2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
38 #include <linux/init.h>
40 #include <rdma/ib_verbs.h>
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_pack.h>
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
46 #include "mthca_memfree.h"
47 #include "mthca_wqe.h"
50 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
51 MTHCA_ACK_REQ_FREQ = 10,
52 MTHCA_FLIGHT_LIMIT = 9,
53 MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */
54 MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */
55 MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */
59 MTHCA_QP_STATE_RST = 0,
60 MTHCA_QP_STATE_INIT = 1,
61 MTHCA_QP_STATE_RTR = 2,
62 MTHCA_QP_STATE_RTS = 3,
63 MTHCA_QP_STATE_SQE = 4,
64 MTHCA_QP_STATE_SQD = 5,
65 MTHCA_QP_STATE_ERR = 6,
66 MTHCA_QP_STATE_DRAINING = 7
78 MTHCA_QP_PM_MIGRATED = 0x3,
79 MTHCA_QP_PM_ARMED = 0x0,
80 MTHCA_QP_PM_REARM = 0x1
84 /* qp_context flags */
85 MTHCA_QP_BIT_DE = 1 << 8,
87 MTHCA_QP_BIT_SRE = 1 << 15,
88 MTHCA_QP_BIT_SWE = 1 << 14,
89 MTHCA_QP_BIT_SAE = 1 << 13,
90 MTHCA_QP_BIT_SIC = 1 << 4,
91 MTHCA_QP_BIT_SSC = 1 << 3,
93 MTHCA_QP_BIT_RRE = 1 << 15,
94 MTHCA_QP_BIT_RWE = 1 << 14,
95 MTHCA_QP_BIT_RAE = 1 << 13,
96 MTHCA_QP_BIT_RIC = 1 << 4,
97 MTHCA_QP_BIT_RSC = 1 << 3
100 struct mthca_qp_path {
109 __be32 sl_tclass_flowlabel;
111 } __attribute__((packed));
113 struct mthca_qp_context {
115 __be32 tavor_sched_queue; /* Reserved on Arbel */
117 u8 rq_size_stride; /* Reserved on Tavor */
118 u8 sq_size_stride; /* Reserved on Tavor */
119 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
124 struct mthca_qp_path pri_path;
125 struct mthca_qp_path alt_path;
132 __be32 next_send_psn;
134 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */
135 __be32 snd_db_index; /* (debugging only entries) */
136 __be32 last_acked_psn;
139 __be32 rnr_nextrecvpsn;
142 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
143 __be32 rcv_db_index; /* (debugging only entries) */
147 __be16 rq_wqe_counter; /* reserved on Tavor */
148 __be16 sq_wqe_counter; /* reserved on Tavor */
150 } __attribute__((packed));
152 struct mthca_qp_param {
153 __be32 opt_param_mask;
155 struct mthca_qp_context context;
157 } __attribute__((packed));
160 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
161 MTHCA_QP_OPTPAR_RRE = 1 << 1,
162 MTHCA_QP_OPTPAR_RAE = 1 << 2,
163 MTHCA_QP_OPTPAR_RWE = 1 << 3,
164 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
165 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
166 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
167 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
168 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
169 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
170 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
171 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
172 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
173 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
174 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
175 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
176 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
179 static const u8 mthca_opcode[] = {
180 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
181 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
182 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
183 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
184 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
185 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
186 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
189 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
191 return qp->qpn >= dev->qp_table.sqp_start &&
192 qp->qpn <= dev->qp_table.sqp_start + 3;
195 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
197 return qp->qpn >= dev->qp_table.sqp_start &&
198 qp->qpn <= dev->qp_table.sqp_start + 1;
201 static void *get_recv_wqe(struct mthca_qp *qp, int n)
204 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
206 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
207 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
210 static void *get_send_wqe(struct mthca_qp *qp, int n)
213 return qp->queue.direct.buf + qp->send_wqe_offset +
214 (n << qp->sq.wqe_shift);
216 return qp->queue.page_list[(qp->send_wqe_offset +
217 (n << qp->sq.wqe_shift)) >>
219 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
223 static void mthca_wq_init(struct mthca_wq *wq)
225 spin_lock_init(&wq->lock);
227 wq->last_comp = wq->max - 1;
232 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
233 enum ib_event_type event_type)
236 struct ib_event event;
238 spin_lock(&dev->qp_table.lock);
239 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
241 atomic_inc(&qp->refcount);
242 spin_unlock(&dev->qp_table.lock);
245 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
249 event.device = &dev->ib_dev;
250 event.event = event_type;
251 event.element.qp = &qp->ibqp;
252 if (qp->ibqp.event_handler)
253 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
255 if (atomic_dec_and_test(&qp->refcount))
259 static int to_mthca_state(enum ib_qp_state ib_state)
262 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
263 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
264 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
265 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
266 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
267 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
268 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
273 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
275 static int to_mthca_st(int transport)
278 case RC: return MTHCA_QP_ST_RC;
279 case UC: return MTHCA_QP_ST_UC;
280 case UD: return MTHCA_QP_ST_UD;
281 case RD: return MTHCA_QP_ST_RD;
282 case MLX: return MTHCA_QP_ST_MLX;
287 static const struct {
289 u32 req_param[NUM_TRANS];
290 u32 opt_param[NUM_TRANS];
291 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
293 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
294 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
296 .trans = MTHCA_TRANS_RST2INIT,
298 [UD] = (IB_QP_PKEY_INDEX |
301 [UC] = (IB_QP_PKEY_INDEX |
304 [RC] = (IB_QP_PKEY_INDEX |
307 [MLX] = (IB_QP_PKEY_INDEX |
310 /* bug-for-bug compatibility with VAPI: */
317 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
318 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
320 .trans = MTHCA_TRANS_INIT2INIT,
322 [UD] = (IB_QP_PKEY_INDEX |
325 [UC] = (IB_QP_PKEY_INDEX |
328 [RC] = (IB_QP_PKEY_INDEX |
331 [MLX] = (IB_QP_PKEY_INDEX |
336 .trans = MTHCA_TRANS_INIT2RTR,
342 IB_QP_MAX_DEST_RD_ATOMIC),
347 IB_QP_MAX_DEST_RD_ATOMIC |
348 IB_QP_MIN_RNR_TIMER),
351 [UD] = (IB_QP_PKEY_INDEX |
353 [UC] = (IB_QP_ALT_PATH |
356 [RC] = (IB_QP_ALT_PATH |
359 [MLX] = (IB_QP_PKEY_INDEX |
365 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
366 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
368 .trans = MTHCA_TRANS_RTR2RTS,
371 [UC] = (IB_QP_SQ_PSN |
372 IB_QP_MAX_QP_RD_ATOMIC),
373 [RC] = (IB_QP_TIMEOUT |
377 IB_QP_MAX_QP_RD_ATOMIC),
378 [MLX] = IB_QP_SQ_PSN,
381 [UD] = (IB_QP_CUR_STATE |
383 [UC] = (IB_QP_CUR_STATE |
387 IB_QP_PATH_MIG_STATE),
388 [RC] = (IB_QP_CUR_STATE |
392 IB_QP_MIN_RNR_TIMER |
393 IB_QP_PATH_MIG_STATE),
394 [MLX] = (IB_QP_CUR_STATE |
400 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
401 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
403 .trans = MTHCA_TRANS_RTS2RTS,
405 [UD] = (IB_QP_CUR_STATE |
407 [UC] = (IB_QP_ACCESS_FLAGS |
409 IB_QP_PATH_MIG_STATE),
410 [RC] = (IB_QP_ACCESS_FLAGS |
412 IB_QP_PATH_MIG_STATE |
413 IB_QP_MIN_RNR_TIMER),
414 [MLX] = (IB_QP_CUR_STATE |
419 .trans = MTHCA_TRANS_RTS2SQD,
423 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
424 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
426 .trans = MTHCA_TRANS_SQD2RTS,
428 [UD] = (IB_QP_CUR_STATE |
430 [UC] = (IB_QP_CUR_STATE |
433 IB_QP_PATH_MIG_STATE),
434 [RC] = (IB_QP_CUR_STATE |
437 IB_QP_MIN_RNR_TIMER |
438 IB_QP_PATH_MIG_STATE),
439 [MLX] = (IB_QP_CUR_STATE |
444 .trans = MTHCA_TRANS_SQD2SQD,
446 [UD] = (IB_QP_PKEY_INDEX |
449 IB_QP_MAX_QP_RD_ATOMIC |
450 IB_QP_MAX_DEST_RD_ATOMIC |
455 IB_QP_PATH_MIG_STATE),
460 IB_QP_MAX_QP_RD_ATOMIC |
461 IB_QP_MAX_DEST_RD_ATOMIC |
466 IB_QP_MIN_RNR_TIMER |
467 IB_QP_PATH_MIG_STATE),
468 [MLX] = (IB_QP_PKEY_INDEX |
474 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
475 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
477 .trans = MTHCA_TRANS_SQERR2RTS,
479 [UD] = (IB_QP_CUR_STATE |
481 [UC] = (IB_QP_CUR_STATE),
482 [RC] = (IB_QP_CUR_STATE |
483 IB_QP_MIN_RNR_TIMER),
484 [MLX] = (IB_QP_CUR_STATE |
490 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
491 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
495 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
498 if (attr_mask & IB_QP_PKEY_INDEX)
499 sqp->pkey_index = attr->pkey_index;
500 if (attr_mask & IB_QP_QKEY)
501 sqp->qkey = attr->qkey;
502 if (attr_mask & IB_QP_SQ_PSN)
503 sqp->send_psn = attr->sq_psn;
506 static void init_port(struct mthca_dev *dev, int port)
510 struct mthca_init_ib_param param;
512 memset(¶m, 0, sizeof param);
514 param.port_width = dev->limits.port_width_cap;
515 param.vl_cap = dev->limits.vl_cap;
516 param.mtu_cap = dev->limits.mtu_cap;
517 param.gid_cap = dev->limits.gid_table_len;
518 param.pkey_cap = dev->limits.pkey_table_len;
520 err = mthca_INIT_IB(dev, ¶m, port, &status);
522 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
524 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
527 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
529 struct mthca_dev *dev = to_mdev(ibqp->device);
530 struct mthca_qp *qp = to_mqp(ibqp);
531 enum ib_qp_state cur_state, new_state;
532 struct mthca_mailbox *mailbox;
533 struct mthca_qp_param *qp_param;
534 struct mthca_qp_context *qp_context;
535 u32 req_param, opt_param;
539 if (attr_mask & IB_QP_CUR_STATE) {
540 if (attr->cur_qp_state != IB_QPS_RTR &&
541 attr->cur_qp_state != IB_QPS_RTS &&
542 attr->cur_qp_state != IB_QPS_SQD &&
543 attr->cur_qp_state != IB_QPS_SQE)
546 cur_state = attr->cur_qp_state;
548 spin_lock_irq(&qp->sq.lock);
549 spin_lock(&qp->rq.lock);
550 cur_state = qp->state;
551 spin_unlock(&qp->rq.lock);
552 spin_unlock_irq(&qp->sq.lock);
555 if (attr_mask & IB_QP_STATE) {
556 if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
558 new_state = attr->qp_state;
560 new_state = cur_state;
562 if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
563 mthca_dbg(dev, "Illegal QP transition "
564 "%d->%d\n", cur_state, new_state);
568 req_param = state_table[cur_state][new_state].req_param[qp->transport];
569 opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
571 if ((req_param & attr_mask) != req_param) {
572 mthca_dbg(dev, "QP transition "
573 "%d->%d missing req attr 0x%08x\n",
574 cur_state, new_state,
575 req_param & ~attr_mask);
579 if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
580 mthca_dbg(dev, "QP transition (transport %d) "
581 "%d->%d has extra attr 0x%08x\n",
583 cur_state, new_state,
584 attr_mask & ~(req_param | opt_param |
589 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
591 return PTR_ERR(mailbox);
592 qp_param = mailbox->buf;
593 qp_context = &qp_param->context;
594 memset(qp_param, 0, sizeof *qp_param);
596 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
597 (to_mthca_st(qp->transport) << 16));
598 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
599 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
600 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
602 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
603 switch (attr->path_mig_state) {
604 case IB_MIG_MIGRATED:
605 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
608 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
611 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
616 /* leave tavor_sched_queue as 0 */
618 if (qp->transport == MLX || qp->transport == UD)
619 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
620 else if (attr_mask & IB_QP_PATH_MTU)
621 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
623 if (mthca_is_memfree(dev)) {
625 qp_context->rq_size_stride = long_log2(qp->rq.max) << 3;
626 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4;
629 qp_context->sq_size_stride = long_log2(qp->sq.max) << 3;
630 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4;
633 /* leave arbel_sched_queue as 0 */
635 if (qp->ibqp.uobject)
636 qp_context->usr_page =
637 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index);
639 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
640 qp_context->local_qpn = cpu_to_be32(qp->qpn);
641 if (attr_mask & IB_QP_DEST_QPN) {
642 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
645 if (qp->transport == MLX)
646 qp_context->pri_path.port_pkey |=
647 cpu_to_be32(to_msqp(qp)->port << 24);
649 if (attr_mask & IB_QP_PORT) {
650 qp_context->pri_path.port_pkey |=
651 cpu_to_be32(attr->port_num << 24);
652 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
656 if (attr_mask & IB_QP_PKEY_INDEX) {
657 qp_context->pri_path.port_pkey |=
658 cpu_to_be32(attr->pkey_index);
659 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
662 if (attr_mask & IB_QP_RNR_RETRY) {
663 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
664 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
667 if (attr_mask & IB_QP_AV) {
668 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
669 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
670 qp_context->pri_path.static_rate = !!attr->ah_attr.static_rate;
671 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
672 qp_context->pri_path.g_mylmc |= 1 << 7;
673 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
674 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
675 qp_context->pri_path.sl_tclass_flowlabel =
676 cpu_to_be32((attr->ah_attr.sl << 28) |
677 (attr->ah_attr.grh.traffic_class << 20) |
678 (attr->ah_attr.grh.flow_label));
679 memcpy(qp_context->pri_path.rgid,
680 attr->ah_attr.grh.dgid.raw, 16);
682 qp_context->pri_path.sl_tclass_flowlabel =
683 cpu_to_be32(attr->ah_attr.sl << 28);
685 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
688 if (attr_mask & IB_QP_TIMEOUT) {
689 qp_context->pri_path.ackto = attr->timeout << 3;
690 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
696 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
697 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
698 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
699 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
700 (MTHCA_FLIGHT_LIMIT << 24) |
704 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
705 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
706 if (attr_mask & IB_QP_RETRY_CNT) {
707 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
708 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
711 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
712 qp_context->params1 |= cpu_to_be32(min(attr->max_rd_atomic ?
713 ffs(attr->max_rd_atomic) - 1 : 0,
715 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
718 if (attr_mask & IB_QP_SQ_PSN)
719 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
720 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
722 if (mthca_is_memfree(dev)) {
723 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
724 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
727 if (attr_mask & IB_QP_ACCESS_FLAGS) {
729 * Only enable RDMA/atomics if we have responder
730 * resources set to a non-zero value.
732 if (qp->resp_depth) {
733 qp_context->params2 |=
734 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
735 MTHCA_QP_BIT_RWE : 0);
736 qp_context->params2 |=
737 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
738 MTHCA_QP_BIT_RRE : 0);
739 qp_context->params2 |=
740 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
741 MTHCA_QP_BIT_RAE : 0);
744 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
745 MTHCA_QP_OPTPAR_RRE |
746 MTHCA_QP_OPTPAR_RAE);
748 qp->atomic_rd_en = attr->qp_access_flags;
751 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
754 if (qp->resp_depth && !attr->max_dest_rd_atomic) {
756 * Lowering our responder resources to zero.
757 * Turn off RDMA/atomics as responder.
758 * (RWE/RRE/RAE in params2 already zero)
760 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
761 MTHCA_QP_OPTPAR_RRE |
762 MTHCA_QP_OPTPAR_RAE);
765 if (!qp->resp_depth && attr->max_dest_rd_atomic) {
767 * Increasing our responder resources from
768 * zero. Turn on RDMA/atomics as appropriate.
770 qp_context->params2 |=
771 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ?
772 MTHCA_QP_BIT_RWE : 0);
773 qp_context->params2 |=
774 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
775 MTHCA_QP_BIT_RRE : 0);
776 qp_context->params2 |=
777 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
778 MTHCA_QP_BIT_RAE : 0);
780 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
781 MTHCA_QP_OPTPAR_RRE |
782 MTHCA_QP_OPTPAR_RAE);
786 1 << rra_max < attr->max_dest_rd_atomic &&
787 rra_max < dev->qp_table.rdb_shift;
791 qp_context->params2 |= cpu_to_be32(rra_max << 21);
792 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
794 qp->resp_depth = attr->max_dest_rd_atomic;
797 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
800 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC);
802 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
803 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
804 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
806 if (attr_mask & IB_QP_RQ_PSN)
807 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
809 qp_context->ra_buff_indx =
810 cpu_to_be32(dev->qp_table.rdb_base +
811 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
812 dev->qp_table.rdb_shift));
814 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
816 if (mthca_is_memfree(dev))
817 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
819 if (attr_mask & IB_QP_QKEY) {
820 qp_context->qkey = cpu_to_be32(attr->qkey);
821 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
825 qp_context->srqn = cpu_to_be32(1 << 24 |
826 to_msrq(ibqp->srq)->srqn);
828 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
829 qp->qpn, 0, mailbox, 0, &status);
831 mthca_warn(dev, "modify QP %d returned status %02x.\n",
832 state_table[cur_state][new_state].trans, status);
837 qp->state = new_state;
839 mthca_free_mailbox(dev, mailbox);
842 store_attrs(to_msqp(qp), attr, attr_mask);
845 * If we moved QP0 to RTR, bring the IB link up; if we moved
846 * QP0 to RESET or ERROR, bring the link back down.
848 if (is_qp0(dev, qp)) {
849 if (cur_state != IB_QPS_RTR &&
850 new_state == IB_QPS_RTR)
851 init_port(dev, to_msqp(qp)->port);
853 if (cur_state != IB_QPS_RESET &&
854 cur_state != IB_QPS_ERR &&
855 (new_state == IB_QPS_RESET ||
856 new_state == IB_QPS_ERR))
857 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
861 * If we moved a kernel QP to RESET, clean up all old CQ
862 * entries and reinitialize the QP.
864 if (!err && new_state == IB_QPS_RESET && !qp->ibqp.uobject) {
865 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
866 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
867 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
868 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
869 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
871 mthca_wq_init(&qp->sq);
872 mthca_wq_init(&qp->rq);
874 if (mthca_is_memfree(dev)) {
884 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
885 * rq.max_gs and sq.max_gs must all be assigned.
886 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
887 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
890 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
897 size = sizeof (struct mthca_next_seg) +
898 qp->rq.max_gs * sizeof (struct mthca_data_seg);
900 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
904 size = sizeof (struct mthca_next_seg) +
905 qp->sq.max_gs * sizeof (struct mthca_data_seg);
906 switch (qp->transport) {
908 size += 2 * sizeof (struct mthca_data_seg);
911 if (mthca_is_memfree(dev))
912 size += sizeof (struct mthca_arbel_ud_seg);
914 size += sizeof (struct mthca_tavor_ud_seg);
917 /* bind seg is as big as atomic + raddr segs */
918 size += sizeof (struct mthca_bind_seg);
921 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
925 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
926 1 << qp->sq.wqe_shift);
929 * If this is a userspace QP, we don't actually have to
930 * allocate anything. All we need is to calculate the WQE
931 * sizes and the send_wqe_offset, so we're done now.
933 if (pd->ibpd.uobject)
936 size = PAGE_ALIGN(qp->send_wqe_offset +
937 (qp->sq.max << qp->sq.wqe_shift));
939 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
944 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE,
945 &qp->queue, &qp->is_direct, pd, 0, &qp->mr);
956 static void mthca_free_wqe_buf(struct mthca_dev *dev,
959 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset +
960 (qp->sq.max << qp->sq.wqe_shift)),
961 &qp->queue, qp->is_direct, &qp->mr);
965 static int mthca_map_memfree(struct mthca_dev *dev,
970 if (mthca_is_memfree(dev)) {
971 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
975 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
979 ret = mthca_table_get(dev, dev->qp_table.rdb_table,
980 qp->qpn << dev->qp_table.rdb_shift);
989 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
992 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
997 static void mthca_unmap_memfree(struct mthca_dev *dev,
1000 mthca_table_put(dev, dev->qp_table.rdb_table,
1001 qp->qpn << dev->qp_table.rdb_shift);
1002 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1003 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1006 static int mthca_alloc_memfree(struct mthca_dev *dev,
1007 struct mthca_qp *qp)
1011 if (mthca_is_memfree(dev)) {
1012 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1013 qp->qpn, &qp->rq.db);
1014 if (qp->rq.db_index < 0)
1017 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1018 qp->qpn, &qp->sq.db);
1019 if (qp->sq.db_index < 0)
1020 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1026 static void mthca_free_memfree(struct mthca_dev *dev,
1027 struct mthca_qp *qp)
1029 if (mthca_is_memfree(dev)) {
1030 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1031 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1035 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1036 struct mthca_pd *pd,
1037 struct mthca_cq *send_cq,
1038 struct mthca_cq *recv_cq,
1039 enum ib_sig_type send_policy,
1040 struct mthca_qp *qp)
1045 atomic_set(&qp->refcount, 1);
1046 init_waitqueue_head(&qp->wait);
1047 qp->state = IB_QPS_RESET;
1048 qp->atomic_rd_en = 0;
1050 qp->sq_policy = send_policy;
1051 mthca_wq_init(&qp->sq);
1052 mthca_wq_init(&qp->rq);
1054 ret = mthca_map_memfree(dev, qp);
1058 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1060 mthca_unmap_memfree(dev, qp);
1065 * If this is a userspace QP, we're done now. The doorbells
1066 * will be allocated and buffers will be initialized in
1069 if (pd->ibpd.uobject)
1072 ret = mthca_alloc_memfree(dev, qp);
1074 mthca_free_wqe_buf(dev, qp);
1075 mthca_unmap_memfree(dev, qp);
1079 if (mthca_is_memfree(dev)) {
1080 struct mthca_next_seg *next;
1081 struct mthca_data_seg *scatter;
1082 int size = (sizeof (struct mthca_next_seg) +
1083 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1085 for (i = 0; i < qp->rq.max; ++i) {
1086 next = get_recv_wqe(qp, i);
1087 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1089 next->ee_nds = cpu_to_be32(size);
1091 for (scatter = (void *) (next + 1);
1092 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1094 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1097 for (i = 0; i < qp->sq.max; ++i) {
1098 next = get_send_wqe(qp, i);
1099 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1101 qp->send_wqe_offset);
1105 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1);
1106 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1);
1111 static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap,
1112 struct mthca_qp *qp)
1114 /* Sanity check QP size before proceeding */
1115 if (cap->max_send_wr > 65536 || cap->max_recv_wr > 65536 ||
1116 cap->max_send_sge > 64 || cap->max_recv_sge > 64)
1119 if (mthca_is_memfree(dev)) {
1120 qp->rq.max = cap->max_recv_wr ?
1121 roundup_pow_of_two(cap->max_recv_wr) : 0;
1122 qp->sq.max = cap->max_send_wr ?
1123 roundup_pow_of_two(cap->max_send_wr) : 0;
1125 qp->rq.max = cap->max_recv_wr;
1126 qp->sq.max = cap->max_send_wr;
1129 qp->rq.max_gs = cap->max_recv_sge;
1130 qp->sq.max_gs = max_t(int, cap->max_send_sge,
1131 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE,
1132 MTHCA_INLINE_CHUNK_SIZE) /
1133 sizeof (struct mthca_data_seg));
1136 * For MLX transport we need 2 extra S/G entries:
1137 * one for the header and one for the checksum at the end
1139 if ((qp->transport == MLX && qp->sq.max_gs + 2 > dev->limits.max_sg) ||
1140 qp->sq.max_gs > dev->limits.max_sg || qp->rq.max_gs > dev->limits.max_sg)
1146 int mthca_alloc_qp(struct mthca_dev *dev,
1147 struct mthca_pd *pd,
1148 struct mthca_cq *send_cq,
1149 struct mthca_cq *recv_cq,
1150 enum ib_qp_type type,
1151 enum ib_sig_type send_policy,
1152 struct ib_qp_cap *cap,
1153 struct mthca_qp *qp)
1157 err = mthca_set_qp_size(dev, cap, qp);
1162 case IB_QPT_RC: qp->transport = RC; break;
1163 case IB_QPT_UC: qp->transport = UC; break;
1164 case IB_QPT_UD: qp->transport = UD; break;
1165 default: return -EINVAL;
1168 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1172 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1175 mthca_free(&dev->qp_table.alloc, qp->qpn);
1179 spin_lock_irq(&dev->qp_table.lock);
1180 mthca_array_set(&dev->qp_table.qp,
1181 qp->qpn & (dev->limits.num_qps - 1), qp);
1182 spin_unlock_irq(&dev->qp_table.lock);
1187 int mthca_alloc_sqp(struct mthca_dev *dev,
1188 struct mthca_pd *pd,
1189 struct mthca_cq *send_cq,
1190 struct mthca_cq *recv_cq,
1191 enum ib_sig_type send_policy,
1192 struct ib_qp_cap *cap,
1195 struct mthca_sqp *sqp)
1197 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1200 err = mthca_set_qp_size(dev, cap, &sqp->qp);
1204 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1205 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1206 &sqp->header_dma, GFP_KERNEL);
1207 if (!sqp->header_buf)
1210 spin_lock_irq(&dev->qp_table.lock);
1211 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1214 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1215 spin_unlock_irq(&dev->qp_table.lock);
1222 sqp->qp.transport = MLX;
1224 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1225 send_policy, &sqp->qp);
1229 atomic_inc(&pd->sqp_count);
1235 * Lock CQs here, so that CQ polling code can do QP lookup
1236 * without taking a lock.
1238 spin_lock_irq(&send_cq->lock);
1239 if (send_cq != recv_cq)
1240 spin_lock(&recv_cq->lock);
1242 spin_lock(&dev->qp_table.lock);
1243 mthca_array_clear(&dev->qp_table.qp, mqpn);
1244 spin_unlock(&dev->qp_table.lock);
1246 if (send_cq != recv_cq)
1247 spin_unlock(&recv_cq->lock);
1248 spin_unlock_irq(&send_cq->lock);
1251 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1252 sqp->header_buf, sqp->header_dma);
1257 void mthca_free_qp(struct mthca_dev *dev,
1258 struct mthca_qp *qp)
1261 struct mthca_cq *send_cq;
1262 struct mthca_cq *recv_cq;
1264 send_cq = to_mcq(qp->ibqp.send_cq);
1265 recv_cq = to_mcq(qp->ibqp.recv_cq);
1268 * Lock CQs here, so that CQ polling code can do QP lookup
1269 * without taking a lock.
1271 spin_lock_irq(&send_cq->lock);
1272 if (send_cq != recv_cq)
1273 spin_lock(&recv_cq->lock);
1275 spin_lock(&dev->qp_table.lock);
1276 mthca_array_clear(&dev->qp_table.qp,
1277 qp->qpn & (dev->limits.num_qps - 1));
1278 spin_unlock(&dev->qp_table.lock);
1280 if (send_cq != recv_cq)
1281 spin_unlock(&recv_cq->lock);
1282 spin_unlock_irq(&send_cq->lock);
1284 atomic_dec(&qp->refcount);
1285 wait_event(qp->wait, !atomic_read(&qp->refcount));
1287 if (qp->state != IB_QPS_RESET)
1288 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1291 * If this is a userspace QP, the buffers, MR, CQs and so on
1292 * will be cleaned up in userspace, so all we have to do is
1293 * unref the mem-free tables and free the QPN in our table.
1295 if (!qp->ibqp.uobject) {
1296 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn,
1297 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1298 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1299 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn,
1300 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1302 mthca_free_memfree(dev, qp);
1303 mthca_free_wqe_buf(dev, qp);
1306 mthca_unmap_memfree(dev, qp);
1308 if (is_sqp(dev, qp)) {
1309 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1310 dma_free_coherent(&dev->pdev->dev,
1311 to_msqp(qp)->header_buf_size,
1312 to_msqp(qp)->header_buf,
1313 to_msqp(qp)->header_dma);
1315 mthca_free(&dev->qp_table.alloc, qp->qpn);
1318 /* Create UD header for an MLX send and build a data segment for it */
1319 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1320 int ind, struct ib_send_wr *wr,
1321 struct mthca_mlx_seg *mlx,
1322 struct mthca_data_seg *data)
1328 ib_ud_header_init(256, /* assume a MAD */
1329 sqp->ud_header.grh_present,
1332 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1335 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1336 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1337 (sqp->ud_header.lrh.destination_lid ==
1338 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) |
1339 (sqp->ud_header.lrh.service_level << 8));
1340 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1343 switch (wr->opcode) {
1345 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1346 sqp->ud_header.immediate_present = 0;
1348 case IB_WR_SEND_WITH_IMM:
1349 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1350 sqp->ud_header.immediate_present = 1;
1351 sqp->ud_header.immediate_data = wr->imm_data;
1357 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1358 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1359 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1360 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1361 if (!sqp->qp.ibqp.qp_num)
1362 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1363 sqp->pkey_index, &pkey);
1365 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1366 wr->wr.ud.pkey_index, &pkey);
1367 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1368 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1369 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1370 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1371 sqp->qkey : wr->wr.ud.remote_qkey);
1372 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1374 header_size = ib_ud_header_pack(&sqp->ud_header,
1376 ind * MTHCA_UD_HEADER_SIZE);
1378 data->byte_count = cpu_to_be32(header_size);
1379 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1380 data->addr = cpu_to_be64(sqp->header_dma +
1381 ind * MTHCA_UD_HEADER_SIZE);
1386 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1387 struct ib_cq *ib_cq)
1390 struct mthca_cq *cq;
1392 cur = wq->head - wq->tail;
1393 if (likely(cur + nreq < wq->max))
1397 spin_lock(&cq->lock);
1398 cur = wq->head - wq->tail;
1399 spin_unlock(&cq->lock);
1401 return cur + nreq >= wq->max;
1404 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1405 struct ib_send_wr **bad_wr)
1407 struct mthca_dev *dev = to_mdev(ibqp->device);
1408 struct mthca_qp *qp = to_mqp(ibqp);
1411 unsigned long flags;
1421 spin_lock_irqsave(&qp->sq.lock, flags);
1423 /* XXX check that state is OK to post send */
1425 ind = qp->sq.next_ind;
1427 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1428 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1429 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1430 " %d max, %d nreq)\n", qp->qpn,
1431 qp->sq.head, qp->sq.tail,
1438 wqe = get_send_wqe(qp, ind);
1439 prev_wqe = qp->sq.last;
1442 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1443 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1444 ((struct mthca_next_seg *) wqe)->flags =
1445 ((wr->send_flags & IB_SEND_SIGNALED) ?
1446 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1447 ((wr->send_flags & IB_SEND_SOLICITED) ?
1448 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1450 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1451 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1452 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1454 wqe += sizeof (struct mthca_next_seg);
1455 size = sizeof (struct mthca_next_seg) / 16;
1457 switch (qp->transport) {
1459 switch (wr->opcode) {
1460 case IB_WR_ATOMIC_CMP_AND_SWP:
1461 case IB_WR_ATOMIC_FETCH_AND_ADD:
1462 ((struct mthca_raddr_seg *) wqe)->raddr =
1463 cpu_to_be64(wr->wr.atomic.remote_addr);
1464 ((struct mthca_raddr_seg *) wqe)->rkey =
1465 cpu_to_be32(wr->wr.atomic.rkey);
1466 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1468 wqe += sizeof (struct mthca_raddr_seg);
1470 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1471 ((struct mthca_atomic_seg *) wqe)->swap_add =
1472 cpu_to_be64(wr->wr.atomic.swap);
1473 ((struct mthca_atomic_seg *) wqe)->compare =
1474 cpu_to_be64(wr->wr.atomic.compare_add);
1476 ((struct mthca_atomic_seg *) wqe)->swap_add =
1477 cpu_to_be64(wr->wr.atomic.compare_add);
1478 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1481 wqe += sizeof (struct mthca_atomic_seg);
1482 size += sizeof (struct mthca_raddr_seg) / 16 +
1483 sizeof (struct mthca_atomic_seg);
1486 case IB_WR_RDMA_WRITE:
1487 case IB_WR_RDMA_WRITE_WITH_IMM:
1488 case IB_WR_RDMA_READ:
1489 ((struct mthca_raddr_seg *) wqe)->raddr =
1490 cpu_to_be64(wr->wr.rdma.remote_addr);
1491 ((struct mthca_raddr_seg *) wqe)->rkey =
1492 cpu_to_be32(wr->wr.rdma.rkey);
1493 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1494 wqe += sizeof (struct mthca_raddr_seg);
1495 size += sizeof (struct mthca_raddr_seg) / 16;
1499 /* No extra segments required for sends */
1506 switch (wr->opcode) {
1507 case IB_WR_RDMA_WRITE:
1508 case IB_WR_RDMA_WRITE_WITH_IMM:
1509 ((struct mthca_raddr_seg *) wqe)->raddr =
1510 cpu_to_be64(wr->wr.rdma.remote_addr);
1511 ((struct mthca_raddr_seg *) wqe)->rkey =
1512 cpu_to_be32(wr->wr.rdma.rkey);
1513 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1514 wqe += sizeof (struct mthca_raddr_seg);
1515 size += sizeof (struct mthca_raddr_seg) / 16;
1519 /* No extra segments required for sends */
1526 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1527 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1528 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1529 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1530 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1531 cpu_to_be32(wr->wr.ud.remote_qpn);
1532 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1533 cpu_to_be32(wr->wr.ud.remote_qkey);
1535 wqe += sizeof (struct mthca_tavor_ud_seg);
1536 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1540 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1541 wqe - sizeof (struct mthca_next_seg),
1547 wqe += sizeof (struct mthca_data_seg);
1548 size += sizeof (struct mthca_data_seg) / 16;
1552 if (wr->num_sge > qp->sq.max_gs) {
1553 mthca_err(dev, "too many gathers\n");
1559 for (i = 0; i < wr->num_sge; ++i) {
1560 ((struct mthca_data_seg *) wqe)->byte_count =
1561 cpu_to_be32(wr->sg_list[i].length);
1562 ((struct mthca_data_seg *) wqe)->lkey =
1563 cpu_to_be32(wr->sg_list[i].lkey);
1564 ((struct mthca_data_seg *) wqe)->addr =
1565 cpu_to_be64(wr->sg_list[i].addr);
1566 wqe += sizeof (struct mthca_data_seg);
1567 size += sizeof (struct mthca_data_seg) / 16;
1570 /* Add one more inline data segment for ICRC */
1571 if (qp->transport == MLX) {
1572 ((struct mthca_data_seg *) wqe)->byte_count =
1573 cpu_to_be32((1 << 31) | 4);
1574 ((u32 *) wqe)[1] = 0;
1575 wqe += sizeof (struct mthca_data_seg);
1576 size += sizeof (struct mthca_data_seg) / 16;
1579 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1581 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1582 mthca_err(dev, "opcode invalid\n");
1588 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1589 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1590 qp->send_wqe_offset) |
1591 mthca_opcode[wr->opcode]);
1593 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1594 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1598 op0 = mthca_opcode[wr->opcode];
1602 if (unlikely(ind >= qp->sq.max))
1610 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1611 qp->send_wqe_offset) | f0 | op0);
1612 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1616 mthca_write64(doorbell,
1617 dev->kar + MTHCA_SEND_DOORBELL,
1618 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1621 qp->sq.next_ind = ind;
1622 qp->sq.head += nreq;
1624 spin_unlock_irqrestore(&qp->sq.lock, flags);
1628 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1629 struct ib_recv_wr **bad_wr)
1631 struct mthca_dev *dev = to_mdev(ibqp->device);
1632 struct mthca_qp *qp = to_mqp(ibqp);
1633 unsigned long flags;
1643 spin_lock_irqsave(&qp->rq.lock, flags);
1645 /* XXX check that state is OK to post receive */
1647 ind = qp->rq.next_ind;
1649 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1650 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1651 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1652 " %d max, %d nreq)\n", qp->qpn,
1653 qp->rq.head, qp->rq.tail,
1660 wqe = get_recv_wqe(qp, ind);
1661 prev_wqe = qp->rq.last;
1664 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1665 ((struct mthca_next_seg *) wqe)->ee_nds =
1666 cpu_to_be32(MTHCA_NEXT_DBD);
1667 ((struct mthca_next_seg *) wqe)->flags = 0;
1669 wqe += sizeof (struct mthca_next_seg);
1670 size = sizeof (struct mthca_next_seg) / 16;
1672 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1678 for (i = 0; i < wr->num_sge; ++i) {
1679 ((struct mthca_data_seg *) wqe)->byte_count =
1680 cpu_to_be32(wr->sg_list[i].length);
1681 ((struct mthca_data_seg *) wqe)->lkey =
1682 cpu_to_be32(wr->sg_list[i].lkey);
1683 ((struct mthca_data_seg *) wqe)->addr =
1684 cpu_to_be64(wr->sg_list[i].addr);
1685 wqe += sizeof (struct mthca_data_seg);
1686 size += sizeof (struct mthca_data_seg) / 16;
1689 qp->wrid[ind] = wr->wr_id;
1691 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1692 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1694 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1695 cpu_to_be32(MTHCA_NEXT_DBD | size);
1701 if (unlikely(ind >= qp->rq.max))
1709 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1710 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1714 mthca_write64(doorbell,
1715 dev->kar + MTHCA_RECEIVE_DOORBELL,
1716 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1719 qp->rq.next_ind = ind;
1720 qp->rq.head += nreq;
1722 spin_unlock_irqrestore(&qp->rq.lock, flags);
1726 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1727 struct ib_send_wr **bad_wr)
1729 struct mthca_dev *dev = to_mdev(ibqp->device);
1730 struct mthca_qp *qp = to_mqp(ibqp);
1733 unsigned long flags;
1743 spin_lock_irqsave(&qp->sq.lock, flags);
1745 /* XXX check that state is OK to post send */
1747 ind = qp->sq.head & (qp->sq.max - 1);
1749 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1750 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1751 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1752 " %d max, %d nreq)\n", qp->qpn,
1753 qp->sq.head, qp->sq.tail,
1760 wqe = get_send_wqe(qp, ind);
1761 prev_wqe = qp->sq.last;
1764 ((struct mthca_next_seg *) wqe)->flags =
1765 ((wr->send_flags & IB_SEND_SIGNALED) ?
1766 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1767 ((wr->send_flags & IB_SEND_SOLICITED) ?
1768 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1770 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1771 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1772 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1774 wqe += sizeof (struct mthca_next_seg);
1775 size = sizeof (struct mthca_next_seg) / 16;
1777 switch (qp->transport) {
1779 switch (wr->opcode) {
1780 case IB_WR_ATOMIC_CMP_AND_SWP:
1781 case IB_WR_ATOMIC_FETCH_AND_ADD:
1782 ((struct mthca_raddr_seg *) wqe)->raddr =
1783 cpu_to_be64(wr->wr.atomic.remote_addr);
1784 ((struct mthca_raddr_seg *) wqe)->rkey =
1785 cpu_to_be32(wr->wr.atomic.rkey);
1786 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1788 wqe += sizeof (struct mthca_raddr_seg);
1790 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1791 ((struct mthca_atomic_seg *) wqe)->swap_add =
1792 cpu_to_be64(wr->wr.atomic.swap);
1793 ((struct mthca_atomic_seg *) wqe)->compare =
1794 cpu_to_be64(wr->wr.atomic.compare_add);
1796 ((struct mthca_atomic_seg *) wqe)->swap_add =
1797 cpu_to_be64(wr->wr.atomic.compare_add);
1798 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1801 wqe += sizeof (struct mthca_atomic_seg);
1802 size += sizeof (struct mthca_raddr_seg) / 16 +
1803 sizeof (struct mthca_atomic_seg);
1806 case IB_WR_RDMA_READ:
1807 case IB_WR_RDMA_WRITE:
1808 case IB_WR_RDMA_WRITE_WITH_IMM:
1809 ((struct mthca_raddr_seg *) wqe)->raddr =
1810 cpu_to_be64(wr->wr.rdma.remote_addr);
1811 ((struct mthca_raddr_seg *) wqe)->rkey =
1812 cpu_to_be32(wr->wr.rdma.rkey);
1813 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1814 wqe += sizeof (struct mthca_raddr_seg);
1815 size += sizeof (struct mthca_raddr_seg) / 16;
1819 /* No extra segments required for sends */
1826 switch (wr->opcode) {
1827 case IB_WR_RDMA_WRITE:
1828 case IB_WR_RDMA_WRITE_WITH_IMM:
1829 ((struct mthca_raddr_seg *) wqe)->raddr =
1830 cpu_to_be64(wr->wr.rdma.remote_addr);
1831 ((struct mthca_raddr_seg *) wqe)->rkey =
1832 cpu_to_be32(wr->wr.rdma.rkey);
1833 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1834 wqe += sizeof (struct mthca_raddr_seg);
1835 size += sizeof (struct mthca_raddr_seg) / 16;
1839 /* No extra segments required for sends */
1846 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1847 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1848 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1849 cpu_to_be32(wr->wr.ud.remote_qpn);
1850 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1851 cpu_to_be32(wr->wr.ud.remote_qkey);
1853 wqe += sizeof (struct mthca_arbel_ud_seg);
1854 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1858 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1859 wqe - sizeof (struct mthca_next_seg),
1865 wqe += sizeof (struct mthca_data_seg);
1866 size += sizeof (struct mthca_data_seg) / 16;
1870 if (wr->num_sge > qp->sq.max_gs) {
1871 mthca_err(dev, "too many gathers\n");
1877 for (i = 0; i < wr->num_sge; ++i) {
1878 ((struct mthca_data_seg *) wqe)->byte_count =
1879 cpu_to_be32(wr->sg_list[i].length);
1880 ((struct mthca_data_seg *) wqe)->lkey =
1881 cpu_to_be32(wr->sg_list[i].lkey);
1882 ((struct mthca_data_seg *) wqe)->addr =
1883 cpu_to_be64(wr->sg_list[i].addr);
1884 wqe += sizeof (struct mthca_data_seg);
1885 size += sizeof (struct mthca_data_seg) / 16;
1888 /* Add one more inline data segment for ICRC */
1889 if (qp->transport == MLX) {
1890 ((struct mthca_data_seg *) wqe)->byte_count =
1891 cpu_to_be32((1 << 31) | 4);
1892 ((u32 *) wqe)[1] = 0;
1893 wqe += sizeof (struct mthca_data_seg);
1894 size += sizeof (struct mthca_data_seg) / 16;
1897 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1899 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1900 mthca_err(dev, "opcode invalid\n");
1906 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1907 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1908 qp->send_wqe_offset) |
1909 mthca_opcode[wr->opcode]);
1911 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1912 cpu_to_be32(MTHCA_NEXT_DBD | size);
1916 op0 = mthca_opcode[wr->opcode];
1920 if (unlikely(ind >= qp->sq.max))
1928 doorbell[0] = cpu_to_be32((nreq << 24) |
1929 ((qp->sq.head & 0xffff) << 8) |
1931 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1933 qp->sq.head += nreq;
1936 * Make sure that descriptors are written before
1940 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1943 * Make sure doorbell record is written before we
1944 * write MMIO send doorbell.
1947 mthca_write64(doorbell,
1948 dev->kar + MTHCA_SEND_DOORBELL,
1949 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1952 spin_unlock_irqrestore(&qp->sq.lock, flags);
1956 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1957 struct ib_recv_wr **bad_wr)
1959 struct mthca_dev *dev = to_mdev(ibqp->device);
1960 struct mthca_qp *qp = to_mqp(ibqp);
1961 unsigned long flags;
1968 spin_lock_irqsave(&qp->rq.lock, flags);
1970 /* XXX check that state is OK to post receive */
1972 ind = qp->rq.head & (qp->rq.max - 1);
1974 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1975 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1976 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1977 " %d max, %d nreq)\n", qp->qpn,
1978 qp->rq.head, qp->rq.tail,
1985 wqe = get_recv_wqe(qp, ind);
1987 ((struct mthca_next_seg *) wqe)->flags = 0;
1989 wqe += sizeof (struct mthca_next_seg);
1991 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1997 for (i = 0; i < wr->num_sge; ++i) {
1998 ((struct mthca_data_seg *) wqe)->byte_count =
1999 cpu_to_be32(wr->sg_list[i].length);
2000 ((struct mthca_data_seg *) wqe)->lkey =
2001 cpu_to_be32(wr->sg_list[i].lkey);
2002 ((struct mthca_data_seg *) wqe)->addr =
2003 cpu_to_be64(wr->sg_list[i].addr);
2004 wqe += sizeof (struct mthca_data_seg);
2007 if (i < qp->rq.max_gs) {
2008 ((struct mthca_data_seg *) wqe)->byte_count = 0;
2009 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
2010 ((struct mthca_data_seg *) wqe)->addr = 0;
2013 qp->wrid[ind] = wr->wr_id;
2016 if (unlikely(ind >= qp->rq.max))
2021 qp->rq.head += nreq;
2024 * Make sure that descriptors are written before
2028 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2031 spin_unlock_irqrestore(&qp->rq.lock, flags);
2035 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2036 int index, int *dbd, __be32 *new_wqe)
2038 struct mthca_next_seg *next;
2041 * For SRQs, all WQEs generate a CQE, so we're always at the
2042 * end of the doorbell chain.
2050 next = get_send_wqe(qp, index);
2052 next = get_recv_wqe(qp, index);
2054 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2055 if (next->ee_nds & cpu_to_be32(0x3f))
2056 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2057 (next->ee_nds & cpu_to_be32(0x3f));
2064 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2070 spin_lock_init(&dev->qp_table.lock);
2073 * We reserve 2 extra QPs per port for the special QPs. The
2074 * special QP for port 1 has to be even, so round up.
2076 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2077 err = mthca_alloc_init(&dev->qp_table.alloc,
2078 dev->limits.num_qps,
2080 dev->qp_table.sqp_start +
2081 MTHCA_MAX_PORTS * 2);
2085 err = mthca_array_init(&dev->qp_table.qp,
2086 dev->limits.num_qps);
2088 mthca_alloc_cleanup(&dev->qp_table.alloc);
2092 for (i = 0; i < 2; ++i) {
2093 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2094 dev->qp_table.sqp_start + i * 2,
2099 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2100 "status %02x, aborting.\n",
2109 for (i = 0; i < 2; ++i)
2110 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2112 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2113 mthca_alloc_cleanup(&dev->qp_table.alloc);
2118 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2123 for (i = 0; i < 2; ++i)
2124 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2126 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2127 mthca_alloc_cleanup(&dev->qp_table.alloc);