.rdb_shift);
1098 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1099 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1102 static int mthca_alloc_memfree(struct mthca_dev *dev,
1103 struct mthca_qp *qp)
1105 if (mthca_is_memfree(dev)) {
1106 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1107 qp->qpn, &qp->rq.db);
1108 if (qp->rq.db_index < 0)
1111 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1112 qp->qpn, &qp->sq.db);
1113 if (qp->sq.db_index < 0) {
1114 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1122 static void mthca_free_memfree(struct mthca_dev *dev,
1123 struct mthca_qp *qp)
1125 if (mthca_is_memfree(dev)) {
1126 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1127 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1131 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1132 struct mthca_pd *pd,
1133 struct mthca_cq *send_cq,
1134 struct mthca_cq *recv_cq,
1135 enum ib_sig_type send_policy,
1136 struct mthca_qp *qp)
1142 init_waitqueue_head(&qp->wait);
1143 mutex_init(&qp->mutex);
1144 qp->state = IB_QPS_RESET;
1145 qp->atomic_rd_en = 0;
1147 qp->sq_policy = send_policy;
1148 mthca_wq_reset(&qp->sq);
1149 mthca_wq_reset(&qp->rq);
1151 spin_lock_init(&qp->sq.lock);
1152 spin_lock_init(&qp->rq.lock);
1154 ret = mthca_map_memfree(dev, qp);
1158 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1160 mthca_unmap_memfree(dev, qp);
1164 mthca_adjust_qp_caps(dev, pd, qp);
1167 * If this is a userspace QP, we're done now. The doorbells
1168 * will be allocated and buffers will be initialized in
1171 if (pd->ibpd.uobject)
1174 ret = mthca_alloc_memfree(dev, qp);
1176 mthca_free_wqe_buf(dev, qp);
1177 mthca_unmap_memfree(dev, qp);
1181 if (mthca_is_memfree(dev)) {
1182 struct mthca_next_seg *next;
1183 struct mthca_data_seg *scatter;
1184 int size = (sizeof (struct mthca_next_seg) +
1185 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1187 for (i = 0; i < qp->rq.max; ++i) {
1188 next = get_recv_wqe(qp, i);
1189 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1191 next->ee_nds = cpu_to_be32(size);
1193 for (scatter = (void *) (next + 1);
1194 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1196 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1199 for (i = 0; i < qp->sq.max; ++i) {
1200 next = get_send_wqe(qp, i);
1201 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1203 qp->send_wqe_offset);
1207 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1208 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1213 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1214 struct mthca_pd *pd, struct mthca_qp *qp)
1216 int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1218 /* Sanity check QP size before proceeding */
1219 if (cap->max_send_wr > dev->limits.max_wqes ||
1220 cap->max_recv_wr > dev->limits.max_wqes ||
1221 cap->max_send_sge > dev->limits.max_sg ||
1222 cap->max_recv_sge > dev->limits.max_sg ||
1223 cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1227 * For MLX transport we need 2 extra S/G entries:
1228 * one for the header and one for the checksum at the end
1230 if (qp->transport == MLX && cap->max_recv_sge + 2 > dev->limits.max_sg)
1233 if (mthca_is_memfree(dev)) {
1234 qp->rq.max = cap->max_recv_wr ?
1235 roundup_pow_of_two(cap->max_recv_wr) : 0;
1236 qp->sq.max = cap->max_send_wr ?
1237 roundup_pow_of_two(cap->max_send_wr) : 0;
1239 qp->rq.max = cap->max_recv_wr;
1240 qp->sq.max = cap->max_send_wr;
1243 qp->rq.max_gs = cap->max_recv_sge;
1244 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1245 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1246 MTHCA_INLINE_CHUNK_SIZE) /
1247 sizeof (struct mthca_data_seg));
1252 int mthca_alloc_qp(struct mthca_dev *dev,
1253 struct mthca_pd *pd,
1254 struct mthca_cq *send_cq,
1255 struct mthca_cq *recv_cq,
1256 enum ib_qp_type type,
1257 enum ib_sig_type send_policy,
1258 struct ib_qp_cap *cap,
1259 struct mthca_qp *qp)
1264 case IB_QPT_RC: qp->transport = RC; break;
1265 case IB_QPT_UC: qp->transport = UC; break;
1266 case IB_QPT_UD: qp->transport = UD; break;
1267 default: return -EINVAL;
1270 err = mthca_set_qp_size(dev, cap, pd, qp);
1274 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1278 /* initialize port to zero for error-catching. */
1281 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1284 mthca_free(&dev->qp_table.alloc, qp->qpn);
1288 spin_lock_irq(&dev->qp_table.lock);
1289 mthca_array_set(&dev->qp_table.qp,
1290 qp->qpn & (dev->limits.num_qps - 1), qp);
1291 spin_unlock_irq(&dev->qp_table.lock);
1296 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1298 if (send_cq == recv_cq)
1299 spin_lock_irq(&send_cq->lock);
1300 else if (send_cq->cqn < recv_cq->cqn) {
1301 spin_lock_irq(&send_cq->lock);
1302 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1304 spin_lock_irq(&recv_cq->lock);
1305 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1309 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1311 if (send_cq == recv_cq)
1312 spin_unlock_irq(&send_cq->lock);
1313 else if (send_cq->cqn < recv_cq->cqn) {
1314 spin_unlock(&recv_cq->lock);
1315 spin_unlock_irq(&send_cq->lock);
1317 spin_unlock(&send_cq->lock);
1318 spin_unlock_irq(&recv_cq->lock);
1322 int mthca_alloc_sqp(struct mthca_dev *dev,
1323 struct mthca_pd *pd,
1324 struct mthca_cq *send_cq,
1325 struct mthca_cq *recv_cq,
1326 enum ib_sig_type send_policy,
1327 struct ib_qp_cap *cap,
1330 struct mthca_sqp *sqp)
1332 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1335 sqp->qp.transport = MLX;
1336 err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1340 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1341 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1342 &sqp->header_dma, GFP_KERNEL);
1343 if (!sqp->header_buf)
1346 spin_lock_irq(&dev->qp_table.lock);
1347 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1350 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1351 spin_unlock_irq(&dev->qp_table.lock);
1356 sqp->qp.port = port;
1358 sqp->qp.transport = MLX;
1360 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1361 send_policy, &sqp->qp);
1365 atomic_inc(&pd->sqp_count);
1371 * Lock CQs here, so that CQ polling code can do QP lookup
1372 * without taking a lock.
1374 mthca_lock_cqs(send_cq, recv_cq);
1376 spin_lock(&dev->qp_table.lock);
1377 mthca_array_clear(&dev->qp_table.qp, mqpn);
1378 spin_unlock(&dev->qp_table.lock);
1380 mthca_unlock_cqs(send_cq, recv_cq);
1383 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1384 sqp->header_buf, sqp->header_dma);
1389 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1393 spin_lock_irq(&dev->qp_table.lock);
1395 spin_unlock_irq(&dev->qp_table.lock);
1400 void mthca_free_qp(struct mthca_dev *dev,
1401 struct mthca_qp *qp)
1404 struct mthca_cq *send_cq;
1405 struct mthca_cq *recv_cq;
1407 send_cq = to_mcq(qp->ibqp.send_cq);
1408 recv_cq = to_mcq(qp->ibqp.recv_cq);
1411 * Lock CQs here, so that CQ polling code can do QP lookup
1412 * without taking a lock.
1414 mthca_lock_cqs(send_cq, recv_cq);
1416 spin_lock(&dev->qp_table.lock);
1417 mthca_array_clear(&dev->qp_table.qp,
1418 qp->qpn & (dev->limits.num_qps - 1));
1420 spin_unlock(&dev->qp_table.lock);
1422 mthca_unlock_cqs(send_cq, recv_cq);
1424 wait_event(qp->wait, !get_qp_refcount(dev, qp));
1426 if (qp->state != IB_QPS_RESET)
1427 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1431 * If this is a userspace QP, the buffers, MR, CQs and so on
1432 * will be cleaned up in userspace, so all we have to do is
1433 * unref the mem-free tables and free the QPN in our table.
1435 if (!qp->ibqp.uobject) {
1436 mthca_cq_clean(dev, recv_cq, qp->qpn,
1437 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1438 if (send_cq != recv_cq)
1439 mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1441 mthca_free_memfree(dev, qp);
1442 mthca_free_wqe_buf(dev, qp);
1445 mthca_unmap_memfree(dev, qp);
1447 if (is_sqp(dev, qp)) {
1448 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1449 dma_free_coherent(&dev->pdev->dev,
1450 to_msqp(qp)->header_buf_size,
1451 to_msqp(qp)->header_buf,
1452 to_msqp(qp)->header_dma);
1454 mthca_free(&dev->qp_table.alloc, qp->qpn);
1457 /* Create UD header for an MLX send and build a data segment for it */
1458 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1459 int ind, struct ib_send_wr *wr,
1460 struct mthca_mlx_seg *mlx,
1461 struct mthca_data_seg *data)
1467 ib_ud_header_init(256, /* assume a MAD */
1468 mthca_ah_grh_present(to_mah(wr->wr.ud.ah)),
1471 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1474 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1475 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1476 (sqp->ud_header.lrh.destination_lid ==
1477 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1478 (sqp->ud_header.lrh.service_level << 8));
1479 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1482 switch (wr->opcode) {
1484 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1485 sqp->ud_header.immediate_present = 0;
1487 case IB_WR_SEND_WITH_IMM:
1488 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1489 sqp->ud_header.immediate_present = 1;
1490 sqp->ud_header.immediate_data = wr->imm_data;
1496 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1497 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1498 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1499 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1500 if (!sqp->qp.ibqp.qp_num)
1501 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1502 sqp->pkey_index, &pkey);
1504 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1505 wr->wr.ud.pkey_index, &pkey);
1506 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1507 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1508 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1509 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1510 sqp->qkey : wr->wr.ud.remote_qkey);
1511 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1513 header_size = ib_ud_header_pack(&sqp->ud_header,
1515 ind * MTHCA_UD_HEADER_SIZE);
1517 data->byte_count = cpu_to_be32(header_size);
1518 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1519 data->addr = cpu_to_be64(sqp->header_dma +
1520 ind * MTHCA_UD_HEADER_SIZE);
1525 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1526 struct ib_cq *ib_cq)
1529 struct mthca_cq *cq;
1531 cur = wq->head - wq->tail;
1532 if (likely(cur + nreq < wq->max))
1536 spin_lock(&cq->lock);
1537 cur = wq->head - wq->tail;
1538 spin_unlock(&cq->lock);
1540 return cur + nreq >= wq->max;
1543 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1544 struct ib_send_wr **bad_wr)
1546 struct mthca_dev *dev = to_mdev(ibqp->device);
1547 struct mthca_qp *qp = to_mqp(ibqp);
1550 unsigned long flags;
1560 spin_lock_irqsave(&qp->sq.lock, flags);
1562 /* XXX check that state is OK to post send */
1564 ind = qp->sq.next_ind;
1566 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1567 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1568 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1569 " %d max, %d nreq)\n", qp->qpn,
1570 qp->sq.head, qp->sq.tail,
1577 wqe = get_send_wqe(qp, ind);
1578 prev_wqe = qp->sq.last;
1581 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1582 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1583 ((struct mthca_next_seg *) wqe)->flags =
1584 ((wr->send_flags & IB_SEND_SIGNALED) ?
1585 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1586 ((wr->send_flags & IB_SEND_SOLICITED) ?
1587 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1589 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1590 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1591 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1593 wqe += sizeof (struct mthca_next_seg);
1594 size = sizeof (struct mthca_next_seg) / 16;
1596 switch (qp->transport) {
1598 switch (wr->opcode) {
1599 case IB_WR_ATOMIC_CMP_AND_SWP:
1600 case IB_WR_ATOMIC_FETCH_AND_ADD:
1601 ((struct mthca_raddr_seg *) wqe)->raddr =
1602 cpu_to_be64(wr->wr.atomic.remote_addr);
1603 ((struct mthca_raddr_seg *) wqe)->rkey =
1604 cpu_to_be32(wr->wr.atomic.rkey);
1605 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1607 wqe += sizeof (struct mthca_raddr_seg);
1609 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1610 ((struct mthca_atomic_seg *) wqe)->swap_add =
1611 cpu_to_be64(wr->wr.atomic.swap);
1612 ((struct mthca_atomic_seg *) wqe)->compare =
1613 cpu_to_be64(wr->wr.atomic.compare_add);
1615 ((struct mthca_atomic_seg *) wqe)->swap_add =
1616 cpu_to_be64(wr->wr.atomic.compare_add);
1617 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1620 wqe += sizeof (struct mthca_atomic_seg);
1621 size += (sizeof (struct mthca_raddr_seg) +
1622 sizeof (struct mthca_atomic_seg)) / 16;
1625 case IB_WR_RDMA_WRITE:
1626 case IB_WR_RDMA_WRITE_WITH_IMM:
1627 case IB_WR_RDMA_READ:
1628 ((struct mthca_raddr_seg *) wqe)->raddr =
1629 cpu_to_be64(wr->wr.rdma.remote_addr);
1630 ((struct mthca_raddr_seg *) wqe)->rkey =
1631 cpu_to_be32(wr->wr.rdma.rkey);
1632 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1633 wqe += sizeof (struct mthca_raddr_seg);
1634 size += sizeof (struct mthca_raddr_seg) / 16;
1638 /* No extra segments required for sends */
1645 switch (wr->opcode) {
1646 case IB_WR_RDMA_WRITE:
1647 case IB_WR_RDMA_WRITE_WITH_IMM:
1648 ((struct mthca_raddr_seg *) wqe)->raddr =
1649 cpu_to_be64(wr->wr.rdma.remote_addr);
1650 ((struct mthca_raddr_seg *) wqe)->rkey =
1651 cpu_to_be32(wr->wr.rdma.rkey);
1652 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1653 wqe += sizeof (struct mthca_raddr_seg);
1654 size += sizeof (struct mthca_raddr_seg) / 16;
1658 /* No extra segments required for sends */
1665 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1666 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1667 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1668 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1669 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1670 cpu_to_be32(wr->wr.ud.remote_qpn);
1671 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1672 cpu_to_be32(wr->wr.ud.remote_qkey);
1674 wqe += sizeof (struct mthca_tavor_ud_seg);
1675 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1679 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1680 wqe - sizeof (struct mthca_next_seg),
1686 wqe += sizeof (struct mthca_data_seg);
1687 size += sizeof (struct mthca_data_seg) / 16;
1691 if (wr->num_sge > qp->sq.max_gs) {
1692 mthca_err(dev, "too many gathers\n");
1698 for (i = 0; i < wr->num_sge; ++i) {
1699 ((struct mthca_data_seg *) wqe)->byte_count =
1700 cpu_to_be32(wr->sg_list[i].length);
1701 ((struct mthca_data_seg *) wqe)->lkey =
1702 cpu_to_be32(wr->sg_list[i].lkey);
1703 ((struct mthca_data_seg *) wqe)->addr =
1704 cpu_to_be64(wr->sg_list[i].addr);
1705 wqe += sizeof (struct mthca_data_seg);
1706 size += sizeof (struct mthca_data_seg) / 16;
1709 /* Add one more inline data segment for ICRC */
1710 if (qp->transport == MLX) {
1711 ((struct mthca_data_seg *) wqe)->byte_count =
1712 cpu_to_be32((1 << 31) | 4);
1713 ((u32 *) wqe)[1] = 0;
1714 wqe += sizeof (struct mthca_data_seg);
1715 size += sizeof (struct mthca_data_seg) / 16;
1718 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1720 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1721 mthca_err(dev, "opcode invalid\n");
1727 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1728 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1729 qp->send_wqe_offset) |
1730 mthca_opcode[wr->opcode]);
1732 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1733 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size |
1734 ((wr->send_flags & IB_SEND_FENCE) ?
1735 MTHCA_NEXT_FENCE : 0));
1739 op0 = mthca_opcode[wr->opcode];
1740 f0 = wr->send_flags & IB_SEND_FENCE ?
1741 MTHCA_SEND_DOORBELL_FENCE : 0;
1745 if (unlikely(ind >= qp->sq.max))
1753 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1754 qp->send_wqe_offset) | f0 | op0);
1755 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1759 mthca_write64(doorbell,
1760 dev->kar + MTHCA_SEND_DOORBELL,
1761 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1763 * Make sure doorbells don't leak out of SQ spinlock
1764 * and reach the HCA out of order:
1769 qp->sq.next_ind = ind;
1770 qp->sq.head += nreq;
1772 spin_unlock_irqrestore(&qp->sq.lock, flags);
1776 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1777 struct ib_recv_wr **bad_wr)
1779 struct mthca_dev *dev = to_mdev(ibqp->device);
1780 struct mthca_qp *qp = to_mqp(ibqp);
1782 unsigned long flags;
1792 spin_lock_irqsave(&qp->rq.lock, flags);
1794 /* XXX check that state is OK to post receive */
1796 ind = qp->rq.next_ind;
1798 for (nreq = 0; wr; wr = wr->next) {
1799 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1800 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1801 " %d max, %d nreq)\n", qp->qpn,
1802 qp->rq.head, qp->rq.tail,
1809 wqe = get_recv_wqe(qp, ind);
1810 prev_wqe = qp->rq.last;
1813 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1814 ((struct mthca_next_seg *) wqe)->ee_nds =
1815 cpu_to_be32(MTHCA_NEXT_DBD);
1816 ((struct mthca_next_seg *) wqe)->flags = 0;
1818 wqe += sizeof (struct mthca_next_seg);
1819 size = sizeof (struct mthca_next_seg) / 16;
1821 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1827 for (i = 0; i < wr->num_sge; ++i) {
1828 ((struct mthca_data_seg *) wqe)->byte_count =
1829 cpu_to_be32(wr->sg_list[i].length);
1830 ((struct mthca_data_seg *) wqe)->lkey =
1831 cpu_to_be32(wr->sg_list[i].lkey);
1832 ((struct mthca_data_seg *) wqe)->addr =
1833 cpu_to_be64(wr->sg_list[i].addr);
1834 wqe += sizeof (struct mthca_data_seg);
1835 size += sizeof (struct mthca_data_seg) / 16;
1838 qp->wrid[ind] = wr->wr_id;
1840 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1841 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1843 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1844 cpu_to_be32(MTHCA_NEXT_DBD | size);
1850 if (unlikely(ind >= qp->rq.max))
1854 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1857 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1858 doorbell[1] = cpu_to_be32(qp->qpn << 8);
1862 mthca_write64(doorbell,
1863 dev->kar + MTHCA_RECEIVE_DOORBELL,
1864 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1866 qp->rq.next_ind = ind;
1867 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1874 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1875 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1879 mthca_write64(doorbell,
1880 dev->kar + MTHCA_RECEIVE_DOORBELL,
1881 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1884 qp->rq.next_ind = ind;
1885 qp->rq.head += nreq;
1888 * Make sure doorbells don't leak out of RQ spinlock and reach
1889 * the HCA out of order:
1893 spin_unlock_irqrestore(&qp->rq.lock, flags);
1897 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1898 struct ib_send_wr **bad_wr)
1900 struct mthca_dev *dev = to_mdev(ibqp->device);
1901 struct mthca_qp *qp = to_mqp(ibqp);
1905 unsigned long flags;
1915 spin_lock_irqsave(&qp->sq.lock, flags);
1917 /* XXX check that state is OK to post send */
1919 ind = qp->sq.head & (qp->sq.max - 1);
1921 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1922 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1925 doorbell[0] = cpu_to_be32((MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1926 ((qp->sq.head & 0xffff) << 8) |
1928 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1930 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1934 * Make sure that descriptors are written before
1938 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1941 * Make sure doorbell record is written before we
1942 * write MMIO send doorbell.
1945 mthca_write64(doorbell,
1946 dev->kar + MTHCA_SEND_DOORBELL,
1947 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1950 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1951 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1952 " %d max, %d nreq)\n", qp->qpn,
1953 qp->sq.head, qp->sq.tail,
1960 wqe = get_send_wqe(qp, ind);
1961 prev_wqe = qp->sq.last;
1964 ((struct mthca_next_seg *) wqe)->flags =
1965 ((wr->send_flags & IB_SEND_SIGNALED) ?
1966 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1967 ((wr->send_flags & IB_SEND_SOLICITED) ?
1968 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1970 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1971 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1972 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1974 wqe += sizeof (struct mthca_next_seg);
1975 size = sizeof (struct mthca_next_seg) / 16;
1977 switch (qp->transport) {
1979 switch (wr->opcode) {
1980 case IB_WR_ATOMIC_CMP_AND_SWP:
1981 case IB_WR_ATOMIC_FETCH_AND_ADD:
1982 ((struct mthca_raddr_seg *) wqe)->raddr =
1983 cpu_to_be64(wr->wr.atomic.remote_addr);
1984 ((struct mthca_raddr_seg *) wqe)->rkey =
1985 cpu_to_be32(wr->wr.atomic.rkey);
1986 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1988 wqe += sizeof (struct mthca_raddr_seg);
1990 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1991 ((struct mthca_atomic_seg *) wqe)->swap_add =
1992 cpu_to_be64(wr->wr.atomic.swap);
1993 ((struct mthca_atomic_seg *) wqe)->compare =
1994 cpu_to_be64(wr->wr.atomic.compare_add);
1996 ((struct mthca_atomic_seg *) wqe)->swap_add =
1997 cpu_to_be64(wr->wr.atomic.compare_add);
1998 ((struct mthca_atomic_seg *) wqe)->compare = 0;
2001 wqe += sizeof (struct mthca_atomic_seg);
2002 size += (sizeof (struct mthca_raddr_seg) +
2003 sizeof (struct mthca_atomic_seg)) / 16;
2006 case IB_WR_RDMA_READ:
2007 case IB_WR_RDMA_WRITE:
2008 case IB_WR_RDMA_WRITE_WITH_IMM:
2009 ((struct mthca_raddr_seg *) wqe)->raddr =
2010 cpu_to_be64(wr->wr.rdma.remote_addr);
2011 ((struct mthca_raddr_seg *) wqe)->rkey =
2012 cpu_to_be32(wr->wr.rdma.rkey);
2013 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
2014 wqe += sizeof (struct mthca_raddr_seg);
2015 size += sizeof (struct mthca_raddr_seg) / 16;
2019 /* No extra segments required for sends */
2026 switch (wr->opcode) {
2027 case IB_WR_RDMA_WRITE:
2028 case IB_WR_RDMA_WRITE_WITH_IMM:
2029 ((struct mthca_raddr_seg *) wqe)->raddr =
2030 cpu_to_be64(wr->wr.rdma.remote_addr);
2031 ((struct mthca_raddr_seg *) wqe)->rkey =
2032 cpu_to_be32(wr->wr.rdma.rkey);
2033 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
2034 wqe += sizeof (struct mthca_raddr_seg);
2035 size += sizeof (struct mthca_raddr_seg) / 16;
2039 /* No extra segments required for sends */
2046 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
2047 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
2048 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
2049 cpu_to_be32(wr->wr.ud.remote_qpn);
2050 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
2051 cpu_to_be32(wr->wr.ud.remote_qkey);
2053 wqe += sizeof (struct mthca_arbel_ud_seg);
2054 size += sizeof (struct mthca_arbel_ud_seg) / 16;
2058 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
2059 wqe - sizeof (struct mthca_next_seg),
2065 wqe += sizeof (struct mthca_data_seg);
2066 size += sizeof (struct mthca_data_seg) / 16;
2070 if (wr->num_sge > qp->sq.max_gs) {
2071 mthca_err(dev, "too many gathers\n");
2077 for (i = 0; i < wr->num_sge; ++i) {
2078 ((struct mthca_data_seg *) wqe)->byte_count =
2079 cpu_to_be32(wr->sg_list[i].length);
2080 ((struct mthca_data_seg *) wqe)->lkey =
2081 cpu_to_be32(wr->sg_list[i].lkey);
2082 ((struct mthca_data_seg *) wqe)->addr =
2083 cpu_to_be64(wr->sg_list[i].addr);
2084 wqe += sizeof (struct mthca_data_seg);
2085 size += sizeof (struct mthca_data_seg) / 16;
2088 /* Add one more inline data segment for ICRC */
2089 if (qp->transport == MLX) {
2090 ((struct mthca_data_seg *) wqe)->byte_count =
2091 cpu_to_be32((1 << 31) | 4);
2092 ((u32 *) wqe)[1] = 0;
2093 wqe += sizeof (struct mthca_data_seg);
2094 size += sizeof (struct mthca_data_seg) / 16;
2097 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2099 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2100 mthca_err(dev, "opcode invalid\n");
2106 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2107 cpu_to_be32(((ind << qp->sq.wqe_shift) +
2108 qp->send_wqe_offset) |
2109 mthca_opcode[wr->opcode]);
2111 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2112 cpu_to_be32(MTHCA_NEXT_DBD | size |
2113 ((wr->send_flags & IB_SEND_FENCE) ?
2114 MTHCA_NEXT_FENCE : 0));
2118 op0 = mthca_opcode[wr->opcode];
2119 f0 = wr->send_flags & IB_SEND_FENCE ?
2120 MTHCA_SEND_DOORBELL_FENCE : 0;
2124 if (unlikely(ind >= qp->sq.max))
2130 doorbell[0] = cpu_to_be32((nreq << 24) |
2131 ((qp->sq.head & 0xffff) << 8) |
2133 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
2135 qp->sq.head += nreq;
2138 * Make sure that descriptors are written before
2142 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2145 * Make sure doorbell record is written before we
2146 * write MMIO send doorbell.
2149 mthca_write64(doorbell,
2150 dev->kar + MTHCA_SEND_DOORBELL,
2151 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2155 * Make sure doorbells don't leak out of SQ spinlock and reach
2156 * the HCA out of order:
2160 spin_unlock_irqrestore(&qp->sq.lock, flags);
2164 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2165 struct ib_recv_wr **bad_wr)
2167 struct mthca_dev *dev = to_mdev(ibqp->device);
2168 struct mthca_qp *qp = to_mqp(ibqp);
2169 unsigned long flags;
2176 spin_lock_irqsave(&qp->rq.lock, flags);
2178 /* XXX check that state is OK to post receive */
2180 ind = qp->rq.head & (qp->rq.max - 1);
2182 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2183 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2184 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2185 " %d max, %d nreq)\n", qp->qpn,
2186 qp->rq.head, qp->rq.tail,
2193 wqe = get_recv_wqe(qp, ind);
2195 ((struct mthca_next_seg *) wqe)->flags = 0;
2197 wqe += sizeof (struct mthca_next_seg);
2199 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2205 for (i = 0; i < wr->num_sge; ++i) {
2206 ((struct mthca_data_seg *) wqe)->byte_count =
2207 cpu_to_be32(wr->sg_list[i].length);
2208 ((struct mthca_data_seg *) wqe)->lkey =
2209 cpu_to_be32(wr->sg_list[i].lkey);
2210 ((struct mthca_data_seg *) wqe)->addr =
2211 cpu_to_be64(wr->sg_list[i].addr);
2212 wqe += sizeof (struct mthca_data_seg);
2215 if (i < qp->rq.max_gs) {
2216 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2217 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2218 ((struct mthca_data_seg *) wqe)->addr = 0;
2221 qp->wrid[ind] = wr->wr_id;
2224 if (unlikely(ind >= qp->rq.max))
2229 qp->rq.head += nreq;
2232 * Make sure that descriptors are written before
2236 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2239 spin_unlock_irqrestore(&qp->rq.lock, flags);
2243 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2244 int index, int *dbd, __be32 *new_wqe)
2246 struct mthca_next_seg *next;
2249 * For SRQs, all WQEs generate a CQE, so we're always at the
2250 * end of the doorbell chain.
2258 next = get_send_wqe(qp, index);
2260 next = get_recv_wqe(qp, index);
2262 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2263 if (next->ee_nds & cpu_to_be32(0x3f))
2264 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2265 (next->ee_nds & cpu_to_be32(0x3f));
2270 int mthca_init_qp_table(struct mthca_dev *dev)
2276 spin_lock_init(&dev->qp_table.lock);
2279 * We reserve 2 extra QPs per port for the special QPs. The
2280 * special QP for port 1 has to be even, so round up.
2282 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2283 err = mthca_alloc_init(&dev->qp_table.alloc,
2284 dev->limits.num_qps,
2286 dev->qp_table.sqp_start +
2287 MTHCA_MAX_PORTS * 2);
2291 err = mthca_array_init(&dev->qp_table.qp,
2292 dev->limits.num_qps);
2294 mthca_alloc_cleanup(&dev->qp_table.alloc);
2298 for (i = 0; i < 2; ++i) {
2299 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2300 dev->qp_table.sqp_start + i * 2,
2305 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2306 "status %02x, aborting.\n",
2315 for (i = 0; i < 2; ++i)
2316 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2318 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2319 mthca_alloc_cleanup(&dev->qp_table.alloc);
2324 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2329 for (i = 0; i < 2; ++i)
2330 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2332 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2333 mthca_alloc_cleanup(&dev->qp_table.alloc);