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
185 MTHCA_INVAL_LKEY = 0x100
188 struct mthca_next_seg {
189 u32 nda_op; /* [31:6] next WQE [4:0] next opcode */
190 u32 ee_nds; /* [31:8] next EE [7] DBD [6] F [5:0] next WQE size */
191 u32 flags; /* [3] CQ [2] Event [1] Solicit */
192 u32 imm; /* immediate data */
195 struct mthca_tavor_ud_seg {
205 struct mthca_arbel_ud_seg {
212 struct mthca_bind_seg {
213 u32 flags; /* [31] Atomic [30] rem write [29] rem read */
221 struct mthca_raddr_seg {
227 struct mthca_atomic_seg {
232 struct mthca_data_seg {
238 struct mthca_mlx_seg {
241 u32 flags; /* [17] VL15 [16] SLR [14:12] static rate
242 [11:8] SL [3] C [2] E */
247 static const u8 mthca_opcode[] = {
248 [IB_WR_SEND] = MTHCA_OPCODE_SEND,
249 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM,
250 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE,
251 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM,
252 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ,
253 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS,
254 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA,
257 static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp)
259 return qp->qpn >= dev->qp_table.sqp_start &&
260 qp->qpn <= dev->qp_table.sqp_start + 3;
263 static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp)
265 return qp->qpn >= dev->qp_table.sqp_start &&
266 qp->qpn <= dev->qp_table.sqp_start + 1;
269 static void *get_recv_wqe(struct mthca_qp *qp, int n)
272 return qp->queue.direct.buf + (n << qp->rq.wqe_shift);
274 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf +
275 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1));
278 static void *get_send_wqe(struct mthca_qp *qp, int n)
281 return qp->queue.direct.buf + qp->send_wqe_offset +
282 (n << qp->sq.wqe_shift);
284 return qp->queue.page_list[(qp->send_wqe_offset +
285 (n << qp->sq.wqe_shift)) >>
287 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) &
291 void mthca_qp_event(struct mthca_dev *dev, u32 qpn,
292 enum ib_event_type event_type)
295 struct ib_event event;
297 spin_lock(&dev->qp_table.lock);
298 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1));
300 atomic_inc(&qp->refcount);
301 spin_unlock(&dev->qp_table.lock);
304 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn);
308 event.device = &dev->ib_dev;
309 event.event = event_type;
310 event.element.qp = &qp->ibqp;
311 if (qp->ibqp.event_handler)
312 qp->ibqp.event_handler(&event, qp->ibqp.qp_context);
314 if (atomic_dec_and_test(&qp->refcount))
318 static int to_mthca_state(enum ib_qp_state ib_state)
321 case IB_QPS_RESET: return MTHCA_QP_STATE_RST;
322 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT;
323 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR;
324 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS;
325 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD;
326 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE;
327 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR;
332 enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS };
334 static int to_mthca_st(int transport)
337 case RC: return MTHCA_QP_ST_RC;
338 case UC: return MTHCA_QP_ST_UC;
339 case UD: return MTHCA_QP_ST_UD;
340 case RD: return MTHCA_QP_ST_RD;
341 case MLX: return MTHCA_QP_ST_MLX;
346 static const struct {
348 u32 req_param[NUM_TRANS];
349 u32 opt_param[NUM_TRANS];
350 } state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
352 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
353 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
355 .trans = MTHCA_TRANS_RST2INIT,
357 [UD] = (IB_QP_PKEY_INDEX |
360 [RC] = (IB_QP_PKEY_INDEX |
363 [MLX] = (IB_QP_PKEY_INDEX |
366 /* bug-for-bug compatibility with VAPI: */
373 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
374 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
376 .trans = MTHCA_TRANS_INIT2INIT,
378 [UD] = (IB_QP_PKEY_INDEX |
381 [RC] = (IB_QP_PKEY_INDEX |
384 [MLX] = (IB_QP_PKEY_INDEX |
389 .trans = MTHCA_TRANS_INIT2RTR,
395 IB_QP_MAX_DEST_RD_ATOMIC |
396 IB_QP_MIN_RNR_TIMER),
399 [UD] = (IB_QP_PKEY_INDEX |
401 [RC] = (IB_QP_ALT_PATH |
404 [MLX] = (IB_QP_PKEY_INDEX |
410 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
411 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
413 .trans = MTHCA_TRANS_RTR2RTS,
416 [RC] = (IB_QP_TIMEOUT |
420 IB_QP_MAX_QP_RD_ATOMIC),
421 [MLX] = IB_QP_SQ_PSN,
424 [UD] = (IB_QP_CUR_STATE |
426 [RC] = (IB_QP_CUR_STATE |
430 IB_QP_MIN_RNR_TIMER |
431 IB_QP_PATH_MIG_STATE),
432 [MLX] = (IB_QP_CUR_STATE |
438 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
439 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
441 .trans = MTHCA_TRANS_RTS2RTS,
443 [UD] = (IB_QP_CUR_STATE |
445 [RC] = (IB_QP_ACCESS_FLAGS |
447 IB_QP_PATH_MIG_STATE |
448 IB_QP_MIN_RNR_TIMER),
449 [MLX] = (IB_QP_CUR_STATE |
454 .trans = MTHCA_TRANS_RTS2SQD,
458 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
459 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
461 .trans = MTHCA_TRANS_SQD2RTS,
463 [UD] = (IB_QP_CUR_STATE |
465 [RC] = (IB_QP_CUR_STATE |
468 IB_QP_MIN_RNR_TIMER |
469 IB_QP_PATH_MIG_STATE),
470 [MLX] = (IB_QP_CUR_STATE |
475 .trans = MTHCA_TRANS_SQD2SQD,
477 [UD] = (IB_QP_PKEY_INDEX |
483 IB_QP_MAX_QP_RD_ATOMIC |
484 IB_QP_MAX_DEST_RD_ATOMIC |
489 IB_QP_MIN_RNR_TIMER |
490 IB_QP_PATH_MIG_STATE),
491 [MLX] = (IB_QP_PKEY_INDEX |
497 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
498 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR },
500 .trans = MTHCA_TRANS_SQERR2RTS,
502 [UD] = (IB_QP_CUR_STATE |
504 [RC] = (IB_QP_CUR_STATE |
505 IB_QP_MIN_RNR_TIMER),
506 [MLX] = (IB_QP_CUR_STATE |
512 [IB_QPS_RESET] = { .trans = MTHCA_TRANS_ANY2RST },
513 [IB_QPS_ERR] = { .trans = MTHCA_TRANS_ANY2ERR }
517 static void store_attrs(struct mthca_sqp *sqp, struct ib_qp_attr *attr,
520 if (attr_mask & IB_QP_PKEY_INDEX)
521 sqp->pkey_index = attr->pkey_index;
522 if (attr_mask & IB_QP_QKEY)
523 sqp->qkey = attr->qkey;
524 if (attr_mask & IB_QP_SQ_PSN)
525 sqp->send_psn = attr->sq_psn;
528 static void init_port(struct mthca_dev *dev, int port)
532 struct mthca_init_ib_param param;
534 memset(¶m, 0, sizeof param);
538 param.vl_cap = dev->limits.vl_cap;
539 param.mtu_cap = dev->limits.mtu_cap;
540 param.gid_cap = dev->limits.gid_table_len;
541 param.pkey_cap = dev->limits.pkey_table_len;
543 err = mthca_INIT_IB(dev, ¶m, port, &status);
545 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err);
547 mthca_warn(dev, "INIT_IB returned status %02x.\n", status);
550 int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask)
552 struct mthca_dev *dev = to_mdev(ibqp->device);
553 struct mthca_qp *qp = to_mqp(ibqp);
554 enum ib_qp_state cur_state, new_state;
555 void *mailbox = NULL;
556 struct mthca_qp_param *qp_param;
557 struct mthca_qp_context *qp_context;
558 u32 req_param, opt_param;
562 if (attr_mask & IB_QP_CUR_STATE) {
563 if (attr->cur_qp_state != IB_QPS_RTR &&
564 attr->cur_qp_state != IB_QPS_RTS &&
565 attr->cur_qp_state != IB_QPS_SQD &&
566 attr->cur_qp_state != IB_QPS_SQE)
569 cur_state = attr->cur_qp_state;
571 spin_lock_irq(&qp->sq.lock);
572 spin_lock(&qp->rq.lock);
573 cur_state = qp->state;
574 spin_unlock(&qp->rq.lock);
575 spin_unlock_irq(&qp->sq.lock);
578 if (attr_mask & IB_QP_STATE) {
579 if (attr->qp_state < 0 || attr->qp_state > IB_QPS_ERR)
581 new_state = attr->qp_state;
583 new_state = cur_state;
585 if (state_table[cur_state][new_state].trans == MTHCA_TRANS_INVALID) {
586 mthca_dbg(dev, "Illegal QP transition "
587 "%d->%d\n", cur_state, new_state);
591 req_param = state_table[cur_state][new_state].req_param[qp->transport];
592 opt_param = state_table[cur_state][new_state].opt_param[qp->transport];
594 if ((req_param & attr_mask) != req_param) {
595 mthca_dbg(dev, "QP transition "
596 "%d->%d missing req attr 0x%08x\n",
597 cur_state, new_state,
598 req_param & ~attr_mask);
602 if (attr_mask & ~(req_param | opt_param | IB_QP_STATE)) {
603 mthca_dbg(dev, "QP transition (transport %d) "
604 "%d->%d has extra attr 0x%08x\n",
606 cur_state, new_state,
607 attr_mask & ~(req_param | opt_param |
612 mailbox = kmalloc(sizeof (*qp_param) + MTHCA_CMD_MAILBOX_EXTRA, GFP_KERNEL);
615 qp_param = MAILBOX_ALIGN(mailbox);
616 qp_context = &qp_param->context;
617 memset(qp_param, 0, sizeof *qp_param);
619 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) |
620 (to_mthca_st(qp->transport) << 16));
621 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE);
622 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
623 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
625 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE);
626 switch (attr->path_mig_state) {
627 case IB_MIG_MIGRATED:
628 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11);
631 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11);
634 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11);
639 /* leave tavor_sched_queue as 0 */
641 if (qp->transport == MLX || qp->transport == UD)
642 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11;
643 else if (attr_mask & IB_QP_PATH_MTU)
644 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31;
646 if (mthca_is_memfree(dev)) {
647 qp_context->rq_size_stride =
648 ((ffs(qp->rq.max) - 1) << 3) | (qp->rq.wqe_shift - 4);
649 qp_context->sq_size_stride =
650 ((ffs(qp->sq.max) - 1) << 3) | (qp->sq.wqe_shift - 4);
653 /* leave arbel_sched_queue as 0 */
655 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index);
656 qp_context->local_qpn = cpu_to_be32(qp->qpn);
657 if (attr_mask & IB_QP_DEST_QPN) {
658 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
661 if (qp->transport == MLX)
662 qp_context->pri_path.port_pkey |=
663 cpu_to_be32(to_msqp(qp)->port << 24);
665 if (attr_mask & IB_QP_PORT) {
666 qp_context->pri_path.port_pkey |=
667 cpu_to_be32(attr->port_num << 24);
668 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM);
672 if (attr_mask & IB_QP_PKEY_INDEX) {
673 qp_context->pri_path.port_pkey |=
674 cpu_to_be32(attr->pkey_index);
675 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX);
678 if (attr_mask & IB_QP_RNR_RETRY) {
679 qp_context->pri_path.rnr_retry = attr->rnr_retry << 5;
680 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY);
683 if (attr_mask & IB_QP_AV) {
684 qp_context->pri_path.g_mylmc = attr->ah_attr.src_path_bits & 0x7f;
685 qp_context->pri_path.rlid = cpu_to_be16(attr->ah_attr.dlid);
686 qp_context->pri_path.static_rate = (!!attr->ah_attr.static_rate) << 3;
687 if (attr->ah_attr.ah_flags & IB_AH_GRH) {
688 qp_context->pri_path.g_mylmc |= 1 << 7;
689 qp_context->pri_path.mgid_index = attr->ah_attr.grh.sgid_index;
690 qp_context->pri_path.hop_limit = attr->ah_attr.grh.hop_limit;
691 qp_context->pri_path.sl_tclass_flowlabel =
692 cpu_to_be32((attr->ah_attr.sl << 28) |
693 (attr->ah_attr.grh.traffic_class << 20) |
694 (attr->ah_attr.grh.flow_label));
695 memcpy(qp_context->pri_path.rgid,
696 attr->ah_attr.grh.dgid.raw, 16);
698 qp_context->pri_path.sl_tclass_flowlabel =
699 cpu_to_be32(attr->ah_attr.sl << 28);
701 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH);
704 if (attr_mask & IB_QP_TIMEOUT) {
705 qp_context->pri_path.ackto = attr->timeout;
706 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT);
712 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num);
713 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */
714 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey);
715 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) |
716 (MTHCA_FLIGHT_LIMIT << 24) |
720 if (qp->sq_policy == IB_SIGNAL_ALL_WR)
721 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC);
722 if (attr_mask & IB_QP_RETRY_CNT) {
723 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
724 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT);
727 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
728 qp_context->params1 |= cpu_to_be32(min(attr->max_dest_rd_atomic ?
729 ffs(attr->max_dest_rd_atomic) - 1 : 0,
731 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX);
734 if (attr_mask & IB_QP_SQ_PSN)
735 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn);
736 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn);
738 if (mthca_is_memfree(dev)) {
739 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset);
740 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index);
743 if (attr_mask & IB_QP_ACCESS_FLAGS) {
745 * Only enable RDMA/atomics if we have responder
746 * resources set to a non-zero value.
748 if (qp->resp_depth) {
749 qp_context->params2 |=
750 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE ?
751 MTHCA_QP_BIT_RWE : 0);
752 qp_context->params2 |=
753 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_READ ?
754 MTHCA_QP_BIT_RRE : 0);
755 qp_context->params2 |=
756 cpu_to_be32(attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC ?
757 MTHCA_QP_BIT_RAE : 0);
760 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
761 MTHCA_QP_OPTPAR_RRE |
762 MTHCA_QP_OPTPAR_RAE);
764 qp->atomic_rd_en = attr->qp_access_flags;
767 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
770 if (qp->resp_depth && !attr->max_rd_atomic) {
772 * Lowering our responder resources to zero.
773 * Turn off RDMA/atomics as responder.
774 * (RWE/RRE/RAE in params2 already zero)
776 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
777 MTHCA_QP_OPTPAR_RRE |
778 MTHCA_QP_OPTPAR_RAE);
781 if (!qp->resp_depth && attr->max_rd_atomic) {
783 * Increasing our responder resources from
784 * zero. Turn on RDMA/atomics as appropriate.
786 qp_context->params2 |=
787 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_WRITE ?
788 MTHCA_QP_BIT_RWE : 0);
789 qp_context->params2 |=
790 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_READ ?
791 MTHCA_QP_BIT_RRE : 0);
792 qp_context->params2 |=
793 cpu_to_be32(qp->atomic_rd_en & IB_ACCESS_REMOTE_ATOMIC ?
794 MTHCA_QP_BIT_RAE : 0);
796 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE |
797 MTHCA_QP_OPTPAR_RRE |
798 MTHCA_QP_OPTPAR_RAE);
802 1 << rra_max < attr->max_rd_atomic &&
803 rra_max < dev->qp_table.rdb_shift;
807 qp_context->params2 |= cpu_to_be32(rra_max << 21);
808 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX);
810 qp->resp_depth = attr->max_rd_atomic;
813 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC);
815 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
816 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
817 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT);
819 if (attr_mask & IB_QP_RQ_PSN)
820 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
822 qp_context->ra_buff_indx =
823 cpu_to_be32(dev->qp_table.rdb_base +
824 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE <<
825 dev->qp_table.rdb_shift));
827 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn);
829 if (mthca_is_memfree(dev))
830 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index);
832 if (attr_mask & IB_QP_QKEY) {
833 qp_context->qkey = cpu_to_be32(attr->qkey);
834 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY);
837 err = mthca_MODIFY_QP(dev, state_table[cur_state][new_state].trans,
838 qp->qpn, 0, qp_param, 0, &status);
840 mthca_warn(dev, "modify QP %d returned status %02x.\n",
841 state_table[cur_state][new_state].trans, status);
846 qp->state = new_state;
851 store_attrs(to_msqp(qp), attr, attr_mask);
854 * If we are moving QP0 to RTR, bring the IB link up; if we
855 * are moving QP0 to RESET or ERROR, bring the link back down.
857 if (is_qp0(dev, qp)) {
858 if (cur_state != IB_QPS_RTR &&
859 new_state == IB_QPS_RTR)
860 init_port(dev, to_msqp(qp)->port);
862 if (cur_state != IB_QPS_RESET &&
863 cur_state != IB_QPS_ERR &&
864 (new_state == IB_QPS_RESET ||
865 new_state == IB_QPS_ERR))
866 mthca_CLOSE_IB(dev, to_msqp(qp)->port, &status);
873 * Allocate and register buffer for WQEs. qp->rq.max, sq.max,
874 * rq.max_gs and sq.max_gs must all be assigned.
875 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and
876 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and
879 static int mthca_alloc_wqe_buf(struct mthca_dev *dev,
887 u64 *dma_list = NULL;
890 size = sizeof (struct mthca_next_seg) +
891 qp->rq.max_gs * sizeof (struct mthca_data_seg);
893 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size;
897 size = sizeof (struct mthca_next_seg) +
898 qp->sq.max_gs * sizeof (struct mthca_data_seg);
899 switch (qp->transport) {
901 size += 2 * sizeof (struct mthca_data_seg);
904 if (mthca_is_memfree(dev))
905 size += sizeof (struct mthca_arbel_ud_seg);
907 size += sizeof (struct mthca_tavor_ud_seg);
910 /* bind seg is as big as atomic + raddr segs */
911 size += sizeof (struct mthca_bind_seg);
914 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size;
918 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift,
919 1 << qp->sq.wqe_shift);
920 size = PAGE_ALIGN(qp->send_wqe_offset +
921 (qp->sq.max << qp->sq.wqe_shift));
923 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64),
928 if (size <= MTHCA_MAX_DIRECT_QP_SIZE) {
931 shift = get_order(size) + PAGE_SHIFT;
934 mthca_dbg(dev, "Creating direct QP of size %d (shift %d)\n",
937 qp->queue.direct.buf = pci_alloc_consistent(dev->pdev, size, &t);
938 if (!qp->queue.direct.buf)
941 pci_unmap_addr_set(&qp->queue.direct, mapping, t);
943 memset(qp->queue.direct.buf, 0, size);
945 while (t & ((1 << shift) - 1)) {
950 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
954 for (i = 0; i < npages; ++i)
955 dma_list[i] = t + i * (1 << shift);
958 npages = size / PAGE_SIZE;
962 mthca_dbg(dev, "Creating indirect QP with %d pages\n", npages);
964 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
968 qp->queue.page_list = kmalloc(npages *
969 sizeof *qp->queue.page_list,
971 if (!qp->queue.page_list)
974 for (i = 0; i < npages; ++i) {
975 qp->queue.page_list[i].buf =
976 pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
977 if (!qp->queue.page_list[i].buf)
980 memset(qp->queue.page_list[i].buf, 0, PAGE_SIZE);
982 pci_unmap_addr_set(&qp->queue.page_list[i], mapping, t);
987 err = mthca_mr_alloc_phys(dev, pd->pd_num, dma_list, shift,
989 MTHCA_MPT_FLAG_LOCAL_READ,
999 pci_free_consistent(dev->pdev, size,
1000 qp->queue.direct.buf,
1001 pci_unmap_addr(&qp->queue.direct, mapping));
1003 for (i = 0; i < npages; ++i) {
1004 if (qp->queue.page_list[i].buf)
1005 pci_free_consistent(dev->pdev, PAGE_SIZE,
1006 qp->queue.page_list[i].buf,
1007 pci_unmap_addr(&qp->queue.page_list[i],
1018 static int mthca_alloc_memfree(struct mthca_dev *dev,
1019 struct mthca_qp *qp)
1023 if (mthca_is_memfree(dev)) {
1024 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn);
1028 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn);
1032 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ,
1033 qp->qpn, &qp->rq.db);
1034 if (qp->rq.db_index < 0) {
1039 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ,
1040 qp->qpn, &qp->sq.db);
1041 if (qp->sq.db_index < 0) {
1050 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1053 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1056 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1061 static void mthca_free_memfree(struct mthca_dev *dev,
1062 struct mthca_qp *qp)
1064 if (mthca_is_memfree(dev)) {
1065 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index);
1066 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index);
1067 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn);
1068 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn);
1072 static void mthca_wq_init(struct mthca_wq* wq)
1074 spin_lock_init(&wq->lock);
1076 wq->last_comp = wq->max - 1;
1082 static int mthca_alloc_qp_common(struct mthca_dev *dev,
1083 struct mthca_pd *pd,
1084 struct mthca_cq *send_cq,
1085 struct mthca_cq *recv_cq,
1086 enum ib_sig_type send_policy,
1087 struct mthca_qp *qp)
1092 atomic_set(&qp->refcount, 1);
1093 qp->state = IB_QPS_RESET;
1094 qp->atomic_rd_en = 0;
1096 qp->sq_policy = send_policy;
1097 mthca_wq_init(&qp->sq);
1098 mthca_wq_init(&qp->rq);
1100 ret = mthca_alloc_memfree(dev, qp);
1104 ret = mthca_alloc_wqe_buf(dev, pd, qp);
1106 mthca_free_memfree(dev, qp);
1110 if (mthca_is_memfree(dev)) {
1111 struct mthca_next_seg *next;
1112 struct mthca_data_seg *scatter;
1113 int size = (sizeof (struct mthca_next_seg) +
1114 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16;
1116 for (i = 0; i < qp->rq.max; ++i) {
1117 next = get_recv_wqe(qp, i);
1118 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) <<
1120 next->ee_nds = cpu_to_be32(size);
1122 for (scatter = (void *) (next + 1);
1123 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift);
1125 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1128 for (i = 0; i < qp->sq.max; ++i) {
1129 next = get_send_wqe(qp, i);
1130 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) <<
1132 qp->send_wqe_offset);
1139 static void mthca_align_qp_size(struct mthca_dev *dev, struct mthca_qp *qp)
1143 if (!mthca_is_memfree(dev))
1146 for (i = 0; 1 << i < qp->rq.max; ++i)
1149 qp->rq.max = 1 << i;
1151 for (i = 0; 1 << i < qp->sq.max; ++i)
1154 qp->sq.max = 1 << i;
1157 int mthca_alloc_qp(struct mthca_dev *dev,
1158 struct mthca_pd *pd,
1159 struct mthca_cq *send_cq,
1160 struct mthca_cq *recv_cq,
1161 enum ib_qp_type type,
1162 enum ib_sig_type send_policy,
1163 struct mthca_qp *qp)
1167 mthca_align_qp_size(dev, qp);
1170 case IB_QPT_RC: qp->transport = RC; break;
1171 case IB_QPT_UC: qp->transport = UC; break;
1172 case IB_QPT_UD: qp->transport = UD; break;
1173 default: return -EINVAL;
1176 qp->qpn = mthca_alloc(&dev->qp_table.alloc);
1180 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1183 mthca_free(&dev->qp_table.alloc, qp->qpn);
1187 spin_lock_irq(&dev->qp_table.lock);
1188 mthca_array_set(&dev->qp_table.qp,
1189 qp->qpn & (dev->limits.num_qps - 1), qp);
1190 spin_unlock_irq(&dev->qp_table.lock);
1195 int mthca_alloc_sqp(struct mthca_dev *dev,
1196 struct mthca_pd *pd,
1197 struct mthca_cq *send_cq,
1198 struct mthca_cq *recv_cq,
1199 enum ib_sig_type send_policy,
1202 struct mthca_sqp *sqp)
1205 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1;
1207 mthca_align_qp_size(dev, &sqp->qp);
1209 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE;
1210 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size,
1211 &sqp->header_dma, GFP_KERNEL);
1212 if (!sqp->header_buf)
1215 spin_lock_irq(&dev->qp_table.lock);
1216 if (mthca_array_get(&dev->qp_table.qp, mqpn))
1219 mthca_array_set(&dev->qp_table.qp, mqpn, sqp);
1220 spin_unlock_irq(&dev->qp_table.lock);
1227 sqp->qp.transport = MLX;
1229 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq,
1230 send_policy, &sqp->qp);
1234 atomic_inc(&pd->sqp_count);
1240 * Lock CQs here, so that CQ polling code can do QP lookup
1241 * without taking a lock.
1243 spin_lock_irq(&send_cq->lock);
1244 if (send_cq != recv_cq)
1245 spin_lock(&recv_cq->lock);
1247 spin_lock(&dev->qp_table.lock);
1248 mthca_array_clear(&dev->qp_table.qp, mqpn);
1249 spin_unlock(&dev->qp_table.lock);
1251 if (send_cq != recv_cq)
1252 spin_unlock(&recv_cq->lock);
1253 spin_unlock_irq(&send_cq->lock);
1256 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size,
1257 sqp->header_buf, sqp->header_dma);
1262 void mthca_free_qp(struct mthca_dev *dev,
1263 struct mthca_qp *qp)
1268 struct mthca_cq *send_cq;
1269 struct mthca_cq *recv_cq;
1271 send_cq = to_mcq(qp->ibqp.send_cq);
1272 recv_cq = to_mcq(qp->ibqp.recv_cq);
1275 * Lock CQs here, so that CQ polling code can do QP lookup
1276 * without taking a lock.
1278 spin_lock_irq(&send_cq->lock);
1279 if (send_cq != recv_cq)
1280 spin_lock(&recv_cq->lock);
1282 spin_lock(&dev->qp_table.lock);
1283 mthca_array_clear(&dev->qp_table.qp,
1284 qp->qpn & (dev->limits.num_qps - 1));
1285 spin_unlock(&dev->qp_table.lock);
1287 if (send_cq != recv_cq)
1288 spin_unlock(&recv_cq->lock);
1289 spin_unlock_irq(&send_cq->lock);
1291 atomic_dec(&qp->refcount);
1292 wait_event(qp->wait, !atomic_read(&qp->refcount));
1294 if (qp->state != IB_QPS_RESET)
1295 mthca_MODIFY_QP(dev, MTHCA_TRANS_ANY2RST, qp->qpn, 0, NULL, 0, &status);
1297 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq)->cqn, qp->qpn);
1298 if (qp->ibqp.send_cq != qp->ibqp.recv_cq)
1299 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq)->cqn, qp->qpn);
1301 mthca_free_mr(dev, &qp->mr);
1303 size = PAGE_ALIGN(qp->send_wqe_offset +
1304 (qp->sq.max << qp->sq.wqe_shift));
1306 if (qp->is_direct) {
1307 pci_free_consistent(dev->pdev, size,
1308 qp->queue.direct.buf,
1309 pci_unmap_addr(&qp->queue.direct, mapping));
1311 for (i = 0; i < size / PAGE_SIZE; ++i) {
1312 pci_free_consistent(dev->pdev, PAGE_SIZE,
1313 qp->queue.page_list[i].buf,
1314 pci_unmap_addr(&qp->queue.page_list[i],
1321 mthca_free_memfree(dev, qp);
1323 if (is_sqp(dev, qp)) {
1324 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count));
1325 dma_free_coherent(&dev->pdev->dev,
1326 to_msqp(qp)->header_buf_size,
1327 to_msqp(qp)->header_buf,
1328 to_msqp(qp)->header_dma);
1330 mthca_free(&dev->qp_table.alloc, qp->qpn);
1333 /* Create UD header for an MLX send and build a data segment for it */
1334 static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp,
1335 int ind, struct ib_send_wr *wr,
1336 struct mthca_mlx_seg *mlx,
1337 struct mthca_data_seg *data)
1342 ib_ud_header_init(256, /* assume a MAD */
1343 sqp->ud_header.grh_present,
1346 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header);
1349 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1);
1350 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) |
1351 (sqp->ud_header.lrh.destination_lid == 0xffff ?
1352 MTHCA_MLX_SLR : 0) |
1353 (sqp->ud_header.lrh.service_level << 8));
1354 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1357 switch (wr->opcode) {
1359 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1360 sqp->ud_header.immediate_present = 0;
1362 case IB_WR_SEND_WITH_IMM:
1363 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1364 sqp->ud_header.immediate_present = 1;
1365 sqp->ud_header.immediate_data = wr->imm_data;
1371 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1372 if (sqp->ud_header.lrh.destination_lid == 0xffff)
1373 sqp->ud_header.lrh.source_lid = 0xffff;
1374 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1375 if (!sqp->qp.ibqp.qp_num)
1376 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1378 &sqp->ud_header.bth.pkey);
1380 ib_get_cached_pkey(&dev->ib_dev, sqp->port,
1381 wr->wr.ud.pkey_index,
1382 &sqp->ud_header.bth.pkey);
1383 cpu_to_be16s(&sqp->ud_header.bth.pkey);
1384 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1385 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1386 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1387 sqp->qkey : wr->wr.ud.remote_qkey);
1388 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1390 header_size = ib_ud_header_pack(&sqp->ud_header,
1392 ind * MTHCA_UD_HEADER_SIZE);
1394 data->byte_count = cpu_to_be32(header_size);
1395 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey);
1396 data->addr = cpu_to_be64(sqp->header_dma +
1397 ind * MTHCA_UD_HEADER_SIZE);
1402 static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq,
1403 struct ib_cq *ib_cq)
1406 struct mthca_cq *cq;
1408 cur = wq->head - wq->tail;
1409 if (likely(cur + nreq < wq->max))
1413 spin_lock(&cq->lock);
1414 cur = wq->head - wq->tail;
1415 spin_unlock(&cq->lock);
1417 return cur + nreq >= wq->max;
1420 int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1421 struct ib_send_wr **bad_wr)
1423 struct mthca_dev *dev = to_mdev(ibqp->device);
1424 struct mthca_qp *qp = to_mqp(ibqp);
1427 unsigned long flags;
1437 spin_lock_irqsave(&qp->sq.lock, flags);
1439 /* XXX check that state is OK to post send */
1441 ind = qp->sq.next_ind;
1443 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1444 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1445 mthca_err(dev, "SQ %06x full (%u head, %u tail,"
1446 " %d max, %d nreq)\n", qp->qpn,
1447 qp->sq.head, qp->sq.tail,
1454 wqe = get_send_wqe(qp, ind);
1455 prev_wqe = qp->sq.last;
1458 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1459 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
1460 ((struct mthca_next_seg *) wqe)->flags =
1461 ((wr->send_flags & IB_SEND_SIGNALED) ?
1462 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) |
1463 ((wr->send_flags & IB_SEND_SOLICITED) ?
1464 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) |
1466 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1467 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1468 ((struct mthca_next_seg *) wqe)->imm = wr->imm_data;
1470 wqe += sizeof (struct mthca_next_seg);
1471 size = sizeof (struct mthca_next_seg) / 16;
1473 switch (qp->transport) {
1475 switch (wr->opcode) {
1476 case IB_WR_ATOMIC_CMP_AND_SWP:
1477 case IB_WR_ATOMIC_FETCH_AND_ADD:
1478 ((struct mthca_raddr_seg *) wqe)->raddr =
1479 cpu_to_be64(wr->wr.atomic.remote_addr);
1480 ((struct mthca_raddr_seg *) wqe)->rkey =
1481 cpu_to_be32(wr->wr.atomic.rkey);
1482 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1484 wqe += sizeof (struct mthca_raddr_seg);
1486 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1487 ((struct mthca_atomic_seg *) wqe)->swap_add =
1488 cpu_to_be64(wr->wr.atomic.swap);
1489 ((struct mthca_atomic_seg *) wqe)->compare =
1490 cpu_to_be64(wr->wr.atomic.compare_add);
1492 ((struct mthca_atomic_seg *) wqe)->swap_add =
1493 cpu_to_be64(wr->wr.atomic.compare_add);
1494 ((struct mthca_atomic_seg *) wqe)->compare = 0;
1497 wqe += sizeof (struct mthca_atomic_seg);
1498 size += sizeof (struct mthca_raddr_seg) / 16 +
1499 sizeof (struct mthca_atomic_seg);
1502 case IB_WR_RDMA_WRITE:
1503 case IB_WR_RDMA_WRITE_WITH_IMM:
1504 case IB_WR_RDMA_READ:
1505 ((struct mthca_raddr_seg *) wqe)->raddr =
1506 cpu_to_be64(wr->wr.rdma.remote_addr);
1507 ((struct mthca_raddr_seg *) wqe)->rkey =
1508 cpu_to_be32(wr->wr.rdma.rkey);
1509 ((struct mthca_raddr_seg *) wqe)->reserved = 0;
1510 wqe += sizeof (struct mthca_raddr_seg);
1511 size += sizeof (struct mthca_raddr_seg) / 16;
1515 /* No extra segments required for sends */
1522 ((struct mthca_tavor_ud_seg *) wqe)->lkey =
1523 cpu_to_be32(to_mah(wr->wr.ud.ah)->key);
1524 ((struct mthca_tavor_ud_seg *) wqe)->av_addr =
1525 cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma);
1526 ((struct mthca_tavor_ud_seg *) wqe)->dqpn =
1527 cpu_to_be32(wr->wr.ud.remote_qpn);
1528 ((struct mthca_tavor_ud_seg *) wqe)->qkey =
1529 cpu_to_be32(wr->wr.ud.remote_qkey);
1531 wqe += sizeof (struct mthca_tavor_ud_seg);
1532 size += sizeof (struct mthca_tavor_ud_seg) / 16;
1536 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1537 wqe - sizeof (struct mthca_next_seg),
1543 wqe += sizeof (struct mthca_data_seg);
1544 size += sizeof (struct mthca_data_seg) / 16;
1548 if (wr->num_sge > qp->sq.max_gs) {
1549 mthca_err(dev, "too many gathers\n");
1555 for (i = 0; i < wr->num_sge; ++i) {
1556 ((struct mthca_data_seg *) wqe)->byte_count =
1557 cpu_to_be32(wr->sg_list[i].length);
1558 ((struct mthca_data_seg *) wqe)->lkey =
1559 cpu_to_be32(wr->sg_list[i].lkey);
1560 ((struct mthca_data_seg *) wqe)->addr =
1561 cpu_to_be64(wr->sg_list[i].addr);
1562 wqe += sizeof (struct mthca_data_seg);
1563 size += sizeof (struct mthca_data_seg) / 16;
1566 /* Add one more inline data segment for ICRC */
1567 if (qp->transport == MLX) {
1568 ((struct mthca_data_seg *) wqe)->byte_count =
1569 cpu_to_be32((1 << 31) | 4);
1570 ((u32 *) wqe)[1] = 0;
1571 wqe += sizeof (struct mthca_data_seg);
1572 size += sizeof (struct mthca_data_seg) / 16;
1575 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1577 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1578 mthca_err(dev, "opcode invalid\n");
1585 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1586 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1587 qp->send_wqe_offset) |
1588 mthca_opcode[wr->opcode]);
1590 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1591 cpu_to_be32((size0 ? 0 : MTHCA_NEXT_DBD) | size);
1596 op0 = mthca_opcode[wr->opcode];
1600 if (unlikely(ind >= qp->sq.max))
1608 doorbell[0] = cpu_to_be32(((qp->sq.next_ind << qp->sq.wqe_shift) +
1609 qp->send_wqe_offset) | f0 | op0);
1610 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1614 mthca_write64(doorbell,
1615 dev->kar + MTHCA_SEND_DOORBELL,
1616 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1619 qp->sq.next_ind = ind;
1620 qp->sq.head += nreq;
1622 spin_unlock_irqrestore(&qp->sq.lock, flags);
1626 int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1627 struct ib_recv_wr **bad_wr)
1629 struct mthca_dev *dev = to_mdev(ibqp->device);
1630 struct mthca_qp *qp = to_mqp(ibqp);
1631 unsigned long flags;
1641 spin_lock_irqsave(&qp->rq.lock, flags);
1643 /* XXX check that state is OK to post receive */
1645 ind = qp->rq.next_ind;
1647 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1648 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1649 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1650 " %d max, %d nreq)\n", qp->qpn,
1651 qp->rq.head, qp->rq.tail,
1658 wqe = get_recv_wqe(qp, ind);
1659 prev_wqe = qp->rq.last;
1662 ((struct mthca_next_seg *) wqe)->nda_op = 0;
1663 ((struct mthca_next_seg *) wqe)->ee_nds =
1664 cpu_to_be32(MTHCA_NEXT_DBD);
1665 ((struct mthca_next_seg *) wqe)->flags = 0;
1667 wqe += sizeof (struct mthca_next_seg);
1668 size = sizeof (struct mthca_next_seg) / 16;
1670 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1676 for (i = 0; i < wr->num_sge; ++i) {
1677 ((struct mthca_data_seg *) wqe)->byte_count =
1678 cpu_to_be32(wr->sg_list[i].length);
1679 ((struct mthca_data_seg *) wqe)->lkey =
1680 cpu_to_be32(wr->sg_list[i].lkey);
1681 ((struct mthca_data_seg *) wqe)->addr =
1682 cpu_to_be64(wr->sg_list[i].addr);
1683 wqe += sizeof (struct mthca_data_seg);
1684 size += sizeof (struct mthca_data_seg) / 16;
1687 qp->wrid[ind] = wr->wr_id;
1689 if (likely(prev_wqe)) {
1690 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1691 cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
1693 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1694 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_WRITE:
1807 case IB_WR_RDMA_WRITE_WITH_IMM:
1808 case IB_WR_RDMA_READ:
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 memcpy(((struct mthca_arbel_ud_seg *) wqe)->av,
1827 to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE);
1828 ((struct mthca_arbel_ud_seg *) wqe)->dqpn =
1829 cpu_to_be32(wr->wr.ud.remote_qpn);
1830 ((struct mthca_arbel_ud_seg *) wqe)->qkey =
1831 cpu_to_be32(wr->wr.ud.remote_qkey);
1833 wqe += sizeof (struct mthca_arbel_ud_seg);
1834 size += sizeof (struct mthca_arbel_ud_seg) / 16;
1838 err = build_mlx_header(dev, to_msqp(qp), ind, wr,
1839 wqe - sizeof (struct mthca_next_seg),
1845 wqe += sizeof (struct mthca_data_seg);
1846 size += sizeof (struct mthca_data_seg) / 16;
1850 if (wr->num_sge > qp->sq.max_gs) {
1851 mthca_err(dev, "too many gathers\n");
1857 for (i = 0; i < wr->num_sge; ++i) {
1858 ((struct mthca_data_seg *) wqe)->byte_count =
1859 cpu_to_be32(wr->sg_list[i].length);
1860 ((struct mthca_data_seg *) wqe)->lkey =
1861 cpu_to_be32(wr->sg_list[i].lkey);
1862 ((struct mthca_data_seg *) wqe)->addr =
1863 cpu_to_be64(wr->sg_list[i].addr);
1864 wqe += sizeof (struct mthca_data_seg);
1865 size += sizeof (struct mthca_data_seg) / 16;
1868 /* Add one more inline data segment for ICRC */
1869 if (qp->transport == MLX) {
1870 ((struct mthca_data_seg *) wqe)->byte_count =
1871 cpu_to_be32((1 << 31) | 4);
1872 ((u32 *) wqe)[1] = 0;
1873 wqe += sizeof (struct mthca_data_seg);
1874 size += sizeof (struct mthca_data_seg) / 16;
1877 qp->wrid[ind + qp->rq.max] = wr->wr_id;
1879 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) {
1880 mthca_err(dev, "opcode invalid\n");
1886 if (likely(prev_wqe)) {
1887 ((struct mthca_next_seg *) prev_wqe)->nda_op =
1888 cpu_to_be32(((ind << qp->sq.wqe_shift) +
1889 qp->send_wqe_offset) |
1890 mthca_opcode[wr->opcode]);
1892 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
1893 cpu_to_be32(MTHCA_NEXT_DBD | size);
1898 op0 = mthca_opcode[wr->opcode];
1902 if (unlikely(ind >= qp->sq.max))
1910 doorbell[0] = cpu_to_be32((nreq << 24) |
1911 ((qp->sq.head & 0xffff) << 8) |
1913 doorbell[1] = cpu_to_be32((qp->qpn << 8) | size0);
1915 qp->sq.head += nreq;
1918 * Make sure that descriptors are written before
1922 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff);
1925 * Make sure doorbell record is written before we
1926 * write MMIO send doorbell.
1929 mthca_write64(doorbell,
1930 dev->kar + MTHCA_SEND_DOORBELL,
1931 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
1934 spin_unlock_irqrestore(&qp->sq.lock, flags);
1938 int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1939 struct ib_recv_wr **bad_wr)
1941 struct mthca_dev *dev = to_mdev(ibqp->device);
1942 struct mthca_qp *qp = to_mqp(ibqp);
1943 unsigned long flags;
1950 spin_lock_irqsave(&qp->rq.lock, flags);
1952 /* XXX check that state is OK to post receive */
1954 ind = qp->rq.head & (qp->rq.max - 1);
1956 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1957 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
1958 mthca_err(dev, "RQ %06x full (%u head, %u tail,"
1959 " %d max, %d nreq)\n", qp->qpn,
1960 qp->rq.head, qp->rq.tail,
1967 wqe = get_recv_wqe(qp, ind);
1969 ((struct mthca_next_seg *) wqe)->flags = 0;
1971 wqe += sizeof (struct mthca_next_seg);
1973 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1979 for (i = 0; i < wr->num_sge; ++i) {
1980 ((struct mthca_data_seg *) wqe)->byte_count =
1981 cpu_to_be32(wr->sg_list[i].length);
1982 ((struct mthca_data_seg *) wqe)->lkey =
1983 cpu_to_be32(wr->sg_list[i].lkey);
1984 ((struct mthca_data_seg *) wqe)->addr =
1985 cpu_to_be64(wr->sg_list[i].addr);
1986 wqe += sizeof (struct mthca_data_seg);
1989 if (i < qp->rq.max_gs) {
1990 ((struct mthca_data_seg *) wqe)->byte_count = 0;
1991 ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
1992 ((struct mthca_data_seg *) wqe)->addr = 0;
1995 qp->wrid[ind] = wr->wr_id;
1998 if (unlikely(ind >= qp->rq.max))
2003 qp->rq.head += nreq;
2006 * Make sure that descriptors are written before
2010 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff);
2013 spin_unlock_irqrestore(&qp->rq.lock, flags);
2017 int mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send,
2018 int index, int *dbd, u32 *new_wqe)
2020 struct mthca_next_seg *next;
2023 next = get_send_wqe(qp, index);
2025 next = get_recv_wqe(qp, index);
2027 if (mthca_is_memfree(dev))
2030 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD));
2031 if (next->ee_nds & cpu_to_be32(0x3f))
2032 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) |
2033 (next->ee_nds & cpu_to_be32(0x3f));
2040 int __devinit mthca_init_qp_table(struct mthca_dev *dev)
2046 spin_lock_init(&dev->qp_table.lock);
2049 * We reserve 2 extra QPs per port for the special QPs. The
2050 * special QP for port 1 has to be even, so round up.
2052 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL;
2053 err = mthca_alloc_init(&dev->qp_table.alloc,
2054 dev->limits.num_qps,
2056 dev->qp_table.sqp_start +
2057 MTHCA_MAX_PORTS * 2);
2061 err = mthca_array_init(&dev->qp_table.qp,
2062 dev->limits.num_qps);
2064 mthca_alloc_cleanup(&dev->qp_table.alloc);
2068 for (i = 0; i < 2; ++i) {
2069 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI,
2070 dev->qp_table.sqp_start + i * 2,
2075 mthca_warn(dev, "CONF_SPECIAL_QP returned "
2076 "status %02x, aborting.\n",
2085 for (i = 0; i < 2; ++i)
2086 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2088 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps);
2089 mthca_alloc_cleanup(&dev->qp_table.alloc);
2094 void __devexit mthca_cleanup_qp_table(struct mthca_dev *dev)
2099 for (i = 0; i < 2; ++i)
2100 mthca_CONF_SPECIAL_QP(dev, i, 0, &status);
2102 mthca_alloc_cleanup(&dev->qp_table.alloc);