2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * $Id: mthca_qp.c 1355 2004-12-17 15:23:43Z roland $
35 #include <linux/init.h>
41 #include "mthca_dev.h"
42 #include "mthca_cmd.h"
43 #include "mthca_memfree.h"
46 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE,
47 MTHCA_ACK_REQ_FREQ = 10,
48 MTHCA_FLIGHT_LIMIT = 9,
49 MTHCA_UD_HEADER_SIZE = 72 /* largest UD header possible */
53 MTHCA_QP_STATE_RST = 0,
54 MTHCA_QP_STATE_INIT = 1,
55 MTHCA_QP_STATE_RTR = 2,
56 MTHCA_QP_STATE_RTS = 3,
57 MTHCA_QP_STATE_SQE = 4,
58 MTHCA_QP_STATE_SQD = 5,
59 MTHCA_QP_STATE_ERR = 6,
60 MTHCA_QP_STATE_DRAINING = 7
72 MTHCA_QP_PM_MIGRATED = 0x3,
73 MTHCA_QP_PM_ARMED = 0x0,
74 MTHCA_QP_PM_REARM = 0x1
78 /* qp_context flags */
79 MTHCA_QP_BIT_DE = 1 << 8,
81 MTHCA_QP_BIT_SRE = 1 << 15,
82 MTHCA_QP_BIT_SWE = 1 << 14,
83 MTHCA_QP_BIT_SAE = 1 << 13,
84 MTHCA_QP_BIT_SIC = 1 << 4,
85 MTHCA_QP_BIT_SSC = 1 << 3,
87 MTHCA_QP_BIT_RRE = 1 << 15,
88 MTHCA_QP_BIT_RWE = 1 << 14,
89 MTHCA_QP_BIT_RAE = 1 << 13,
90 MTHCA_QP_BIT_RIC = 1 << 4,
91 MTHCA_QP_BIT_RSC = 1 << 3
94 struct mthca_qp_path {
103 u32 sl_tclass_flowlabel;
105 } __attribute__((packed));
107 struct mthca_qp_context {
109 u32 tavor_sched_queue; /* Reserved on Arbel */
111 u8 rq_size_stride; /* Reserved on Tavor */
112 u8 sq_size_stride; /* Reserved on Tavor */
113 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */
118 struct mthca_qp_path pri_path;
119 struct mthca_qp_path alt_path;
128 u32 snd_wqe_base_l; /* Next send WQE on Tavor */
129 u32 snd_db_index; /* (debugging only entries) */
136 u32 rcv_wqe_base_l; /* Next recv WQE on Tavor */
137 u32 rcv_db_index; /* (debugging only entries) */
141 u16 rq_wqe_counter; /* reserved on Tavor */
142 u16 sq_wqe_counter; /* reserved on Tavor */
144 } __attribute__((packed));
146 struct mthca_qp_param {
149 struct mthca_qp_context context;
151 } __attribute__((packed));
154 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0,
155 MTHCA_QP_OPTPAR_RRE = 1 << 1,
156 MTHCA_QP_OPTPAR_RAE = 1 << 2,
157 MTHCA_QP_OPTPAR_RWE = 1 << 3,
158 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4,
159 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5,
160 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6,
161 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7,
162 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8,
163 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9,
164 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10,
165 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11,
166 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12,
167 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13,
168 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14,
169 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15,
170 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16
174 MTHCA_NEXT_DBD = 1 << 7,
175 MTHCA_NEXT_FENCE = 1 << 6,
176 MTHCA_NEXT_CQ_UPDATE = 1 << 3,
177 MTHCA_NEXT_EVENT_GEN = 1 << 2,
178 MTHCA_NEXT_SOLICIT = 1 << 1,
180 MTHCA_MLX_VL15 = 1 << 17,
181 MTHCA_MLX_SLR = 1 << 16
184 struct mthca_next_seg {
185 u32 nda_op; /* [31:6] next WQE [4:0] next opcode */
186 u32 ee_nds; /* [31:8] next EE [7] DBD [6] F [5:0] next WQE size */
187 u32 flags; /* [3] CQ [2] Event [1] Solicit */
188 u32 imm; /* immediate data */
191 struct mthca_tavor_ud_seg {
201 struct mthca_arbel_ud_seg {
208 struct mthca_bind_seg {
209 u32 flags; /* [31] Atomic [30] rem write [29] rem read */
217 struct mthca_raddr_seg {
223 struct mthca_atomic_seg {
228 struct mthca_data_seg {
234 struct mthca_mlx_seg {
237 u32 flags; /* [17] VL15 [16] SLR [14:12] static rate
238 [11:8] SL [3] C [2] E */
243 static const u8 mthca_opcode[] = {
244 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
245 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
246 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
247 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
248 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
249 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
250 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
253 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
255 return qp->qpn >= dev->qp_table.sqp_start &&
256 qp->qpn <= dev->qp_table.sqp_start + 3;
259 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
261 return qp->qpn >= dev->qp_table.sqp_start &&
262 qp->qpn <= dev->qp_table.sqp_start + 1;
265 static void *get_recv_wqe(struct mthca_qp *qp, int n)
268 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
270 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
271 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
274 static void *get_send_wqe(struct mthca_qp *qp, int n)
277 return qp->queue.direct.buf + qp->send_wqe_offset +
278 (n << qp->sq.wqe_shift);
280 return qp->queue.page_list[(qp->send_wqe_offset +
281 (n << qp->sq.wqe_shift)) >>
283 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
287 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
288 enum ib_event_type event_type)
291 struct ib_event event;
293 spin_lock(&dev->qp_table.lock);
294 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
296 atomic_inc(&qp->refcount);
297 spin_unlock(&dev->qp_table.lock);
300 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
304 event.device = &dev->ib_dev;
305 event.event = event_type;
306 event.element.qp = &qp->ibqp;
307 if (qp->ibqp.event_handler)
308 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
310 if (atomic_dec_and_test(&qp->refcount))
314 static int to_mthca_state(enum ib_qp_state ib_state)
317 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
318 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
319 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
320 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
321 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
322 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
323 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
328 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
330 static int to_mthca_st(int transport)
333 case RC: return MTHCA_QP_ST_RC;
334 case UC: return MTHCA_QP_ST_UC;
335 case UD: return MTHCA_QP_ST_UD;
336 case RD: return MTHCA_QP_ST_RD;
337 case MLX: return MTHCA_QP_ST_MLX;
342 static const struct {
344 u32 req_param[NUM_TRANS];
345 u32 opt_param[NUM_TRANS];
346 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
348 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
349 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
351 .trans = MTHCA_TRANS_RST2INIT,
353 [UD] = (IB_QP_PKEY_INDEX |
356 [RC] = (IB_QP_PKEY_INDEX |
359 [MLX] = (IB_QP_PKEY_INDEX |
362 /* bug-for-bug compatibility with VAPI: */
369 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
370 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
372 .trans = MTHCA_TRANS_INIT2INIT,
374 [UD] = (IB_QP_PKEY_INDEX |
377 [RC] = (IB_QP_PKEY_INDEX |
380 [MLX] = (IB_QP_PKEY_INDEX |
385 .trans = MTHCA_TRANS_INIT2RTR,
391 IB_QP_MAX_DEST_RD_ATOMIC |
392 IB_QP_MIN_RNR_TIMER),
395 [UD] = (IB_QP_PKEY_INDEX |
397 [RC] = (IB_QP_ALT_PATH |
400 [MLX] = (IB_QP_PKEY_INDEX |
406 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
407 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
409 .trans = MTHCA_TRANS_RTR2RTS,
412 [RC] = (IB_QP_TIMEOUT |
416 IB_QP_MAX_QP_RD_ATOMIC),
417 [MLX] = IB_QP_SQ_PSN,
420 [UD] = (IB_QP_CUR_STATE |
422 [RC] = (IB_QP_CUR_STATE |
426 IB_QP_MIN_RNR_TIMER |
427 IB_QP_PATH_MIG_STATE),
428 [MLX] = (IB_QP_CUR_STATE |
434 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
435 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
437 .trans = MTHCA_TRANS_RTS2RTS,
439 [UD] = (IB_QP_CUR_STATE |
441 [RC] = (IB_QP_ACCESS_FLAGS |
443 IB_QP_PATH_MIG_STATE |
444 IB_QP_MIN_RNR_TIMER),
445 [MLX] = (IB_QP_CUR_STATE |
450 .trans = MTHCA_TRANS_RTS2SQD,
454 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
455 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
457 .trans = MTHCA_TRANS_SQD2RTS,
459 [UD] = (IB_QP_CUR_STATE |
461 [RC] = (IB_QP_CUR_STATE |
464 IB_QP_MIN_RNR_TIMER |
465 IB_QP_PATH_MIG_STATE),
466 [MLX] = (IB_QP_CUR_STATE |
471 .trans = MTHCA_TRANS_SQD2SQD,
473 [UD] = (IB_QP_PKEY_INDEX |
479 IB_QP_MAX_QP_RD_ATOMIC |
480 IB_QP_MAX_DEST_RD_ATOMIC |
485 IB_QP_MIN_RNR_TIMER |
486 IB_QP_PATH_MIG_STATE),
487 [MLX] = (IB_QP_PKEY_INDEX |
493 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
494 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
496 .trans = MTHCA_TRANS_SQERR2RTS,
498 [UD] = (IB_QP_CUR_STATE |
500 [RC] = (IB_QP_CUR_STATE |
501 IB_QP_MIN_RNR_TIMER),
502 [MLX] = (IB_QP_CUR_STATE |
508 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
509 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
513 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
516 if (attr_mask & IB_QP_PKEY_INDEX)
517 sqp->pkey_index = attr->pkey_index;
518 if (attr_mask & IB_QP_QKEY)
519 sqp->qkey = attr->qkey;
520 if (attr_mask & IB_QP_SQ_PSN)
521 sqp->send_psn = attr->sq_psn;
524 static void init_port(struct mthca_dev *dev, int port)
528 struct mthca_init_ib_param param;
530 memset(¶m, 0, sizeof param);
534 param.vl_cap = dev->limits.vl_cap;
535 param.mtu_cap = dev->limits.mtu_cap;
536 param.gid_cap = dev->limits.gid_table_len;
537 param.pkey_cap = dev->limits.pkey_table_len;
539 err = mthca_INIT_IB(dev, ¶m, port, &status);
541 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
543 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
546 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
548 struct mthca_dev *dev = to_mdev(ibqp->device);
549 struct mthca_qp *qp = to_mqp(ibqp);
550 enum ib_qp_state cur_state, new_state;
551 void *mailbox = NULL;
552 struct mthca_qp_param *qp_param;
553 struct mthca_qp_context *qp_context;
554 u32 req_param, opt_param;
558 if (attr_mask & IB_QP_CUR_STATE) {
559 if (attr->cur_qp_state != IB_QPS_RTR &&
560 attr->cur_qp_state != IB_QPS_RTS &&
561 attr->cur_qp_state != IB_QPS_SQD &&
562 attr->cur_qp_state != IB_QPS_SQE)
565 cur_state = attr->cur_qp_state;
567 spin_lock_irq(&qp->sq.lock);
568 spin_lock(&qp->rq.lock);
569 cur_state = qp->state;
570 spin_unlock(&qp->rq.lock);
571 spin_unlock_irq(&qp->sq.lock);
574 if (attr_mask & IB_QP_STATE) {
575 if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
577 new_state = attr->qp_state;
579 new_state = cur_state;
581 if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
582 mthca_dbg(dev, "Illegal QP transition "
583 "%d->%d\n", cur_state, new_state);
587 req_param = state_table[cur_state][new_state].req_param[qp->transport];
588 opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
590 if ((req_param & attr_mask) != req_param) {
591 mthca_dbg(dev, "QP transition "
592 "%d->%d missing req attr 0x%08x\n",
593 cur_state, new_state,
594 req_param & ~attr_mask);
598 if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
599 mthca_dbg(dev, "QP transition (transport %d) "
600 "%d->%d has extra attr 0x%08x\n",
602 cur_state, new_state,
603 attr_mask & ~(req_param | opt_param |
608 mailbox = kmalloc(sizeof (*qp_param) + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
611 qp_param = MAILBOX_ALIGN(mailbox);
612 qp_context = &qp_param->context;
613 memset(qp_param, 0, sizeof *qp_param);
615 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
616 (to_mthca_st(qp->transport) << 16));
617 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
618 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
619 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
621 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
622 switch (attr->path_mig_state) {
623 case IB_MIG_MIGRATED:
624 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
627 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
630 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
635 /* leave tavor_sched_queue as 0 */
637 if (qp->transport == MLX || qp->transport == UD)
638 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
639 else if (attr_mask & IB_QP_PATH_MTU)
640 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
642 if (mthca_is_memfree(dev)) {
643 qp_context->rq_size_stride =
644 ((ffs(qp->rq.max) - 1) << 3) | (qp->rq.wqe_shift - 4);
645 qp_context->sq_size_stride =
646 ((ffs(qp->sq.max) - 1) << 3) | (qp->sq.wqe_shift - 4);
649 /* leave arbel_sched_queue as 0 */
651 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
652 qp_context->local_qpn = cpu_to_be32(qp->qpn);
653 if (attr_mask & IB_QP_DEST_QPN) {
654 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
657 if (qp->transport == MLX)
658 qp_context->pri_path.port_pkey |=
659 cpu_to_be32(to_msqp(qp)->port << 24);
661 if (attr_mask & IB_QP_PORT) {
662 qp_context->pri_path.port_pkey |=
663 cpu_to_be32(attr->port_num << 24);
664 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
668 if (attr_mask & IB_QP_PKEY_INDEX) {
669 qp_context->pri_path.port_pkey |=
670 cpu_to_be32(attr->pkey_index);
671 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
674 if (attr_mask & IB_QP_RNR_RETRY) {
675 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
676 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
679 if (attr_mask & IB_QP_AV) {
680 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
681 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
682 qp_context->pri_path.static_rate = (!!attr->ah_attr.static_rate) << 3;
683 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
684 qp_context->pri_path.g_mylmc |= 1 << 7;
685 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
686 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
687 qp_context->pri_path.sl_tclass_flowlabel =
688 cpu_to_be32((attr->ah_attr.sl << 28) |
689 (attr->ah_attr.grh.traffic_class << 20) |
690 (attr->ah_attr.grh.flow_label));
691 memcpy(qp_context->pri_path.rgid,
692 attr->ah_attr.grh.dgid.raw, 16);
694 qp_context->pri_path.sl_tclass_flowlabel =
695 cpu_to_be32(attr->ah_attr.sl << 28);
697 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
700 if (attr_mask & IB_QP_TIMEOUT) {
701 qp_context->pri_path.ackto = attr->timeout;
702 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
708 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
709 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
710 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
711 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
712 (MTHCA_FLIGHT_LIMIT << 24) |
716 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
717 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
718 if (attr_mask & IB_QP_RETRY_CNT) {
719 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
720 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
723 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
724 qp_context->params1 |= cpu_to_be32(min(attr->max_dest_rd_atomic ?
725 ffs(attr->max_dest_rd_atomic) - 1 : 0,
727 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
730 if (attr_mask & IB_QP_SQ_PSN)
731 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
732 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
734 if (mthca_is_memfree(dev)) {
735 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
736 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
739 if (attr_mask & IB_QP_ACCESS_FLAGS) {
741 * Only enable RDMA/atomics if we have responder
742 * resources set to a non-zero value.
744 if (qp->resp_depth) {
745 qp_context->params2 |=
746 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
747 MTHCA_QP_BIT_RWE : 0);
748 qp_context->params2 |=
749 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
750 MTHCA_QP_BIT_RRE : 0);
751 qp_context->params2 |=
752 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
753 MTHCA_QP_BIT_RAE : 0);
756 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
757 MTHCA_QP_OPTPAR_RRE |
758 MTHCA_QP_OPTPAR_RAE);
760 qp->atomic_rd_en = attr->qp_access_flags;
763 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
766 if (qp->resp_depth && !attr->max_rd_atomic) {
768 * Lowering our responder resources to zero.
769 * Turn off RDMA/atomics as responder.
770 * (RWE/RRE/RAE in params2 already zero)
772 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
773 MTHCA_QP_OPTPAR_RRE |
774 MTHCA_QP_OPTPAR_RAE);
777 if (!qp->resp_depth && attr->max_rd_atomic) {
779 * Increasing our responder resources from
780 * zero. Turn on RDMA/atomics as appropriate.
782 qp_context->params2 |=
783 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ?
784 MTHCA_QP_BIT_RWE : 0);
785 qp_context->params2 |=
786 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
787 MTHCA_QP_BIT_RRE : 0);
788 qp_context->params2 |=
789 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
790 MTHCA_QP_BIT_RAE : 0);
792 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
793 MTHCA_QP_OPTPAR_RRE |
794 MTHCA_QP_OPTPAR_RAE);
798 1 << rra_max < attr->max_rd_atomic &&
799 rra_max < dev->qp_table.rdb_shift;
803 qp_context->params2 |= cpu_to_be32(rra_max << 21);
804 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
806 qp->resp_depth = attr->max_rd_atomic;
809 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
811 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
812 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
813 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
815 if (attr_mask & IB_QP_RQ_PSN)
816 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
818 qp_context->ra_buff_indx =
819 cpu_to_be32(dev->qp_table.rdb_base +
820 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
821 dev->qp_table.rdb_shift));
823 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
825 if (mthca_is_memfree(dev))
826 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
828 if (attr_mask & IB_QP_QKEY) {
829 qp_context->qkey = cpu_to_be32(attr->qkey);
830 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
833 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
834 qp->qpn, 0, qp_param, 0, &status);
836 mthca_warn(dev, "modify QP %d returned status %02x.\n",
837 state_table[cur_state][new_state].trans, status);
842 qp->state = new_state;
847 store_attrs(to_msqp(qp), attr, attr_mask);
850 * If we are moving QP0 to RTR, bring the IB link up; if we
851 * are moving QP0 to RESET or ERROR, bring the link back down.
853 if (is_qp0(dev, qp)) {
854 if (cur_state != IB_QPS_RTR &&
855 new_state == IB_QPS_RTR)
856 init_port(dev, to_msqp(qp)->port);
858 if (cur_state != IB_QPS_RESET &&
859 cur_state != IB_QPS_ERR &&
860 (new_state == IB_QPS_RESET ||
861 new_state == IB_QPS_ERR))
862 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
869 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
870 * rq.max_gs and sq.max_gs must all be assigned.
871 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
872 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
875 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
883 u64 *dma_list = NULL;
886 size = sizeof (struct mthca_next_seg) +
887 qp->rq.max_gs * sizeof (struct mthca_data_seg);
889 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
893 size = sizeof (struct mthca_next_seg) +
894 qp->sq.max_gs * sizeof (struct mthca_data_seg);
895 switch (qp->transport) {
897 size += 2 * sizeof (struct mthca_data_seg);
900 if (mthca_is_memfree(dev))
901 size += sizeof (struct mthca_arbel_ud_seg);
903 size += sizeof (struct mthca_tavor_ud_seg);
906 /* bind seg is as big as atomic + raddr segs */
907 size += sizeof (struct mthca_bind_seg);
910 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
914 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
915 1 << qp->sq.wqe_shift);
916 size = PAGE_ALIGN(qp->send_wqe_offset +
917 (qp->sq.max << qp->sq.wqe_shift));
919 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
924 if (size <= MTHCA_MAX_DIRECT_QP_SIZE) {
927 shift = get_order(size) + PAGE_SHIFT;
930 mthca_dbg(dev, "Creating direct QP of size %d (shift %d)\n",
933 qp->queue.direct.buf = pci_alloc_consistent(dev->pdev, size, &t);
934 if (!qp->queue.direct.buf)
937 pci_unmap_addr_set(&qp->queue.direct, mapping, t);
939 memset(qp->queue.direct.buf, 0, size);
941 while (t & ((1 << shift) - 1)) {
946 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
950 for (i = 0; i < npages; ++i)
951 dma_list[i] = t + i * (1 << shift);
954 npages = size / PAGE_SIZE;
958 mthca_dbg(dev, "Creating indirect QP with %d pages\n", npages);
960 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
964 qp->queue.page_list = kmalloc(npages *
965 sizeof *qp->queue.page_list,
967 if (!qp->queue.page_list)
970 for (i = 0; i < npages; ++i) {
971 qp->queue.page_list[i].buf =
972 pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
973 if (!qp->queue.page_list[i].buf)
976 memset(qp->queue.page_list[i].buf, 0, PAGE_SIZE);
978 pci_unmap_addr_set(&qp->queue.page_list[i], mapping, t);
983 err = mthca_mr_alloc_phys(dev, pd->pd_num, dma_list, shift,
985 MTHCA_MPT_FLAG_LOCAL_READ,
995 pci_free_consistent(dev->pdev, size,
996 qp->queue.direct.buf,
997 pci_unmap_addr(&qp->queue.direct, mapping));
999 for (i = 0; i < npages; ++i) {
1000 if (qp->queue.page_list[i].buf)
1001 pci_free_consistent(dev->pdev, PAGE_SIZE,
1002 qp->queue.page_list[i].buf,
1003 pci_unmap_addr(&qp->queue.page_list[i],
1014 static int mthca_alloc_memfree(struct mthca_dev *dev,
1015 struct mthca_qp *qp)
1019 if (mthca_is_memfree(dev)) {
1020 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1024 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1028 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1029 qp->qpn, &qp->rq.db);
1030 if (qp->rq.db_index < 0) {
1035 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1036 qp->qpn, &qp->sq.db);
1037 if (qp->sq.db_index < 0) {
1046 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1049 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1052 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1057 static void mthca_free_memfree(struct mthca_dev *dev,
1058 struct mthca_qp *qp)
1060 if (mthca_is_memfree(dev)) {
1061 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1062 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1063 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1064 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1068 static void mthca_wq_init(struct mthca_wq* wq)
1070 spin_lock_init(&wq->lock);
1072 wq->last_comp = wq->max - 1;
1078 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1079 struct mthca_pd *pd,
1080 struct mthca_cq *send_cq,
1081 struct mthca_cq *recv_cq,
1082 enum ib_sig_type send_policy,
1083 struct mthca_qp *qp)
1085 struct mthca_next_seg *wqe;
1089 atomic_set(&qp->refcount, 1);
1090 qp->state = IB_QPS_RESET;
1091 qp->atomic_rd_en = 0;
1093 qp->sq_policy = send_policy;
1094 mthca_wq_init(&qp->sq);
1095 mthca_wq_init(&qp->rq);
1097 ret = mthca_alloc_memfree(dev, qp);
1101 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1103 mthca_free_memfree(dev, qp);
1107 if (mthca_is_memfree(dev)) {
1108 for (i = 0; i < qp->rq.max; ++i) {
1109 wqe = get_recv_wqe(qp, i);
1110 wqe->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1112 wqe->ee_nds = cpu_to_be32(1 << (qp->rq.wqe_shift - 4));
1115 for (i = 0; i < qp->sq.max; ++i) {
1116 wqe = get_send_wqe(qp, i);
1117 wqe->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1119 qp->send_wqe_offset);
1126 static void mthca_align_qp_size(struct mthca_dev *dev, struct mthca_qp *qp)
1130 if (!mthca_is_memfree(dev))
1133 for (i = 0; 1 << i < qp->rq.max; ++i)
1136 qp->rq.max = 1 << i;
1138 for (i = 0; 1 << i < qp->sq.max; ++i)
1141 qp->sq.max = 1 << i;
1144 int mthca_alloc_qp(struct mthca_dev *dev,
1145 struct mthca_pd *pd,
1146 struct mthca_cq *send_cq,
1147 struct mthca_cq *recv_cq,
1148 enum ib_qp_type type,
1149 enum ib_sig_type send_policy,
1150 struct mthca_qp *qp)
1154 mthca_align_qp_size(dev, qp);
1157 case IB_QPT_RC: qp->transport = RC; break;
1158 case IB_QPT_UC: qp->transport = UC; break;
1159 case IB_QPT_UD: qp->transport = UD; break;
1160 default: return -EINVAL;
1163 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1167 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1170 mthca_free(&dev->qp_table.alloc, qp->qpn);
1174 spin_lock_irq(&dev->qp_table.lock);
1175 mthca_array_set(&dev->qp_table.qp,
1176 qp->qpn & (dev->limits.num_qps - 1), qp);
1177 spin_unlock_irq(&dev->qp_table.lock);
1182 int mthca_alloc_sqp(struct mthca_dev *dev,
1183 struct mthca_pd *pd,
1184 struct mthca_cq *send_cq,
1185 struct mthca_cq *recv_cq,
1186 enum ib_sig_type send_policy,
1189 struct mthca_sqp *sqp)
1192 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1194 mthca_align_qp_size(dev, &sqp->qp);
1196 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1197 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1198 &sqp->header_dma, GFP_KERNEL);
1199 if (!sqp->header_buf)
1202 spin_lock_irq(&dev->qp_table.lock);
1203 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1206 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1207 spin_unlock_irq(&dev->qp_table.lock);
1214 sqp->qp.transport = MLX;
1216 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1217 send_policy, &sqp->qp);
1221 atomic_inc(&pd->sqp_count);
1227 * Lock CQs here, so that CQ polling code can do QP lookup
1228 * without taking a lock.
1230 spin_lock_irq(&send_cq->lock);
1231 if (send_cq != recv_cq)
1232 spin_lock(&recv_cq->lock);
1234 spin_lock(&dev->qp_table.lock);
1235 mthca_array_clear(&dev->qp_table.qp, mqpn);
1236 spin_unlock(&dev->qp_table.lock);
1238 if (send_cq != recv_cq)
1239 spin_unlock(&recv_cq->lock);
1240 spin_unlock_irq(&send_cq->lock);
1243 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1244 sqp->header_buf, sqp->header_dma);
1249 void mthca_free_qp(struct mthca_dev *dev,
1250 struct mthca_qp *qp)
1255 struct mthca_cq *send_cq;
1256 struct mthca_cq *recv_cq;
1258 send_cq = to_mcq(qp->ibqp.send_cq);
1259 recv_cq = to_mcq(qp->ibqp.recv_cq);
1262 * Lock CQs here, so that CQ polling code can do QP lookup
1263 * without taking a lock.
1265 spin_lock_irq(&send_cq->lock);
1266 if (send_cq != recv_cq)
1267 spin_lock(&recv_cq->lock);
1269 spin_lock(&dev->qp_table.lock);
1270 mthca_array_clear(&dev->qp_table.qp,
1271 qp->qpn & (dev->limits.num_qps - 1));
1272 spin_unlock(&dev->qp_table.lock);
1274 if (send_cq != recv_cq)
1275 spin_unlock(&recv_cq->lock);
1276 spin_unlock_irq(&send_cq->lock);
1278 atomic_dec(&qp->refcount);
1279 wait_event(qp->wait, !atomic_read(&qp->refcount));
1281 if (qp->state != IB_QPS_RESET)
1282 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1284 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
1285 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1286 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
1288 mthca_free_mr(dev, &qp->mr);
1290 size = PAGE_ALIGN(qp->send_wqe_offset +
1291 (qp->sq.max << qp->sq.wqe_shift));
1293 if (qp->is_direct) {
1294 pci_free_consistent(dev->pdev, size,
1295 qp->queue.direct.buf,
1296 pci_unmap_addr(&qp->queue.direct, mapping));
1298 for (i = 0; i < size / PAGE_SIZE; ++i) {
1299 pci_free_consistent(dev->pdev, PAGE_SIZE,
1300 qp->queue.page_list[i].buf,
1301 pci_unmap_addr(&qp->queue.page_list[i],
1308 mthca_free_memfree(dev, qp);
1310 if (is_sqp(dev, qp)) {
1311 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1312 dma_free_coherent(&dev->pdev->dev,
1313 to_msqp(qp)->header_buf_size,
1314 to_msqp(qp)->header_buf,
1315 to_msqp(qp)->header_dma);
1317 mthca_free(&dev->qp_table.alloc, qp->qpn);
1320 /* Create UD header for an MLX send and build a data segment for it */
1321 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1322 int ind, struct ib_send_wr *wr,
1323 struct mthca_mlx_seg *mlx,
1324 struct mthca_data_seg *data)
1329 ib_ud_header_init(256, /* assume a MAD */
1330 sqp->ud_header.grh_present,
1333 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1336 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1337 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1338 (sqp->ud_header.lrh.destination_lid == 0xffff ?
1339 MTHCA_MLX_SLR : 0) |
1340 (sqp->ud_header.lrh.service_level << 8));
1341 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1344 switch (wr->opcode) {
1346 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1347 sqp->ud_header.immediate_present = 0;
1349 case IB_WR_SEND_WITH_IMM:
1350 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1351 sqp->ud_header.immediate_present = 1;
1352 sqp->ud_header.immediate_data = wr->imm_data;
1358 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1359 if (sqp->ud_header.lrh.destination_lid == 0xffff)
1360 sqp->ud_header.lrh.source_lid = 0xffff;
1361 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1362 if (!sqp->qp.ibqp.qp_num)
1363 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1365 &sqp->ud_header.bth.pkey);
1367 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1368 wr->wr.ud.pkey_index,
1369 &sqp->ud_header.bth.pkey);
1370 cpu_to_be16s(&sqp->ud_header.bth.pkey);
1371 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1372 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1373 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1374 sqp->qkey : wr->wr.ud.remote_qkey);
1375 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1377 header_size = ib_ud_header_pack(&sqp->ud_header,
1379 ind * MTHCA_UD_HEADER_SIZE);
1381 data->byte_count = cpu_to_be32(header_size);
1382 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1383 data->addr = cpu_to_be64(sqp->header_dma +
1384 ind * MTHCA_UD_HEADER_SIZE);
1389 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1390 struct ib_cq *ib_cq)
1393 struct mthca_cq *cq;
1395 cur = wq->head - wq->tail;
1396 if (likely(cur + nreq < wq->max))
1400 spin_lock(&cq->lock);
1401 cur = wq->head - wq->tail;
1402 spin_unlock(&cq->lock);
1404 return cur + nreq >= wq->max;
1407 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1408 struct ib_send_wr **bad_wr)
1410 struct mthca_dev *dev = to_mdev(ibqp->device);
1411 struct mthca_qp *qp = to_mqp(ibqp);
1414 unsigned long flags;
1424 spin_lock_irqsave(&qp->sq.lock, flags);
1426 /* XXX check that state is OK to post send */
1428 ind = qp->sq.next_ind;
1430 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1431 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1432 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1433 " %d max, %d nreq)\n", qp->qpn,
1434 qp->sq.head, qp->sq.tail,
1441 wqe = get_send_wqe(qp, ind);
1442 prev_wqe = qp->sq.last;
1445 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1446 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1447 ((struct mthca_next_seg *) wqe)->flags =
1448 ((wr->send_flags & IB_SEND_SIGNALED) ?
1449 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1450 ((wr->send_flags & IB_SEND_SOLICITED) ?
1451 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1453 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1454 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1455 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1457 wqe += sizeof (struct mthca_next_seg);
1458 size = sizeof (struct mthca_next_seg) / 16;
1460 switch (qp->transport) {
1462 switch (wr->opcode) {
1463 case IB_WR_ATOMIC_CMP_AND_SWP:
1464 case IB_WR_ATOMIC_FETCH_AND_ADD:
1465 ((struct mthca_raddr_seg *) wqe)->raddr =
1466 cpu_to_be64(wr->wr.atomic.remote_addr);
1467 ((struct mthca_raddr_seg *) wqe)->rkey =
1468 cpu_to_be32(wr->wr.atomic.rkey);
1469 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1471 wqe += sizeof (struct mthca_raddr_seg);
1473 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1474 ((struct mthca_atomic_seg *) wqe)->swap_add =
1475 cpu_to_be64(wr->wr.atomic.swap);
1476 ((struct mthca_atomic_seg *) wqe)->compare =
1477 cpu_to_be64(wr->wr.atomic.compare_add);
1479 ((struct mthca_atomic_seg *) wqe)->swap_add =
1480 cpu_to_be64(wr->wr.atomic.compare_add);
1481 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1484 wqe += sizeof (struct mthca_atomic_seg);
1485 size += sizeof (struct mthca_raddr_seg) / 16 +
1486 sizeof (struct mthca_atomic_seg);
1489 case IB_WR_RDMA_WRITE:
1490 case IB_WR_RDMA_WRITE_WITH_IMM:
1491 case IB_WR_RDMA_READ:
1492 ((struct mthca_raddr_seg *) wqe)->raddr =
1493 cpu_to_be64(wr->wr.rdma.remote_addr);
1494 ((struct mthca_raddr_seg *) wqe)->rkey =
1495 cpu_to_be32(wr->wr.rdma.rkey);
1496 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1497 wqe += sizeof (struct mthca_raddr_seg);
1498 size += sizeof (struct mthca_raddr_seg) / 16;
1502 /* No extra segments required for sends */
1509 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1510 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1511 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1512 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1513 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1514 cpu_to_be32(wr->wr.ud.remote_qpn);
1515 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1516 cpu_to_be32(wr->wr.ud.remote_qkey);
1518 wqe += sizeof (struct mthca_tavor_ud_seg);
1519 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1523 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1524 wqe - sizeof (struct mthca_next_seg),
1530 wqe += sizeof (struct mthca_data_seg);
1531 size += sizeof (struct mthca_data_seg) / 16;
1535 if (wr->num_sge > qp->sq.max_gs) {
1536 mthca_err(dev, "too many gathers\n");
1542 for (i = 0; i < wr->num_sge; ++i) {
1543 ((struct mthca_data_seg *) wqe)->byte_count =
1544 cpu_to_be32(wr->sg_list[i].length);
1545 ((struct mthca_data_seg *) wqe)->lkey =
1546 cpu_to_be32(wr->sg_list[i].lkey);
1547 ((struct mthca_data_seg *) wqe)->addr =
1548 cpu_to_be64(wr->sg_list[i].addr);
1549 wqe += sizeof (struct mthca_data_seg);
1550 size += sizeof (struct mthca_data_seg) / 16;
1553 /* Add one more inline data segment for ICRC */
1554 if (qp->transport == MLX) {
1555 ((struct mthca_data_seg *) wqe)->byte_count =
1556 cpu_to_be32((1 << 31) | 4);
1557 ((u32 *) wqe)[1] = 0;
1558 wqe += sizeof (struct mthca_data_seg);
1559 size += sizeof (struct mthca_data_seg) / 16;
1562 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1564 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1565 mthca_err(dev, "opcode invalid\n");
1572 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1573 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1574 qp->send_wqe_offset) |
1575 mthca_opcode[wr->opcode]);
1577 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1578 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1583 op0 = mthca_opcode[wr->opcode];
1587 if (unlikely(ind >= qp->sq.max))
1595 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1596 qp->send_wqe_offset) | f0 | op0);
1597 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1601 mthca_write64(doorbell,
1602 dev->kar + MTHCA_SEND_DOORBELL,
1603 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1606 qp->sq.next_ind = ind;
1607 qp->sq.head += nreq;
1609 spin_unlock_irqrestore(&qp->sq.lock, flags);
1613 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1614 struct ib_recv_wr **bad_wr)
1616 struct mthca_dev *dev = to_mdev(ibqp->device);
1617 struct mthca_qp *qp = to_mqp(ibqp);
1618 unsigned long flags;
1628 spin_lock_irqsave(&qp->rq.lock, flags);
1630 /* XXX check that state is OK to post receive */
1632 ind = qp->rq.next_ind;
1634 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1635 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1636 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1637 " %d max, %d nreq)\n", qp->qpn,
1638 qp->rq.head, qp->rq.tail,
1645 wqe = get_recv_wqe(qp, ind);
1646 prev_wqe = qp->rq.last;
1649 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1650 ((struct mthca_next_seg *) wqe)->ee_nds =
1651 cpu_to_be32(MTHCA_NEXT_DBD);
1652 ((struct mthca_next_seg *) wqe)->flags = 0;
1654 wqe += sizeof (struct mthca_next_seg);
1655 size = sizeof (struct mthca_next_seg) / 16;
1657 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1663 for (i = 0; i < wr->num_sge; ++i) {
1664 ((struct mthca_data_seg *) wqe)->byte_count =
1665 cpu_to_be32(wr->sg_list[i].length);
1666 ((struct mthca_data_seg *) wqe)->lkey =
1667 cpu_to_be32(wr->sg_list[i].lkey);
1668 ((struct mthca_data_seg *) wqe)->addr =
1669 cpu_to_be64(wr->sg_list[i].addr);
1670 wqe += sizeof (struct mthca_data_seg);
1671 size += sizeof (struct mthca_data_seg) / 16;
1674 qp->wrid[ind] = wr->wr_id;
1676 if (likely(prev_wqe)) {
1677 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1678 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1680 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1681 cpu_to_be32(MTHCA_NEXT_DBD | size);
1688 if (unlikely(ind >= qp->rq.max))
1696 doorbell[0] = cpu_to_be32((qp->rq.next_ind << qp->rq.wqe_shift) | size0);
1697 doorbell[1] = cpu_to_be32((qp->qpn << 8) | nreq);
1701 mthca_write64(doorbell,
1702 dev->kar + MTHCA_RECEIVE_DOORBELL,
1703 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1706 qp->rq.next_ind = ind;
1707 qp->rq.head += nreq;
1709 spin_unlock_irqrestore(&qp->rq.lock, flags);
1713 int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1714 struct ib_send_wr **bad_wr)
1716 struct mthca_dev *dev = to_mdev(ibqp->device);
1717 struct mthca_qp *qp = to_mqp(ibqp);
1720 unsigned long flags;
1730 spin_lock_irqsave(&qp->sq.lock, flags);
1732 /* XXX check that state is OK to post send */
1734 ind = qp->sq.head & (qp->sq.max - 1);
1736 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1737 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1738 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1739 " %d max, %d nreq)\n", qp->qpn,
1740 qp->sq.head, qp->sq.tail,
1747 wqe = get_send_wqe(qp, ind);
1748 prev_wqe = qp->sq.last;
1751 ((struct mthca_next_seg *) wqe)->flags =
1752 ((wr->send_flags & IB_SEND_SIGNALED) ?
1753 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1754 ((wr->send_flags & IB_SEND_SOLICITED) ?
1755 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1757 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1758 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1759 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1761 wqe += sizeof (struct mthca_next_seg);
1762 size = sizeof (struct mthca_next_seg) / 16;
1764 switch (qp->transport) {
1766 switch (wr->opcode) {
1767 case IB_WR_ATOMIC_CMP_AND_SWP:
1768 case IB_WR_ATOMIC_FETCH_AND_ADD:
1769 ((struct mthca_raddr_seg *) wqe)->raddr =
1770 cpu_to_be64(wr->wr.atomic.remote_addr);
1771 ((struct mthca_raddr_seg *) wqe)->rkey =
1772 cpu_to_be32(wr->wr.atomic.rkey);
1773 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1775 wqe += sizeof (struct mthca_raddr_seg);
1777 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1778 ((struct mthca_atomic_seg *) wqe)->swap_add =
1779 cpu_to_be64(wr->wr.atomic.swap);
1780 ((struct mthca_atomic_seg *) wqe)->compare =
1781 cpu_to_be64(wr->wr.atomic.compare_add);
1783 ((struct mthca_atomic_seg *) wqe)->swap_add =
1784 cpu_to_be64(wr->wr.atomic.compare_add);
1785 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1788 wqe += sizeof (struct mthca_atomic_seg);
1789 size += sizeof (struct mthca_raddr_seg) / 16 +
1790 sizeof (struct mthca_atomic_seg);
1793 case IB_WR_RDMA_WRITE:
1794 case IB_WR_RDMA_WRITE_WITH_IMM:
1795 case IB_WR_RDMA_READ:
1796 ((struct mthca_raddr_seg *) wqe)->raddr =
1797 cpu_to_be64(wr->wr.rdma.remote_addr);
1798 ((struct mthca_raddr_seg *) wqe)->rkey =
1799 cpu_to_be32(wr->wr.rdma.rkey);
1800 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1801 wqe += sizeof (struct mthca_raddr_seg);
1802 size += sizeof (struct mthca_raddr_seg) / 16;
1806 /* No extra segments required for sends */
1813 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1814 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1815 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1816 cpu_to_be32(wr->wr.ud.remote_qpn);
1817 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1818 cpu_to_be32(wr->wr.ud.remote_qkey);
1820 wqe += sizeof (struct mthca_arbel_ud_seg);
1821 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1825 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1826 wqe - sizeof (struct mthca_next_seg),
1832 wqe += sizeof (struct mthca_data_seg);
1833 size += sizeof (struct mthca_data_seg) / 16;
1837 if (wr->num_sge > qp->sq.max_gs) {
1838 mthca_err(dev, "too many gathers\n");
1844 for (i = 0; i < wr->num_sge; ++i) {
1845 ((struct mthca_data_seg *) wqe)->byte_count =
1846 cpu_to_be32(wr->sg_list[i].length);
1847 ((struct mthca_data_seg *) wqe)->lkey =
1848 cpu_to_be32(wr->sg_list[i].lkey);
1849 ((struct mthca_data_seg *) wqe)->addr =
1850 cpu_to_be64(wr->sg_list[i].addr);
1851 wqe += sizeof (struct mthca_data_seg);
1852 size += sizeof (struct mthca_data_seg) / 16;
1855 /* Add one more inline data segment for ICRC */
1856 if (qp->transport == MLX) {
1857 ((struct mthca_data_seg *) wqe)->byte_count =
1858 cpu_to_be32((1 << 31) | 4);
1859 ((u32 *) wqe)[1] = 0;
1860 wqe += sizeof (struct mthca_data_seg);
1861 size += sizeof (struct mthca_data_seg) / 16;
1864 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1866 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1867 mthca_err(dev, "opcode invalid\n");
1873 if (likely(prev_wqe)) {
1874 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1875 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1876 qp->send_wqe_offset) |
1877 mthca_opcode[wr->opcode]);
1879 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1880 cpu_to_be32(MTHCA_NEXT_DBD | size);
1885 op0 = mthca_opcode[wr->opcode];
1889 if (unlikely(ind >= qp->sq.max))
1897 doorbell[0] = cpu_to_be32((nreq << 24) |
1898 ((qp->sq.head & 0xffff) << 8) |
1900 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1902 qp->sq.head += nreq;
1905 * Make sure that descriptors are written before
1909 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1912 * Make sure doorbell record is written before we
1913 * write MMIO send doorbell.
1916 mthca_write64(doorbell,
1917 dev->kar + MTHCA_SEND_DOORBELL,
1918 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1921 spin_unlock_irqrestore(&qp->sq.lock, flags);
1925 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1926 struct ib_recv_wr **bad_wr)
1928 struct mthca_dev *dev = to_mdev(ibqp->device);
1929 struct mthca_qp *qp = to_mqp(ibqp);
1930 unsigned long flags;
1937 spin_lock_irqsave(&qp->rq.lock, flags);
1939 /* XXX check that state is OK to post receive */
1941 ind = qp->rq.head & (qp->rq.max - 1);
1943 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1944 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1945 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1946 " %d max, %d nreq)\n", qp->qpn,
1947 qp->rq.head, qp->rq.tail,
1954 wqe = get_recv_wqe(qp, ind);
1956 ((struct mthca_next_seg *) wqe)->flags = 0;
1958 wqe += sizeof (struct mthca_next_seg);
1960 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1966 for (i = 0; i < wr->num_sge; ++i) {
1967 ((struct mthca_data_seg *) wqe)->byte_count =
1968 cpu_to_be32(wr->sg_list[i].length);
1969 ((struct mthca_data_seg *) wqe)->lkey =
1970 cpu_to_be32(wr->sg_list[i].lkey);
1971 ((struct mthca_data_seg *) wqe)->addr =
1972 cpu_to_be64(wr->sg_list[i].addr);
1973 wqe += sizeof (struct mthca_data_seg);
1976 if (i < qp->rq.max_gs) {
1977 ((struct mthca_data_seg *) wqe)->byte_count = 0;
1978 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(0x100);
1979 ((struct mthca_data_seg *) wqe)->addr = 0;
1982 qp->wrid[ind] = wr->wr_id;
1985 if (unlikely(ind >= qp->rq.max))
1990 qp->rq.head += nreq;
1993 * Make sure that descriptors are written before
1997 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2000 spin_unlock_irqrestore(&qp->rq.lock, flags);
2004 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2005 int index, int *dbd, u32 *new_wqe)
2007 struct mthca_next_seg *next;
2010 next = get_send_wqe(qp, index);
2012 next = get_recv_wqe(qp, index);
2014 if (mthca_is_memfree(dev))
2017 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2018 if (next->ee_nds & cpu_to_be32(0x3f))
2019 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2020 (next->ee_nds & cpu_to_be32(0x3f));
2027 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2033 spin_lock_init(&dev->qp_table.lock);
2036 * We reserve 2 extra QPs per port for the special QPs. The
2037 * special QP for port 1 has to be even, so round up.
2039 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2040 err = mthca_alloc_init(&dev->qp_table.alloc,
2041 dev->limits.num_qps,
2043 dev->qp_table.sqp_start +
2044 MTHCA_MAX_PORTS * 2);
2048 err = mthca_array_init(&dev->qp_table.qp,
2049 dev->limits.num_qps);
2051 mthca_alloc_cleanup(&dev->qp_table.alloc);
2055 for (i = 0; i < 2; ++i) {
2056 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2057 dev->qp_table.sqp_start + i * 2,
2062 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2063 "status %02x, aborting.\n",
2072 for (i = 0; i < 2; ++i)
2073 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2075 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2076 mthca_alloc_cleanup(&dev->qp_table.alloc);
2081 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2086 for (i = 0; i < 2; ++i)
2087 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2089 mthca_alloc_cleanup(&dev->qp_table.alloc);