2 * Copyright (c) 2013-2015, Mellanox Technologies. 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 <linux/module.h>
34 #include <rdma/ib_umem.h>
38 /* not supported currently */
39 static int wq_signature;
42 MLX5_IB_ACK_REQ_FREQ = 8,
46 MLX5_IB_DEFAULT_SCHED_QUEUE = 0x83,
47 MLX5_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
48 MLX5_IB_LINK_TYPE_IB = 0,
49 MLX5_IB_LINK_TYPE_ETH = 1
53 MLX5_IB_SQ_STRIDE = 6,
54 MLX5_IB_CACHE_LINE_SIZE = 64,
57 static const u32 mlx5_ib_opcode[] = {
58 [IB_WR_SEND] = MLX5_OPCODE_SEND,
59 [IB_WR_SEND_WITH_IMM] = MLX5_OPCODE_SEND_IMM,
60 [IB_WR_RDMA_WRITE] = MLX5_OPCODE_RDMA_WRITE,
61 [IB_WR_RDMA_WRITE_WITH_IMM] = MLX5_OPCODE_RDMA_WRITE_IMM,
62 [IB_WR_RDMA_READ] = MLX5_OPCODE_RDMA_READ,
63 [IB_WR_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_CS,
64 [IB_WR_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_FA,
65 [IB_WR_SEND_WITH_INV] = MLX5_OPCODE_SEND_INVAL,
66 [IB_WR_LOCAL_INV] = MLX5_OPCODE_UMR,
67 [IB_WR_FAST_REG_MR] = MLX5_OPCODE_UMR,
68 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_MASKED_CS,
69 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_MASKED_FA,
70 [MLX5_IB_WR_UMR] = MLX5_OPCODE_UMR,
74 static int is_qp0(enum ib_qp_type qp_type)
76 return qp_type == IB_QPT_SMI;
79 static int is_qp1(enum ib_qp_type qp_type)
81 return qp_type == IB_QPT_GSI;
84 static int is_sqp(enum ib_qp_type qp_type)
86 return is_qp0(qp_type) || is_qp1(qp_type);
89 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
91 return mlx5_buf_offset(&qp->buf, offset);
94 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
96 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
99 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
101 return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
105 * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
107 * @qp: QP to copy from.
108 * @send: copy from the send queue when non-zero, use the receive queue
110 * @wqe_index: index to start copying from. For send work queues, the
111 * wqe_index is in units of MLX5_SEND_WQE_BB.
112 * For receive work queue, it is the number of work queue
113 * element in the queue.
114 * @buffer: destination buffer.
115 * @length: maximum number of bytes to copy.
117 * Copies at least a single WQE, but may copy more data.
119 * Return: the number of bytes copied, or an error code.
121 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
122 void *buffer, u32 length)
124 struct ib_device *ibdev = qp->ibqp.device;
125 struct mlx5_ib_dev *dev = to_mdev(ibdev);
126 struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
129 struct ib_umem *umem = qp->umem;
130 u32 first_copy_length;
134 if (wq->wqe_cnt == 0) {
135 mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
140 offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
141 wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
143 if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
146 if (offset > umem->length ||
147 (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
150 first_copy_length = min_t(u32, offset + length, wq_end) - offset;
151 ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
156 struct mlx5_wqe_ctrl_seg *ctrl = buffer;
157 int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
159 wqe_length = ds * MLX5_WQE_DS_UNITS;
161 wqe_length = 1 << wq->wqe_shift;
164 if (wqe_length <= first_copy_length)
165 return first_copy_length;
167 ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
168 wqe_length - first_copy_length);
175 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
177 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
178 struct ib_event event;
180 if (type == MLX5_EVENT_TYPE_PATH_MIG)
181 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
183 if (ibqp->event_handler) {
184 event.device = ibqp->device;
185 event.element.qp = ibqp;
187 case MLX5_EVENT_TYPE_PATH_MIG:
188 event.event = IB_EVENT_PATH_MIG;
190 case MLX5_EVENT_TYPE_COMM_EST:
191 event.event = IB_EVENT_COMM_EST;
193 case MLX5_EVENT_TYPE_SQ_DRAINED:
194 event.event = IB_EVENT_SQ_DRAINED;
196 case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
197 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
199 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
200 event.event = IB_EVENT_QP_FATAL;
202 case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
203 event.event = IB_EVENT_PATH_MIG_ERR;
205 case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
206 event.event = IB_EVENT_QP_REQ_ERR;
208 case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
209 event.event = IB_EVENT_QP_ACCESS_ERR;
212 pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
216 ibqp->event_handler(&event, ibqp->qp_context);
220 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
221 int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
226 /* Sanity check RQ size before proceeding */
227 if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
233 qp->rq.wqe_shift = 0;
236 qp->rq.wqe_cnt = ucmd->rq_wqe_count;
237 qp->rq.wqe_shift = ucmd->rq_wqe_shift;
238 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
239 qp->rq.max_post = qp->rq.wqe_cnt;
241 wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
242 wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
243 wqe_size = roundup_pow_of_two(wqe_size);
244 wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
245 wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
246 qp->rq.wqe_cnt = wq_size / wqe_size;
247 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
248 mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
250 MLX5_CAP_GEN(dev->mdev,
254 qp->rq.wqe_shift = ilog2(wqe_size);
255 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
256 qp->rq.max_post = qp->rq.wqe_cnt;
263 static int sq_overhead(enum ib_qp_type qp_type)
269 size += sizeof(struct mlx5_wqe_xrc_seg);
272 size += sizeof(struct mlx5_wqe_ctrl_seg) +
273 sizeof(struct mlx5_wqe_atomic_seg) +
274 sizeof(struct mlx5_wqe_raddr_seg);
281 size += sizeof(struct mlx5_wqe_ctrl_seg) +
282 sizeof(struct mlx5_wqe_raddr_seg) +
283 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
284 sizeof(struct mlx5_mkey_seg);
290 size += sizeof(struct mlx5_wqe_ctrl_seg) +
291 sizeof(struct mlx5_wqe_datagram_seg);
294 case MLX5_IB_QPT_REG_UMR:
295 size += sizeof(struct mlx5_wqe_ctrl_seg) +
296 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
297 sizeof(struct mlx5_mkey_seg);
307 static int calc_send_wqe(struct ib_qp_init_attr *attr)
312 size = sq_overhead(attr->qp_type);
316 if (attr->cap.max_inline_data) {
317 inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
318 attr->cap.max_inline_data;
321 size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
322 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
323 ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
324 return MLX5_SIG_WQE_SIZE;
326 return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
329 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
330 struct mlx5_ib_qp *qp)
335 if (!attr->cap.max_send_wr)
338 wqe_size = calc_send_wqe(attr);
339 mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
343 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
344 mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
345 wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
349 qp->max_inline_data = wqe_size - sq_overhead(attr->qp_type) -
350 sizeof(struct mlx5_wqe_inline_seg);
351 attr->cap.max_inline_data = qp->max_inline_data;
353 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
354 qp->signature_en = true;
356 wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
357 qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
358 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
359 mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
361 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
364 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
365 qp->sq.max_gs = attr->cap.max_send_sge;
366 qp->sq.max_post = wq_size / wqe_size;
367 attr->cap.max_send_wr = qp->sq.max_post;
372 static int set_user_buf_size(struct mlx5_ib_dev *dev,
373 struct mlx5_ib_qp *qp,
374 struct mlx5_ib_create_qp *ucmd)
376 int desc_sz = 1 << qp->sq.wqe_shift;
378 if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
379 mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
380 desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
384 if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
385 mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
386 ucmd->sq_wqe_count, ucmd->sq_wqe_count);
390 qp->sq.wqe_cnt = ucmd->sq_wqe_count;
392 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
393 mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
395 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
399 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
400 (qp->sq.wqe_cnt << 6);
405 static int qp_has_rq(struct ib_qp_init_attr *attr)
407 if (attr->qp_type == IB_QPT_XRC_INI ||
408 attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
409 attr->qp_type == MLX5_IB_QPT_REG_UMR ||
410 !attr->cap.max_recv_wr)
416 static int first_med_uuar(void)
421 static int next_uuar(int n)
425 while (((n % 4) & 2))
431 static int num_med_uuar(struct mlx5_uuar_info *uuari)
435 n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
436 uuari->num_low_latency_uuars - 1;
438 return n >= 0 ? n : 0;
441 static int max_uuari(struct mlx5_uuar_info *uuari)
443 return uuari->num_uars * 4;
446 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
452 med = num_med_uuar(uuari);
453 for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
462 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
466 for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
467 if (!test_bit(i, uuari->bitmap)) {
468 set_bit(i, uuari->bitmap);
477 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
479 int minidx = first_med_uuar();
482 for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
483 if (uuari->count[i] < uuari->count[minidx])
487 uuari->count[minidx]++;
491 static int alloc_uuar(struct mlx5_uuar_info *uuari,
492 enum mlx5_ib_latency_class lat)
496 mutex_lock(&uuari->lock);
498 case MLX5_IB_LATENCY_CLASS_LOW:
500 uuari->count[uuarn]++;
503 case MLX5_IB_LATENCY_CLASS_MEDIUM:
507 uuarn = alloc_med_class_uuar(uuari);
510 case MLX5_IB_LATENCY_CLASS_HIGH:
514 uuarn = alloc_high_class_uuar(uuari);
517 case MLX5_IB_LATENCY_CLASS_FAST_PATH:
521 mutex_unlock(&uuari->lock);
526 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
528 clear_bit(uuarn, uuari->bitmap);
529 --uuari->count[uuarn];
532 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
534 clear_bit(uuarn, uuari->bitmap);
535 --uuari->count[uuarn];
538 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
540 int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
541 int high_uuar = nuuars - uuari->num_low_latency_uuars;
543 mutex_lock(&uuari->lock);
545 --uuari->count[uuarn];
549 if (uuarn < high_uuar) {
550 free_med_class_uuar(uuari, uuarn);
554 free_high_class_uuar(uuari, uuarn);
557 mutex_unlock(&uuari->lock);
560 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
563 case IB_QPS_RESET: return MLX5_QP_STATE_RST;
564 case IB_QPS_INIT: return MLX5_QP_STATE_INIT;
565 case IB_QPS_RTR: return MLX5_QP_STATE_RTR;
566 case IB_QPS_RTS: return MLX5_QP_STATE_RTS;
567 case IB_QPS_SQD: return MLX5_QP_STATE_SQD;
568 case IB_QPS_SQE: return MLX5_QP_STATE_SQER;
569 case IB_QPS_ERR: return MLX5_QP_STATE_ERR;
574 static int to_mlx5_st(enum ib_qp_type type)
577 case IB_QPT_RC: return MLX5_QP_ST_RC;
578 case IB_QPT_UC: return MLX5_QP_ST_UC;
579 case IB_QPT_UD: return MLX5_QP_ST_UD;
580 case MLX5_IB_QPT_REG_UMR: return MLX5_QP_ST_REG_UMR;
582 case IB_QPT_XRC_TGT: return MLX5_QP_ST_XRC;
583 case IB_QPT_SMI: return MLX5_QP_ST_QP0;
584 case IB_QPT_GSI: return MLX5_QP_ST_QP1;
585 case IB_QPT_RAW_IPV6: return MLX5_QP_ST_RAW_IPV6;
586 case IB_QPT_RAW_ETHERTYPE: return MLX5_QP_ST_RAW_ETHERTYPE;
587 case IB_QPT_RAW_PACKET:
589 default: return -EINVAL;
593 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
595 return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
598 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
599 struct mlx5_ib_qp *qp, struct ib_udata *udata,
600 struct mlx5_create_qp_mbox_in **in,
601 struct mlx5_ib_create_qp_resp *resp, int *inlen)
603 struct mlx5_ib_ucontext *context;
604 struct mlx5_ib_create_qp ucmd;
613 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
615 mlx5_ib_dbg(dev, "copy failed\n");
619 context = to_mucontext(pd->uobject->context);
621 * TBD: should come from the verbs when we have the API
623 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
625 mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
626 mlx5_ib_dbg(dev, "reverting to medium latency\n");
627 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
629 mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
630 mlx5_ib_dbg(dev, "reverting to high latency\n");
631 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
633 mlx5_ib_warn(dev, "uuar allocation failed\n");
639 uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
640 mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
643 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
644 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
646 err = set_user_buf_size(dev, qp, &ucmd);
650 if (ucmd.buf_addr && qp->buf_size) {
651 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
653 if (IS_ERR(qp->umem)) {
654 mlx5_ib_dbg(dev, "umem_get failed\n");
655 err = PTR_ERR(qp->umem);
663 mlx5_ib_cont_pages(qp->umem, ucmd.buf_addr, &npages, &page_shift,
665 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift, &offset);
667 mlx5_ib_warn(dev, "bad offset\n");
670 mlx5_ib_dbg(dev, "addr 0x%llx, size %d, npages %d, page_shift %d, ncont %d, offset %d\n",
671 ucmd.buf_addr, qp->buf_size, npages, page_shift, ncont, offset);
674 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
675 *in = mlx5_vzalloc(*inlen);
681 mlx5_ib_populate_pas(dev, qp->umem, page_shift, (*in)->pas, 0);
682 (*in)->ctx.log_pg_sz_remote_qpn =
683 cpu_to_be32((page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
684 (*in)->ctx.params2 = cpu_to_be32(offset << 6);
686 (*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
687 resp->uuar_index = uuarn;
690 err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
692 mlx5_ib_dbg(dev, "map failed\n");
696 err = ib_copy_to_udata(udata, resp, sizeof(*resp));
698 mlx5_ib_dbg(dev, "copy failed\n");
701 qp->create_type = MLX5_QP_USER;
706 mlx5_ib_db_unmap_user(context, &qp->db);
713 ib_umem_release(qp->umem);
716 free_uuar(&context->uuari, uuarn);
720 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp)
722 struct mlx5_ib_ucontext *context;
724 context = to_mucontext(pd->uobject->context);
725 mlx5_ib_db_unmap_user(context, &qp->db);
727 ib_umem_release(qp->umem);
728 free_uuar(&context->uuari, qp->uuarn);
731 static int create_kernel_qp(struct mlx5_ib_dev *dev,
732 struct ib_qp_init_attr *init_attr,
733 struct mlx5_ib_qp *qp,
734 struct mlx5_create_qp_mbox_in **in, int *inlen)
736 enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
737 struct mlx5_uuar_info *uuari;
742 uuari = &dev->mdev->priv.uuari;
743 if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN | IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK))
746 if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
747 lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
749 uuarn = alloc_uuar(uuari, lc);
751 mlx5_ib_dbg(dev, "\n");
755 qp->bf = &uuari->bfs[uuarn];
756 uar_index = qp->bf->uar->index;
758 err = calc_sq_size(dev, init_attr, qp);
760 mlx5_ib_dbg(dev, "err %d\n", err);
765 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
766 qp->buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
768 err = mlx5_buf_alloc(dev->mdev, qp->buf_size, &qp->buf);
770 mlx5_ib_dbg(dev, "err %d\n", err);
774 qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
775 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * qp->buf.npages;
776 *in = mlx5_vzalloc(*inlen);
781 (*in)->ctx.qp_counter_set_usr_page = cpu_to_be32(uar_index);
782 (*in)->ctx.log_pg_sz_remote_qpn =
783 cpu_to_be32((qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT) << 24);
784 /* Set "fast registration enabled" for all kernel QPs */
785 (*in)->ctx.params1 |= cpu_to_be32(1 << 11);
786 (*in)->ctx.sq_crq_size |= cpu_to_be16(1 << 4);
788 mlx5_fill_page_array(&qp->buf, (*in)->pas);
790 err = mlx5_db_alloc(dev->mdev, &qp->db);
792 mlx5_ib_dbg(dev, "err %d\n", err);
796 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
797 qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
798 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
799 qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
800 qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
802 if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
803 !qp->sq.w_list || !qp->sq.wqe_head) {
807 qp->create_type = MLX5_QP_KERNEL;
812 mlx5_db_free(dev->mdev, &qp->db);
813 kfree(qp->sq.wqe_head);
814 kfree(qp->sq.w_list);
816 kfree(qp->sq.wr_data);
823 mlx5_buf_free(dev->mdev, &qp->buf);
826 free_uuar(&dev->mdev->priv.uuari, uuarn);
830 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
832 mlx5_db_free(dev->mdev, &qp->db);
833 kfree(qp->sq.wqe_head);
834 kfree(qp->sq.w_list);
836 kfree(qp->sq.wr_data);
838 mlx5_buf_free(dev->mdev, &qp->buf);
839 free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
842 static __be32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
844 if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
845 (attr->qp_type == IB_QPT_XRC_INI))
846 return cpu_to_be32(MLX5_SRQ_RQ);
847 else if (!qp->has_rq)
848 return cpu_to_be32(MLX5_ZERO_LEN_RQ);
850 return cpu_to_be32(MLX5_NON_ZERO_RQ);
853 static int is_connected(enum ib_qp_type qp_type)
855 if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
861 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
862 struct ib_qp_init_attr *init_attr,
863 struct ib_udata *udata, struct mlx5_ib_qp *qp)
865 struct mlx5_ib_resources *devr = &dev->devr;
866 struct mlx5_core_dev *mdev = dev->mdev;
867 struct mlx5_ib_create_qp_resp resp;
868 struct mlx5_create_qp_mbox_in *in;
869 struct mlx5_ib_create_qp ucmd;
870 int inlen = sizeof(*in);
873 mlx5_ib_odp_create_qp(qp);
875 mutex_init(&qp->mutex);
876 spin_lock_init(&qp->sq.lock);
877 spin_lock_init(&qp->rq.lock);
879 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
880 if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
881 mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
884 qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
888 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
889 qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
891 if (pd && pd->uobject) {
892 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
893 mlx5_ib_dbg(dev, "copy failed\n");
897 qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
898 qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
900 qp->wq_sig = !!wq_signature;
903 qp->has_rq = qp_has_rq(init_attr);
904 err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
905 qp, (pd && pd->uobject) ? &ucmd : NULL);
907 mlx5_ib_dbg(dev, "err %d\n", err);
914 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
915 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
916 if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
917 ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
918 mlx5_ib_dbg(dev, "invalid rq params\n");
921 if (ucmd.sq_wqe_count > max_wqes) {
922 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
923 ucmd.sq_wqe_count, max_wqes);
926 err = create_user_qp(dev, pd, qp, udata, &in, &resp, &inlen);
928 mlx5_ib_dbg(dev, "err %d\n", err);
930 err = create_kernel_qp(dev, init_attr, qp, &in, &inlen);
932 mlx5_ib_dbg(dev, "err %d\n", err);
934 qp->pa_lkey = to_mpd(pd)->pa_lkey;
940 in = mlx5_vzalloc(sizeof(*in));
944 qp->create_type = MLX5_QP_EMPTY;
947 if (is_sqp(init_attr->qp_type))
948 qp->port = init_attr->port_num;
950 in->ctx.flags = cpu_to_be32(to_mlx5_st(init_attr->qp_type) << 16 |
951 MLX5_QP_PM_MIGRATED << 11);
953 if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
954 in->ctx.flags_pd = cpu_to_be32(to_mpd(pd ? pd : devr->p0)->pdn);
956 in->ctx.flags_pd = cpu_to_be32(MLX5_QP_LAT_SENSITIVE);
959 in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_ENABLE_SIG);
961 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
962 in->ctx.flags_pd |= cpu_to_be32(MLX5_QP_BLOCK_MCAST);
964 if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
968 rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
969 scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
972 in->ctx.cs_res = MLX5_RES_SCAT_DATA64_CQE;
974 in->ctx.cs_res = MLX5_RES_SCAT_DATA32_CQE;
976 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
978 in->ctx.cs_req = MLX5_REQ_SCAT_DATA64_CQE;
980 in->ctx.cs_req = MLX5_REQ_SCAT_DATA32_CQE;
984 if (qp->rq.wqe_cnt) {
985 in->ctx.rq_size_stride = (qp->rq.wqe_shift - 4);
986 in->ctx.rq_size_stride |= ilog2(qp->rq.wqe_cnt) << 3;
989 in->ctx.rq_type_srqn = get_rx_type(qp, init_attr);
992 in->ctx.sq_crq_size |= cpu_to_be16(ilog2(qp->sq.wqe_cnt) << 11);
994 in->ctx.sq_crq_size |= cpu_to_be16(0x8000);
996 /* Set default resources */
997 switch (init_attr->qp_type) {
999 in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1000 in->ctx.cqn_send = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1001 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1002 in->ctx.xrcd = cpu_to_be32(to_mxrcd(init_attr->xrcd)->xrcdn);
1004 case IB_QPT_XRC_INI:
1005 in->ctx.cqn_recv = cpu_to_be32(to_mcq(devr->c0)->mcq.cqn);
1006 in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1007 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1010 if (init_attr->srq) {
1011 in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x0)->xrcdn);
1012 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(init_attr->srq)->msrq.srqn);
1014 in->ctx.xrcd = cpu_to_be32(to_mxrcd(devr->x1)->xrcdn);
1015 in->ctx.rq_type_srqn |= cpu_to_be32(to_msrq(devr->s0)->msrq.srqn);
1019 if (init_attr->send_cq)
1020 in->ctx.cqn_send = cpu_to_be32(to_mcq(init_attr->send_cq)->mcq.cqn);
1022 if (init_attr->recv_cq)
1023 in->ctx.cqn_recv = cpu_to_be32(to_mcq(init_attr->recv_cq)->mcq.cqn);
1025 in->ctx.db_rec_addr = cpu_to_be64(qp->db.dma);
1027 err = mlx5_core_create_qp(dev->mdev, &qp->mqp, in, inlen);
1029 mlx5_ib_dbg(dev, "create qp failed\n");
1034 /* Hardware wants QPN written in big-endian order (after
1035 * shifting) for send doorbell. Precompute this value to save
1036 * a little bit when posting sends.
1038 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
1040 qp->mqp.event = mlx5_ib_qp_event;
1045 if (qp->create_type == MLX5_QP_USER)
1046 destroy_qp_user(pd, qp);
1047 else if (qp->create_type == MLX5_QP_KERNEL)
1048 destroy_qp_kernel(dev, qp);
1054 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1055 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1059 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1060 spin_lock_irq(&send_cq->lock);
1061 spin_lock_nested(&recv_cq->lock,
1062 SINGLE_DEPTH_NESTING);
1063 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1064 spin_lock_irq(&send_cq->lock);
1065 __acquire(&recv_cq->lock);
1067 spin_lock_irq(&recv_cq->lock);
1068 spin_lock_nested(&send_cq->lock,
1069 SINGLE_DEPTH_NESTING);
1072 spin_lock_irq(&send_cq->lock);
1073 __acquire(&recv_cq->lock);
1075 } else if (recv_cq) {
1076 spin_lock_irq(&recv_cq->lock);
1077 __acquire(&send_cq->lock);
1079 __acquire(&send_cq->lock);
1080 __acquire(&recv_cq->lock);
1084 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1085 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1089 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1090 spin_unlock(&recv_cq->lock);
1091 spin_unlock_irq(&send_cq->lock);
1092 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1093 __release(&recv_cq->lock);
1094 spin_unlock_irq(&send_cq->lock);
1096 spin_unlock(&send_cq->lock);
1097 spin_unlock_irq(&recv_cq->lock);
1100 __release(&recv_cq->lock);
1101 spin_unlock_irq(&send_cq->lock);
1103 } else if (recv_cq) {
1104 __release(&send_cq->lock);
1105 spin_unlock_irq(&recv_cq->lock);
1107 __release(&recv_cq->lock);
1108 __release(&send_cq->lock);
1112 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1114 return to_mpd(qp->ibqp.pd);
1117 static void get_cqs(struct mlx5_ib_qp *qp,
1118 struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1120 switch (qp->ibqp.qp_type) {
1121 case IB_QPT_XRC_TGT:
1125 case MLX5_IB_QPT_REG_UMR:
1126 case IB_QPT_XRC_INI:
1127 *send_cq = to_mcq(qp->ibqp.send_cq);
1136 case IB_QPT_RAW_IPV6:
1137 case IB_QPT_RAW_ETHERTYPE:
1138 *send_cq = to_mcq(qp->ibqp.send_cq);
1139 *recv_cq = to_mcq(qp->ibqp.recv_cq);
1142 case IB_QPT_RAW_PACKET:
1151 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1153 struct mlx5_ib_cq *send_cq, *recv_cq;
1154 struct mlx5_modify_qp_mbox_in *in;
1157 in = kzalloc(sizeof(*in), GFP_KERNEL);
1161 if (qp->state != IB_QPS_RESET) {
1162 mlx5_ib_qp_disable_pagefaults(qp);
1163 if (mlx5_core_qp_modify(dev->mdev, to_mlx5_state(qp->state),
1164 MLX5_QP_STATE_RST, in, 0, &qp->mqp))
1165 mlx5_ib_warn(dev, "mlx5_ib: modify QP %06x to RESET failed\n",
1169 get_cqs(qp, &send_cq, &recv_cq);
1171 if (qp->create_type == MLX5_QP_KERNEL) {
1172 mlx5_ib_lock_cqs(send_cq, recv_cq);
1173 __mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1174 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1175 if (send_cq != recv_cq)
1176 __mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1177 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1180 err = mlx5_core_destroy_qp(dev->mdev, &qp->mqp);
1182 mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n", qp->mqp.qpn);
1186 if (qp->create_type == MLX5_QP_KERNEL)
1187 destroy_qp_kernel(dev, qp);
1188 else if (qp->create_type == MLX5_QP_USER)
1189 destroy_qp_user(&get_pd(qp)->ibpd, qp);
1192 static const char *ib_qp_type_str(enum ib_qp_type type)
1196 return "IB_QPT_SMI";
1198 return "IB_QPT_GSI";
1205 case IB_QPT_RAW_IPV6:
1206 return "IB_QPT_RAW_IPV6";
1207 case IB_QPT_RAW_ETHERTYPE:
1208 return "IB_QPT_RAW_ETHERTYPE";
1209 case IB_QPT_XRC_INI:
1210 return "IB_QPT_XRC_INI";
1211 case IB_QPT_XRC_TGT:
1212 return "IB_QPT_XRC_TGT";
1213 case IB_QPT_RAW_PACKET:
1214 return "IB_QPT_RAW_PACKET";
1215 case MLX5_IB_QPT_REG_UMR:
1216 return "MLX5_IB_QPT_REG_UMR";
1219 return "Invalid QP type";
1223 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1224 struct ib_qp_init_attr *init_attr,
1225 struct ib_udata *udata)
1227 struct mlx5_ib_dev *dev;
1228 struct mlx5_ib_qp *qp;
1233 dev = to_mdev(pd->device);
1235 /* being cautious here */
1236 if (init_attr->qp_type != IB_QPT_XRC_TGT &&
1237 init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
1238 pr_warn("%s: no PD for transport %s\n", __func__,
1239 ib_qp_type_str(init_attr->qp_type));
1240 return ERR_PTR(-EINVAL);
1242 dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
1245 switch (init_attr->qp_type) {
1246 case IB_QPT_XRC_TGT:
1247 case IB_QPT_XRC_INI:
1248 if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
1249 mlx5_ib_dbg(dev, "XRC not supported\n");
1250 return ERR_PTR(-ENOSYS);
1252 init_attr->recv_cq = NULL;
1253 if (init_attr->qp_type == IB_QPT_XRC_TGT) {
1254 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1255 init_attr->send_cq = NULL;
1264 case MLX5_IB_QPT_REG_UMR:
1265 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1267 return ERR_PTR(-ENOMEM);
1269 err = create_qp_common(dev, pd, init_attr, udata, qp);
1271 mlx5_ib_dbg(dev, "create_qp_common failed\n");
1273 return ERR_PTR(err);
1276 if (is_qp0(init_attr->qp_type))
1277 qp->ibqp.qp_num = 0;
1278 else if (is_qp1(init_attr->qp_type))
1279 qp->ibqp.qp_num = 1;
1281 qp->ibqp.qp_num = qp->mqp.qpn;
1283 mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
1284 qp->ibqp.qp_num, qp->mqp.qpn, to_mcq(init_attr->recv_cq)->mcq.cqn,
1285 to_mcq(init_attr->send_cq)->mcq.cqn);
1291 case IB_QPT_RAW_IPV6:
1292 case IB_QPT_RAW_ETHERTYPE:
1293 case IB_QPT_RAW_PACKET:
1296 mlx5_ib_dbg(dev, "unsupported qp type %d\n",
1297 init_attr->qp_type);
1298 /* Don't support raw QPs */
1299 return ERR_PTR(-EINVAL);
1305 int mlx5_ib_destroy_qp(struct ib_qp *qp)
1307 struct mlx5_ib_dev *dev = to_mdev(qp->device);
1308 struct mlx5_ib_qp *mqp = to_mqp(qp);
1310 destroy_qp_common(dev, mqp);
1317 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
1320 u32 hw_access_flags = 0;
1324 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1325 dest_rd_atomic = attr->max_dest_rd_atomic;
1327 dest_rd_atomic = qp->resp_depth;
1329 if (attr_mask & IB_QP_ACCESS_FLAGS)
1330 access_flags = attr->qp_access_flags;
1332 access_flags = qp->atomic_rd_en;
1334 if (!dest_rd_atomic)
1335 access_flags &= IB_ACCESS_REMOTE_WRITE;
1337 if (access_flags & IB_ACCESS_REMOTE_READ)
1338 hw_access_flags |= MLX5_QP_BIT_RRE;
1339 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1340 hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
1341 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1342 hw_access_flags |= MLX5_QP_BIT_RWE;
1344 return cpu_to_be32(hw_access_flags);
1348 MLX5_PATH_FLAG_FL = 1 << 0,
1349 MLX5_PATH_FLAG_FREE_AR = 1 << 1,
1350 MLX5_PATH_FLAG_COUNTER = 1 << 2,
1353 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
1355 if (rate == IB_RATE_PORT_CURRENT) {
1357 } else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
1360 while (rate != IB_RATE_2_5_GBPS &&
1361 !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
1362 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
1366 return rate + MLX5_STAT_RATE_OFFSET;
1369 static int mlx5_set_path(struct mlx5_ib_dev *dev, const struct ib_ah_attr *ah,
1370 struct mlx5_qp_path *path, u8 port, int attr_mask,
1371 u32 path_flags, const struct ib_qp_attr *attr)
1375 path->fl = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
1376 path->free_ar = (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x80 : 0;
1378 if (attr_mask & IB_QP_PKEY_INDEX)
1379 path->pkey_index = attr->pkey_index;
1381 path->grh_mlid = ah->src_path_bits & 0x7f;
1382 path->rlid = cpu_to_be16(ah->dlid);
1384 if (ah->ah_flags & IB_AH_GRH) {
1385 if (ah->grh.sgid_index >=
1386 dev->mdev->port_caps[port - 1].gid_table_len) {
1387 pr_err("sgid_index (%u) too large. max is %d\n",
1389 dev->mdev->port_caps[port - 1].gid_table_len);
1392 path->grh_mlid |= 1 << 7;
1393 path->mgid_index = ah->grh.sgid_index;
1394 path->hop_limit = ah->grh.hop_limit;
1395 path->tclass_flowlabel =
1396 cpu_to_be32((ah->grh.traffic_class << 20) |
1397 (ah->grh.flow_label));
1398 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1401 err = ib_rate_to_mlx5(dev, ah->static_rate);
1404 path->static_rate = err;
1407 if (attr_mask & IB_QP_TIMEOUT)
1408 path->ackto_lt = attr->timeout << 3;
1410 path->sl = ah->sl & 0xf;
1415 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
1416 [MLX5_QP_STATE_INIT] = {
1417 [MLX5_QP_STATE_INIT] = {
1418 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
1419 MLX5_QP_OPTPAR_RAE |
1420 MLX5_QP_OPTPAR_RWE |
1421 MLX5_QP_OPTPAR_PKEY_INDEX |
1422 MLX5_QP_OPTPAR_PRI_PORT,
1423 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
1424 MLX5_QP_OPTPAR_PKEY_INDEX |
1425 MLX5_QP_OPTPAR_PRI_PORT,
1426 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
1427 MLX5_QP_OPTPAR_Q_KEY |
1428 MLX5_QP_OPTPAR_PRI_PORT,
1430 [MLX5_QP_STATE_RTR] = {
1431 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1432 MLX5_QP_OPTPAR_RRE |
1433 MLX5_QP_OPTPAR_RAE |
1434 MLX5_QP_OPTPAR_RWE |
1435 MLX5_QP_OPTPAR_PKEY_INDEX,
1436 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1437 MLX5_QP_OPTPAR_RWE |
1438 MLX5_QP_OPTPAR_PKEY_INDEX,
1439 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
1440 MLX5_QP_OPTPAR_Q_KEY,
1441 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX |
1442 MLX5_QP_OPTPAR_Q_KEY,
1443 [MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1444 MLX5_QP_OPTPAR_RRE |
1445 MLX5_QP_OPTPAR_RAE |
1446 MLX5_QP_OPTPAR_RWE |
1447 MLX5_QP_OPTPAR_PKEY_INDEX,
1450 [MLX5_QP_STATE_RTR] = {
1451 [MLX5_QP_STATE_RTS] = {
1452 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1453 MLX5_QP_OPTPAR_RRE |
1454 MLX5_QP_OPTPAR_RAE |
1455 MLX5_QP_OPTPAR_RWE |
1456 MLX5_QP_OPTPAR_PM_STATE |
1457 MLX5_QP_OPTPAR_RNR_TIMEOUT,
1458 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
1459 MLX5_QP_OPTPAR_RWE |
1460 MLX5_QP_OPTPAR_PM_STATE,
1461 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1464 [MLX5_QP_STATE_RTS] = {
1465 [MLX5_QP_STATE_RTS] = {
1466 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
1467 MLX5_QP_OPTPAR_RAE |
1468 MLX5_QP_OPTPAR_RWE |
1469 MLX5_QP_OPTPAR_RNR_TIMEOUT |
1470 MLX5_QP_OPTPAR_PM_STATE |
1471 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1472 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
1473 MLX5_QP_OPTPAR_PM_STATE |
1474 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
1475 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY |
1476 MLX5_QP_OPTPAR_SRQN |
1477 MLX5_QP_OPTPAR_CQN_RCV,
1480 [MLX5_QP_STATE_SQER] = {
1481 [MLX5_QP_STATE_RTS] = {
1482 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
1483 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
1484 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
1485 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT |
1486 MLX5_QP_OPTPAR_RWE |
1487 MLX5_QP_OPTPAR_RAE |
1493 static int ib_nr_to_mlx5_nr(int ib_mask)
1498 case IB_QP_CUR_STATE:
1500 case IB_QP_EN_SQD_ASYNC_NOTIFY:
1502 case IB_QP_ACCESS_FLAGS:
1503 return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
1505 case IB_QP_PKEY_INDEX:
1506 return MLX5_QP_OPTPAR_PKEY_INDEX;
1508 return MLX5_QP_OPTPAR_PRI_PORT;
1510 return MLX5_QP_OPTPAR_Q_KEY;
1512 return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
1513 MLX5_QP_OPTPAR_PRI_PORT;
1514 case IB_QP_PATH_MTU:
1517 return MLX5_QP_OPTPAR_ACK_TIMEOUT;
1518 case IB_QP_RETRY_CNT:
1519 return MLX5_QP_OPTPAR_RETRY_COUNT;
1520 case IB_QP_RNR_RETRY:
1521 return MLX5_QP_OPTPAR_RNR_RETRY;
1524 case IB_QP_MAX_QP_RD_ATOMIC:
1525 return MLX5_QP_OPTPAR_SRA_MAX;
1526 case IB_QP_ALT_PATH:
1527 return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
1528 case IB_QP_MIN_RNR_TIMER:
1529 return MLX5_QP_OPTPAR_RNR_TIMEOUT;
1532 case IB_QP_MAX_DEST_RD_ATOMIC:
1533 return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
1534 MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
1535 case IB_QP_PATH_MIG_STATE:
1536 return MLX5_QP_OPTPAR_PM_STATE;
1539 case IB_QP_DEST_QPN:
1545 static int ib_mask_to_mlx5_opt(int ib_mask)
1550 for (i = 0; i < 8 * sizeof(int); i++) {
1551 if ((1 << i) & ib_mask)
1552 result |= ib_nr_to_mlx5_nr(1 << i);
1558 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
1559 const struct ib_qp_attr *attr, int attr_mask,
1560 enum ib_qp_state cur_state, enum ib_qp_state new_state)
1562 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1563 struct mlx5_ib_qp *qp = to_mqp(ibqp);
1564 struct mlx5_ib_cq *send_cq, *recv_cq;
1565 struct mlx5_qp_context *context;
1566 struct mlx5_modify_qp_mbox_in *in;
1567 struct mlx5_ib_pd *pd;
1568 enum mlx5_qp_state mlx5_cur, mlx5_new;
1569 enum mlx5_qp_optpar optpar;
1574 in = kzalloc(sizeof(*in), GFP_KERNEL);
1579 err = to_mlx5_st(ibqp->qp_type);
1583 context->flags = cpu_to_be32(err << 16);
1585 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
1586 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1588 switch (attr->path_mig_state) {
1589 case IB_MIG_MIGRATED:
1590 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
1593 context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
1596 context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
1601 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) {
1602 context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
1603 } else if (ibqp->qp_type == IB_QPT_UD ||
1604 ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
1605 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1606 } else if (attr_mask & IB_QP_PATH_MTU) {
1607 if (attr->path_mtu < IB_MTU_256 ||
1608 attr->path_mtu > IB_MTU_4096) {
1609 mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
1613 context->mtu_msgmax = (attr->path_mtu << 5) |
1614 (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
1617 if (attr_mask & IB_QP_DEST_QPN)
1618 context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
1620 if (attr_mask & IB_QP_PKEY_INDEX)
1621 context->pri_path.pkey_index = attr->pkey_index;
1623 /* todo implement counter_index functionality */
1625 if (is_sqp(ibqp->qp_type))
1626 context->pri_path.port = qp->port;
1628 if (attr_mask & IB_QP_PORT)
1629 context->pri_path.port = attr->port_num;
1631 if (attr_mask & IB_QP_AV) {
1632 err = mlx5_set_path(dev, &attr->ah_attr, &context->pri_path,
1633 attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
1634 attr_mask, 0, attr);
1639 if (attr_mask & IB_QP_TIMEOUT)
1640 context->pri_path.ackto_lt |= attr->timeout << 3;
1642 if (attr_mask & IB_QP_ALT_PATH) {
1643 err = mlx5_set_path(dev, &attr->alt_ah_attr, &context->alt_path,
1644 attr->alt_port_num, attr_mask, 0, attr);
1650 get_cqs(qp, &send_cq, &recv_cq);
1652 context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
1653 context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
1654 context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
1655 context->params1 = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
1657 if (attr_mask & IB_QP_RNR_RETRY)
1658 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1660 if (attr_mask & IB_QP_RETRY_CNT)
1661 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1663 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1664 if (attr->max_rd_atomic)
1666 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1669 if (attr_mask & IB_QP_SQ_PSN)
1670 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1672 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1673 if (attr->max_dest_rd_atomic)
1675 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1678 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
1679 context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
1681 if (attr_mask & IB_QP_MIN_RNR_TIMER)
1682 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1684 if (attr_mask & IB_QP_RQ_PSN)
1685 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1687 if (attr_mask & IB_QP_QKEY)
1688 context->qkey = cpu_to_be32(attr->qkey);
1690 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1691 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1693 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
1694 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1699 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1700 context->sq_crq_size |= cpu_to_be16(1 << 4);
1703 mlx5_cur = to_mlx5_state(cur_state);
1704 mlx5_new = to_mlx5_state(new_state);
1705 mlx5_st = to_mlx5_st(ibqp->qp_type);
1709 /* If moving to a reset or error state, we must disable page faults on
1710 * this QP and flush all current page faults. Otherwise a stale page
1711 * fault may attempt to work on this QP after it is reset and moved
1712 * again to RTS, and may cause the driver and the device to get out of
1714 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1715 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1716 mlx5_ib_qp_disable_pagefaults(qp);
1718 optpar = ib_mask_to_mlx5_opt(attr_mask);
1719 optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
1720 in->optparam = cpu_to_be32(optpar);
1721 err = mlx5_core_qp_modify(dev->mdev, to_mlx5_state(cur_state),
1722 to_mlx5_state(new_state), in, sqd_event,
1727 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1728 mlx5_ib_qp_enable_pagefaults(qp);
1730 qp->state = new_state;
1732 if (attr_mask & IB_QP_ACCESS_FLAGS)
1733 qp->atomic_rd_en = attr->qp_access_flags;
1734 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1735 qp->resp_depth = attr->max_dest_rd_atomic;
1736 if (attr_mask & IB_QP_PORT)
1737 qp->port = attr->port_num;
1738 if (attr_mask & IB_QP_ALT_PATH)
1739 qp->alt_port = attr->alt_port_num;
1742 * If we moved a kernel QP to RESET, clean up all old CQ
1743 * entries and reinitialize the QP.
1745 if (new_state == IB_QPS_RESET && !ibqp->uobject) {
1746 mlx5_ib_cq_clean(recv_cq, qp->mqp.qpn,
1747 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1748 if (send_cq != recv_cq)
1749 mlx5_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1755 qp->sq.cur_post = 0;
1756 qp->sq.last_poll = 0;
1757 qp->db.db[MLX5_RCV_DBR] = 0;
1758 qp->db.db[MLX5_SND_DBR] = 0;
1766 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1767 int attr_mask, struct ib_udata *udata)
1769 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
1770 struct mlx5_ib_qp *qp = to_mqp(ibqp);
1771 enum ib_qp_state cur_state, new_state;
1775 mutex_lock(&qp->mutex);
1777 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1778 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1780 if (ibqp->qp_type != MLX5_IB_QPT_REG_UMR &&
1781 !ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask,
1782 IB_LINK_LAYER_UNSPECIFIED))
1785 if ((attr_mask & IB_QP_PORT) &&
1786 (attr->port_num == 0 ||
1787 attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports)))
1790 if (attr_mask & IB_QP_PKEY_INDEX) {
1791 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1792 if (attr->pkey_index >=
1793 dev->mdev->port_caps[port - 1].pkey_table_len)
1797 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1798 attr->max_rd_atomic >
1799 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp)))
1802 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1803 attr->max_dest_rd_atomic >
1804 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp)))
1807 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1812 err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1815 mutex_unlock(&qp->mutex);
1819 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
1821 struct mlx5_ib_cq *cq;
1824 cur = wq->head - wq->tail;
1825 if (likely(cur + nreq < wq->max_post))
1829 spin_lock(&cq->lock);
1830 cur = wq->head - wq->tail;
1831 spin_unlock(&cq->lock);
1833 return cur + nreq >= wq->max_post;
1836 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
1837 u64 remote_addr, u32 rkey)
1839 rseg->raddr = cpu_to_be64(remote_addr);
1840 rseg->rkey = cpu_to_be32(rkey);
1844 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
1845 struct ib_send_wr *wr)
1847 memcpy(&dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof(struct mlx5_av));
1848 dseg->av.dqp_dct = cpu_to_be32(wr->wr.ud.remote_qpn | MLX5_EXTENDED_UD_AV);
1849 dseg->av.key.qkey.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
1852 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
1854 dseg->byte_count = cpu_to_be32(sg->length);
1855 dseg->lkey = cpu_to_be32(sg->lkey);
1856 dseg->addr = cpu_to_be64(sg->addr);
1859 static __be16 get_klm_octo(int npages)
1861 return cpu_to_be16(ALIGN(npages, 8) / 2);
1864 static __be64 frwr_mkey_mask(void)
1868 result = MLX5_MKEY_MASK_LEN |
1869 MLX5_MKEY_MASK_PAGE_SIZE |
1870 MLX5_MKEY_MASK_START_ADDR |
1871 MLX5_MKEY_MASK_EN_RINVAL |
1872 MLX5_MKEY_MASK_KEY |
1878 MLX5_MKEY_MASK_SMALL_FENCE |
1879 MLX5_MKEY_MASK_FREE;
1881 return cpu_to_be64(result);
1884 static __be64 sig_mkey_mask(void)
1888 result = MLX5_MKEY_MASK_LEN |
1889 MLX5_MKEY_MASK_PAGE_SIZE |
1890 MLX5_MKEY_MASK_START_ADDR |
1891 MLX5_MKEY_MASK_EN_SIGERR |
1892 MLX5_MKEY_MASK_EN_RINVAL |
1893 MLX5_MKEY_MASK_KEY |
1898 MLX5_MKEY_MASK_SMALL_FENCE |
1899 MLX5_MKEY_MASK_FREE |
1900 MLX5_MKEY_MASK_BSF_EN;
1902 return cpu_to_be64(result);
1905 static void set_frwr_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1906 struct ib_send_wr *wr, int li)
1908 memset(umr, 0, sizeof(*umr));
1911 umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
1912 umr->flags = 1 << 7;
1916 umr->flags = (1 << 5); /* fail if not free */
1917 umr->klm_octowords = get_klm_octo(wr->wr.fast_reg.page_list_len);
1918 umr->mkey_mask = frwr_mkey_mask();
1921 static __be64 get_umr_reg_mr_mask(void)
1925 result = MLX5_MKEY_MASK_LEN |
1926 MLX5_MKEY_MASK_PAGE_SIZE |
1927 MLX5_MKEY_MASK_START_ADDR |
1931 MLX5_MKEY_MASK_KEY |
1935 MLX5_MKEY_MASK_FREE;
1937 return cpu_to_be64(result);
1940 static __be64 get_umr_unreg_mr_mask(void)
1944 result = MLX5_MKEY_MASK_FREE;
1946 return cpu_to_be64(result);
1949 static __be64 get_umr_update_mtt_mask(void)
1953 result = MLX5_MKEY_MASK_FREE;
1955 return cpu_to_be64(result);
1958 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
1959 struct ib_send_wr *wr)
1961 struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
1963 memset(umr, 0, sizeof(*umr));
1965 if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
1966 umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
1968 umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
1970 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
1971 umr->klm_octowords = get_klm_octo(umrwr->npages);
1972 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
1973 umr->mkey_mask = get_umr_update_mtt_mask();
1974 umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
1975 umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
1977 umr->mkey_mask = get_umr_reg_mr_mask();
1980 umr->mkey_mask = get_umr_unreg_mr_mask();
1984 umr->flags |= MLX5_UMR_INLINE;
1987 static u8 get_umr_flags(int acc)
1989 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC : 0) |
1990 (acc & IB_ACCESS_REMOTE_WRITE ? MLX5_PERM_REMOTE_WRITE : 0) |
1991 (acc & IB_ACCESS_REMOTE_READ ? MLX5_PERM_REMOTE_READ : 0) |
1992 (acc & IB_ACCESS_LOCAL_WRITE ? MLX5_PERM_LOCAL_WRITE : 0) |
1993 MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
1996 static void set_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr,
1999 memset(seg, 0, sizeof(*seg));
2001 seg->status = MLX5_MKEY_STATUS_FREE;
2005 seg->flags = get_umr_flags(wr->wr.fast_reg.access_flags) |
2006 MLX5_ACCESS_MODE_MTT;
2007 *writ = seg->flags & (MLX5_PERM_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE);
2008 seg->qpn_mkey7_0 = cpu_to_be32((wr->wr.fast_reg.rkey & 0xff) | 0xffffff00);
2009 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
2010 seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
2011 seg->len = cpu_to_be64(wr->wr.fast_reg.length);
2012 seg->xlt_oct_size = cpu_to_be32((wr->wr.fast_reg.page_list_len + 1) / 2);
2013 seg->log2_page_size = wr->wr.fast_reg.page_shift;
2016 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
2018 struct mlx5_umr_wr *umrwr = (struct mlx5_umr_wr *)&wr->wr.fast_reg;
2020 memset(seg, 0, sizeof(*seg));
2021 if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
2022 seg->status = MLX5_MKEY_STATUS_FREE;
2026 seg->flags = convert_access(umrwr->access_flags);
2027 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
2028 seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
2029 seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
2031 seg->len = cpu_to_be64(umrwr->length);
2032 seg->log2_page_size = umrwr->page_shift;
2033 seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
2034 mlx5_mkey_variant(umrwr->mkey));
2037 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
2038 struct ib_send_wr *wr,
2039 struct mlx5_core_dev *mdev,
2040 struct mlx5_ib_pd *pd,
2043 struct mlx5_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2044 u64 *page_list = wr->wr.fast_reg.page_list->page_list;
2045 u64 perm = MLX5_EN_RD | (writ ? MLX5_EN_WR : 0);
2048 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++)
2049 mfrpl->mapped_page_list[i] = cpu_to_be64(page_list[i] | perm);
2050 dseg->addr = cpu_to_be64(mfrpl->map);
2051 dseg->byte_count = cpu_to_be32(ALIGN(sizeof(u64) * wr->wr.fast_reg.page_list_len, 64));
2052 dseg->lkey = cpu_to_be32(pd->pa_lkey);
2055 static __be32 send_ieth(struct ib_send_wr *wr)
2057 switch (wr->opcode) {
2058 case IB_WR_SEND_WITH_IMM:
2059 case IB_WR_RDMA_WRITE_WITH_IMM:
2060 return wr->ex.imm_data;
2062 case IB_WR_SEND_WITH_INV:
2063 return cpu_to_be32(wr->ex.invalidate_rkey);
2070 static u8 calc_sig(void *wqe, int size)
2076 for (i = 0; i < size; i++)
2082 static u8 wq_sig(void *wqe)
2084 return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
2087 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
2090 struct mlx5_wqe_inline_seg *seg;
2091 void *qend = qp->sq.qend;
2099 wqe += sizeof(*seg);
2100 for (i = 0; i < wr->num_sge; i++) {
2101 addr = (void *)(unsigned long)(wr->sg_list[i].addr);
2102 len = wr->sg_list[i].length;
2105 if (unlikely(inl > qp->max_inline_data))
2108 if (unlikely(wqe + len > qend)) {
2110 memcpy(wqe, addr, copy);
2113 wqe = mlx5_get_send_wqe(qp, 0);
2115 memcpy(wqe, addr, len);
2119 seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
2121 *sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
2126 static u16 prot_field_size(enum ib_signature_type type)
2129 case IB_SIG_TYPE_T10_DIF:
2130 return MLX5_DIF_SIZE;
2136 static u8 bs_selector(int block_size)
2138 switch (block_size) {
2139 case 512: return 0x1;
2140 case 520: return 0x2;
2141 case 4096: return 0x3;
2142 case 4160: return 0x4;
2143 case 1073741824: return 0x5;
2148 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
2149 struct mlx5_bsf_inl *inl)
2151 /* Valid inline section and allow BSF refresh */
2152 inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
2153 MLX5_BSF_REFRESH_DIF);
2154 inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
2155 inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
2156 /* repeating block */
2157 inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
2158 inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
2159 MLX5_DIF_CRC : MLX5_DIF_IPCS;
2161 if (domain->sig.dif.ref_remap)
2162 inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
2164 if (domain->sig.dif.app_escape) {
2165 if (domain->sig.dif.ref_escape)
2166 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
2168 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
2171 inl->dif_app_bitmask_check =
2172 cpu_to_be16(domain->sig.dif.apptag_check_mask);
2175 static int mlx5_set_bsf(struct ib_mr *sig_mr,
2176 struct ib_sig_attrs *sig_attrs,
2177 struct mlx5_bsf *bsf, u32 data_size)
2179 struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
2180 struct mlx5_bsf_basic *basic = &bsf->basic;
2181 struct ib_sig_domain *mem = &sig_attrs->mem;
2182 struct ib_sig_domain *wire = &sig_attrs->wire;
2184 memset(bsf, 0, sizeof(*bsf));
2186 /* Basic + Extended + Inline */
2187 basic->bsf_size_sbs = 1 << 7;
2188 /* Input domain check byte mask */
2189 basic->check_byte_mask = sig_attrs->check_mask;
2190 basic->raw_data_size = cpu_to_be32(data_size);
2193 switch (sig_attrs->mem.sig_type) {
2194 case IB_SIG_TYPE_NONE:
2196 case IB_SIG_TYPE_T10_DIF:
2197 basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
2198 basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
2199 mlx5_fill_inl_bsf(mem, &bsf->m_inl);
2206 switch (sig_attrs->wire.sig_type) {
2207 case IB_SIG_TYPE_NONE:
2209 case IB_SIG_TYPE_T10_DIF:
2210 if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
2211 mem->sig_type == wire->sig_type) {
2212 /* Same block structure */
2213 basic->bsf_size_sbs |= 1 << 4;
2214 if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
2215 basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
2216 if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
2217 basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
2218 if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
2219 basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
2221 basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
2223 basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
2224 mlx5_fill_inl_bsf(wire, &bsf->w_inl);
2233 static int set_sig_data_segment(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2234 void **seg, int *size)
2236 struct ib_sig_attrs *sig_attrs = wr->wr.sig_handover.sig_attrs;
2237 struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2238 struct mlx5_bsf *bsf;
2239 u32 data_len = wr->sg_list->length;
2240 u32 data_key = wr->sg_list->lkey;
2241 u64 data_va = wr->sg_list->addr;
2245 if (!wr->wr.sig_handover.prot ||
2246 (data_key == wr->wr.sig_handover.prot->lkey &&
2247 data_va == wr->wr.sig_handover.prot->addr &&
2248 data_len == wr->wr.sig_handover.prot->length)) {
2250 * Source domain doesn't contain signature information
2251 * or data and protection are interleaved in memory.
2252 * So need construct:
2253 * ------------------
2255 * ------------------
2257 * ------------------
2259 struct mlx5_klm *data_klm = *seg;
2261 data_klm->bcount = cpu_to_be32(data_len);
2262 data_klm->key = cpu_to_be32(data_key);
2263 data_klm->va = cpu_to_be64(data_va);
2264 wqe_size = ALIGN(sizeof(*data_klm), 64);
2267 * Source domain contains signature information
2268 * So need construct a strided block format:
2269 * ---------------------------
2270 * | stride_block_ctrl |
2271 * ---------------------------
2273 * ---------------------------
2275 * ---------------------------
2277 * ---------------------------
2279 struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
2280 struct mlx5_stride_block_entry *data_sentry;
2281 struct mlx5_stride_block_entry *prot_sentry;
2282 u32 prot_key = wr->wr.sig_handover.prot->lkey;
2283 u64 prot_va = wr->wr.sig_handover.prot->addr;
2284 u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
2288 data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
2289 prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
2291 prot_size = prot_field_size(sig_attrs->mem.sig_type);
2293 pr_err("Bad block size given: %u\n", block_size);
2296 sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
2298 sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
2299 sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
2300 sblock_ctrl->num_entries = cpu_to_be16(2);
2302 data_sentry->bcount = cpu_to_be16(block_size);
2303 data_sentry->key = cpu_to_be32(data_key);
2304 data_sentry->va = cpu_to_be64(data_va);
2305 data_sentry->stride = cpu_to_be16(block_size);
2307 prot_sentry->bcount = cpu_to_be16(prot_size);
2308 prot_sentry->key = cpu_to_be32(prot_key);
2309 prot_sentry->va = cpu_to_be64(prot_va);
2310 prot_sentry->stride = cpu_to_be16(prot_size);
2312 wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
2313 sizeof(*prot_sentry), 64);
2317 *size += wqe_size / 16;
2318 if (unlikely((*seg == qp->sq.qend)))
2319 *seg = mlx5_get_send_wqe(qp, 0);
2322 ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
2326 *seg += sizeof(*bsf);
2327 *size += sizeof(*bsf) / 16;
2328 if (unlikely((*seg == qp->sq.qend)))
2329 *seg = mlx5_get_send_wqe(qp, 0);
2334 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
2335 struct ib_send_wr *wr, u32 nelements,
2336 u32 length, u32 pdn)
2338 struct ib_mr *sig_mr = wr->wr.sig_handover.sig_mr;
2339 u32 sig_key = sig_mr->rkey;
2340 u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
2342 memset(seg, 0, sizeof(*seg));
2344 seg->flags = get_umr_flags(wr->wr.sig_handover.access_flags) |
2345 MLX5_ACCESS_MODE_KLM;
2346 seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
2347 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
2348 MLX5_MKEY_BSF_EN | pdn);
2349 seg->len = cpu_to_be64(length);
2350 seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
2351 seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
2354 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
2355 struct ib_send_wr *wr, u32 nelements)
2357 memset(umr, 0, sizeof(*umr));
2359 umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
2360 umr->klm_octowords = get_klm_octo(nelements);
2361 umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
2362 umr->mkey_mask = sig_mkey_mask();
2366 static int set_sig_umr_wr(struct ib_send_wr *wr, struct mlx5_ib_qp *qp,
2367 void **seg, int *size)
2369 struct mlx5_ib_mr *sig_mr = to_mmr(wr->wr.sig_handover.sig_mr);
2370 u32 pdn = get_pd(qp)->pdn;
2372 int region_len, ret;
2374 if (unlikely(wr->num_sge != 1) ||
2375 unlikely(wr->wr.sig_handover.access_flags &
2376 IB_ACCESS_REMOTE_ATOMIC) ||
2377 unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
2378 unlikely(!sig_mr->sig->sig_status_checked))
2381 /* length of the protected region, data + protection */
2382 region_len = wr->sg_list->length;
2383 if (wr->wr.sig_handover.prot &&
2384 (wr->wr.sig_handover.prot->lkey != wr->sg_list->lkey ||
2385 wr->wr.sig_handover.prot->addr != wr->sg_list->addr ||
2386 wr->wr.sig_handover.prot->length != wr->sg_list->length))
2387 region_len += wr->wr.sig_handover.prot->length;
2390 * KLM octoword size - if protection was provided
2391 * then we use strided block format (3 octowords),
2392 * else we use single KLM (1 octoword)
2394 klm_oct_size = wr->wr.sig_handover.prot ? 3 : 1;
2396 set_sig_umr_segment(*seg, wr, klm_oct_size);
2397 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2398 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2399 if (unlikely((*seg == qp->sq.qend)))
2400 *seg = mlx5_get_send_wqe(qp, 0);
2402 set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
2403 *seg += sizeof(struct mlx5_mkey_seg);
2404 *size += sizeof(struct mlx5_mkey_seg) / 16;
2405 if (unlikely((*seg == qp->sq.qend)))
2406 *seg = mlx5_get_send_wqe(qp, 0);
2408 ret = set_sig_data_segment(wr, qp, seg, size);
2412 sig_mr->sig->sig_status_checked = false;
2416 static int set_psv_wr(struct ib_sig_domain *domain,
2417 u32 psv_idx, void **seg, int *size)
2419 struct mlx5_seg_set_psv *psv_seg = *seg;
2421 memset(psv_seg, 0, sizeof(*psv_seg));
2422 psv_seg->psv_num = cpu_to_be32(psv_idx);
2423 switch (domain->sig_type) {
2424 case IB_SIG_TYPE_NONE:
2426 case IB_SIG_TYPE_T10_DIF:
2427 psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
2428 domain->sig.dif.app_tag);
2429 psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
2432 pr_err("Bad signature type given.\n");
2436 *seg += sizeof(*psv_seg);
2437 *size += sizeof(*psv_seg) / 16;
2442 static int set_frwr_li_wr(void **seg, struct ib_send_wr *wr, int *size,
2443 struct mlx5_core_dev *mdev, struct mlx5_ib_pd *pd, struct mlx5_ib_qp *qp)
2448 li = wr->opcode == IB_WR_LOCAL_INV ? 1 : 0;
2449 if (unlikely(wr->send_flags & IB_SEND_INLINE))
2452 set_frwr_umr_segment(*seg, wr, li);
2453 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2454 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2455 if (unlikely((*seg == qp->sq.qend)))
2456 *seg = mlx5_get_send_wqe(qp, 0);
2457 set_mkey_segment(*seg, wr, li, &writ);
2458 *seg += sizeof(struct mlx5_mkey_seg);
2459 *size += sizeof(struct mlx5_mkey_seg) / 16;
2460 if (unlikely((*seg == qp->sq.qend)))
2461 *seg = mlx5_get_send_wqe(qp, 0);
2463 if (unlikely(wr->wr.fast_reg.page_list_len >
2464 wr->wr.fast_reg.page_list->max_page_list_len))
2467 set_frwr_pages(*seg, wr, mdev, pd, writ);
2468 *seg += sizeof(struct mlx5_wqe_data_seg);
2469 *size += (sizeof(struct mlx5_wqe_data_seg) / 16);
2474 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
2480 pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
2481 for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
2482 if ((i & 0xf) == 0) {
2483 void *buf = mlx5_get_send_wqe(qp, tidx);
2484 tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
2488 pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
2489 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
2490 be32_to_cpu(p[j + 3]));
2494 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
2495 unsigned bytecnt, struct mlx5_ib_qp *qp)
2497 while (bytecnt > 0) {
2498 __iowrite64_copy(dst++, src++, 8);
2499 __iowrite64_copy(dst++, src++, 8);
2500 __iowrite64_copy(dst++, src++, 8);
2501 __iowrite64_copy(dst++, src++, 8);
2502 __iowrite64_copy(dst++, src++, 8);
2503 __iowrite64_copy(dst++, src++, 8);
2504 __iowrite64_copy(dst++, src++, 8);
2505 __iowrite64_copy(dst++, src++, 8);
2507 if (unlikely(src == qp->sq.qend))
2508 src = mlx5_get_send_wqe(qp, 0);
2512 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
2514 if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
2515 wr->send_flags & IB_SEND_FENCE))
2516 return MLX5_FENCE_MODE_STRONG_ORDERING;
2518 if (unlikely(fence)) {
2519 if (wr->send_flags & IB_SEND_FENCE)
2520 return MLX5_FENCE_MODE_SMALL_AND_FENCE;
2529 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
2530 struct mlx5_wqe_ctrl_seg **ctrl,
2531 struct ib_send_wr *wr, unsigned *idx,
2532 int *size, int nreq)
2536 if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq))) {
2541 *idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
2542 *seg = mlx5_get_send_wqe(qp, *idx);
2544 *(uint32_t *)(*seg + 8) = 0;
2545 (*ctrl)->imm = send_ieth(wr);
2546 (*ctrl)->fm_ce_se = qp->sq_signal_bits |
2547 (wr->send_flags & IB_SEND_SIGNALED ?
2548 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
2549 (wr->send_flags & IB_SEND_SOLICITED ?
2550 MLX5_WQE_CTRL_SOLICITED : 0);
2552 *seg += sizeof(**ctrl);
2553 *size = sizeof(**ctrl) / 16;
2558 static void finish_wqe(struct mlx5_ib_qp *qp,
2559 struct mlx5_wqe_ctrl_seg *ctrl,
2560 u8 size, unsigned idx, u64 wr_id,
2561 int nreq, u8 fence, u8 next_fence,
2566 ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
2567 mlx5_opcode | ((u32)opmod << 24));
2568 ctrl->qpn_ds = cpu_to_be32(size | (qp->mqp.qpn << 8));
2569 ctrl->fm_ce_se |= fence;
2570 qp->fm_cache = next_fence;
2571 if (unlikely(qp->wq_sig))
2572 ctrl->signature = wq_sig(ctrl);
2574 qp->sq.wrid[idx] = wr_id;
2575 qp->sq.w_list[idx].opcode = mlx5_opcode;
2576 qp->sq.wqe_head[idx] = qp->sq.head + nreq;
2577 qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
2578 qp->sq.w_list[idx].next = qp->sq.cur_post;
2582 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2583 struct ib_send_wr **bad_wr)
2585 struct mlx5_wqe_ctrl_seg *ctrl = NULL; /* compiler warning */
2586 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2587 struct mlx5_core_dev *mdev = dev->mdev;
2588 struct mlx5_ib_qp *qp = to_mqp(ibqp);
2589 struct mlx5_ib_mr *mr;
2590 struct mlx5_wqe_data_seg *dpseg;
2591 struct mlx5_wqe_xrc_seg *xrc;
2592 struct mlx5_bf *bf = qp->bf;
2593 int uninitialized_var(size);
2594 void *qend = qp->sq.qend;
2595 unsigned long flags;
2606 spin_lock_irqsave(&qp->sq.lock, flags);
2608 for (nreq = 0; wr; nreq++, wr = wr->next) {
2609 if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
2610 mlx5_ib_warn(dev, "\n");
2616 fence = qp->fm_cache;
2617 num_sge = wr->num_sge;
2618 if (unlikely(num_sge > qp->sq.max_gs)) {
2619 mlx5_ib_warn(dev, "\n");
2625 err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
2627 mlx5_ib_warn(dev, "\n");
2633 switch (ibqp->qp_type) {
2634 case IB_QPT_XRC_INI:
2636 xrc->xrc_srqn = htonl(wr->xrc_remote_srq_num);
2637 seg += sizeof(*xrc);
2638 size += sizeof(*xrc) / 16;
2641 switch (wr->opcode) {
2642 case IB_WR_RDMA_READ:
2643 case IB_WR_RDMA_WRITE:
2644 case IB_WR_RDMA_WRITE_WITH_IMM:
2645 set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2647 seg += sizeof(struct mlx5_wqe_raddr_seg);
2648 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2651 case IB_WR_ATOMIC_CMP_AND_SWP:
2652 case IB_WR_ATOMIC_FETCH_AND_ADD:
2653 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2654 mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
2659 case IB_WR_LOCAL_INV:
2660 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2661 qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
2662 ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
2663 err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2665 mlx5_ib_warn(dev, "\n");
2672 case IB_WR_FAST_REG_MR:
2673 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2674 qp->sq.wr_data[idx] = IB_WR_FAST_REG_MR;
2675 ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2676 err = set_frwr_li_wr(&seg, wr, &size, mdev, to_mpd(ibqp->pd), qp);
2678 mlx5_ib_warn(dev, "\n");
2685 case IB_WR_REG_SIG_MR:
2686 qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
2687 mr = to_mmr(wr->wr.sig_handover.sig_mr);
2689 ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
2690 err = set_sig_umr_wr(wr, qp, &seg, &size);
2692 mlx5_ib_warn(dev, "\n");
2697 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2698 nreq, get_fence(fence, wr),
2699 next_fence, MLX5_OPCODE_UMR);
2701 * SET_PSV WQEs are not signaled and solicited
2704 wr->send_flags &= ~IB_SEND_SIGNALED;
2705 wr->send_flags |= IB_SEND_SOLICITED;
2706 err = begin_wqe(qp, &seg, &ctrl, wr,
2709 mlx5_ib_warn(dev, "\n");
2715 err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->mem,
2716 mr->sig->psv_memory.psv_idx, &seg,
2719 mlx5_ib_warn(dev, "\n");
2724 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2725 nreq, get_fence(fence, wr),
2726 next_fence, MLX5_OPCODE_SET_PSV);
2727 err = begin_wqe(qp, &seg, &ctrl, wr,
2730 mlx5_ib_warn(dev, "\n");
2736 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
2737 err = set_psv_wr(&wr->wr.sig_handover.sig_attrs->wire,
2738 mr->sig->psv_wire.psv_idx, &seg,
2741 mlx5_ib_warn(dev, "\n");
2746 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
2747 nreq, get_fence(fence, wr),
2748 next_fence, MLX5_OPCODE_SET_PSV);
2758 switch (wr->opcode) {
2759 case IB_WR_RDMA_WRITE:
2760 case IB_WR_RDMA_WRITE_WITH_IMM:
2761 set_raddr_seg(seg, wr->wr.rdma.remote_addr,
2763 seg += sizeof(struct mlx5_wqe_raddr_seg);
2764 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
2775 set_datagram_seg(seg, wr);
2776 seg += sizeof(struct mlx5_wqe_datagram_seg);
2777 size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
2778 if (unlikely((seg == qend)))
2779 seg = mlx5_get_send_wqe(qp, 0);
2782 case MLX5_IB_QPT_REG_UMR:
2783 if (wr->opcode != MLX5_IB_WR_UMR) {
2785 mlx5_ib_warn(dev, "bad opcode\n");
2788 qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
2789 ctrl->imm = cpu_to_be32(wr->wr.fast_reg.rkey);
2790 set_reg_umr_segment(seg, wr);
2791 seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
2792 size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
2793 if (unlikely((seg == qend)))
2794 seg = mlx5_get_send_wqe(qp, 0);
2795 set_reg_mkey_segment(seg, wr);
2796 seg += sizeof(struct mlx5_mkey_seg);
2797 size += sizeof(struct mlx5_mkey_seg) / 16;
2798 if (unlikely((seg == qend)))
2799 seg = mlx5_get_send_wqe(qp, 0);
2806 if (wr->send_flags & IB_SEND_INLINE && num_sge) {
2807 int uninitialized_var(sz);
2809 err = set_data_inl_seg(qp, wr, seg, &sz);
2810 if (unlikely(err)) {
2811 mlx5_ib_warn(dev, "\n");
2819 for (i = 0; i < num_sge; i++) {
2820 if (unlikely(dpseg == qend)) {
2821 seg = mlx5_get_send_wqe(qp, 0);
2824 if (likely(wr->sg_list[i].length)) {
2825 set_data_ptr_seg(dpseg, wr->sg_list + i);
2826 size += sizeof(struct mlx5_wqe_data_seg) / 16;
2832 finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
2833 get_fence(fence, wr), next_fence,
2834 mlx5_ib_opcode[wr->opcode]);
2837 dump_wqe(qp, idx, size);
2842 qp->sq.head += nreq;
2844 /* Make sure that descriptors are written before
2845 * updating doorbell record and ringing the doorbell
2849 qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
2851 /* Make sure doorbell record is visible to the HCA before
2852 * we hit doorbell */
2856 spin_lock(&bf->lock);
2858 __acquire(&bf->lock);
2861 if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
2862 mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
2865 mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
2866 MLX5_GET_DOORBELL_LOCK(&bf->lock32));
2867 /* Make sure doorbells don't leak out of SQ spinlock
2868 * and reach the HCA out of order.
2872 bf->offset ^= bf->buf_size;
2874 spin_unlock(&bf->lock);
2876 __release(&bf->lock);
2879 spin_unlock_irqrestore(&qp->sq.lock, flags);
2884 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
2886 sig->signature = calc_sig(sig, size);
2889 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2890 struct ib_recv_wr **bad_wr)
2892 struct mlx5_ib_qp *qp = to_mqp(ibqp);
2893 struct mlx5_wqe_data_seg *scat;
2894 struct mlx5_rwqe_sig *sig;
2895 unsigned long flags;
2901 spin_lock_irqsave(&qp->rq.lock, flags);
2903 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2905 for (nreq = 0; wr; nreq++, wr = wr->next) {
2906 if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2912 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2918 scat = get_recv_wqe(qp, ind);
2922 for (i = 0; i < wr->num_sge; i++)
2923 set_data_ptr_seg(scat + i, wr->sg_list + i);
2925 if (i < qp->rq.max_gs) {
2926 scat[i].byte_count = 0;
2927 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
2932 sig = (struct mlx5_rwqe_sig *)scat;
2933 set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
2936 qp->rq.wrid[ind] = wr->wr_id;
2938 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2943 qp->rq.head += nreq;
2945 /* Make sure that descriptors are written before
2950 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2953 spin_unlock_irqrestore(&qp->rq.lock, flags);
2958 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
2960 switch (mlx5_state) {
2961 case MLX5_QP_STATE_RST: return IB_QPS_RESET;
2962 case MLX5_QP_STATE_INIT: return IB_QPS_INIT;
2963 case MLX5_QP_STATE_RTR: return IB_QPS_RTR;
2964 case MLX5_QP_STATE_RTS: return IB_QPS_RTS;
2965 case MLX5_QP_STATE_SQ_DRAINING:
2966 case MLX5_QP_STATE_SQD: return IB_QPS_SQD;
2967 case MLX5_QP_STATE_SQER: return IB_QPS_SQE;
2968 case MLX5_QP_STATE_ERR: return IB_QPS_ERR;
2973 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
2975 switch (mlx5_mig_state) {
2976 case MLX5_QP_PM_ARMED: return IB_MIG_ARMED;
2977 case MLX5_QP_PM_REARM: return IB_MIG_REARM;
2978 case MLX5_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
2983 static int to_ib_qp_access_flags(int mlx5_flags)
2987 if (mlx5_flags & MLX5_QP_BIT_RRE)
2988 ib_flags |= IB_ACCESS_REMOTE_READ;
2989 if (mlx5_flags & MLX5_QP_BIT_RWE)
2990 ib_flags |= IB_ACCESS_REMOTE_WRITE;
2991 if (mlx5_flags & MLX5_QP_BIT_RAE)
2992 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2997 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2998 struct mlx5_qp_path *path)
3000 struct mlx5_core_dev *dev = ibdev->mdev;
3002 memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
3003 ib_ah_attr->port_num = path->port;
3005 if (ib_ah_attr->port_num == 0 ||
3006 ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
3009 ib_ah_attr->sl = path->sl & 0xf;
3011 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
3012 ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
3013 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
3014 ib_ah_attr->ah_flags = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
3015 if (ib_ah_attr->ah_flags) {
3016 ib_ah_attr->grh.sgid_index = path->mgid_index;
3017 ib_ah_attr->grh.hop_limit = path->hop_limit;
3018 ib_ah_attr->grh.traffic_class =
3019 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3020 ib_ah_attr->grh.flow_label =
3021 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3022 memcpy(ib_ah_attr->grh.dgid.raw,
3023 path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
3027 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3028 struct ib_qp_init_attr *qp_init_attr)
3030 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3031 struct mlx5_ib_qp *qp = to_mqp(ibqp);
3032 struct mlx5_query_qp_mbox_out *outb;
3033 struct mlx5_qp_context *context;
3037 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
3039 * Wait for any outstanding page faults, in case the user frees memory
3040 * based upon this query's result.
3042 flush_workqueue(mlx5_ib_page_fault_wq);
3045 mutex_lock(&qp->mutex);
3046 outb = kzalloc(sizeof(*outb), GFP_KERNEL);
3051 context = &outb->ctx;
3052 err = mlx5_core_qp_query(dev->mdev, &qp->mqp, outb, sizeof(*outb));
3056 mlx5_state = be32_to_cpu(context->flags) >> 28;
3058 qp->state = to_ib_qp_state(mlx5_state);
3059 qp_attr->qp_state = qp->state;
3060 qp_attr->path_mtu = context->mtu_msgmax >> 5;
3061 qp_attr->path_mig_state =
3062 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
3063 qp_attr->qkey = be32_to_cpu(context->qkey);
3064 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
3065 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
3066 qp_attr->dest_qp_num = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
3067 qp_attr->qp_access_flags =
3068 to_ib_qp_access_flags(be32_to_cpu(context->params2));
3070 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3071 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
3072 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
3073 qp_attr->alt_pkey_index = context->alt_path.pkey_index & 0x7f;
3074 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
3077 qp_attr->pkey_index = context->pri_path.pkey_index & 0x7f;
3078 qp_attr->port_num = context->pri_path.port;
3080 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3081 qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
3083 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
3085 qp_attr->max_dest_rd_atomic =
3086 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
3087 qp_attr->min_rnr_timer =
3088 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
3089 qp_attr->timeout = context->pri_path.ackto_lt >> 3;
3090 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
3091 qp_attr->rnr_retry = (be32_to_cpu(context->params1) >> 13) & 0x7;
3092 qp_attr->alt_timeout = context->alt_path.ackto_lt >> 3;
3093 qp_attr->cur_qp_state = qp_attr->qp_state;
3094 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
3095 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
3097 if (!ibqp->uobject) {
3098 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
3099 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3101 qp_attr->cap.max_send_wr = 0;
3102 qp_attr->cap.max_send_sge = 0;
3105 /* We don't support inline sends for kernel QPs (yet), and we
3106 * don't know what userspace's value should be.
3108 qp_attr->cap.max_inline_data = 0;
3110 qp_init_attr->cap = qp_attr->cap;
3112 qp_init_attr->create_flags = 0;
3113 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3114 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3116 qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
3117 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3123 mutex_unlock(&qp->mutex);
3127 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
3128 struct ib_ucontext *context,
3129 struct ib_udata *udata)
3131 struct mlx5_ib_dev *dev = to_mdev(ibdev);
3132 struct mlx5_ib_xrcd *xrcd;
3135 if (!MLX5_CAP_GEN(dev->mdev, xrc))
3136 return ERR_PTR(-ENOSYS);
3138 xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
3140 return ERR_PTR(-ENOMEM);
3142 err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
3145 return ERR_PTR(-ENOMEM);
3148 return &xrcd->ibxrcd;
3151 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
3153 struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
3154 u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
3157 err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
3159 mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);