2 * Copyright (c) 2007 Cisco Systems, Inc. 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
33 #include <rdma/ib_cache.h>
34 #include <rdma/ib_pack.h>
36 #include <linux/mlx4/qp.h>
42 MLX4_IB_ACK_REQ_FREQ = 8,
46 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83,
47 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f
52 * Largest possible UD header: send with GRH and immediate data.
54 MLX4_IB_UD_HEADER_SIZE = 72
62 struct ib_ud_header ud_header;
63 u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
66 static const __be32 mlx4_ib_opcode[] = {
67 [IB_WR_SEND] = __constant_cpu_to_be32(MLX4_OPCODE_SEND),
68 [IB_WR_SEND_WITH_IMM] = __constant_cpu_to_be32(MLX4_OPCODE_SEND_IMM),
69 [IB_WR_RDMA_WRITE] = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
70 [IB_WR_RDMA_WRITE_WITH_IMM] = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
71 [IB_WR_RDMA_READ] = __constant_cpu_to_be32(MLX4_OPCODE_RDMA_READ),
72 [IB_WR_ATOMIC_CMP_AND_SWP] = __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
73 [IB_WR_ATOMIC_FETCH_AND_ADD] = __constant_cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
76 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
78 return container_of(mqp, struct mlx4_ib_sqp, qp);
81 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
83 return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
84 qp->mqp.qpn <= dev->dev->caps.sqp_start + 3;
87 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
89 return qp->mqp.qpn >= dev->dev->caps.sqp_start &&
90 qp->mqp.qpn <= dev->dev->caps.sqp_start + 1;
93 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
95 if (qp->buf.nbufs == 1)
96 return qp->buf.u.direct.buf + offset;
98 return qp->buf.u.page_list[offset >> PAGE_SHIFT].buf +
99 (offset & (PAGE_SIZE - 1));
102 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
104 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
107 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
109 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
112 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
114 struct ib_event event;
115 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
117 if (type == MLX4_EVENT_TYPE_PATH_MIG)
118 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
120 if (ibqp->event_handler) {
121 event.device = ibqp->device;
122 event.element.qp = ibqp;
124 case MLX4_EVENT_TYPE_PATH_MIG:
125 event.event = IB_EVENT_PATH_MIG;
127 case MLX4_EVENT_TYPE_COMM_EST:
128 event.event = IB_EVENT_COMM_EST;
130 case MLX4_EVENT_TYPE_SQ_DRAINED:
131 event.event = IB_EVENT_SQ_DRAINED;
133 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
134 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
136 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
137 event.event = IB_EVENT_QP_FATAL;
139 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
140 event.event = IB_EVENT_PATH_MIG_ERR;
142 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
143 event.event = IB_EVENT_QP_REQ_ERR;
145 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
146 event.event = IB_EVENT_QP_ACCESS_ERR;
149 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
150 "on QP %06x\n", type, qp->qpn);
154 ibqp->event_handler(&event, ibqp->qp_context);
158 static int send_wqe_overhead(enum ib_qp_type type)
161 * UD WQEs must have a datagram segment.
162 * RC and UC WQEs might have a remote address segment.
163 * MLX WQEs need two extra inline data segments (for the UD
164 * header and space for the ICRC).
168 return sizeof (struct mlx4_wqe_ctrl_seg) +
169 sizeof (struct mlx4_wqe_datagram_seg);
171 return sizeof (struct mlx4_wqe_ctrl_seg) +
172 sizeof (struct mlx4_wqe_raddr_seg);
174 return sizeof (struct mlx4_wqe_ctrl_seg) +
175 sizeof (struct mlx4_wqe_atomic_seg) +
176 sizeof (struct mlx4_wqe_raddr_seg);
179 return sizeof (struct mlx4_wqe_ctrl_seg) +
180 ALIGN(MLX4_IB_UD_HEADER_SIZE +
181 sizeof (struct mlx4_wqe_inline_seg),
182 sizeof (struct mlx4_wqe_data_seg)) +
184 sizeof (struct mlx4_wqe_inline_seg),
185 sizeof (struct mlx4_wqe_data_seg));
187 return sizeof (struct mlx4_wqe_ctrl_seg);
191 static int set_qp_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
192 enum ib_qp_type type, struct mlx4_ib_qp *qp)
194 /* Sanity check QP size before proceeding */
195 if (cap->max_send_wr > dev->dev->caps.max_wqes ||
196 cap->max_recv_wr > dev->dev->caps.max_wqes ||
197 cap->max_send_sge > dev->dev->caps.max_sq_sg ||
198 cap->max_recv_sge > dev->dev->caps.max_rq_sg ||
199 cap->max_inline_data + send_wqe_overhead(type) +
200 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
204 * For MLX transport we need 2 extra S/G entries:
205 * one for the header and one for the checksum at the end
207 if ((type == IB_QPT_SMI || type == IB_QPT_GSI) &&
208 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
211 qp->rq.max = cap->max_recv_wr ? roundup_pow_of_two(cap->max_recv_wr) : 0;
212 qp->sq.max = cap->max_send_wr ? roundup_pow_of_two(cap->max_send_wr) : 0;
214 qp->rq.wqe_shift = ilog2(roundup_pow_of_two(cap->max_recv_sge *
215 sizeof (struct mlx4_wqe_data_seg)));
216 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof (struct mlx4_wqe_data_seg);
218 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(max(cap->max_send_sge *
219 sizeof (struct mlx4_wqe_data_seg),
220 cap->max_inline_data +
221 sizeof (struct mlx4_wqe_inline_seg)) +
222 send_wqe_overhead(type)));
223 qp->sq.max_gs = ((1 << qp->sq.wqe_shift) - send_wqe_overhead(type)) /
224 sizeof (struct mlx4_wqe_data_seg);
226 qp->buf_size = (qp->rq.max << qp->rq.wqe_shift) +
227 (qp->sq.max << qp->sq.wqe_shift);
228 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
230 qp->sq.offset = qp->rq.max << qp->rq.wqe_shift;
232 qp->rq.offset = qp->sq.max << qp->sq.wqe_shift;
236 cap->max_send_wr = qp->sq.max;
237 cap->max_recv_wr = qp->rq.max;
238 cap->max_send_sge = qp->sq.max_gs;
239 cap->max_recv_sge = qp->rq.max_gs;
240 cap->max_inline_data = (1 << qp->sq.wqe_shift) - send_wqe_overhead(type) -
241 sizeof (struct mlx4_wqe_inline_seg);
246 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
247 struct ib_qp_init_attr *init_attr,
248 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp)
250 struct mlx4_wqe_ctrl_seg *ctrl;
254 mutex_init(&qp->mutex);
255 spin_lock_init(&qp->sq.lock);
256 spin_lock_init(&qp->rq.lock);
258 qp->state = IB_QPS_RESET;
259 qp->atomic_rd_en = 0;
267 err = set_qp_size(dev, &init_attr->cap, init_attr->qp_type, qp);
272 struct mlx4_ib_create_qp ucmd;
274 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
279 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
281 if (IS_ERR(qp->umem)) {
282 err = PTR_ERR(qp->umem);
286 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
287 ilog2(qp->umem->page_size), &qp->mtt);
291 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
295 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
296 ucmd.db_addr, &qp->db);
300 err = mlx4_ib_db_alloc(dev, &qp->db, 0);
306 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) {
311 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
316 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf);
320 for (i = 0; i < qp->sq.max; ++i) {
321 ctrl = get_send_wqe(qp, i);
322 ctrl->owner_opcode = cpu_to_be32(1 << 31);
325 qp->sq.wrid = kmalloc(qp->sq.max * sizeof (u64), GFP_KERNEL);
326 qp->rq.wrid = kmalloc(qp->rq.max * sizeof (u64), GFP_KERNEL);
328 if (!qp->sq.wrid || !qp->rq.wrid) {
333 /* We don't support inline sends for kernel QPs (yet) */
334 init_attr->cap.max_inline_data = 0;
337 err = mlx4_qp_alloc(dev->dev, sqpn, &qp->mqp);
342 * Hardware wants QPN written in big-endian order (after
343 * shifting) for send doorbell. Precompute this value to save
344 * a little bit when posting sends.
346 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
348 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
349 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
351 qp->sq_signal_bits = 0;
353 qp->mqp.event = mlx4_ib_qp_event;
359 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
366 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
370 ib_umem_release(qp->umem);
372 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
376 mlx4_ib_db_free(dev, &qp->db);
382 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
385 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
386 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
387 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
388 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
389 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
390 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
391 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
396 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
398 if (send_cq == recv_cq)
399 spin_lock_irq(&send_cq->lock);
400 else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
401 spin_lock_irq(&send_cq->lock);
402 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
404 spin_lock_irq(&recv_cq->lock);
405 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
409 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
411 if (send_cq == recv_cq)
412 spin_unlock_irq(&send_cq->lock);
413 else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
414 spin_unlock(&recv_cq->lock);
415 spin_unlock_irq(&send_cq->lock);
417 spin_unlock(&send_cq->lock);
418 spin_unlock_irq(&recv_cq->lock);
422 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
425 struct mlx4_ib_cq *send_cq, *recv_cq;
427 if (qp->state != IB_QPS_RESET)
428 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
429 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
430 printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n",
433 send_cq = to_mcq(qp->ibqp.send_cq);
434 recv_cq = to_mcq(qp->ibqp.recv_cq);
436 mlx4_ib_lock_cqs(send_cq, recv_cq);
439 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
440 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
441 if (send_cq != recv_cq)
442 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
445 mlx4_qp_remove(dev->dev, &qp->mqp);
447 mlx4_ib_unlock_cqs(send_cq, recv_cq);
449 mlx4_qp_free(dev->dev, &qp->mqp);
450 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
453 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
455 ib_umem_release(qp->umem);
459 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
460 mlx4_ib_db_free(dev, &qp->db);
464 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
465 struct ib_qp_init_attr *init_attr,
466 struct ib_udata *udata)
468 struct mlx4_ib_dev *dev = to_mdev(pd->device);
469 struct mlx4_ib_sqp *sqp;
470 struct mlx4_ib_qp *qp;
473 switch (init_attr->qp_type) {
478 qp = kmalloc(sizeof *qp, GFP_KERNEL);
480 return ERR_PTR(-ENOMEM);
482 err = create_qp_common(dev, pd, init_attr, udata, 0, qp);
488 qp->ibqp.qp_num = qp->mqp.qpn;
495 /* Userspace is not allowed to create special QPs: */
497 return ERR_PTR(-EINVAL);
499 sqp = kmalloc(sizeof *sqp, GFP_KERNEL);
501 return ERR_PTR(-ENOMEM);
505 err = create_qp_common(dev, pd, init_attr, udata,
506 dev->dev->caps.sqp_start +
507 (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) +
508 init_attr->port_num - 1,
515 qp->port = init_attr->port_num;
516 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
521 /* Don't support raw QPs */
522 return ERR_PTR(-EINVAL);
528 int mlx4_ib_destroy_qp(struct ib_qp *qp)
530 struct mlx4_ib_dev *dev = to_mdev(qp->device);
531 struct mlx4_ib_qp *mqp = to_mqp(qp);
533 if (is_qp0(dev, mqp))
534 mlx4_CLOSE_PORT(dev->dev, mqp->port);
536 destroy_qp_common(dev, mqp, !!qp->pd->uobject);
538 if (is_sqp(dev, mqp))
546 static void init_port(struct mlx4_ib_dev *dev, int port)
548 struct mlx4_init_port_param param;
551 memset(¶m, 0, sizeof param);
553 param.port_width_cap = dev->dev->caps.port_width_cap;
554 param.vl_cap = dev->dev->caps.vl_cap;
555 param.mtu = ib_mtu_enum_to_int(dev->dev->caps.mtu_cap);
556 param.max_gid = dev->dev->caps.gid_table_len;
557 param.max_pkey = dev->dev->caps.pkey_table_len;
559 err = mlx4_INIT_PORT(dev->dev, ¶m, port);
561 printk(KERN_WARNING "INIT_PORT failed, return code %d.\n", err);
564 static int to_mlx4_st(enum ib_qp_type type)
567 case IB_QPT_RC: return MLX4_QP_ST_RC;
568 case IB_QPT_UC: return MLX4_QP_ST_UC;
569 case IB_QPT_UD: return MLX4_QP_ST_UD;
571 case IB_QPT_GSI: return MLX4_QP_ST_MLX;
576 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
581 u32 hw_access_flags = 0;
583 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
584 dest_rd_atomic = attr->max_dest_rd_atomic;
586 dest_rd_atomic = qp->resp_depth;
588 if (attr_mask & IB_QP_ACCESS_FLAGS)
589 access_flags = attr->qp_access_flags;
591 access_flags = qp->atomic_rd_en;
594 access_flags &= IB_ACCESS_REMOTE_WRITE;
596 if (access_flags & IB_ACCESS_REMOTE_READ)
597 hw_access_flags |= MLX4_QP_BIT_RRE;
598 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
599 hw_access_flags |= MLX4_QP_BIT_RAE;
600 if (access_flags & IB_ACCESS_REMOTE_WRITE)
601 hw_access_flags |= MLX4_QP_BIT_RWE;
603 return cpu_to_be32(hw_access_flags);
606 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
609 if (attr_mask & IB_QP_PKEY_INDEX)
610 sqp->pkey_index = attr->pkey_index;
611 if (attr_mask & IB_QP_QKEY)
612 sqp->qkey = attr->qkey;
613 if (attr_mask & IB_QP_SQ_PSN)
614 sqp->send_psn = attr->sq_psn;
617 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
619 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
622 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
623 struct mlx4_qp_path *path, u8 port)
625 path->grh_mylmc = ah->src_path_bits & 0x7f;
626 path->rlid = cpu_to_be16(ah->dlid);
627 if (ah->static_rate) {
628 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
629 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
630 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
633 path->static_rate = 0;
634 path->counter_index = 0xff;
636 if (ah->ah_flags & IB_AH_GRH) {
637 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len) {
638 printk(KERN_ERR "sgid_index (%u) too large. max is %d\n",
639 ah->grh.sgid_index, dev->dev->caps.gid_table_len - 1);
643 path->grh_mylmc |= 1 << 7;
644 path->mgid_index = ah->grh.sgid_index;
645 path->hop_limit = ah->grh.hop_limit;
646 path->tclass_flowlabel =
647 cpu_to_be32((ah->grh.traffic_class << 20) |
648 (ah->grh.flow_label));
649 memcpy(path->rgid, ah->grh.dgid.raw, 16);
652 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
653 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
658 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
659 const struct ib_qp_attr *attr, int attr_mask,
660 enum ib_qp_state cur_state, enum ib_qp_state new_state)
662 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
663 struct mlx4_ib_qp *qp = to_mqp(ibqp);
664 struct mlx4_qp_context *context;
665 enum mlx4_qp_optpar optpar = 0;
669 context = kzalloc(sizeof *context, GFP_KERNEL);
673 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
674 (to_mlx4_st(ibqp->qp_type) << 16));
675 context->flags |= cpu_to_be32(1 << 8); /* DE? */
677 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
678 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
680 optpar |= MLX4_QP_OPTPAR_PM_STATE;
681 switch (attr->path_mig_state) {
682 case IB_MIG_MIGRATED:
683 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
686 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
689 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
694 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
695 ibqp->qp_type == IB_QPT_UD)
696 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
697 else if (attr_mask & IB_QP_PATH_MTU) {
698 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
699 printk(KERN_ERR "path MTU (%u) is invalid\n",
703 context->mtu_msgmax = (attr->path_mtu << 5) | 31;
707 context->rq_size_stride = ilog2(qp->rq.max) << 3;
708 context->rq_size_stride |= qp->rq.wqe_shift - 4;
711 context->sq_size_stride = ilog2(qp->sq.max) << 3;
712 context->sq_size_stride |= qp->sq.wqe_shift - 4;
714 if (qp->ibqp.uobject)
715 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
717 context->usr_page = cpu_to_be32(dev->priv_uar.index);
719 if (attr_mask & IB_QP_DEST_QPN)
720 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
722 if (attr_mask & IB_QP_PORT) {
723 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
724 !(attr_mask & IB_QP_AV)) {
725 mlx4_set_sched(&context->pri_path, attr->port_num);
726 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
730 if (attr_mask & IB_QP_PKEY_INDEX) {
731 context->pri_path.pkey_index = attr->pkey_index;
732 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
735 if (attr_mask & IB_QP_RNR_RETRY) {
736 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
737 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
740 if (attr_mask & IB_QP_AV) {
741 if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path,
742 attr_mask & IB_QP_PORT ? attr->port_num : qp->port)) {
747 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
748 MLX4_QP_OPTPAR_SCHED_QUEUE);
751 if (attr_mask & IB_QP_TIMEOUT) {
752 context->pri_path.ackto = attr->timeout << 3;
753 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
756 if (attr_mask & IB_QP_ALT_PATH) {
757 if (attr->alt_pkey_index >= dev->dev->caps.pkey_table_len)
760 if (attr->alt_port_num == 0 ||
761 attr->alt_port_num > dev->dev->caps.num_ports)
764 if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
768 context->alt_path.pkey_index = attr->alt_pkey_index;
769 context->alt_path.ackto = attr->alt_timeout << 3;
770 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
773 context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pdn);
774 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
775 if (attr_mask & IB_QP_RETRY_CNT) {
776 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
777 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
780 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
781 if (attr->max_rd_atomic)
783 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
784 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
787 if (attr_mask & IB_QP_SQ_PSN)
788 context->next_send_psn = cpu_to_be32(attr->sq_psn);
790 context->cqn_send = cpu_to_be32(to_mcq(ibqp->send_cq)->mcq.cqn);
792 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
793 if (attr->max_dest_rd_atomic)
795 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
796 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
799 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
800 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
801 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
805 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
807 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
808 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
809 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
811 if (attr_mask & IB_QP_RQ_PSN)
812 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
814 context->cqn_recv = cpu_to_be32(to_mcq(ibqp->recv_cq)->mcq.cqn);
816 if (attr_mask & IB_QP_QKEY) {
817 context->qkey = cpu_to_be32(attr->qkey);
818 optpar |= MLX4_QP_OPTPAR_Q_KEY;
822 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
824 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
825 context->db_rec_addr = cpu_to_be64(qp->db.dma);
827 if (cur_state == IB_QPS_INIT &&
828 new_state == IB_QPS_RTR &&
829 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
830 ibqp->qp_type == IB_QPT_UD)) {
831 context->pri_path.sched_queue = (qp->port - 1) << 6;
833 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
835 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
838 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
839 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
844 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
845 to_mlx4_state(new_state), context, optpar,
846 sqd_event, &qp->mqp);
850 qp->state = new_state;
852 if (attr_mask & IB_QP_ACCESS_FLAGS)
853 qp->atomic_rd_en = attr->qp_access_flags;
854 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
855 qp->resp_depth = attr->max_dest_rd_atomic;
856 if (attr_mask & IB_QP_PORT)
857 qp->port = attr->port_num;
858 if (attr_mask & IB_QP_ALT_PATH)
859 qp->alt_port = attr->alt_port_num;
862 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
865 * If we moved QP0 to RTR, bring the IB link up; if we moved
866 * QP0 to RESET or ERROR, bring the link back down.
868 if (is_qp0(dev, qp)) {
869 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
870 init_port(dev, qp->port);
872 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
873 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
874 mlx4_CLOSE_PORT(dev->dev, qp->port);
878 * If we moved a kernel QP to RESET, clean up all old CQ
879 * entries and reinitialize the QP.
881 if (new_state == IB_QPS_RESET && !ibqp->uobject) {
882 mlx4_ib_cq_clean(to_mcq(ibqp->recv_cq), qp->mqp.qpn,
883 ibqp->srq ? to_msrq(ibqp->srq): NULL);
884 if (ibqp->send_cq != ibqp->recv_cq)
885 mlx4_ib_cq_clean(to_mcq(ibqp->send_cq), qp->mqp.qpn, NULL);
899 static const struct ib_qp_attr mlx4_ib_qp_attr = { .port_num = 1 };
900 static const int mlx4_ib_qp_attr_mask_table[IB_QPT_UD + 1] = {
901 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
904 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
907 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
910 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
912 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
916 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
917 int attr_mask, struct ib_udata *udata)
919 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
920 struct mlx4_ib_qp *qp = to_mqp(ibqp);
921 enum ib_qp_state cur_state, new_state;
924 mutex_lock(&qp->mutex);
926 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
927 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
929 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask))
932 if ((attr_mask & IB_QP_PKEY_INDEX) &&
933 attr->pkey_index >= dev->dev->caps.pkey_table_len) {
937 if ((attr_mask & IB_QP_PORT) &&
938 (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) {
942 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
943 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
947 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
948 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
952 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
957 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_ERR) {
958 err = __mlx4_ib_modify_qp(ibqp, &mlx4_ib_qp_attr,
959 mlx4_ib_qp_attr_mask_table[ibqp->qp_type],
960 IB_QPS_RESET, IB_QPS_INIT);
963 cur_state = IB_QPS_INIT;
966 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
969 mutex_unlock(&qp->mutex);
973 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
976 struct ib_device *ib_dev = &to_mdev(sqp->qp.ibqp.device)->ib_dev;
977 struct mlx4_wqe_mlx_seg *mlx = wqe;
978 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
979 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
986 for (i = 0; i < wr->num_sge; ++i)
987 send_size += wr->sg_list[i].length;
989 ib_ud_header_init(send_size, mlx4_ib_ah_grh_present(ah), &sqp->ud_header);
991 sqp->ud_header.lrh.service_level =
992 be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
993 sqp->ud_header.lrh.destination_lid = ah->av.dlid;
994 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.g_slid & 0x7f);
995 if (mlx4_ib_ah_grh_present(ah)) {
996 sqp->ud_header.grh.traffic_class =
997 (be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
998 sqp->ud_header.grh.flow_label =
999 ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
1000 sqp->ud_header.grh.hop_limit = ah->av.hop_limit;
1001 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
1002 ah->av.gid_index, &sqp->ud_header.grh.source_gid);
1003 memcpy(sqp->ud_header.grh.destination_gid.raw,
1007 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1008 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
1009 (sqp->ud_header.lrh.destination_lid ==
1010 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
1011 (sqp->ud_header.lrh.service_level << 8));
1012 mlx->rlid = sqp->ud_header.lrh.destination_lid;
1014 switch (wr->opcode) {
1016 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1017 sqp->ud_header.immediate_present = 0;
1019 case IB_WR_SEND_WITH_IMM:
1020 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1021 sqp->ud_header.immediate_present = 1;
1022 sqp->ud_header.immediate_data = wr->imm_data;
1028 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0;
1029 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
1030 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
1031 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1032 if (!sqp->qp.ibqp.qp_num)
1033 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
1035 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
1036 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1037 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1038 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1039 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
1040 sqp->qkey : wr->wr.ud.remote_qkey);
1041 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
1043 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
1046 printk(KERN_ERR "built UD header of size %d:\n", header_size);
1047 for (i = 0; i < header_size / 4; ++i) {
1049 printk(" [%02x] ", i * 4);
1051 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
1052 if ((i + 1) % 8 == 0)
1058 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
1059 memcpy(inl + 1, sqp->header_buf, header_size);
1061 return ALIGN(sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
1064 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1067 struct mlx4_ib_cq *cq;
1069 cur = wq->head - wq->tail;
1070 if (likely(cur + nreq < wq->max))
1074 spin_lock(&cq->lock);
1075 cur = wq->head - wq->tail;
1076 spin_unlock(&cq->lock);
1078 return cur + nreq >= wq->max;
1081 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1082 struct ib_send_wr **bad_wr)
1084 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1086 struct mlx4_wqe_ctrl_seg *ctrl;
1087 unsigned long flags;
1094 spin_lock_irqsave(&qp->rq.lock, flags);
1098 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1099 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
1105 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
1111 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.max - 1));
1112 qp->sq.wrid[ind & (qp->sq.max - 1)] = wr->wr_id;
1115 (wr->send_flags & IB_SEND_SIGNALED ?
1116 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
1117 (wr->send_flags & IB_SEND_SOLICITED ?
1118 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
1121 if (wr->opcode == IB_WR_SEND_WITH_IMM ||
1122 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
1123 ctrl->imm = wr->imm_data;
1127 wqe += sizeof *ctrl;
1128 size = sizeof *ctrl / 16;
1130 switch (ibqp->qp_type) {
1133 switch (wr->opcode) {
1134 case IB_WR_ATOMIC_CMP_AND_SWP:
1135 case IB_WR_ATOMIC_FETCH_AND_ADD:
1136 ((struct mlx4_wqe_raddr_seg *) wqe)->raddr =
1137 cpu_to_be64(wr->wr.atomic.remote_addr);
1138 ((struct mlx4_wqe_raddr_seg *) wqe)->rkey =
1139 cpu_to_be32(wr->wr.atomic.rkey);
1140 ((struct mlx4_wqe_raddr_seg *) wqe)->reserved = 0;
1142 wqe += sizeof (struct mlx4_wqe_raddr_seg);
1144 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
1145 ((struct mlx4_wqe_atomic_seg *) wqe)->swap_add =
1146 cpu_to_be64(wr->wr.atomic.swap);
1147 ((struct mlx4_wqe_atomic_seg *) wqe)->compare =
1148 cpu_to_be64(wr->wr.atomic.compare_add);
1150 ((struct mlx4_wqe_atomic_seg *) wqe)->swap_add =
1151 cpu_to_be64(wr->wr.atomic.compare_add);
1152 ((struct mlx4_wqe_atomic_seg *) wqe)->compare = 0;
1155 wqe += sizeof (struct mlx4_wqe_atomic_seg);
1156 size += (sizeof (struct mlx4_wqe_raddr_seg) +
1157 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
1161 case IB_WR_RDMA_READ:
1162 case IB_WR_RDMA_WRITE:
1163 case IB_WR_RDMA_WRITE_WITH_IMM:
1164 ((struct mlx4_wqe_raddr_seg *) wqe)->raddr =
1165 cpu_to_be64(wr->wr.rdma.remote_addr);
1166 ((struct mlx4_wqe_raddr_seg *) wqe)->rkey =
1167 cpu_to_be32(wr->wr.rdma.rkey);
1168 ((struct mlx4_wqe_raddr_seg *) wqe)->reserved = 0;
1170 wqe += sizeof (struct mlx4_wqe_raddr_seg);
1171 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
1176 /* No extra segments required for sends */
1182 memcpy(((struct mlx4_wqe_datagram_seg *) wqe)->av,
1183 &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
1184 ((struct mlx4_wqe_datagram_seg *) wqe)->dqpn =
1185 cpu_to_be32(wr->wr.ud.remote_qpn);
1186 ((struct mlx4_wqe_datagram_seg *) wqe)->qkey =
1187 cpu_to_be32(wr->wr.ud.remote_qkey);
1189 wqe += sizeof (struct mlx4_wqe_datagram_seg);
1190 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
1195 err = build_mlx_header(to_msqp(qp), wr, ctrl);
1210 for (i = 0; i < wr->num_sge; ++i) {
1211 ((struct mlx4_wqe_data_seg *) wqe)->byte_count =
1212 cpu_to_be32(wr->sg_list[i].length);
1213 ((struct mlx4_wqe_data_seg *) wqe)->lkey =
1214 cpu_to_be32(wr->sg_list[i].lkey);
1215 ((struct mlx4_wqe_data_seg *) wqe)->addr =
1216 cpu_to_be64(wr->sg_list[i].addr);
1218 wqe += sizeof (struct mlx4_wqe_data_seg);
1219 size += sizeof (struct mlx4_wqe_data_seg) / 16;
1222 /* Add one more inline data segment for ICRC for MLX sends */
1223 if (qp->ibqp.qp_type == IB_QPT_SMI || qp->ibqp.qp_type == IB_QPT_GSI) {
1224 ((struct mlx4_wqe_inline_seg *) wqe)->byte_count =
1225 cpu_to_be32((1 << 31) | 4);
1226 ((u32 *) wqe)[1] = 0;
1227 wqe += sizeof (struct mlx4_wqe_data_seg);
1228 size += sizeof (struct mlx4_wqe_data_seg) / 16;
1231 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
1232 MLX4_WQE_CTRL_FENCE : 0) | size;
1235 * Make sure descriptor is fully written before
1236 * setting ownership bit (because HW can start
1237 * executing as soon as we do).
1241 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
1246 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
1247 (ind & qp->sq.max ? cpu_to_be32(1 << 31) : 0);
1254 qp->sq.head += nreq;
1257 * Make sure that descriptors are written before
1262 writel(qp->doorbell_qpn,
1263 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
1266 * Make sure doorbells don't leak out of SQ spinlock
1267 * and reach the HCA out of order.
1272 spin_unlock_irqrestore(&qp->rq.lock, flags);
1277 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1278 struct ib_recv_wr **bad_wr)
1280 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1281 struct mlx4_wqe_data_seg *scat;
1282 unsigned long flags;
1288 spin_lock_irqsave(&qp->rq.lock, flags);
1290 ind = qp->rq.head & (qp->rq.max - 1);
1292 for (nreq = 0; wr; ++nreq, wr = wr->next) {
1293 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.send_cq)) {
1299 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
1305 scat = get_recv_wqe(qp, ind);
1307 for (i = 0; i < wr->num_sge; ++i) {
1308 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
1309 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
1310 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
1313 if (i < qp->rq.max_gs) {
1314 scat[i].byte_count = 0;
1315 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
1319 qp->rq.wrid[ind] = wr->wr_id;
1321 ind = (ind + 1) & (qp->rq.max - 1);
1326 qp->rq.head += nreq;
1329 * Make sure that descriptors are written before
1334 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
1337 spin_unlock_irqrestore(&qp->rq.lock, flags);