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
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
42 #include <rdma/ib_verbs.h>
43 #include <rdma/ib_cache.h>
44 #include <rdma/ib_pack.h>
45 #include <rdma/uverbs_ioctl.h>
47 #include "mthca_dev.h"
48 #include "mthca_cmd.h"
49 #include "mthca_memfree.h"
50 #include "mthca_wqe.h"
53 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
54 MTHCA_ACK_REQ_FREQ = 10,
55 MTHCA_FLIGHT_LIMIT = 9,
56 MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */
57 MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */
58 MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */
62 MTHCA_QP_STATE_RST = 0,
63 MTHCA_QP_STATE_INIT = 1,
64 MTHCA_QP_STATE_RTR = 2,
65 MTHCA_QP_STATE_RTS = 3,
66 MTHCA_QP_STATE_SQE = 4,
67 MTHCA_QP_STATE_SQD = 5,
68 MTHCA_QP_STATE_ERR = 6,
69 MTHCA_QP_STATE_DRAINING = 7
81 MTHCA_QP_PM_MIGRATED = 0x3,
82 MTHCA_QP_PM_ARMED = 0x0,
83 MTHCA_QP_PM_REARM = 0x1
87 /* qp_context flags */
88 MTHCA_QP_BIT_DE = 1 << 8,
90 MTHCA_QP_BIT_SRE = 1 << 15,
91 MTHCA_QP_BIT_SWE = 1 << 14,
92 MTHCA_QP_BIT_SAE = 1 << 13,
93 MTHCA_QP_BIT_SIC = 1 << 4,
94 MTHCA_QP_BIT_SSC = 1 << 3,
96 MTHCA_QP_BIT_RRE = 1 << 15,
97 MTHCA_QP_BIT_RWE = 1 << 14,
98 MTHCA_QP_BIT_RAE = 1 << 13,
99 MTHCA_QP_BIT_RIC = 1 << 4,
100 MTHCA_QP_BIT_RSC = 1 << 3
104 MTHCA_SEND_DOORBELL_FENCE = 1 << 5
107 struct mthca_qp_path {
116 __be32 sl_tclass_flowlabel;
118 } __attribute__((packed));
120 struct mthca_qp_context {
122 __be32 tavor_sched_queue; /* Reserved on Arbel */
124 u8 rq_size_stride; /* Reserved on Tavor */
125 u8 sq_size_stride; /* Reserved on Tavor */
126 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
131 struct mthca_qp_path pri_path;
132 struct mthca_qp_path alt_path;
139 __be32 next_send_psn;
141 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
142 __be32 snd_db_index; /* (debugging only entries) */
143 __be32 last_acked_psn;
146 __be32 rnr_nextrecvpsn;
149 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
150 __be32 rcv_db_index; /* (debugging only entries) */
154 __be16 rq_wqe_counter; /* reserved on Tavor */
155 __be16 sq_wqe_counter; /* reserved on Tavor */
157 } __attribute__((packed));
159 struct mthca_qp_param {
160 __be32 opt_param_mask;
162 struct mthca_qp_context context;
164 } __attribute__((packed));
167 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
168 MTHCA_QP_OPTPAR_RRE = 1 << 1,
169 MTHCA_QP_OPTPAR_RAE = 1 << 2,
170 MTHCA_QP_OPTPAR_RWE = 1 << 3,
171 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
172 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
173 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
174 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
175 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
176 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
177 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
178 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
179 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
180 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
181 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
182 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
183 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
186 static const u8 mthca_opcode[] = {
187 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
188 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
189 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
190 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
191 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
192 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
193 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
196 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
198 return qp->qpn >= dev->qp_table.sqp_start &&
199 qp->qpn <= dev->qp_table.sqp_start + 3;
202 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
204 return qp->qpn >= dev->qp_table.sqp_start &&
205 qp->qpn <= dev->qp_table.sqp_start + 1;
208 static void *get_recv_wqe(struct mthca_qp *qp, int n)
211 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
213 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
214 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
217 static void *get_send_wqe(struct mthca_qp *qp, int n)
220 return qp->queue.direct.buf + qp->send_wqe_offset +
221 (n << qp->sq.wqe_shift);
223 return qp->queue.page_list[(qp->send_wqe_offset +
224 (n << qp->sq.wqe_shift)) >>
226 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
230 static void mthca_wq_reset(struct mthca_wq *wq)
233 wq->last_comp = wq->max - 1;
238 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
239 enum ib_event_type event_type)
242 struct ib_event event;
244 spin_lock(&dev->qp_table.lock);
245 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
248 spin_unlock(&dev->qp_table.lock);
251 mthca_warn(dev, "Async event %d for bogus QP %08x\n",
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, const 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)
313 struct mthca_init_ib_param param;
315 memset(¶m, 0, sizeof param);
317 param.port_width = dev->limits.port_width_cap;
318 param.vl_cap = dev->limits.vl_cap;
319 param.mtu_cap = dev->limits.mtu_cap;
320 param.gid_cap = dev->limits.gid_table_len;
321 param.pkey_cap = dev->limits.pkey_table_len;
323 err = mthca_INIT_IB(dev, ¶m, port);
325 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
328 static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr,
333 u32 hw_access_flags = 0;
335 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
336 dest_rd_atomic = attr->max_dest_rd_atomic;
338 dest_rd_atomic = qp->resp_depth;
340 if (attr_mask & IB_QP_ACCESS_FLAGS)
341 access_flags = attr->qp_access_flags;
343 access_flags = qp->atomic_rd_en;
346 access_flags &= IB_ACCESS_REMOTE_WRITE;
348 if (access_flags & IB_ACCESS_REMOTE_READ)
349 hw_access_flags |= MTHCA_QP_BIT_RRE;
350 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
351 hw_access_flags |= MTHCA_QP_BIT_RAE;
352 if (access_flags & IB_ACCESS_REMOTE_WRITE)
353 hw_access_flags |= MTHCA_QP_BIT_RWE;
355 return cpu_to_be32(hw_access_flags);
358 static inline enum ib_qp_state to_ib_qp_state(int mthca_state)
360 switch (mthca_state) {
361 case MTHCA_QP_STATE_RST: return IB_QPS_RESET;
362 case MTHCA_QP_STATE_INIT: return IB_QPS_INIT;
363 case MTHCA_QP_STATE_RTR: return IB_QPS_RTR;
364 case MTHCA_QP_STATE_RTS: return IB_QPS_RTS;
365 case MTHCA_QP_STATE_DRAINING:
366 case MTHCA_QP_STATE_SQD: return IB_QPS_SQD;
367 case MTHCA_QP_STATE_SQE: return IB_QPS_SQE;
368 case MTHCA_QP_STATE_ERR: return IB_QPS_ERR;
373 static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state)
375 switch (mthca_mig_state) {
376 case 0: return IB_MIG_ARMED;
377 case 1: return IB_MIG_REARM;
378 case 3: return IB_MIG_MIGRATED;
383 static int to_ib_qp_access_flags(int mthca_flags)
387 if (mthca_flags & MTHCA_QP_BIT_RRE)
388 ib_flags |= IB_ACCESS_REMOTE_READ;
389 if (mthca_flags & MTHCA_QP_BIT_RWE)
390 ib_flags |= IB_ACCESS_REMOTE_WRITE;
391 if (mthca_flags & MTHCA_QP_BIT_RAE)
392 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
397 static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr,
398 struct mthca_qp_path *path)
400 memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
401 ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3;
403 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports)
406 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
407 ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28;
408 ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f;
409 ib_ah_attr->static_rate = mthca_rate_to_ib(dev,
410 path->static_rate & 0xf,
411 ib_ah_attr->port_num);
412 ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
413 if (ib_ah_attr->ah_flags) {
414 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1);
415 ib_ah_attr->grh.hop_limit = path->hop_limit;
416 ib_ah_attr->grh.traffic_class =
417 (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff;
418 ib_ah_attr->grh.flow_label =
419 be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff;
420 memcpy(ib_ah_attr->grh.dgid.raw,
421 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
425 int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
426 struct ib_qp_init_attr *qp_init_attr)
428 struct mthca_dev *dev = to_mdev(ibqp->device);
429 struct mthca_qp *qp = to_mqp(ibqp);
431 struct mthca_mailbox *mailbox = NULL;
432 struct mthca_qp_param *qp_param;
433 struct mthca_qp_context *context;
436 mutex_lock(&qp->mutex);
438 if (qp->state == IB_QPS_RESET) {
439 qp_attr->qp_state = IB_QPS_RESET;
443 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
444 if (IS_ERR(mailbox)) {
445 err = PTR_ERR(mailbox);
449 err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox);
451 mthca_warn(dev, "QUERY_QP failed (%d)\n", err);
455 qp_param = mailbox->buf;
456 context = &qp_param->context;
457 mthca_state = be32_to_cpu(context->flags) >> 28;
459 qp->state = to_ib_qp_state(mthca_state);
460 qp_attr->qp_state = qp->state;
461 qp_attr->path_mtu = context->mtu_msgmax >> 5;
462 qp_attr->path_mig_state =
463 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
464 qp_attr->qkey = be32_to_cpu(context->qkey);
465 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
466 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
467 qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff;
468 qp_attr->qp_access_flags =
469 to_ib_qp_access_flags(be32_to_cpu(context->params2));
471 if (qp->transport == RC || qp->transport == UC) {
472 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
473 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
474 qp_attr->alt_pkey_index =
475 be32_to_cpu(context->alt_path.port_pkey) & 0x7f;
476 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
479 qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f;
481 (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3;
483 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
484 qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING;
486 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
488 qp_attr->max_dest_rd_atomic =
489 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
490 qp_attr->min_rnr_timer =
491 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
492 qp_attr->timeout = context->pri_path.ackto >> 3;
493 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
494 qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5;
495 qp_attr->alt_timeout = context->alt_path.ackto >> 3;
498 qp_attr->cur_qp_state = qp_attr->qp_state;
499 qp_attr->cap.max_send_wr = qp->sq.max;
500 qp_attr->cap.max_recv_wr = qp->rq.max;
501 qp_attr->cap.max_send_sge = qp->sq.max_gs;
502 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
503 qp_attr->cap.max_inline_data = qp->max_inline_data;
505 qp_init_attr->cap = qp_attr->cap;
506 qp_init_attr->sq_sig_type = qp->sq_policy;
509 mthca_free_mailbox(dev, mailbox);
512 mutex_unlock(&qp->mutex);
516 static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah,
517 struct mthca_qp_path *path, u8 port)
519 path->g_mylmc = ah->src_path_bits & 0x7f;
520 path->rlid = cpu_to_be16(ah->dlid);
521 path->static_rate = mthca_get_rate(dev, ah->static_rate, port);
523 if (ah->ah_flags & IB_AH_GRH) {
524 if (ah->grh.sgid_index >= dev->limits.gid_table_len) {
525 mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n",
526 ah->grh.sgid_index, dev->limits.gid_table_len-1);
530 path->g_mylmc |= 1 << 7;
531 path->mgid_index = ah->grh.sgid_index;
532 path->hop_limit = ah->grh.hop_limit;
533 path->sl_tclass_flowlabel =
534 cpu_to_be32((ah->sl << 28) |
535 (ah->grh.traffic_class << 20) |
536 (ah->grh.flow_label));
537 memcpy(path->rgid, ah->grh.dgid.raw, 16);
539 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28);
544 static int __mthca_modify_qp(struct ib_qp *ibqp,
545 const struct ib_qp_attr *attr, int attr_mask,
546 enum ib_qp_state cur_state,
547 enum ib_qp_state new_state,
548 struct ib_udata *udata)
550 struct mthca_dev *dev = to_mdev(ibqp->device);
551 struct mthca_qp *qp = to_mqp(ibqp);
552 struct mthca_ucontext *context = rdma_udata_to_drv_context(
553 udata, struct mthca_ucontext, ibucontext);
554 struct mthca_mailbox *mailbox;
555 struct mthca_qp_param *qp_param;
556 struct mthca_qp_context *qp_context;
560 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
561 if (IS_ERR(mailbox)) {
562 err = PTR_ERR(mailbox);
565 qp_param = mailbox->buf;
566 qp_context = &qp_param->context;
567 memset(qp_param, 0, sizeof *qp_param);
569 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
570 (to_mthca_st(qp->transport) << 16));
571 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
572 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
573 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
575 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
576 switch (attr->path_mig_state) {
577 case IB_MIG_MIGRATED:
578 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
581 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
584 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
589 /* leave tavor_sched_queue as 0 */
591 if (qp->transport == MLX || qp->transport == UD)
592 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
593 else if (attr_mask & IB_QP_PATH_MTU) {
594 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) {
595 mthca_dbg(dev, "path MTU (%u) is invalid\n",
599 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
602 if (mthca_is_memfree(dev)) {
604 qp_context->rq_size_stride = ilog2(qp->rq.max) << 3;
605 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
608 qp_context->sq_size_stride = ilog2(qp->sq.max) << 3;
609 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
612 /* leave arbel_sched_queue as 0 */
614 if (qp->ibqp.uobject)
615 qp_context->usr_page = cpu_to_be32(context->uar.index);
617 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
618 qp_context->local_qpn = cpu_to_be32(qp->qpn);
619 if (attr_mask & IB_QP_DEST_QPN) {
620 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
623 if (qp->transport == MLX)
624 qp_context->pri_path.port_pkey |=
625 cpu_to_be32(qp->port << 24);
627 if (attr_mask & IB_QP_PORT) {
628 qp_context->pri_path.port_pkey |=
629 cpu_to_be32(attr->port_num << 24);
630 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
634 if (attr_mask & IB_QP_PKEY_INDEX) {
635 qp_context->pri_path.port_pkey |=
636 cpu_to_be32(attr->pkey_index);
637 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
640 if (attr_mask & IB_QP_RNR_RETRY) {
641 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry =
642 attr->rnr_retry << 5;
643 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY |
644 MTHCA_QP_OPTPAR_ALT_RNR_RETRY);
647 if (attr_mask & IB_QP_AV) {
648 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path,
649 attr_mask & IB_QP_PORT ? attr->port_num : qp->port))
652 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
655 if (ibqp->qp_type == IB_QPT_RC &&
656 cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
657 u8 sched_queue = ibqp->uobject ? 0x2 : 0x1;
659 if (mthca_is_memfree(dev))
660 qp_context->rlkey_arbel_sched_queue |= sched_queue;
662 qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue);
664 qp_param->opt_param_mask |=
665 cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE);
668 if (attr_mask & IB_QP_TIMEOUT) {
669 qp_context->pri_path.ackto = attr->timeout << 3;
670 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
673 if (attr_mask & IB_QP_ALT_PATH) {
674 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) {
675 mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n",
676 attr->alt_pkey_index, dev->limits.pkey_table_len-1);
680 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) {
681 mthca_dbg(dev, "Alternate port number (%u) is invalid\n",
686 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path,
687 attr->alt_ah_attr.port_num))
690 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index |
691 attr->alt_port_num << 24);
692 qp_context->alt_path.ackto = attr->alt_timeout << 3;
693 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH);
697 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
698 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
699 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
700 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
701 (MTHCA_FLIGHT_LIMIT << 24) |
703 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
704 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
705 if (attr_mask & IB_QP_RETRY_CNT) {
706 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
707 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
710 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
711 if (attr->max_rd_atomic) {
712 qp_context->params1 |=
713 cpu_to_be32(MTHCA_QP_BIT_SRE |
715 qp_context->params1 |=
716 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
718 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
721 if (attr_mask & IB_QP_SQ_PSN)
722 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
723 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
725 if (mthca_is_memfree(dev)) {
726 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
727 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
730 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
731 if (attr->max_dest_rd_atomic)
732 qp_context->params2 |=
733 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
735 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
738 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
739 qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask);
740 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
741 MTHCA_QP_OPTPAR_RRE |
742 MTHCA_QP_OPTPAR_RAE);
745 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
748 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
750 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
751 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
752 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
754 if (attr_mask & IB_QP_RQ_PSN)
755 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
757 qp_context->ra_buff_indx =
758 cpu_to_be32(dev->qp_table.rdb_base +
759 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
760 dev->qp_table.rdb_shift));
762 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
764 if (mthca_is_memfree(dev))
765 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
767 if (attr_mask & IB_QP_QKEY) {
768 qp_context->qkey = cpu_to_be32(attr->qkey);
769 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
773 qp_context->srqn = cpu_to_be32(1 << 24 |
774 to_msrq(ibqp->srq)->srqn);
776 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
777 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY &&
778 attr->en_sqd_async_notify)
781 err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0,
784 mthca_warn(dev, "modify QP %d->%d returned %d.\n",
785 cur_state, new_state, err);
789 qp->state = new_state;
790 if (attr_mask & IB_QP_ACCESS_FLAGS)
791 qp->atomic_rd_en = attr->qp_access_flags;
792 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
793 qp->resp_depth = attr->max_dest_rd_atomic;
794 if (attr_mask & IB_QP_PORT)
795 qp->port = attr->port_num;
796 if (attr_mask & IB_QP_ALT_PATH)
797 qp->alt_port = attr->alt_port_num;
800 store_attrs(to_msqp(qp), attr, attr_mask);
803 * If we moved QP0 to RTR, bring the IB link up; if we moved
804 * QP0 to RESET or ERROR, bring the link back down.
806 if (is_qp0(dev, qp)) {
807 if (cur_state != IB_QPS_RTR &&
808 new_state == IB_QPS_RTR)
809 init_port(dev, qp->port);
811 if (cur_state != IB_QPS_RESET &&
812 cur_state != IB_QPS_ERR &&
813 (new_state == IB_QPS_RESET ||
814 new_state == IB_QPS_ERR))
815 mthca_CLOSE_IB(dev, qp->port);
819 * If we moved a kernel QP to RESET, clean up all old CQ
820 * entries and reinitialize the QP.
822 if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
823 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn,
824 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
825 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
826 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL);
828 mthca_wq_reset(&qp->sq);
829 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
831 mthca_wq_reset(&qp->rq);
832 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
834 if (mthca_is_memfree(dev)) {
841 mthca_free_mailbox(dev, mailbox);
846 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
847 struct ib_udata *udata)
849 struct mthca_dev *dev = to_mdev(ibqp->device);
850 struct mthca_qp *qp = to_mqp(ibqp);
851 enum ib_qp_state cur_state, new_state;
854 mutex_lock(&qp->mutex);
855 if (attr_mask & IB_QP_CUR_STATE) {
856 cur_state = attr->cur_qp_state;
858 spin_lock_irq(&qp->sq.lock);
859 spin_lock(&qp->rq.lock);
860 cur_state = qp->state;
861 spin_unlock(&qp->rq.lock);
862 spin_unlock_irq(&qp->sq.lock);
865 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
867 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) {
868 mthca_dbg(dev, "Bad QP transition (transport %d) "
869 "%d->%d with attr 0x%08x\n",
870 qp->transport, cur_state, new_state,
875 if ((attr_mask & IB_QP_PKEY_INDEX) &&
876 attr->pkey_index >= dev->limits.pkey_table_len) {
877 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n",
878 attr->pkey_index, dev->limits.pkey_table_len-1);
882 if ((attr_mask & IB_QP_PORT) &&
883 (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) {
884 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num);
888 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
889 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) {
890 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n",
891 attr->max_rd_atomic, dev->limits.max_qp_init_rdma);
895 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
896 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) {
897 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n",
898 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift);
902 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
907 err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state,
911 mutex_unlock(&qp->mutex);
915 static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz)
918 * Calculate the maximum size of WQE s/g segments, excluding
919 * the next segment and other non-data segments.
921 int max_data_size = desc_sz - sizeof (struct mthca_next_seg);
923 switch (qp->transport) {
925 max_data_size -= 2 * sizeof (struct mthca_data_seg);
929 if (mthca_is_memfree(dev))
930 max_data_size -= sizeof (struct mthca_arbel_ud_seg);
932 max_data_size -= sizeof (struct mthca_tavor_ud_seg);
936 max_data_size -= sizeof (struct mthca_raddr_seg);
940 return max_data_size;
943 static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size)
945 /* We don't support inline data for kernel QPs (yet). */
946 return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0;
949 static void mthca_adjust_qp_caps(struct mthca_dev *dev,
953 int max_data_size = mthca_max_data_size(dev, qp,
954 min(dev->limits.max_desc_sz,
955 1 << qp->sq.wqe_shift));
957 qp->max_inline_data = mthca_max_inline_data(pd, max_data_size);
959 qp->sq.max_gs = min_t(int, dev->limits.max_sg,
960 max_data_size / sizeof (struct mthca_data_seg));
961 qp->rq.max_gs = min_t(int, dev->limits.max_sg,
962 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) -
963 sizeof (struct mthca_next_seg)) /
964 sizeof (struct mthca_data_seg));
968 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
969 * rq.max_gs and sq.max_gs must all be assigned.
970 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
971 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
974 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
977 struct ib_udata *udata)
982 size = sizeof (struct mthca_next_seg) +
983 qp->rq.max_gs * sizeof (struct mthca_data_seg);
985 if (size > dev->limits.max_desc_sz)
988 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
992 size = qp->sq.max_gs * sizeof (struct mthca_data_seg);
993 switch (qp->transport) {
995 size += 2 * sizeof (struct mthca_data_seg);
999 size += mthca_is_memfree(dev) ?
1000 sizeof (struct mthca_arbel_ud_seg) :
1001 sizeof (struct mthca_tavor_ud_seg);
1005 size += sizeof (struct mthca_raddr_seg);
1009 size += sizeof (struct mthca_raddr_seg);
1011 * An atomic op will require an atomic segment, a
1012 * remote address segment and one scatter entry.
1014 size = max_t(int, size,
1015 sizeof (struct mthca_atomic_seg) +
1016 sizeof (struct mthca_raddr_seg) +
1017 sizeof (struct mthca_data_seg));
1024 /* Make sure that we have enough space for a bind request */
1025 size = max_t(int, size, sizeof (struct mthca_bind_seg));
1027 size += sizeof (struct mthca_next_seg);
1029 if (size > dev->limits.max_desc_sz)
1032 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
1036 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
1037 1 << qp->sq.wqe_shift);
1040 * If this is a userspace QP, we don't actually have to
1041 * allocate anything. All we need is to calculate the WQE
1042 * sizes and the send_wqe_offset, so we're done now.
1047 size = PAGE_ALIGN(qp->send_wqe_offset +
1048 (qp->sq.max << qp->sq.wqe_shift));
1050 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
1055 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
1056 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
1067 static void mthca_free_wqe_buf(struct mthca_dev *dev,
1068 struct mthca_qp *qp)
1070 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
1071 (qp->sq.max << qp->sq.wqe_shift)),
1072 &qp->queue, qp->is_direct, &qp->mr);
1076 static int mthca_map_memfree(struct mthca_dev *dev,
1077 struct mthca_qp *qp)
1081 if (mthca_is_memfree(dev)) {
1082 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1086 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1090 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
1091 qp->qpn << dev->qp_table.rdb_shift);
1099 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1102 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1107 static void mthca_unmap_memfree(struct mthca_dev *dev,
1108 struct mthca_qp *qp)
1110 mthca_table_put(dev, dev->qp_table.rdb_table,
1111 qp->qpn << dev->qp_table.rdb_shift);
1112 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1113 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1116 static int mthca_alloc_memfree(struct mthca_dev *dev,
1117 struct mthca_qp *qp)
1119 if (mthca_is_memfree(dev)) {
1120 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1121 qp->qpn, &qp->rq.db);
1122 if (qp->rq.db_index < 0)
1125 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1126 qp->qpn, &qp->sq.db);
1127 if (qp->sq.db_index < 0) {
1128 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1136 static void mthca_free_memfree(struct mthca_dev *dev,
1137 struct mthca_qp *qp)
1139 if (mthca_is_memfree(dev)) {
1140 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1141 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1145 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1146 struct mthca_pd *pd,
1147 struct mthca_cq *send_cq,
1148 struct mthca_cq *recv_cq,
1149 enum ib_sig_type send_policy,
1150 struct mthca_qp *qp,
1151 struct ib_udata *udata)
1155 struct mthca_next_seg *next;
1158 init_waitqueue_head(&qp->wait);
1159 mutex_init(&qp->mutex);
1160 qp->state = IB_QPS_RESET;
1161 qp->atomic_rd_en = 0;
1163 qp->sq_policy = send_policy;
1164 mthca_wq_reset(&qp->sq);
1165 mthca_wq_reset(&qp->rq);
1167 spin_lock_init(&qp->sq.lock);
1168 spin_lock_init(&qp->rq.lock);
1170 ret = mthca_map_memfree(dev, qp);
1174 ret = mthca_alloc_wqe_buf(dev, pd, qp, udata);
1176 mthca_unmap_memfree(dev, qp);
1180 mthca_adjust_qp_caps(dev, pd, qp);
1183 * If this is a userspace QP, we're done now. The doorbells
1184 * will be allocated and buffers will be initialized in
1190 ret = mthca_alloc_memfree(dev, qp);
1192 mthca_free_wqe_buf(dev, qp);
1193 mthca_unmap_memfree(dev, qp);
1197 if (mthca_is_memfree(dev)) {
1198 struct mthca_data_seg *scatter;
1199 int size = (sizeof (struct mthca_next_seg) +
1200 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1202 for (i = 0; i < qp->rq.max; ++i) {
1203 next = get_recv_wqe(qp, i);
1204 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1206 next->ee_nds = cpu_to_be32(size);
1208 for (scatter = (void *) (next + 1);
1209 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1211 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1214 for (i = 0; i < qp->sq.max; ++i) {
1215 next = get_send_wqe(qp, i);
1216 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1218 qp->send_wqe_offset);
1221 for (i = 0; i < qp->rq.max; ++i) {
1222 next = get_recv_wqe(qp, i);
1223 next->nda_op = htonl((((i + 1) % qp->rq.max) <<
1224 qp->rq.wqe_shift) | 1);
1228 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1229 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1234 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1235 struct mthca_pd *pd, struct mthca_qp *qp)
1237 int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz);
1239 /* Sanity check QP size before proceeding */
1240 if (cap->max_send_wr > dev->limits.max_wqes ||
1241 cap->max_recv_wr > dev->limits.max_wqes ||
1242 cap->max_send_sge > dev->limits.max_sg ||
1243 cap->max_recv_sge > dev->limits.max_sg ||
1244 cap->max_inline_data > mthca_max_inline_data(pd, max_data_size))
1248 * For MLX transport we need 2 extra send gather entries:
1249 * one for the header and one for the checksum at the end
1251 if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg)
1254 if (mthca_is_memfree(dev)) {
1255 qp->rq.max = cap->max_recv_wr ?
1256 roundup_pow_of_two(cap->max_recv_wr) : 0;
1257 qp->sq.max = cap->max_send_wr ?
1258 roundup_pow_of_two(cap->max_send_wr) : 0;
1260 qp->rq.max = cap->max_recv_wr;
1261 qp->sq.max = cap->max_send_wr;
1264 qp->rq.max_gs = cap->max_recv_sge;
1265 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1266 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1267 MTHCA_INLINE_CHUNK_SIZE) /
1268 sizeof (struct mthca_data_seg));
1273 int mthca_alloc_qp(struct mthca_dev *dev,
1274 struct mthca_pd *pd,
1275 struct mthca_cq *send_cq,
1276 struct mthca_cq *recv_cq,
1277 enum ib_qp_type type,
1278 enum ib_sig_type send_policy,
1279 struct ib_qp_cap *cap,
1280 struct mthca_qp *qp,
1281 struct ib_udata *udata)
1286 case IB_QPT_RC: qp->transport = RC; break;
1287 case IB_QPT_UC: qp->transport = UC; break;
1288 case IB_QPT_UD: qp->transport = UD; break;
1289 default: return -EINVAL;
1292 err = mthca_set_qp_size(dev, cap, pd, qp);
1296 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1300 /* initialize port to zero for error-catching. */
1303 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1304 send_policy, qp, udata);
1306 mthca_free(&dev->qp_table.alloc, qp->qpn);
1310 spin_lock_irq(&dev->qp_table.lock);
1311 mthca_array_set(&dev->qp_table.qp,
1312 qp->qpn & (dev->limits.num_qps - 1), qp);
1313 spin_unlock_irq(&dev->qp_table.lock);
1318 static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1319 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1321 if (send_cq == recv_cq) {
1322 spin_lock_irq(&send_cq->lock);
1323 __acquire(&recv_cq->lock);
1324 } else if (send_cq->cqn < recv_cq->cqn) {
1325 spin_lock_irq(&send_cq->lock);
1326 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
1328 spin_lock_irq(&recv_cq->lock);
1329 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
1333 static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq)
1334 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1336 if (send_cq == recv_cq) {
1337 __release(&recv_cq->lock);
1338 spin_unlock_irq(&send_cq->lock);
1339 } else if (send_cq->cqn < recv_cq->cqn) {
1340 spin_unlock(&recv_cq->lock);
1341 spin_unlock_irq(&send_cq->lock);
1343 spin_unlock(&send_cq->lock);
1344 spin_unlock_irq(&recv_cq->lock);
1348 int mthca_alloc_sqp(struct mthca_dev *dev,
1349 struct mthca_pd *pd,
1350 struct mthca_cq *send_cq,
1351 struct mthca_cq *recv_cq,
1352 enum ib_sig_type send_policy,
1353 struct ib_qp_cap *cap,
1356 struct mthca_sqp *sqp,
1357 struct ib_udata *udata)
1359 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1362 sqp->qp.transport = MLX;
1363 err = mthca_set_qp_size(dev, cap, pd, &sqp->qp);
1367 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1368 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1369 &sqp->header_dma, GFP_KERNEL);
1370 if (!sqp->header_buf)
1373 spin_lock_irq(&dev->qp_table.lock);
1374 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1377 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1378 spin_unlock_irq(&dev->qp_table.lock);
1383 sqp->qp.port = port;
1385 sqp->qp.transport = MLX;
1387 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1388 send_policy, &sqp->qp, udata);
1392 atomic_inc(&pd->sqp_count);
1398 * Lock CQs here, so that CQ polling code can do QP lookup
1399 * without taking a lock.
1401 mthca_lock_cqs(send_cq, recv_cq);
1403 spin_lock(&dev->qp_table.lock);
1404 mthca_array_clear(&dev->qp_table.qp, mqpn);
1405 spin_unlock(&dev->qp_table.lock);
1407 mthca_unlock_cqs(send_cq, recv_cq);
1410 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1411 sqp->header_buf, sqp->header_dma);
1416 static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp)
1420 spin_lock_irq(&dev->qp_table.lock);
1422 spin_unlock_irq(&dev->qp_table.lock);
1427 void mthca_free_qp(struct mthca_dev *dev,
1428 struct mthca_qp *qp)
1430 struct mthca_cq *send_cq;
1431 struct mthca_cq *recv_cq;
1433 send_cq = to_mcq(qp->ibqp.send_cq);
1434 recv_cq = to_mcq(qp->ibqp.recv_cq);
1437 * Lock CQs here, so that CQ polling code can do QP lookup
1438 * without taking a lock.
1440 mthca_lock_cqs(send_cq, recv_cq);
1442 spin_lock(&dev->qp_table.lock);
1443 mthca_array_clear(&dev->qp_table.qp,
1444 qp->qpn & (dev->limits.num_qps - 1));
1446 spin_unlock(&dev->qp_table.lock);
1448 mthca_unlock_cqs(send_cq, recv_cq);
1450 wait_event(qp->wait, !get_qp_refcount(dev, qp));
1452 if (qp->state != IB_QPS_RESET)
1453 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0,
1457 * If this is a userspace QP, the buffers, MR, CQs and so on
1458 * will be cleaned up in userspace, so all we have to do is
1459 * unref the mem-free tables and free the QPN in our table.
1461 if (!qp->ibqp.uobject) {
1462 mthca_cq_clean(dev, recv_cq, qp->qpn,
1463 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1464 if (send_cq != recv_cq)
1465 mthca_cq_clean(dev, send_cq, qp->qpn, NULL);
1467 mthca_free_memfree(dev, qp);
1468 mthca_free_wqe_buf(dev, qp);
1471 mthca_unmap_memfree(dev, qp);
1473 if (is_sqp(dev, qp)) {
1474 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1475 dma_free_coherent(&dev->pdev->dev,
1476 to_msqp(qp)->header_buf_size,
1477 to_msqp(qp)->header_buf,
1478 to_msqp(qp)->header_dma);
1480 mthca_free(&dev->qp_table.alloc, qp->qpn);
1483 /* Create UD header for an MLX send and build a data segment for it */
1484 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1485 int ind, const struct ib_ud_wr *wr,
1486 struct mthca_mlx_seg *mlx,
1487 struct mthca_data_seg *data)
1493 ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0,
1494 mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0,
1497 err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header);
1500 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1501 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1502 (sqp->ud_header.lrh.destination_lid ==
1503 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1504 (sqp->ud_header.lrh.service_level << 8));
1505 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1508 switch (wr->wr.opcode) {
1510 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1511 sqp->ud_header.immediate_present = 0;
1513 case IB_WR_SEND_WITH_IMM:
1514 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1515 sqp->ud_header.immediate_present = 1;
1516 sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
1522 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1523 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1524 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1525 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
1526 if (!sqp->qp.ibqp.qp_num)
1527 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1528 sqp->pkey_index, &pkey);
1530 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port,
1531 wr->pkey_index, &pkey);
1532 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1533 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
1534 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1535 sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
1536 sqp->qkey : wr->remote_qkey);
1537 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1539 header_size = ib_ud_header_pack(&sqp->ud_header,
1541 ind * MTHCA_UD_HEADER_SIZE);
1543 data->byte_count = cpu_to_be32(header_size);
1544 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1545 data->addr = cpu_to_be64(sqp->header_dma +
1546 ind * MTHCA_UD_HEADER_SIZE);
1551 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1552 struct ib_cq *ib_cq)
1555 struct mthca_cq *cq;
1557 cur = wq->head - wq->tail;
1558 if (likely(cur + nreq < wq->max))
1562 spin_lock(&cq->lock);
1563 cur = wq->head - wq->tail;
1564 spin_unlock(&cq->lock);
1566 return cur + nreq >= wq->max;
1569 static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg,
1570 u64 remote_addr, u32 rkey)
1572 rseg->raddr = cpu_to_be64(remote_addr);
1573 rseg->rkey = cpu_to_be32(rkey);
1577 static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg,
1578 const struct ib_atomic_wr *wr)
1580 if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1581 aseg->swap_add = cpu_to_be64(wr->swap);
1582 aseg->compare = cpu_to_be64(wr->compare_add);
1584 aseg->swap_add = cpu_to_be64(wr->compare_add);
1590 static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg,
1591 const struct ib_ud_wr *wr)
1593 useg->lkey = cpu_to_be32(to_mah(wr->ah)->key);
1594 useg->av_addr = cpu_to_be64(to_mah(wr->ah)->avdma);
1595 useg->dqpn = cpu_to_be32(wr->remote_qpn);
1596 useg->qkey = cpu_to_be32(wr->remote_qkey);
1600 static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg,
1601 const struct ib_ud_wr *wr)
1603 memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE);
1604 useg->dqpn = cpu_to_be32(wr->remote_qpn);
1605 useg->qkey = cpu_to_be32(wr->remote_qkey);
1608 int mthca_tavor_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1609 const struct ib_send_wr **bad_wr)
1611 struct mthca_dev *dev = to_mdev(ibqp->device);
1612 struct mthca_qp *qp = to_mqp(ibqp);
1615 unsigned long flags;
1621 * f0 and size0 are only used if nreq != 0, and they will
1622 * always be initialized the first time through the main loop
1623 * before nreq is incremented. So nreq cannot become non-zero
1624 * without initializing f0 and size0, and they are in fact
1625 * never used uninitialized.
1627 int uninitialized_var(size0);
1628 u32 uninitialized_var(f0);
1632 spin_lock_irqsave(&qp->sq.lock, flags);
1634 /* XXX check that state is OK to post send */
1636 ind = qp->sq.next_ind;
1638 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1639 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1640 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1641 " %d max, %d nreq)\n", qp->qpn,
1642 qp->sq.head, qp->sq.tail,
1649 wqe = get_send_wqe(qp, ind);
1650 prev_wqe = qp->sq.last;
1653 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1654 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1655 ((struct mthca_next_seg *) wqe)->flags =
1656 ((wr->send_flags & IB_SEND_SIGNALED) ?
1657 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1658 ((wr->send_flags & IB_SEND_SOLICITED) ?
1659 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1661 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1662 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1663 ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
1665 wqe += sizeof (struct mthca_next_seg);
1666 size = sizeof (struct mthca_next_seg) / 16;
1668 switch (qp->transport) {
1670 switch (wr->opcode) {
1671 case IB_WR_ATOMIC_CMP_AND_SWP:
1672 case IB_WR_ATOMIC_FETCH_AND_ADD:
1673 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
1674 atomic_wr(wr)->rkey);
1675 wqe += sizeof (struct mthca_raddr_seg);
1677 set_atomic_seg(wqe, atomic_wr(wr));
1678 wqe += sizeof (struct mthca_atomic_seg);
1679 size += (sizeof (struct mthca_raddr_seg) +
1680 sizeof (struct mthca_atomic_seg)) / 16;
1683 case IB_WR_RDMA_WRITE:
1684 case IB_WR_RDMA_WRITE_WITH_IMM:
1685 case IB_WR_RDMA_READ:
1686 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1688 wqe += sizeof (struct mthca_raddr_seg);
1689 size += sizeof (struct mthca_raddr_seg) / 16;
1693 /* No extra segments required for sends */
1700 switch (wr->opcode) {
1701 case IB_WR_RDMA_WRITE:
1702 case IB_WR_RDMA_WRITE_WITH_IMM:
1703 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
1705 wqe += sizeof (struct mthca_raddr_seg);
1706 size += sizeof (struct mthca_raddr_seg) / 16;
1710 /* No extra segments required for sends */
1717 set_tavor_ud_seg(wqe, ud_wr(wr));
1718 wqe += sizeof (struct mthca_tavor_ud_seg);
1719 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1723 err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
1724 wqe - sizeof (struct mthca_next_seg),
1730 wqe += sizeof (struct mthca_data_seg);
1731 size += sizeof (struct mthca_data_seg) / 16;
1735 if (wr->num_sge > qp->sq.max_gs) {
1736 mthca_err(dev, "too many gathers\n");
1742 for (i = 0; i < wr->num_sge; ++i) {
1743 mthca_set_data_seg(wqe, wr->sg_list + i);
1744 wqe += sizeof (struct mthca_data_seg);
1745 size += sizeof (struct mthca_data_seg) / 16;
1748 /* Add one more inline data segment for ICRC */
1749 if (qp->transport == MLX) {
1750 ((struct mthca_data_seg *) wqe)->byte_count =
1751 cpu_to_be32((1 << 31) | 4);
1752 ((u32 *) wqe)[1] = 0;
1753 wqe += sizeof (struct mthca_data_seg);
1754 size += sizeof (struct mthca_data_seg) / 16;
1757 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1759 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1760 mthca_err(dev, "opcode invalid\n");
1766 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1767 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1768 qp->send_wqe_offset) |
1769 mthca_opcode[wr->opcode]);
1771 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1772 cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size |
1773 ((wr->send_flags & IB_SEND_FENCE) ?
1774 MTHCA_NEXT_FENCE : 0));
1778 op0 = mthca_opcode[wr->opcode];
1779 f0 = wr->send_flags & IB_SEND_FENCE ?
1780 MTHCA_SEND_DOORBELL_FENCE : 0;
1784 if (unlikely(ind >= qp->sq.max))
1792 mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) +
1793 qp->send_wqe_offset) | f0 | op0,
1794 (qp->qpn << 8) | size0,
1795 dev->kar + MTHCA_SEND_DOORBELL,
1796 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1798 * Make sure doorbells don't leak out of SQ spinlock
1799 * and reach the HCA out of order:
1804 qp->sq.next_ind = ind;
1805 qp->sq.head += nreq;
1807 spin_unlock_irqrestore(&qp->sq.lock, flags);
1811 int mthca_tavor_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1812 const struct ib_recv_wr **bad_wr)
1814 struct mthca_dev *dev = to_mdev(ibqp->device);
1815 struct mthca_qp *qp = to_mqp(ibqp);
1816 unsigned long flags;
1822 * size0 is only used if nreq != 0, and it will always be
1823 * initialized the first time through the main loop before
1824 * nreq is incremented. So nreq cannot become non-zero
1825 * without initializing size0, and it is in fact never used
1828 int uninitialized_var(size0);
1833 spin_lock_irqsave(&qp->rq.lock, flags);
1835 /* XXX check that state is OK to post receive */
1837 ind = qp->rq.next_ind;
1839 for (nreq = 0; wr; wr = wr->next) {
1840 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1841 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1842 " %d max, %d nreq)\n", qp->qpn,
1843 qp->rq.head, qp->rq.tail,
1850 wqe = get_recv_wqe(qp, ind);
1851 prev_wqe = qp->rq.last;
1854 ((struct mthca_next_seg *) wqe)->ee_nds =
1855 cpu_to_be32(MTHCA_NEXT_DBD);
1856 ((struct mthca_next_seg *) wqe)->flags = 0;
1858 wqe += sizeof (struct mthca_next_seg);
1859 size = sizeof (struct mthca_next_seg) / 16;
1861 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1867 for (i = 0; i < wr->num_sge; ++i) {
1868 mthca_set_data_seg(wqe, wr->sg_list + i);
1869 wqe += sizeof (struct mthca_data_seg);
1870 size += sizeof (struct mthca_data_seg) / 16;
1873 qp->wrid[ind] = wr->wr_id;
1875 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1876 cpu_to_be32(MTHCA_NEXT_DBD | size);
1882 if (unlikely(ind >= qp->rq.max))
1886 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
1891 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1892 qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL,
1893 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1895 qp->rq.next_ind = ind;
1896 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB;
1904 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0,
1905 qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL,
1906 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1909 qp->rq.next_ind = ind;
1910 qp->rq.head += nreq;
1913 * Make sure doorbells don't leak out of RQ spinlock and reach
1914 * the HCA out of order:
1918 spin_unlock_irqrestore(&qp->rq.lock, flags);
1922 int mthca_arbel_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
1923 const struct ib_send_wr **bad_wr)
1925 struct mthca_dev *dev = to_mdev(ibqp->device);
1926 struct mthca_qp *qp = to_mqp(ibqp);
1930 unsigned long flags;
1936 * f0 and size0 are only used if nreq != 0, and they will
1937 * always be initialized the first time through the main loop
1938 * before nreq is incremented. So nreq cannot become non-zero
1939 * without initializing f0 and size0, and they are in fact
1940 * never used uninitialized.
1942 int uninitialized_var(size0);
1943 u32 uninitialized_var(f0);
1947 spin_lock_irqsave(&qp->sq.lock, flags);
1949 /* XXX check that state is OK to post send */
1951 ind = qp->sq.head & (qp->sq.max - 1);
1953 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1954 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) {
1957 dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) |
1958 ((qp->sq.head & 0xffff) << 8) | f0 | op0;
1960 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB;
1963 * Make sure that descriptors are written before
1967 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1970 * Make sure doorbell record is written before we
1971 * write MMIO send doorbell.
1975 mthca_write64(dbhi, (qp->qpn << 8) | size0,
1976 dev->kar + MTHCA_SEND_DOORBELL,
1977 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1980 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1981 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1982 " %d max, %d nreq)\n", qp->qpn,
1983 qp->sq.head, qp->sq.tail,
1990 wqe = get_send_wqe(qp, ind);
1991 prev_wqe = qp->sq.last;
1994 ((struct mthca_next_seg *) wqe)->flags =
1995 ((wr->send_flags & IB_SEND_SIGNALED) ?
1996 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1997 ((wr->send_flags & IB_SEND_SOLICITED) ?
1998 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1999 ((wr->send_flags & IB_SEND_IP_CSUM) ?
2000 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) |
2002 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
2003 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
2004 ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data;
2006 wqe += sizeof (struct mthca_next_seg);
2007 size = sizeof (struct mthca_next_seg) / 16;
2009 switch (qp->transport) {
2011 switch (wr->opcode) {
2012 case IB_WR_ATOMIC_CMP_AND_SWP:
2013 case IB_WR_ATOMIC_FETCH_AND_ADD:
2014 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
2015 atomic_wr(wr)->rkey);
2016 wqe += sizeof (struct mthca_raddr_seg);
2018 set_atomic_seg(wqe, atomic_wr(wr));
2019 wqe += sizeof (struct mthca_atomic_seg);
2020 size += (sizeof (struct mthca_raddr_seg) +
2021 sizeof (struct mthca_atomic_seg)) / 16;
2024 case IB_WR_RDMA_READ:
2025 case IB_WR_RDMA_WRITE:
2026 case IB_WR_RDMA_WRITE_WITH_IMM:
2027 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2029 wqe += sizeof (struct mthca_raddr_seg);
2030 size += sizeof (struct mthca_raddr_seg) / 16;
2034 /* No extra segments required for sends */
2041 switch (wr->opcode) {
2042 case IB_WR_RDMA_WRITE:
2043 case IB_WR_RDMA_WRITE_WITH_IMM:
2044 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
2046 wqe += sizeof (struct mthca_raddr_seg);
2047 size += sizeof (struct mthca_raddr_seg) / 16;
2051 /* No extra segments required for sends */
2058 set_arbel_ud_seg(wqe, ud_wr(wr));
2059 wqe += sizeof (struct mthca_arbel_ud_seg);
2060 size += sizeof (struct mthca_arbel_ud_seg) / 16;
2064 err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr),
2065 wqe - sizeof (struct mthca_next_seg),
2071 wqe += sizeof (struct mthca_data_seg);
2072 size += sizeof (struct mthca_data_seg) / 16;
2076 if (wr->num_sge > qp->sq.max_gs) {
2077 mthca_err(dev, "too many gathers\n");
2083 for (i = 0; i < wr->num_sge; ++i) {
2084 mthca_set_data_seg(wqe, wr->sg_list + i);
2085 wqe += sizeof (struct mthca_data_seg);
2086 size += sizeof (struct mthca_data_seg) / 16;
2089 /* Add one more inline data segment for ICRC */
2090 if (qp->transport == MLX) {
2091 ((struct mthca_data_seg *) wqe)->byte_count =
2092 cpu_to_be32((1 << 31) | 4);
2093 ((u32 *) wqe)[1] = 0;
2094 wqe += sizeof (struct mthca_data_seg);
2095 size += sizeof (struct mthca_data_seg) / 16;
2098 qp->wrid[ind + qp->rq.max] = wr->wr_id;
2100 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
2101 mthca_err(dev, "opcode invalid\n");
2107 ((struct mthca_next_seg *) prev_wqe)->nda_op =
2108 cpu_to_be32(((ind << qp->sq.wqe_shift) +
2109 qp->send_wqe_offset) |
2110 mthca_opcode[wr->opcode]);
2112 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
2113 cpu_to_be32(MTHCA_NEXT_DBD | size |
2114 ((wr->send_flags & IB_SEND_FENCE) ?
2115 MTHCA_NEXT_FENCE : 0));
2119 op0 = mthca_opcode[wr->opcode];
2120 f0 = wr->send_flags & IB_SEND_FENCE ?
2121 MTHCA_SEND_DOORBELL_FENCE : 0;
2125 if (unlikely(ind >= qp->sq.max))
2131 dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0;
2133 qp->sq.head += nreq;
2136 * Make sure that descriptors are written before
2140 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
2143 * Make sure doorbell record is written before we
2144 * write MMIO send doorbell.
2148 mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL,
2149 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
2153 * Make sure doorbells don't leak out of SQ spinlock and reach
2154 * the HCA out of order:
2158 spin_unlock_irqrestore(&qp->sq.lock, flags);
2162 int mthca_arbel_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
2163 const struct ib_recv_wr **bad_wr)
2165 struct mthca_dev *dev = to_mdev(ibqp->device);
2166 struct mthca_qp *qp = to_mqp(ibqp);
2167 unsigned long flags;
2174 spin_lock_irqsave(&qp->rq.lock, flags);
2176 /* XXX check that state is OK to post receive */
2178 ind = qp->rq.head & (qp->rq.max - 1);
2180 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2181 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2182 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
2183 " %d max, %d nreq)\n", qp->qpn,
2184 qp->rq.head, qp->rq.tail,
2191 wqe = get_recv_wqe(qp, ind);
2193 ((struct mthca_next_seg *) wqe)->flags = 0;
2195 wqe += sizeof (struct mthca_next_seg);
2197 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2203 for (i = 0; i < wr->num_sge; ++i) {
2204 mthca_set_data_seg(wqe, wr->sg_list + i);
2205 wqe += sizeof (struct mthca_data_seg);
2208 if (i < qp->rq.max_gs)
2209 mthca_set_data_seg_inval(wqe);
2211 qp->wrid[ind] = wr->wr_id;
2214 if (unlikely(ind >= qp->rq.max))
2219 qp->rq.head += nreq;
2222 * Make sure that descriptors are written before
2226 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2229 spin_unlock_irqrestore(&qp->rq.lock, flags);
2233 void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2234 int index, int *dbd, __be32 *new_wqe)
2236 struct mthca_next_seg *next;
2239 * For SRQs, all receive WQEs generate a CQE, so we're always
2240 * at the end of the doorbell chain.
2242 if (qp->ibqp.srq && !is_send) {
2248 next = get_send_wqe(qp, index);
2250 next = get_recv_wqe(qp, index);
2252 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2253 if (next->ee_nds & cpu_to_be32(0x3f))
2254 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2255 (next->ee_nds & cpu_to_be32(0x3f));
2260 int mthca_init_qp_table(struct mthca_dev *dev)
2265 spin_lock_init(&dev->qp_table.lock);
2268 * We reserve 2 extra QPs per port for the special QPs. The
2269 * special QP for port 1 has to be even, so round up.
2271 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2272 err = mthca_alloc_init(&dev->qp_table.alloc,
2273 dev->limits.num_qps,
2275 dev->qp_table.sqp_start +
2276 MTHCA_MAX_PORTS * 2);
2280 err = mthca_array_init(&dev->qp_table.qp,
2281 dev->limits.num_qps);
2283 mthca_alloc_cleanup(&dev->qp_table.alloc);
2287 for (i = 0; i < 2; ++i) {
2288 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2289 dev->qp_table.sqp_start + i * 2);
2291 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2292 "%d, aborting.\n", err);
2299 for (i = 0; i < 2; ++i)
2300 mthca_CONF_SPECIAL_QP(dev, i, 0);
2302 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2303 mthca_alloc_cleanup(&dev->qp_table.alloc);
2308 void mthca_cleanup_qp_table(struct mthca_dev *dev)
2312 for (i = 0; i < 2; ++i)
2313 mthca_CONF_SPECIAL_QP(dev, i, 0);
2315 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2316 mthca_alloc_cleanup(&dev->qp_table.alloc);