2 * Copyright (c) 2013-2015, Mellanox Technologies, Ltd. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <linux/module.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/ib_cache.h>
31 #include <rdma/ib_user_verbs.h>
34 /* not supported currently */
35 static int wq_signature;
38 MLX5_IB_ACK_REQ_FREQ = 8,
42 MLX5_IB_DEFAULT_SCHED_QUEUE = 0x83,
43 MLX5_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
44 MLX5_IB_LINK_TYPE_IB = 0,
45 MLX5_IB_LINK_TYPE_ETH = 1
49 MLX5_IB_SQ_STRIDE = 6,
52 static const u32 mlx5_ib_opcode[] = {
53 [IB_WR_SEND] = MLX5_OPCODE_SEND,
54 [IB_WR_LSO] = MLX5_OPCODE_LSO,
55 [IB_WR_SEND_WITH_IMM] = MLX5_OPCODE_SEND_IMM,
56 [IB_WR_RDMA_WRITE] = MLX5_OPCODE_RDMA_WRITE,
57 [IB_WR_RDMA_WRITE_WITH_IMM] = MLX5_OPCODE_RDMA_WRITE_IMM,
58 [IB_WR_RDMA_READ] = MLX5_OPCODE_RDMA_READ,
59 [IB_WR_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_CS,
60 [IB_WR_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_FA,
61 [IB_WR_SEND_WITH_INV] = MLX5_OPCODE_SEND_INVAL,
62 [IB_WR_LOCAL_INV] = MLX5_OPCODE_UMR,
63 [IB_WR_REG_MR] = MLX5_OPCODE_UMR,
64 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_MASKED_CS,
65 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_MASKED_FA,
66 [MLX5_IB_WR_UMR] = MLX5_OPCODE_UMR,
69 struct mlx5_wqe_eth_pad {
73 enum raw_qp_set_mask_map {
74 MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID = 1UL << 0,
77 struct mlx5_modify_raw_qp_param {
80 u32 set_mask; /* raw_qp_set_mask_map */
84 static void get_cqs(enum ib_qp_type qp_type,
85 struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
86 struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
88 static int is_qp0(enum ib_qp_type qp_type)
90 return qp_type == IB_QPT_SMI;
93 static int is_sqp(enum ib_qp_type qp_type)
95 return is_qp0(qp_type) || is_qp1(qp_type);
98 static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
100 return mlx5_buf_offset(&qp->buf, offset);
103 static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
105 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
108 void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
110 return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
114 * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
116 * @qp: QP to copy from.
117 * @send: copy from the send queue when non-zero, use the receive queue
119 * @wqe_index: index to start copying from. For send work queues, the
120 * wqe_index is in units of MLX5_SEND_WQE_BB.
121 * For receive work queue, it is the number of work queue
122 * element in the queue.
123 * @buffer: destination buffer.
124 * @length: maximum number of bytes to copy.
126 * Copies at least a single WQE, but may copy more data.
128 * Return: the number of bytes copied, or an error code.
130 int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
131 void *buffer, u32 length,
132 struct mlx5_ib_qp_base *base)
134 struct ib_device *ibdev = qp->ibqp.device;
135 struct mlx5_ib_dev *dev = to_mdev(ibdev);
136 struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
139 struct ib_umem *umem = base->ubuffer.umem;
140 u32 first_copy_length;
144 if (wq->wqe_cnt == 0) {
145 mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
150 offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
151 wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
153 if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
156 if (offset > umem->length ||
157 (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
160 first_copy_length = min_t(u32, offset + length, wq_end) - offset;
161 ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
166 struct mlx5_wqe_ctrl_seg *ctrl = buffer;
167 int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
169 wqe_length = ds * MLX5_WQE_DS_UNITS;
171 wqe_length = 1 << wq->wqe_shift;
174 if (wqe_length <= first_copy_length)
175 return first_copy_length;
177 ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
178 wqe_length - first_copy_length);
185 static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
187 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
188 struct ib_event event;
190 if (type == MLX5_EVENT_TYPE_PATH_MIG) {
191 /* This event is only valid for trans_qps */
192 to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
195 if (ibqp->event_handler) {
196 event.device = ibqp->device;
197 event.element.qp = ibqp;
199 case MLX5_EVENT_TYPE_PATH_MIG:
200 event.event = IB_EVENT_PATH_MIG;
202 case MLX5_EVENT_TYPE_COMM_EST:
203 event.event = IB_EVENT_COMM_EST;
205 case MLX5_EVENT_TYPE_SQ_DRAINED:
206 event.event = IB_EVENT_SQ_DRAINED;
208 case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
209 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
211 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
212 event.event = IB_EVENT_QP_FATAL;
214 case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
215 event.event = IB_EVENT_PATH_MIG_ERR;
217 case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
218 event.event = IB_EVENT_QP_REQ_ERR;
220 case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
221 event.event = IB_EVENT_QP_ACCESS_ERR;
224 pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
228 ibqp->event_handler(&event, ibqp->qp_context);
232 static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
233 int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
238 /* Sanity check RQ size before proceeding */
239 if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
245 qp->rq.wqe_shift = 0;
246 cap->max_recv_wr = 0;
247 cap->max_recv_sge = 0;
250 qp->rq.wqe_cnt = ucmd->rq_wqe_count;
251 qp->rq.wqe_shift = ucmd->rq_wqe_shift;
252 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
253 qp->rq.max_post = qp->rq.wqe_cnt;
255 wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
256 wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
257 wqe_size = roundup_pow_of_two(wqe_size);
258 wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
259 wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
260 qp->rq.wqe_cnt = wq_size / wqe_size;
261 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
262 mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
264 MLX5_CAP_GEN(dev->mdev,
268 qp->rq.wqe_shift = ilog2(wqe_size);
269 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
270 qp->rq.max_post = qp->rq.wqe_cnt;
277 static int sq_overhead(struct ib_qp_init_attr *attr)
281 switch (attr->qp_type) {
283 size += sizeof(struct mlx5_wqe_xrc_seg);
286 size += sizeof(struct mlx5_wqe_ctrl_seg) +
287 max(sizeof(struct mlx5_wqe_atomic_seg) +
288 sizeof(struct mlx5_wqe_raddr_seg),
289 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
290 sizeof(struct mlx5_mkey_seg));
297 size += sizeof(struct mlx5_wqe_ctrl_seg) +
298 max(sizeof(struct mlx5_wqe_raddr_seg),
299 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
300 sizeof(struct mlx5_mkey_seg));
304 if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
305 size += sizeof(struct mlx5_wqe_eth_pad) +
306 sizeof(struct mlx5_wqe_eth_seg);
309 case MLX5_IB_QPT_HW_GSI:
310 size += sizeof(struct mlx5_wqe_ctrl_seg) +
311 sizeof(struct mlx5_wqe_datagram_seg);
314 case MLX5_IB_QPT_REG_UMR:
315 size += sizeof(struct mlx5_wqe_ctrl_seg) +
316 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
317 sizeof(struct mlx5_mkey_seg);
327 static int calc_send_wqe(struct ib_qp_init_attr *attr)
332 size = sq_overhead(attr);
336 if (attr->cap.max_inline_data) {
337 inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
338 attr->cap.max_inline_data;
341 size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
342 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
343 ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
344 return MLX5_SIG_WQE_SIZE;
346 return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
349 static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
350 struct mlx5_ib_qp *qp)
355 if (!attr->cap.max_send_wr)
358 wqe_size = calc_send_wqe(attr);
359 mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
363 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
364 mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
365 wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
369 qp->max_inline_data = wqe_size - sq_overhead(attr) -
370 sizeof(struct mlx5_wqe_inline_seg);
371 attr->cap.max_inline_data = qp->max_inline_data;
373 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
374 qp->signature_en = true;
376 wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
377 qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
378 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
379 mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
381 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
384 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
385 qp->sq.max_gs = attr->cap.max_send_sge;
386 qp->sq.max_post = wq_size / wqe_size;
387 attr->cap.max_send_wr = qp->sq.max_post;
392 static int set_user_buf_size(struct mlx5_ib_dev *dev,
393 struct mlx5_ib_qp *qp,
394 struct mlx5_ib_create_qp *ucmd,
395 struct mlx5_ib_qp_base *base,
396 struct ib_qp_init_attr *attr)
398 int desc_sz = 1 << qp->sq.wqe_shift;
400 if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
401 mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
402 desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
406 if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
407 mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
408 ucmd->sq_wqe_count, ucmd->sq_wqe_count);
412 qp->sq.wqe_cnt = ucmd->sq_wqe_count;
414 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
415 mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
417 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
421 if (attr->qp_type == IB_QPT_RAW_PACKET) {
422 base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
423 qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
425 base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
426 (qp->sq.wqe_cnt << 6);
432 static int qp_has_rq(struct ib_qp_init_attr *attr)
434 if (attr->qp_type == IB_QPT_XRC_INI ||
435 attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
436 attr->qp_type == MLX5_IB_QPT_REG_UMR ||
437 !attr->cap.max_recv_wr)
443 static int first_med_uuar(void)
448 static int next_uuar(int n)
452 while (((n % 4) & 2))
458 static int num_med_uuar(struct mlx5_uuar_info *uuari)
462 n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
463 uuari->num_low_latency_uuars - 1;
465 return n >= 0 ? n : 0;
468 static int max_uuari(struct mlx5_uuar_info *uuari)
470 return uuari->num_uars * 4;
473 static int first_hi_uuar(struct mlx5_uuar_info *uuari)
479 med = num_med_uuar(uuari);
480 for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
489 static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
493 for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
494 if (!test_bit(i, uuari->bitmap)) {
495 set_bit(i, uuari->bitmap);
504 static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
506 int minidx = first_med_uuar();
509 for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
510 if (uuari->count[i] < uuari->count[minidx])
514 uuari->count[minidx]++;
518 static int alloc_uuar(struct mlx5_uuar_info *uuari,
519 enum mlx5_ib_latency_class lat)
523 mutex_lock(&uuari->lock);
525 case MLX5_IB_LATENCY_CLASS_LOW:
527 uuari->count[uuarn]++;
530 case MLX5_IB_LATENCY_CLASS_MEDIUM:
534 uuarn = alloc_med_class_uuar(uuari);
537 case MLX5_IB_LATENCY_CLASS_HIGH:
541 uuarn = alloc_high_class_uuar(uuari);
544 case MLX5_IB_LATENCY_CLASS_FAST_PATH:
548 mutex_unlock(&uuari->lock);
553 static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
555 clear_bit(uuarn, uuari->bitmap);
556 --uuari->count[uuarn];
559 static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
561 clear_bit(uuarn, uuari->bitmap);
562 --uuari->count[uuarn];
565 static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
567 int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
568 int high_uuar = nuuars - uuari->num_low_latency_uuars;
570 mutex_lock(&uuari->lock);
572 --uuari->count[uuarn];
576 if (uuarn < high_uuar) {
577 free_med_class_uuar(uuari, uuarn);
581 free_high_class_uuar(uuari, uuarn);
584 mutex_unlock(&uuari->lock);
587 static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
590 case IB_QPS_RESET: return MLX5_QP_STATE_RST;
591 case IB_QPS_INIT: return MLX5_QP_STATE_INIT;
592 case IB_QPS_RTR: return MLX5_QP_STATE_RTR;
593 case IB_QPS_RTS: return MLX5_QP_STATE_RTS;
594 case IB_QPS_SQD: return MLX5_QP_STATE_SQD;
595 case IB_QPS_SQE: return MLX5_QP_STATE_SQER;
596 case IB_QPS_ERR: return MLX5_QP_STATE_ERR;
601 static int to_mlx5_st(enum ib_qp_type type)
604 case IB_QPT_RC: return MLX5_QP_ST_RC;
605 case IB_QPT_UC: return MLX5_QP_ST_UC;
606 case IB_QPT_UD: return MLX5_QP_ST_UD;
607 case MLX5_IB_QPT_REG_UMR: return MLX5_QP_ST_REG_UMR;
609 case IB_QPT_XRC_TGT: return MLX5_QP_ST_XRC;
610 case IB_QPT_SMI: return MLX5_QP_ST_QP0;
611 case MLX5_IB_QPT_HW_GSI: return MLX5_QP_ST_QP1;
612 case IB_QPT_RAW_IPV6: return MLX5_QP_ST_RAW_IPV6;
613 case IB_QPT_RAW_PACKET:
614 case IB_QPT_RAW_ETHERTYPE: return MLX5_QP_ST_RAW_ETHERTYPE;
616 default: return -EINVAL;
620 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
621 struct mlx5_ib_cq *recv_cq);
622 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
623 struct mlx5_ib_cq *recv_cq);
625 static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
627 return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
630 static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
632 unsigned long addr, size_t size,
633 struct ib_umem **umem,
634 int *npages, int *page_shift, int *ncont,
639 *umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
641 mlx5_ib_dbg(dev, "umem_get failed\n");
642 return PTR_ERR(*umem);
645 mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
647 err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
649 mlx5_ib_warn(dev, "bad offset\n");
653 mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
654 addr, size, *npages, *page_shift, *ncont, *offset);
659 ib_umem_release(*umem);
665 static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
667 struct mlx5_ib_ucontext *context;
669 context = to_mucontext(pd->uobject->context);
670 mlx5_ib_db_unmap_user(context, &rwq->db);
672 ib_umem_release(rwq->umem);
675 static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
676 struct mlx5_ib_rwq *rwq,
677 struct mlx5_ib_create_wq *ucmd)
679 struct mlx5_ib_ucontext *context;
689 context = to_mucontext(pd->uobject->context);
690 rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
691 rwq->buf_size, 0, 0);
692 if (IS_ERR(rwq->umem)) {
693 mlx5_ib_dbg(dev, "umem_get failed\n");
694 err = PTR_ERR(rwq->umem);
698 mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, &npages, &page_shift,
700 err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
701 &rwq->rq_page_offset);
703 mlx5_ib_warn(dev, "bad offset\n");
707 rwq->rq_num_pas = ncont;
708 rwq->page_shift = page_shift;
709 rwq->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
710 rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
712 mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
713 (unsigned long long)ucmd->buf_addr, rwq->buf_size,
714 npages, page_shift, ncont, offset);
716 err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
718 mlx5_ib_dbg(dev, "map failed\n");
722 rwq->create_type = MLX5_WQ_USER;
726 ib_umem_release(rwq->umem);
730 static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
731 struct mlx5_ib_qp *qp, struct ib_udata *udata,
732 struct ib_qp_init_attr *attr,
734 struct mlx5_ib_create_qp_resp *resp, int *inlen,
735 struct mlx5_ib_qp_base *base)
737 struct mlx5_ib_ucontext *context;
738 struct mlx5_ib_create_qp ucmd;
739 struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
750 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
752 mlx5_ib_dbg(dev, "copy failed\n");
756 context = to_mucontext(pd->uobject->context);
758 * TBD: should come from the verbs when we have the API
760 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
761 /* In CROSS_CHANNEL CQ and QP must use the same UAR */
762 uuarn = MLX5_CROSS_CHANNEL_UUAR;
764 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
766 mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
767 mlx5_ib_dbg(dev, "reverting to medium latency\n");
768 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
770 mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
771 mlx5_ib_dbg(dev, "reverting to high latency\n");
772 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
774 mlx5_ib_warn(dev, "uuar allocation failed\n");
781 uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
782 mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
785 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
786 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
788 err = set_user_buf_size(dev, qp, &ucmd, base, attr);
792 if (ucmd.buf_addr && ubuffer->buf_size) {
793 ubuffer->buf_addr = ucmd.buf_addr;
794 err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
796 &ubuffer->umem, &npages, &page_shift,
801 ubuffer->umem = NULL;
804 *inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
805 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
806 *in = mlx5_vzalloc(*inlen);
812 pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
814 mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
816 qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
818 MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
819 MLX5_SET(qpc, qpc, page_offset, offset);
821 MLX5_SET(qpc, qpc, uar_page, uar_index);
822 resp->uuar_index = uuarn;
825 err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
827 mlx5_ib_dbg(dev, "map failed\n");
831 err = ib_copy_to_udata(udata, resp, sizeof(*resp));
833 mlx5_ib_dbg(dev, "copy failed\n");
836 qp->create_type = MLX5_QP_USER;
841 mlx5_ib_db_unmap_user(context, &qp->db);
848 ib_umem_release(ubuffer->umem);
851 free_uuar(&context->uuari, uuarn);
855 static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
856 struct mlx5_ib_qp_base *base)
858 struct mlx5_ib_ucontext *context;
860 context = to_mucontext(pd->uobject->context);
861 mlx5_ib_db_unmap_user(context, &qp->db);
862 if (base->ubuffer.umem)
863 ib_umem_release(base->ubuffer.umem);
864 free_uuar(&context->uuari, qp->uuarn);
867 static int create_kernel_qp(struct mlx5_ib_dev *dev,
868 struct ib_qp_init_attr *init_attr,
869 struct mlx5_ib_qp *qp,
870 u32 **in, int *inlen,
871 struct mlx5_ib_qp_base *base)
873 enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
874 struct mlx5_uuar_info *uuari;
880 uuari = &dev->mdev->priv.uuari;
881 if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
882 IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
883 IB_QP_CREATE_IPOIB_UD_LSO |
884 mlx5_ib_create_qp_sqpn_qp1()))
887 if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
888 lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
890 uuarn = alloc_uuar(uuari, lc);
892 mlx5_ib_dbg(dev, "\n");
896 qp->bf = &uuari->bfs[uuarn];
897 uar_index = qp->bf->uar->index;
899 err = calc_sq_size(dev, init_attr, qp);
901 mlx5_ib_dbg(dev, "err %d\n", err);
906 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
907 base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
909 err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size,
910 2 * PAGE_SIZE, &qp->buf);
912 mlx5_ib_dbg(dev, "err %d\n", err);
916 qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
917 *inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
918 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
919 *in = mlx5_vzalloc(*inlen);
925 qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
926 MLX5_SET(qpc, qpc, uar_page, uar_index);
927 MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
929 /* Set "fast registration enabled" for all kernel QPs */
930 MLX5_SET(qpc, qpc, fre, 1);
931 MLX5_SET(qpc, qpc, rlky, 1);
933 if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
934 MLX5_SET(qpc, qpc, deth_sqpn, 1);
935 qp->flags |= MLX5_IB_QP_SQPN_QP1;
938 mlx5_fill_page_array(&qp->buf,
939 (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
941 err = mlx5_db_alloc(dev->mdev, &qp->db);
943 mlx5_ib_dbg(dev, "err %d\n", err);
947 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
948 qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
949 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
950 qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
951 qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
953 if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
954 !qp->sq.w_list || !qp->sq.wqe_head) {
958 qp->create_type = MLX5_QP_KERNEL;
963 mlx5_db_free(dev->mdev, &qp->db);
964 kfree(qp->sq.wqe_head);
965 kfree(qp->sq.w_list);
967 kfree(qp->sq.wr_data);
974 mlx5_buf_free(dev->mdev, &qp->buf);
977 free_uuar(&dev->mdev->priv.uuari, uuarn);
981 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
983 mlx5_db_free(dev->mdev, &qp->db);
984 kfree(qp->sq.wqe_head);
985 kfree(qp->sq.w_list);
987 kfree(qp->sq.wr_data);
989 mlx5_buf_free(dev->mdev, &qp->buf);
990 free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
993 static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
995 if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
996 (attr->qp_type == IB_QPT_XRC_INI))
998 else if (!qp->has_rq)
999 return MLX5_ZERO_LEN_RQ;
1001 return MLX5_NON_ZERO_RQ;
1004 static int is_connected(enum ib_qp_type qp_type)
1006 if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1012 static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1013 struct mlx5_ib_sq *sq, u32 tdn)
1015 u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
1016 void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1018 MLX5_SET(tisc, tisc, transport_domain, tdn);
1019 return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1022 static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1023 struct mlx5_ib_sq *sq)
1025 mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1028 static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1029 struct mlx5_ib_sq *sq, void *qpin,
1032 struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1036 void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1045 err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1046 &sq->ubuffer.umem, &npages, &page_shift,
1051 inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1052 in = mlx5_vzalloc(inlen);
1058 sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1059 MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1060 MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1061 MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1062 MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1063 MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1064 MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1066 wq = MLX5_ADDR_OF(sqc, sqc, wq);
1067 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1068 MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1069 MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1070 MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1071 MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1072 MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1073 MLX5_SET(wq, wq, log_wq_pg_sz, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1074 MLX5_SET(wq, wq, page_offset, offset);
1076 pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1077 mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1079 err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1089 ib_umem_release(sq->ubuffer.umem);
1090 sq->ubuffer.umem = NULL;
1095 static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1096 struct mlx5_ib_sq *sq)
1098 mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1099 ib_umem_release(sq->ubuffer.umem);
1102 static int get_rq_pas_size(void *qpc)
1104 u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1105 u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1106 u32 log_rq_size = MLX5_GET(qpc, qpc, log_rq_size);
1107 u32 page_offset = MLX5_GET(qpc, qpc, page_offset);
1108 u32 po_quanta = 1 << (log_page_size - 6);
1109 u32 rq_sz = 1 << (log_rq_size + 4 + log_rq_stride);
1110 u32 page_size = 1 << log_page_size;
1111 u32 rq_sz_po = rq_sz + (page_offset * po_quanta);
1112 u32 rq_num_pas = (rq_sz_po + page_size - 1) / page_size;
1114 return rq_num_pas * sizeof(u64);
1117 static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1118 struct mlx5_ib_rq *rq, void *qpin)
1120 struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
1126 void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1129 u32 rq_pas_size = get_rq_pas_size(qpc);
1131 inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1132 in = mlx5_vzalloc(inlen);
1136 rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1137 MLX5_SET(rqc, rqc, vlan_strip_disable, 1);
1138 MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
1139 MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1140 MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1141 MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1142 MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1144 if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1145 MLX5_SET(rqc, rqc, scatter_fcs, 1);
1147 wq = MLX5_ADDR_OF(rqc, rqc, wq);
1148 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1149 MLX5_SET(wq, wq, end_padding_mode,
1150 MLX5_GET(qpc, qpc, end_padding_mode));
1151 MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1152 MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1153 MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1154 MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1155 MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1156 MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1158 pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1159 qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1160 memcpy(pas, qp_pas, rq_pas_size);
1162 err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1169 static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1170 struct mlx5_ib_rq *rq)
1172 mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1175 static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1176 struct mlx5_ib_rq *rq, u32 tdn)
1183 inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1184 in = mlx5_vzalloc(inlen);
1188 tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1189 MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1190 MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1191 MLX5_SET(tirc, tirc, transport_domain, tdn);
1193 err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1200 static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1201 struct mlx5_ib_rq *rq)
1203 mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1206 static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1210 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1211 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1212 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1213 struct ib_uobject *uobj = pd->uobject;
1214 struct ib_ucontext *ucontext = uobj->context;
1215 struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1217 u32 tdn = mucontext->tdn;
1219 if (qp->sq.wqe_cnt) {
1220 err = create_raw_packet_qp_tis(dev, sq, tdn);
1224 err = create_raw_packet_qp_sq(dev, sq, in, pd);
1226 goto err_destroy_tis;
1228 sq->base.container_mibqp = qp;
1231 if (qp->rq.wqe_cnt) {
1232 rq->base.container_mibqp = qp;
1234 err = create_raw_packet_qp_rq(dev, rq, in);
1236 goto err_destroy_sq;
1239 err = create_raw_packet_qp_tir(dev, rq, tdn);
1241 goto err_destroy_rq;
1244 qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1250 destroy_raw_packet_qp_rq(dev, rq);
1252 if (!qp->sq.wqe_cnt)
1254 destroy_raw_packet_qp_sq(dev, sq);
1256 destroy_raw_packet_qp_tis(dev, sq);
1261 static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1262 struct mlx5_ib_qp *qp)
1264 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1265 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1266 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1268 if (qp->rq.wqe_cnt) {
1269 destroy_raw_packet_qp_tir(dev, rq);
1270 destroy_raw_packet_qp_rq(dev, rq);
1273 if (qp->sq.wqe_cnt) {
1274 destroy_raw_packet_qp_sq(dev, sq);
1275 destroy_raw_packet_qp_tis(dev, sq);
1279 static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1280 struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1282 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1283 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1287 sq->doorbell = &qp->db;
1288 rq->doorbell = &qp->db;
1291 static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1293 mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1296 static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1298 struct ib_qp_init_attr *init_attr,
1299 struct ib_udata *udata)
1301 struct ib_uobject *uobj = pd->uobject;
1302 struct ib_ucontext *ucontext = uobj->context;
1303 struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1304 struct mlx5_ib_create_qp_resp resp = {};
1310 u32 selected_fields = 0;
1311 size_t min_resp_len;
1312 u32 tdn = mucontext->tdn;
1313 struct mlx5_ib_create_qp_rss ucmd = {};
1314 size_t required_cmd_sz;
1316 if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1319 if (init_attr->create_flags || init_attr->send_cq)
1322 min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1323 if (udata->outlen < min_resp_len)
1326 required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1327 if (udata->inlen < required_cmd_sz) {
1328 mlx5_ib_dbg(dev, "invalid inlen\n");
1332 if (udata->inlen > sizeof(ucmd) &&
1333 !ib_is_udata_cleared(udata, sizeof(ucmd),
1334 udata->inlen - sizeof(ucmd))) {
1335 mlx5_ib_dbg(dev, "inlen is not supported\n");
1339 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1340 mlx5_ib_dbg(dev, "copy failed\n");
1344 if (ucmd.comp_mask) {
1345 mlx5_ib_dbg(dev, "invalid comp mask\n");
1349 if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1350 mlx5_ib_dbg(dev, "invalid reserved\n");
1354 err = ib_copy_to_udata(udata, &resp, min_resp_len);
1356 mlx5_ib_dbg(dev, "copy failed\n");
1360 inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1361 in = mlx5_vzalloc(inlen);
1365 tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
1366 MLX5_SET(tirc, tirc, disp_type,
1367 MLX5_TIRC_DISP_TYPE_INDIRECT);
1368 MLX5_SET(tirc, tirc, indirect_table,
1369 init_attr->rwq_ind_tbl->ind_tbl_num);
1370 MLX5_SET(tirc, tirc, transport_domain, tdn);
1372 hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1373 switch (ucmd.rx_hash_function) {
1374 case MLX5_RX_HASH_FUNC_TOEPLITZ:
1376 void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1377 size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1379 if (len != ucmd.rx_key_len) {
1384 MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FUNC_TOEPLITZ);
1385 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1386 memcpy(rss_key, ucmd.rx_hash_key, len);
1394 if (!ucmd.rx_hash_fields_mask) {
1395 /* special case when this TIR serves as steering entry without hashing */
1396 if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1402 if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1403 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1404 ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1405 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1410 /* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1411 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1412 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1413 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1414 MLX5_L3_PROT_TYPE_IPV4);
1415 else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1416 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1417 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1418 MLX5_L3_PROT_TYPE_IPV6);
1420 if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1421 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1422 ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1423 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1428 /* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1429 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1430 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1431 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1432 MLX5_L4_PROT_TYPE_TCP);
1433 else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1434 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1435 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1436 MLX5_L4_PROT_TYPE_UDP);
1438 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1439 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1440 selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1442 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1443 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1444 selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1446 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1447 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1448 selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1450 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1451 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1452 selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1454 MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1457 err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1463 /* qpn is reserved for that QP */
1464 qp->trans_qp.base.mqp.qpn = 0;
1465 qp->flags |= MLX5_IB_QP_RSS;
1473 static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1474 struct ib_qp_init_attr *init_attr,
1475 struct ib_udata *udata, struct mlx5_ib_qp *qp)
1477 struct mlx5_ib_resources *devr = &dev->devr;
1478 int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
1479 struct mlx5_core_dev *mdev = dev->mdev;
1480 struct mlx5_ib_create_qp_resp resp;
1481 struct mlx5_ib_cq *send_cq;
1482 struct mlx5_ib_cq *recv_cq;
1483 unsigned long flags;
1484 u32 uidx = MLX5_IB_DEFAULT_UIDX;
1485 struct mlx5_ib_create_qp ucmd;
1486 struct mlx5_ib_qp_base *base;
1491 base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1492 &qp->raw_packet_qp.rq.base :
1495 if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1496 mlx5_ib_odp_create_qp(qp);
1498 mutex_init(&qp->mutex);
1499 spin_lock_init(&qp->sq.lock);
1500 spin_lock_init(&qp->rq.lock);
1502 if (init_attr->rwq_ind_tbl) {
1506 err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1510 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
1511 if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
1512 mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1515 qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1519 if (init_attr->create_flags &
1520 (IB_QP_CREATE_CROSS_CHANNEL |
1521 IB_QP_CREATE_MANAGED_SEND |
1522 IB_QP_CREATE_MANAGED_RECV)) {
1523 if (!MLX5_CAP_GEN(mdev, cd)) {
1524 mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1527 if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1528 qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1529 if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1530 qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1531 if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1532 qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1535 if (init_attr->qp_type == IB_QPT_UD &&
1536 (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1537 if (!MLX5_CAP_GEN(mdev, ipoib_ipoib_offloads)) {
1538 mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1542 if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1543 if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1544 mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1547 if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1548 !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1549 mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1552 qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1555 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1556 qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1558 if (pd && pd->uobject) {
1559 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1560 mlx5_ib_dbg(dev, "copy failed\n");
1564 err = get_qp_user_index(to_mucontext(pd->uobject->context),
1565 &ucmd, udata->inlen, &uidx);
1569 qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1570 qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1572 qp->wq_sig = !!wq_signature;
1575 qp->has_rq = qp_has_rq(init_attr);
1576 err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1577 qp, (pd && pd->uobject) ? &ucmd : NULL);
1579 mlx5_ib_dbg(dev, "err %d\n", err);
1586 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
1587 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1588 if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1589 ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1590 mlx5_ib_dbg(dev, "invalid rq params\n");
1593 if (ucmd.sq_wqe_count > max_wqes) {
1594 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
1595 ucmd.sq_wqe_count, max_wqes);
1598 if (init_attr->create_flags &
1599 mlx5_ib_create_qp_sqpn_qp1()) {
1600 mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1603 err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1604 &resp, &inlen, base);
1606 mlx5_ib_dbg(dev, "err %d\n", err);
1608 err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1611 mlx5_ib_dbg(dev, "err %d\n", err);
1617 in = mlx5_vzalloc(inlen);
1621 qp->create_type = MLX5_QP_EMPTY;
1624 if (is_sqp(init_attr->qp_type))
1625 qp->port = init_attr->port_num;
1627 qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1629 MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1630 MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
1632 if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
1633 MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
1635 MLX5_SET(qpc, qpc, latency_sensitive, 1);
1639 MLX5_SET(qpc, qpc, wq_signature, 1);
1641 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
1642 MLX5_SET(qpc, qpc, block_lb_mc, 1);
1644 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
1645 MLX5_SET(qpc, qpc, cd_master, 1);
1646 if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
1647 MLX5_SET(qpc, qpc, cd_slave_send, 1);
1648 if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
1649 MLX5_SET(qpc, qpc, cd_slave_receive, 1);
1651 if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1655 rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1656 scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1659 MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
1661 MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
1663 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1665 MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
1667 MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
1671 if (qp->rq.wqe_cnt) {
1672 MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1673 MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
1676 MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
1679 MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
1681 MLX5_SET(qpc, qpc, no_sq, 1);
1683 /* Set default resources */
1684 switch (init_attr->qp_type) {
1685 case IB_QPT_XRC_TGT:
1686 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1687 MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1688 MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1689 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
1691 case IB_QPT_XRC_INI:
1692 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1693 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1694 MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s0)->msrq.srqn);
1697 if (init_attr->srq) {
1698 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1699 MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(init_attr->srq)->msrq.srqn);
1701 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1702 MLX5_SET(qpc, qpc, srqn_rmpn, to_msrq(devr->s1)->msrq.srqn);
1706 if (init_attr->send_cq)
1707 MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
1709 if (init_attr->recv_cq)
1710 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
1712 MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
1714 /* 0xffffff means we ask to work with cqe version 0 */
1715 if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
1716 MLX5_SET(qpc, qpc, user_index, uidx);
1718 /* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1719 if (init_attr->qp_type == IB_QPT_UD &&
1720 (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
1721 MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1722 qp->flags |= MLX5_IB_QP_LSO;
1725 if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1726 qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1727 raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1728 err = create_raw_packet_qp(dev, qp, in, pd);
1730 err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1734 mlx5_ib_dbg(dev, "create qp failed\n");
1740 base->container_mibqp = qp;
1741 base->mqp.event = mlx5_ib_qp_event;
1743 get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1744 &send_cq, &recv_cq);
1745 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1746 mlx5_ib_lock_cqs(send_cq, recv_cq);
1747 /* Maintain device to QPs access, needed for further handling via reset
1750 list_add_tail(&qp->qps_list, &dev->qp_list);
1751 /* Maintain CQ to QPs access, needed for further handling via reset flow
1754 list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1756 list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1757 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1758 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1763 if (qp->create_type == MLX5_QP_USER)
1764 destroy_qp_user(pd, qp, base);
1765 else if (qp->create_type == MLX5_QP_KERNEL)
1766 destroy_qp_kernel(dev, qp);
1772 static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1773 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1777 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1778 spin_lock(&send_cq->lock);
1779 spin_lock_nested(&recv_cq->lock,
1780 SINGLE_DEPTH_NESTING);
1781 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1782 spin_lock(&send_cq->lock);
1783 __acquire(&recv_cq->lock);
1785 spin_lock(&recv_cq->lock);
1786 spin_lock_nested(&send_cq->lock,
1787 SINGLE_DEPTH_NESTING);
1790 spin_lock(&send_cq->lock);
1791 __acquire(&recv_cq->lock);
1793 } else if (recv_cq) {
1794 spin_lock(&recv_cq->lock);
1795 __acquire(&send_cq->lock);
1797 __acquire(&send_cq->lock);
1798 __acquire(&recv_cq->lock);
1802 static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1803 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1807 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1808 spin_unlock(&recv_cq->lock);
1809 spin_unlock(&send_cq->lock);
1810 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1811 __release(&recv_cq->lock);
1812 spin_unlock(&send_cq->lock);
1814 spin_unlock(&send_cq->lock);
1815 spin_unlock(&recv_cq->lock);
1818 __release(&recv_cq->lock);
1819 spin_unlock(&send_cq->lock);
1821 } else if (recv_cq) {
1822 __release(&send_cq->lock);
1823 spin_unlock(&recv_cq->lock);
1825 __release(&recv_cq->lock);
1826 __release(&send_cq->lock);
1830 static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1832 return to_mpd(qp->ibqp.pd);
1835 static void get_cqs(enum ib_qp_type qp_type,
1836 struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
1837 struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1840 case IB_QPT_XRC_TGT:
1844 case MLX5_IB_QPT_REG_UMR:
1845 case IB_QPT_XRC_INI:
1846 *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1851 case MLX5_IB_QPT_HW_GSI:
1855 case IB_QPT_RAW_IPV6:
1856 case IB_QPT_RAW_ETHERTYPE:
1857 case IB_QPT_RAW_PACKET:
1858 *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1859 *recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
1870 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1871 const struct mlx5_modify_raw_qp_param *raw_qp_param,
1872 u8 lag_tx_affinity);
1874 static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1876 struct mlx5_ib_cq *send_cq, *recv_cq;
1877 struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
1878 unsigned long flags;
1881 if (qp->ibqp.rwq_ind_tbl) {
1882 destroy_rss_raw_qp_tir(dev, qp);
1886 base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1887 &qp->raw_packet_qp.rq.base :
1890 if (qp->state != IB_QPS_RESET) {
1891 if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1892 mlx5_ib_qp_disable_pagefaults(qp);
1893 err = mlx5_core_qp_modify(dev->mdev,
1894 MLX5_CMD_OP_2RST_QP, 0,
1897 struct mlx5_modify_raw_qp_param raw_qp_param = {
1898 .operation = MLX5_CMD_OP_2RST_QP
1901 err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
1904 mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
1908 get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
1909 &send_cq, &recv_cq);
1911 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1912 mlx5_ib_lock_cqs(send_cq, recv_cq);
1913 /* del from lists under both locks above to protect reset flow paths */
1914 list_del(&qp->qps_list);
1916 list_del(&qp->cq_send_list);
1919 list_del(&qp->cq_recv_list);
1921 if (qp->create_type == MLX5_QP_KERNEL) {
1922 __mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
1923 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1924 if (send_cq != recv_cq)
1925 __mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1928 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1929 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1931 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1932 destroy_raw_packet_qp(dev, qp);
1934 err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1936 mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1940 if (qp->create_type == MLX5_QP_KERNEL)
1941 destroy_qp_kernel(dev, qp);
1942 else if (qp->create_type == MLX5_QP_USER)
1943 destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
1946 static const char *ib_qp_type_str(enum ib_qp_type type)
1950 return "IB_QPT_SMI";
1952 return "IB_QPT_GSI";
1959 case IB_QPT_RAW_IPV6:
1960 return "IB_QPT_RAW_IPV6";
1961 case IB_QPT_RAW_ETHERTYPE:
1962 return "IB_QPT_RAW_ETHERTYPE";
1963 case IB_QPT_XRC_INI:
1964 return "IB_QPT_XRC_INI";
1965 case IB_QPT_XRC_TGT:
1966 return "IB_QPT_XRC_TGT";
1967 case IB_QPT_RAW_PACKET:
1968 return "IB_QPT_RAW_PACKET";
1969 case MLX5_IB_QPT_REG_UMR:
1970 return "MLX5_IB_QPT_REG_UMR";
1973 return "Invalid QP type";
1977 struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1978 struct ib_qp_init_attr *init_attr,
1979 struct ib_udata *udata)
1981 struct mlx5_ib_dev *dev;
1982 struct mlx5_ib_qp *qp;
1987 dev = to_mdev(pd->device);
1989 if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1991 mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
1992 return ERR_PTR(-EINVAL);
1993 } else if (!to_mucontext(pd->uobject->context)->cqe_version) {
1994 mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
1995 return ERR_PTR(-EINVAL);
1999 /* being cautious here */
2000 if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2001 init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2002 pr_warn("%s: no PD for transport %s\n", __func__,
2003 ib_qp_type_str(init_attr->qp_type));
2004 return ERR_PTR(-EINVAL);
2006 dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
2009 switch (init_attr->qp_type) {
2010 case IB_QPT_XRC_TGT:
2011 case IB_QPT_XRC_INI:
2012 if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
2013 mlx5_ib_dbg(dev, "XRC not supported\n");
2014 return ERR_PTR(-ENOSYS);
2016 init_attr->recv_cq = NULL;
2017 if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2018 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2019 init_attr->send_cq = NULL;
2023 case IB_QPT_RAW_PACKET:
2028 case MLX5_IB_QPT_HW_GSI:
2029 case MLX5_IB_QPT_REG_UMR:
2030 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2032 return ERR_PTR(-ENOMEM);
2034 err = create_qp_common(dev, pd, init_attr, udata, qp);
2036 mlx5_ib_dbg(dev, "create_qp_common failed\n");
2038 return ERR_PTR(err);
2041 if (is_qp0(init_attr->qp_type))
2042 qp->ibqp.qp_num = 0;
2043 else if (is_qp1(init_attr->qp_type))
2044 qp->ibqp.qp_num = 1;
2046 qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
2048 mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
2049 qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2050 init_attr->recv_cq ? to_mcq(init_attr->recv_cq)->mcq.cqn : -1,
2051 init_attr->send_cq ? to_mcq(init_attr->send_cq)->mcq.cqn : -1);
2053 qp->trans_qp.xrcdn = xrcdn;
2058 return mlx5_ib_gsi_create_qp(pd, init_attr);
2060 case IB_QPT_RAW_IPV6:
2061 case IB_QPT_RAW_ETHERTYPE:
2064 mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2065 init_attr->qp_type);
2066 /* Don't support raw QPs */
2067 return ERR_PTR(-EINVAL);
2073 int mlx5_ib_destroy_qp(struct ib_qp *qp)
2075 struct mlx5_ib_dev *dev = to_mdev(qp->device);
2076 struct mlx5_ib_qp *mqp = to_mqp(qp);
2078 if (unlikely(qp->qp_type == IB_QPT_GSI))
2079 return mlx5_ib_gsi_destroy_qp(qp);
2081 destroy_qp_common(dev, mqp);
2088 static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
2091 u32 hw_access_flags = 0;
2095 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2096 dest_rd_atomic = attr->max_dest_rd_atomic;
2098 dest_rd_atomic = qp->trans_qp.resp_depth;
2100 if (attr_mask & IB_QP_ACCESS_FLAGS)
2101 access_flags = attr->qp_access_flags;
2103 access_flags = qp->trans_qp.atomic_rd_en;
2105 if (!dest_rd_atomic)
2106 access_flags &= IB_ACCESS_REMOTE_WRITE;
2108 if (access_flags & IB_ACCESS_REMOTE_READ)
2109 hw_access_flags |= MLX5_QP_BIT_RRE;
2110 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
2111 hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
2112 if (access_flags & IB_ACCESS_REMOTE_WRITE)
2113 hw_access_flags |= MLX5_QP_BIT_RWE;
2115 return cpu_to_be32(hw_access_flags);
2119 MLX5_PATH_FLAG_FL = 1 << 0,
2120 MLX5_PATH_FLAG_FREE_AR = 1 << 1,
2121 MLX5_PATH_FLAG_COUNTER = 1 << 2,
2124 static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2126 if (rate == IB_RATE_PORT_CURRENT) {
2128 } else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
2131 while (rate != IB_RATE_2_5_GBPS &&
2132 !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
2133 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
2137 return rate + MLX5_STAT_RATE_OFFSET;
2140 static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2141 struct mlx5_ib_sq *sq, u8 sl)
2148 inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2149 in = mlx5_vzalloc(inlen);
2153 MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2155 tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2156 MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2158 err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2165 static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2166 struct mlx5_ib_sq *sq, u8 tx_affinity)
2173 inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2174 in = mlx5_vzalloc(inlen);
2178 MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2180 tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2181 MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2183 err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2190 static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2191 const struct ib_ah_attr *ah,
2192 struct mlx5_qp_path *path, u8 port, int attr_mask,
2193 u32 path_flags, const struct ib_qp_attr *attr,
2196 enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2198 enum ib_gid_type gid_type;
2200 if (attr_mask & IB_QP_PKEY_INDEX)
2201 path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2204 if (ah->ah_flags & IB_AH_GRH) {
2205 if (ah->grh.sgid_index >=
2206 dev->mdev->port_caps[port - 1].gid_table_len) {
2207 pr_err("sgid_index (%u) too large. max is %d\n",
2209 dev->mdev->port_caps[port - 1].gid_table_len);
2214 if (ll == IB_LINK_LAYER_ETHERNET) {
2215 if (!(ah->ah_flags & IB_AH_GRH))
2217 err = mlx5_get_roce_gid_type(dev, port, ah->grh.sgid_index,
2221 memcpy(path->rmac, ah->dmac, sizeof(ah->dmac));
2222 path->udp_sport = mlx5_get_roce_udp_sport(dev, port,
2223 ah->grh.sgid_index);
2224 path->dci_cfi_prio_sl = (ah->sl & 0x7) << 4;
2225 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
2226 path->ecn_dscp = (ah->grh.traffic_class >> 2) & 0x3f;
2228 path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2230 (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2231 path->rlid = cpu_to_be16(ah->dlid);
2232 path->grh_mlid = ah->src_path_bits & 0x7f;
2233 if (ah->ah_flags & IB_AH_GRH)
2234 path->grh_mlid |= 1 << 7;
2235 path->dci_cfi_prio_sl = ah->sl & 0xf;
2238 if (ah->ah_flags & IB_AH_GRH) {
2239 path->mgid_index = ah->grh.sgid_index;
2240 path->hop_limit = ah->grh.hop_limit;
2241 path->tclass_flowlabel =
2242 cpu_to_be32((ah->grh.traffic_class << 20) |
2243 (ah->grh.flow_label));
2244 memcpy(path->rgid, ah->grh.dgid.raw, 16);
2247 err = ib_rate_to_mlx5(dev, ah->static_rate);
2250 path->static_rate = err;
2253 if (attr_mask & IB_QP_TIMEOUT)
2254 path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
2256 if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2257 return modify_raw_packet_eth_prio(dev->mdev,
2258 &qp->raw_packet_qp.sq,
2264 static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2265 [MLX5_QP_STATE_INIT] = {
2266 [MLX5_QP_STATE_INIT] = {
2267 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
2268 MLX5_QP_OPTPAR_RAE |
2269 MLX5_QP_OPTPAR_RWE |
2270 MLX5_QP_OPTPAR_PKEY_INDEX |
2271 MLX5_QP_OPTPAR_PRI_PORT,
2272 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
2273 MLX5_QP_OPTPAR_PKEY_INDEX |
2274 MLX5_QP_OPTPAR_PRI_PORT,
2275 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
2276 MLX5_QP_OPTPAR_Q_KEY |
2277 MLX5_QP_OPTPAR_PRI_PORT,
2279 [MLX5_QP_STATE_RTR] = {
2280 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2281 MLX5_QP_OPTPAR_RRE |
2282 MLX5_QP_OPTPAR_RAE |
2283 MLX5_QP_OPTPAR_RWE |
2284 MLX5_QP_OPTPAR_PKEY_INDEX,
2285 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2286 MLX5_QP_OPTPAR_RWE |
2287 MLX5_QP_OPTPAR_PKEY_INDEX,
2288 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
2289 MLX5_QP_OPTPAR_Q_KEY,
2290 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX |
2291 MLX5_QP_OPTPAR_Q_KEY,
2292 [MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2293 MLX5_QP_OPTPAR_RRE |
2294 MLX5_QP_OPTPAR_RAE |
2295 MLX5_QP_OPTPAR_RWE |
2296 MLX5_QP_OPTPAR_PKEY_INDEX,
2299 [MLX5_QP_STATE_RTR] = {
2300 [MLX5_QP_STATE_RTS] = {
2301 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2302 MLX5_QP_OPTPAR_RRE |
2303 MLX5_QP_OPTPAR_RAE |
2304 MLX5_QP_OPTPAR_RWE |
2305 MLX5_QP_OPTPAR_PM_STATE |
2306 MLX5_QP_OPTPAR_RNR_TIMEOUT,
2307 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2308 MLX5_QP_OPTPAR_RWE |
2309 MLX5_QP_OPTPAR_PM_STATE,
2310 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2313 [MLX5_QP_STATE_RTS] = {
2314 [MLX5_QP_STATE_RTS] = {
2315 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
2316 MLX5_QP_OPTPAR_RAE |
2317 MLX5_QP_OPTPAR_RWE |
2318 MLX5_QP_OPTPAR_RNR_TIMEOUT |
2319 MLX5_QP_OPTPAR_PM_STATE |
2320 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2321 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
2322 MLX5_QP_OPTPAR_PM_STATE |
2323 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
2324 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY |
2325 MLX5_QP_OPTPAR_SRQN |
2326 MLX5_QP_OPTPAR_CQN_RCV,
2329 [MLX5_QP_STATE_SQER] = {
2330 [MLX5_QP_STATE_RTS] = {
2331 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2332 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
2333 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
2334 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT |
2335 MLX5_QP_OPTPAR_RWE |
2336 MLX5_QP_OPTPAR_RAE |
2342 static int ib_nr_to_mlx5_nr(int ib_mask)
2347 case IB_QP_CUR_STATE:
2349 case IB_QP_EN_SQD_ASYNC_NOTIFY:
2351 case IB_QP_ACCESS_FLAGS:
2352 return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2354 case IB_QP_PKEY_INDEX:
2355 return MLX5_QP_OPTPAR_PKEY_INDEX;
2357 return MLX5_QP_OPTPAR_PRI_PORT;
2359 return MLX5_QP_OPTPAR_Q_KEY;
2361 return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2362 MLX5_QP_OPTPAR_PRI_PORT;
2363 case IB_QP_PATH_MTU:
2366 return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2367 case IB_QP_RETRY_CNT:
2368 return MLX5_QP_OPTPAR_RETRY_COUNT;
2369 case IB_QP_RNR_RETRY:
2370 return MLX5_QP_OPTPAR_RNR_RETRY;
2373 case IB_QP_MAX_QP_RD_ATOMIC:
2374 return MLX5_QP_OPTPAR_SRA_MAX;
2375 case IB_QP_ALT_PATH:
2376 return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2377 case IB_QP_MIN_RNR_TIMER:
2378 return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2381 case IB_QP_MAX_DEST_RD_ATOMIC:
2382 return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2383 MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2384 case IB_QP_PATH_MIG_STATE:
2385 return MLX5_QP_OPTPAR_PM_STATE;
2388 case IB_QP_DEST_QPN:
2394 static int ib_mask_to_mlx5_opt(int ib_mask)
2399 for (i = 0; i < 8 * sizeof(int); i++) {
2400 if ((1 << i) & ib_mask)
2401 result |= ib_nr_to_mlx5_nr(1 << i);
2407 static int modify_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
2408 struct mlx5_ib_rq *rq, int new_state,
2409 const struct mlx5_modify_raw_qp_param *raw_qp_param)
2416 inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2417 in = mlx5_vzalloc(inlen);
2421 MLX5_SET(modify_rq_in, in, rqn, rq->base.mqp.qpn);
2422 MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2424 rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2425 MLX5_SET(rqc, rqc, state, new_state);
2427 if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2428 if (MLX5_CAP_GEN(dev->mdev, modify_rq_counters_set_id)) {
2429 MLX5_SET64(modify_rq_in, in, modify_bitmask,
2430 MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2431 MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2433 pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2437 err = mlx5_core_modify_rq(dev->mdev, in, inlen);
2441 rq->state = new_state;
2448 static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2449 struct mlx5_ib_sq *sq, int new_state)
2456 inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2457 in = mlx5_vzalloc(inlen);
2461 MLX5_SET(modify_sq_in, in, sqn, sq->base.mqp.qpn);
2462 MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2464 sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2465 MLX5_SET(sqc, sqc, state, new_state);
2467 err = mlx5_core_modify_sq(dev, in, inlen);
2471 sq->state = new_state;
2478 static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2479 const struct mlx5_modify_raw_qp_param *raw_qp_param,
2482 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2483 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2484 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2489 switch (raw_qp_param->operation) {
2490 case MLX5_CMD_OP_RST2INIT_QP:
2491 rq_state = MLX5_RQC_STATE_RDY;
2492 sq_state = MLX5_SQC_STATE_RDY;
2494 case MLX5_CMD_OP_2ERR_QP:
2495 rq_state = MLX5_RQC_STATE_ERR;
2496 sq_state = MLX5_SQC_STATE_ERR;
2498 case MLX5_CMD_OP_2RST_QP:
2499 rq_state = MLX5_RQC_STATE_RST;
2500 sq_state = MLX5_SQC_STATE_RST;
2502 case MLX5_CMD_OP_INIT2INIT_QP:
2503 case MLX5_CMD_OP_INIT2RTR_QP:
2504 case MLX5_CMD_OP_RTR2RTS_QP:
2505 case MLX5_CMD_OP_RTS2RTS_QP:
2506 if (raw_qp_param->set_mask)
2515 if (qp->rq.wqe_cnt) {
2516 err = modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param);
2521 if (qp->sq.wqe_cnt) {
2523 err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2529 return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
2535 static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2536 const struct ib_qp_attr *attr, int attr_mask,
2537 enum ib_qp_state cur_state, enum ib_qp_state new_state)
2539 static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2540 [MLX5_QP_STATE_RST] = {
2541 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2542 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2543 [MLX5_QP_STATE_INIT] = MLX5_CMD_OP_RST2INIT_QP,
2545 [MLX5_QP_STATE_INIT] = {
2546 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2547 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2548 [MLX5_QP_STATE_INIT] = MLX5_CMD_OP_INIT2INIT_QP,
2549 [MLX5_QP_STATE_RTR] = MLX5_CMD_OP_INIT2RTR_QP,
2551 [MLX5_QP_STATE_RTR] = {
2552 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2553 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2554 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_RTR2RTS_QP,
2556 [MLX5_QP_STATE_RTS] = {
2557 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2558 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2559 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_RTS2RTS_QP,
2561 [MLX5_QP_STATE_SQD] = {
2562 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2563 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2565 [MLX5_QP_STATE_SQER] = {
2566 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2567 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2568 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_SQERR2RTS_QP,
2570 [MLX5_QP_STATE_ERR] = {
2571 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2572 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2576 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2577 struct mlx5_ib_qp *qp = to_mqp(ibqp);
2578 struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
2579 struct mlx5_ib_cq *send_cq, *recv_cq;
2580 struct mlx5_qp_context *context;
2581 struct mlx5_ib_pd *pd;
2582 struct mlx5_ib_port *mibport = NULL;
2583 enum mlx5_qp_state mlx5_cur, mlx5_new;
2584 enum mlx5_qp_optpar optpar;
2590 context = kzalloc(sizeof(*context), GFP_KERNEL);
2594 err = to_mlx5_st(ibqp->qp_type);
2596 mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
2600 context->flags = cpu_to_be32(err << 16);
2602 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2603 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2605 switch (attr->path_mig_state) {
2606 case IB_MIG_MIGRATED:
2607 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2610 context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2613 context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2618 if (is_sqp(ibqp->qp_type)) {
2619 context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2620 } else if (ibqp->qp_type == IB_QPT_UD ||
2621 ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2622 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2623 } else if (attr_mask & IB_QP_PATH_MTU) {
2624 if (attr->path_mtu < IB_MTU_256 ||
2625 attr->path_mtu > IB_MTU_4096) {
2626 mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2630 context->mtu_msgmax = (attr->path_mtu << 5) |
2631 (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
2634 if (attr_mask & IB_QP_DEST_QPN)
2635 context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2637 if (attr_mask & IB_QP_PKEY_INDEX)
2638 context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
2640 /* todo implement counter_index functionality */
2642 if (is_sqp(ibqp->qp_type))
2643 context->pri_path.port = qp->port;
2645 if (attr_mask & IB_QP_PORT)
2646 context->pri_path.port = attr->port_num;
2648 if (attr_mask & IB_QP_AV) {
2649 err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
2650 attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
2651 attr_mask, 0, attr, false);
2656 if (attr_mask & IB_QP_TIMEOUT)
2657 context->pri_path.ackto_lt |= attr->timeout << 3;
2659 if (attr_mask & IB_QP_ALT_PATH) {
2660 err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2663 attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2670 get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2671 &send_cq, &recv_cq);
2673 context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2674 context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2675 context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2676 context->params1 = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2678 if (attr_mask & IB_QP_RNR_RETRY)
2679 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2681 if (attr_mask & IB_QP_RETRY_CNT)
2682 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2684 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2685 if (attr->max_rd_atomic)
2687 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2690 if (attr_mask & IB_QP_SQ_PSN)
2691 context->next_send_psn = cpu_to_be32(attr->sq_psn);
2693 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2694 if (attr->max_dest_rd_atomic)
2696 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2699 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2700 context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2702 if (attr_mask & IB_QP_MIN_RNR_TIMER)
2703 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2705 if (attr_mask & IB_QP_RQ_PSN)
2706 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2708 if (attr_mask & IB_QP_QKEY)
2709 context->qkey = cpu_to_be32(attr->qkey);
2711 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2712 context->db_rec_addr = cpu_to_be64(qp->db.dma);
2714 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
2715 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2720 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2721 u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2723 mibport = &dev->port[port_num];
2724 context->qp_counter_set_usr_page |=
2725 cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
2728 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2729 context->sq_crq_size |= cpu_to_be16(1 << 4);
2731 if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2732 context->deth_sqpn = cpu_to_be32(1);
2734 mlx5_cur = to_mlx5_state(cur_state);
2735 mlx5_new = to_mlx5_state(new_state);
2736 mlx5_st = to_mlx5_st(ibqp->qp_type);
2740 /* If moving to a reset or error state, we must disable page faults on
2741 * this QP and flush all current page faults. Otherwise a stale page
2742 * fault may attempt to work on this QP after it is reset and moved
2743 * again to RTS, and may cause the driver and the device to get out of
2745 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2746 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2747 (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2748 mlx5_ib_qp_disable_pagefaults(qp);
2750 if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2751 !optab[mlx5_cur][mlx5_new])
2754 op = optab[mlx5_cur][mlx5_new];
2755 optpar = ib_mask_to_mlx5_opt(attr_mask);
2756 optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
2758 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2759 struct mlx5_modify_raw_qp_param raw_qp_param = {};
2761 raw_qp_param.operation = op;
2762 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2763 raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2764 raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2766 err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
2768 err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
2775 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2776 (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
2777 mlx5_ib_qp_enable_pagefaults(qp);
2779 qp->state = new_state;
2781 if (attr_mask & IB_QP_ACCESS_FLAGS)
2782 qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
2783 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2784 qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
2785 if (attr_mask & IB_QP_PORT)
2786 qp->port = attr->port_num;
2787 if (attr_mask & IB_QP_ALT_PATH)
2788 qp->trans_qp.alt_port = attr->alt_port_num;
2791 * If we moved a kernel QP to RESET, clean up all old CQ
2792 * entries and reinitialize the QP.
2794 if (new_state == IB_QPS_RESET && !ibqp->uobject) {
2795 mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
2796 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2797 if (send_cq != recv_cq)
2798 mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
2804 qp->sq.cur_post = 0;
2805 qp->sq.last_poll = 0;
2806 qp->db.db[MLX5_RCV_DBR] = 0;
2807 qp->db.db[MLX5_SND_DBR] = 0;
2815 int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2816 int attr_mask, struct ib_udata *udata)
2818 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2819 struct mlx5_ib_qp *qp = to_mqp(ibqp);
2820 enum ib_qp_type qp_type;
2821 enum ib_qp_state cur_state, new_state;
2824 enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
2826 if (ibqp->rwq_ind_tbl)
2829 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2830 return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2832 qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2833 IB_QPT_GSI : ibqp->qp_type;
2835 mutex_lock(&qp->mutex);
2837 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2838 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2840 if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2841 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2842 ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2845 if (qp_type != MLX5_IB_QPT_REG_UMR &&
2846 !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
2847 mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2848 cur_state, new_state, ibqp->qp_type, attr_mask);
2852 if ((attr_mask & IB_QP_PORT) &&
2853 (attr->port_num == 0 ||
2854 attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2855 mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2856 attr->port_num, dev->num_ports);
2860 if (attr_mask & IB_QP_PKEY_INDEX) {
2861 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2862 if (attr->pkey_index >=
2863 dev->mdev->port_caps[port - 1].pkey_table_len) {
2864 mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2870 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2871 attr->max_rd_atomic >
2872 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2873 mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2874 attr->max_rd_atomic);
2878 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2879 attr->max_dest_rd_atomic >
2880 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2881 mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2882 attr->max_dest_rd_atomic);
2886 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2891 err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2894 mutex_unlock(&qp->mutex);
2898 static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2900 struct mlx5_ib_cq *cq;
2903 cur = wq->head - wq->tail;
2904 if (likely(cur + nreq < wq->max_post))
2908 spin_lock(&cq->lock);
2909 cur = wq->head - wq->tail;
2910 spin_unlock(&cq->lock);
2912 return cur + nreq >= wq->max_post;
2915 static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2916 u64 remote_addr, u32 rkey)
2918 rseg->raddr = cpu_to_be64(remote_addr);
2919 rseg->rkey = cpu_to_be32(rkey);
2923 static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2924 struct ib_send_wr *wr, void *qend,
2925 struct mlx5_ib_qp *qp, int *size)
2929 memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2931 if (wr->send_flags & IB_SEND_IP_CSUM)
2932 eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2933 MLX5_ETH_WQE_L4_CSUM;
2935 seg += sizeof(struct mlx5_wqe_eth_seg);
2936 *size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2938 if (wr->opcode == IB_WR_LSO) {
2939 struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2940 int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2941 u64 left, leftlen, copysz;
2942 void *pdata = ud_wr->header;
2945 eseg->mss = cpu_to_be16(ud_wr->mss);
2946 eseg->inline_hdr_sz = cpu_to_be16(left);
2949 * check if there is space till the end of queue, if yes,
2950 * copy all in one shot, otherwise copy till the end of queue,
2951 * rollback and than the copy the left
2953 leftlen = qend - (void *)eseg->inline_hdr_start;
2954 copysz = min_t(u64, leftlen, left);
2956 memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2958 if (likely(copysz > size_of_inl_hdr_start)) {
2959 seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2960 *size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2963 if (unlikely(copysz < left)) { /* the last wqe in the queue */
2964 seg = mlx5_get_send_wqe(qp, 0);
2967 memcpy(seg, pdata, left);
2968 seg += ALIGN(left, 16);
2969 *size += ALIGN(left, 16) / 16;
2976 static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
2977 struct ib_send_wr *wr)
2979 memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
2980 dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
2981 dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
2984 static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
2986 dseg->byte_count = cpu_to_be32(sg->length);
2987 dseg->lkey = cpu_to_be32(sg->lkey);
2988 dseg->addr = cpu_to_be64(sg->addr);
2991 static __be16 get_klm_octo(int npages)
2993 return cpu_to_be16(ALIGN(npages, 8) / 2);
2996 static __be64 frwr_mkey_mask(void)
3000 result = MLX5_MKEY_MASK_LEN |
3001 MLX5_MKEY_MASK_PAGE_SIZE |
3002 MLX5_MKEY_MASK_START_ADDR |
3003 MLX5_MKEY_MASK_EN_RINVAL |
3004 MLX5_MKEY_MASK_KEY |
3010 MLX5_MKEY_MASK_SMALL_FENCE |
3011 MLX5_MKEY_MASK_FREE;
3013 return cpu_to_be64(result);
3016 static __be64 sig_mkey_mask(void)
3020 result = MLX5_MKEY_MASK_LEN |
3021 MLX5_MKEY_MASK_PAGE_SIZE |
3022 MLX5_MKEY_MASK_START_ADDR |
3023 MLX5_MKEY_MASK_EN_SIGERR |
3024 MLX5_MKEY_MASK_EN_RINVAL |
3025 MLX5_MKEY_MASK_KEY |
3030 MLX5_MKEY_MASK_SMALL_FENCE |
3031 MLX5_MKEY_MASK_FREE |
3032 MLX5_MKEY_MASK_BSF_EN;
3034 return cpu_to_be64(result);
3037 static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3038 struct mlx5_ib_mr *mr)
3040 int ndescs = mr->ndescs;
3042 memset(umr, 0, sizeof(*umr));
3044 if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3045 /* KLMs take twice the size of MTTs */
3048 umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3049 umr->klm_octowords = get_klm_octo(ndescs);
3050 umr->mkey_mask = frwr_mkey_mask();
3053 static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
3055 memset(umr, 0, sizeof(*umr));
3056 umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3057 umr->flags = 1 << 7;
3060 static __be64 get_umr_reg_mr_mask(void)
3064 result = MLX5_MKEY_MASK_LEN |
3065 MLX5_MKEY_MASK_PAGE_SIZE |
3066 MLX5_MKEY_MASK_START_ADDR |
3070 MLX5_MKEY_MASK_KEY |
3074 MLX5_MKEY_MASK_FREE;
3076 return cpu_to_be64(result);
3079 static __be64 get_umr_unreg_mr_mask(void)
3083 result = MLX5_MKEY_MASK_FREE;
3085 return cpu_to_be64(result);
3088 static __be64 get_umr_update_mtt_mask(void)
3092 result = MLX5_MKEY_MASK_FREE;
3094 return cpu_to_be64(result);
3097 static __be64 get_umr_update_translation_mask(void)
3101 result = MLX5_MKEY_MASK_LEN |
3102 MLX5_MKEY_MASK_PAGE_SIZE |
3103 MLX5_MKEY_MASK_START_ADDR |
3104 MLX5_MKEY_MASK_KEY |
3105 MLX5_MKEY_MASK_FREE;
3107 return cpu_to_be64(result);
3110 static __be64 get_umr_update_access_mask(void)
3114 result = MLX5_MKEY_MASK_LW |
3118 MLX5_MKEY_MASK_KEY |
3119 MLX5_MKEY_MASK_FREE;
3121 return cpu_to_be64(result);
3124 static __be64 get_umr_update_pd_mask(void)
3128 result = MLX5_MKEY_MASK_PD |
3129 MLX5_MKEY_MASK_KEY |
3130 MLX5_MKEY_MASK_FREE;
3132 return cpu_to_be64(result);
3135 static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3136 struct ib_send_wr *wr)
3138 struct mlx5_umr_wr *umrwr = umr_wr(wr);
3140 memset(umr, 0, sizeof(*umr));
3142 if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3143 umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3145 umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3147 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
3148 umr->klm_octowords = get_klm_octo(umrwr->npages);
3149 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3150 umr->mkey_mask = get_umr_update_mtt_mask();
3151 umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3152 umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
3154 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3155 umr->mkey_mask |= get_umr_update_translation_mask();
3156 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3157 umr->mkey_mask |= get_umr_update_access_mask();
3158 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3159 umr->mkey_mask |= get_umr_update_pd_mask();
3160 if (!umr->mkey_mask)
3161 umr->mkey_mask = get_umr_reg_mr_mask();
3163 umr->mkey_mask = get_umr_unreg_mr_mask();
3167 umr->flags |= MLX5_UMR_INLINE;
3170 static u8 get_umr_flags(int acc)
3172 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC : 0) |
3173 (acc & IB_ACCESS_REMOTE_WRITE ? MLX5_PERM_REMOTE_WRITE : 0) |
3174 (acc & IB_ACCESS_REMOTE_READ ? MLX5_PERM_REMOTE_READ : 0) |
3175 (acc & IB_ACCESS_LOCAL_WRITE ? MLX5_PERM_LOCAL_WRITE : 0) |
3176 MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
3179 static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3180 struct mlx5_ib_mr *mr,
3181 u32 key, int access)
3183 int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3185 memset(seg, 0, sizeof(*seg));
3187 if (mr->access_mode == MLX5_ACCESS_MODE_MTT)
3188 seg->log2_page_size = ilog2(mr->ibmr.page_size);
3189 else if (mr->access_mode == MLX5_ACCESS_MODE_KLM)
3190 /* KLMs take twice the size of MTTs */
3193 seg->flags = get_umr_flags(access) | mr->access_mode;
3194 seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3195 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3196 seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3197 seg->len = cpu_to_be64(mr->ibmr.length);
3198 seg->xlt_oct_size = cpu_to_be32(ndescs);
3201 static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
3203 memset(seg, 0, sizeof(*seg));
3204 seg->status = MLX5_MKEY_STATUS_FREE;
3207 static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
3209 struct mlx5_umr_wr *umrwr = umr_wr(wr);
3211 memset(seg, 0, sizeof(*seg));
3212 if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
3213 seg->status = MLX5_MKEY_STATUS_FREE;
3217 seg->flags = convert_access(umrwr->access_flags);
3218 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
3220 seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
3221 seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3223 seg->len = cpu_to_be64(umrwr->length);
3224 seg->log2_page_size = umrwr->page_shift;
3225 seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
3226 mlx5_mkey_variant(umrwr->mkey));
3229 static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3230 struct mlx5_ib_mr *mr,
3231 struct mlx5_ib_pd *pd)
3233 int bcount = mr->desc_size * mr->ndescs;
3235 dseg->addr = cpu_to_be64(mr->desc_map);
3236 dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3237 dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3240 static __be32 send_ieth(struct ib_send_wr *wr)
3242 switch (wr->opcode) {
3243 case IB_WR_SEND_WITH_IMM:
3244 case IB_WR_RDMA_WRITE_WITH_IMM:
3245 return wr->ex.imm_data;
3247 case IB_WR_SEND_WITH_INV:
3248 return cpu_to_be32(wr->ex.invalidate_rkey);
3255 static u8 calc_sig(void *wqe, int size)
3261 for (i = 0; i < size; i++)
3267 static u8 wq_sig(void *wqe)
3269 return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3272 static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
3275 struct mlx5_wqe_inline_seg *seg;
3276 void *qend = qp->sq.qend;
3284 wqe += sizeof(*seg);
3285 for (i = 0; i < wr->num_sge; i++) {
3286 addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3287 len = wr->sg_list[i].length;
3290 if (unlikely(inl > qp->max_inline_data))
3293 if (unlikely(wqe + len > qend)) {
3295 memcpy(wqe, addr, copy);
3298 wqe = mlx5_get_send_wqe(qp, 0);
3300 memcpy(wqe, addr, len);
3304 seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3306 *sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3311 static u16 prot_field_size(enum ib_signature_type type)
3314 case IB_SIG_TYPE_T10_DIF:
3315 return MLX5_DIF_SIZE;
3321 static u8 bs_selector(int block_size)
3323 switch (block_size) {
3324 case 512: return 0x1;
3325 case 520: return 0x2;
3326 case 4096: return 0x3;
3327 case 4160: return 0x4;
3328 case 1073741824: return 0x5;
3333 static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3334 struct mlx5_bsf_inl *inl)
3336 /* Valid inline section and allow BSF refresh */
3337 inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3338 MLX5_BSF_REFRESH_DIF);
3339 inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3340 inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
3341 /* repeating block */
3342 inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3343 inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3344 MLX5_DIF_CRC : MLX5_DIF_IPCS;
3346 if (domain->sig.dif.ref_remap)
3347 inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
3349 if (domain->sig.dif.app_escape) {
3350 if (domain->sig.dif.ref_escape)
3351 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3353 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
3356 inl->dif_app_bitmask_check =
3357 cpu_to_be16(domain->sig.dif.apptag_check_mask);
3360 static int mlx5_set_bsf(struct ib_mr *sig_mr,
3361 struct ib_sig_attrs *sig_attrs,
3362 struct mlx5_bsf *bsf, u32 data_size)
3364 struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3365 struct mlx5_bsf_basic *basic = &bsf->basic;
3366 struct ib_sig_domain *mem = &sig_attrs->mem;
3367 struct ib_sig_domain *wire = &sig_attrs->wire;
3369 memset(bsf, 0, sizeof(*bsf));
3371 /* Basic + Extended + Inline */
3372 basic->bsf_size_sbs = 1 << 7;
3373 /* Input domain check byte mask */
3374 basic->check_byte_mask = sig_attrs->check_mask;
3375 basic->raw_data_size = cpu_to_be32(data_size);
3378 switch (sig_attrs->mem.sig_type) {
3379 case IB_SIG_TYPE_NONE:
3381 case IB_SIG_TYPE_T10_DIF:
3382 basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3383 basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3384 mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3391 switch (sig_attrs->wire.sig_type) {
3392 case IB_SIG_TYPE_NONE:
3394 case IB_SIG_TYPE_T10_DIF:
3395 if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
3396 mem->sig_type == wire->sig_type) {
3397 /* Same block structure */
3398 basic->bsf_size_sbs |= 1 << 4;
3399 if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
3400 basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
3401 if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
3402 basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
3403 if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
3404 basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
3406 basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3408 basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
3409 mlx5_fill_inl_bsf(wire, &bsf->w_inl);
3418 static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3419 struct mlx5_ib_qp *qp, void **seg, int *size)
3421 struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3422 struct ib_mr *sig_mr = wr->sig_mr;
3423 struct mlx5_bsf *bsf;
3424 u32 data_len = wr->wr.sg_list->length;
3425 u32 data_key = wr->wr.sg_list->lkey;
3426 u64 data_va = wr->wr.sg_list->addr;
3431 (data_key == wr->prot->lkey &&
3432 data_va == wr->prot->addr &&
3433 data_len == wr->prot->length)) {
3435 * Source domain doesn't contain signature information
3436 * or data and protection are interleaved in memory.
3437 * So need construct:
3438 * ------------------
3440 * ------------------
3442 * ------------------
3444 struct mlx5_klm *data_klm = *seg;
3446 data_klm->bcount = cpu_to_be32(data_len);
3447 data_klm->key = cpu_to_be32(data_key);
3448 data_klm->va = cpu_to_be64(data_va);
3449 wqe_size = ALIGN(sizeof(*data_klm), 64);
3452 * Source domain contains signature information
3453 * So need construct a strided block format:
3454 * ---------------------------
3455 * | stride_block_ctrl |
3456 * ---------------------------
3458 * ---------------------------
3460 * ---------------------------
3462 * ---------------------------
3464 struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3465 struct mlx5_stride_block_entry *data_sentry;
3466 struct mlx5_stride_block_entry *prot_sentry;
3467 u32 prot_key = wr->prot->lkey;
3468 u64 prot_va = wr->prot->addr;
3469 u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3473 data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3474 prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3476 prot_size = prot_field_size(sig_attrs->mem.sig_type);
3478 pr_err("Bad block size given: %u\n", block_size);
3481 sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3483 sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3484 sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3485 sblock_ctrl->num_entries = cpu_to_be16(2);
3487 data_sentry->bcount = cpu_to_be16(block_size);
3488 data_sentry->key = cpu_to_be32(data_key);
3489 data_sentry->va = cpu_to_be64(data_va);
3490 data_sentry->stride = cpu_to_be16(block_size);
3492 prot_sentry->bcount = cpu_to_be16(prot_size);
3493 prot_sentry->key = cpu_to_be32(prot_key);
3494 prot_sentry->va = cpu_to_be64(prot_va);
3495 prot_sentry->stride = cpu_to_be16(prot_size);
3497 wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3498 sizeof(*prot_sentry), 64);
3502 *size += wqe_size / 16;
3503 if (unlikely((*seg == qp->sq.qend)))
3504 *seg = mlx5_get_send_wqe(qp, 0);
3507 ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3511 *seg += sizeof(*bsf);
3512 *size += sizeof(*bsf) / 16;
3513 if (unlikely((*seg == qp->sq.qend)))
3514 *seg = mlx5_get_send_wqe(qp, 0);
3519 static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
3520 struct ib_sig_handover_wr *wr, u32 nelements,
3521 u32 length, u32 pdn)
3523 struct ib_mr *sig_mr = wr->sig_mr;
3524 u32 sig_key = sig_mr->rkey;
3525 u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
3527 memset(seg, 0, sizeof(*seg));
3529 seg->flags = get_umr_flags(wr->access_flags) |
3530 MLX5_ACCESS_MODE_KLM;
3531 seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
3532 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
3533 MLX5_MKEY_BSF_EN | pdn);
3534 seg->len = cpu_to_be64(length);
3535 seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3536 seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3539 static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3542 memset(umr, 0, sizeof(*umr));
3544 umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3545 umr->klm_octowords = get_klm_octo(nelements);
3546 umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3547 umr->mkey_mask = sig_mkey_mask();
3551 static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
3552 void **seg, int *size)
3554 struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3555 struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
3556 u32 pdn = get_pd(qp)->pdn;
3558 int region_len, ret;
3560 if (unlikely(wr->wr.num_sge != 1) ||
3561 unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
3562 unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3563 unlikely(!sig_mr->sig->sig_status_checked))
3566 /* length of the protected region, data + protection */
3567 region_len = wr->wr.sg_list->length;
3569 (wr->prot->lkey != wr->wr.sg_list->lkey ||
3570 wr->prot->addr != wr->wr.sg_list->addr ||
3571 wr->prot->length != wr->wr.sg_list->length))
3572 region_len += wr->prot->length;
3575 * KLM octoword size - if protection was provided
3576 * then we use strided block format (3 octowords),
3577 * else we use single KLM (1 octoword)
3579 klm_oct_size = wr->prot ? 3 : 1;
3581 set_sig_umr_segment(*seg, klm_oct_size);
3582 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3583 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3584 if (unlikely((*seg == qp->sq.qend)))
3585 *seg = mlx5_get_send_wqe(qp, 0);
3587 set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3588 *seg += sizeof(struct mlx5_mkey_seg);
3589 *size += sizeof(struct mlx5_mkey_seg) / 16;
3590 if (unlikely((*seg == qp->sq.qend)))
3591 *seg = mlx5_get_send_wqe(qp, 0);
3593 ret = set_sig_data_segment(wr, qp, seg, size);
3597 sig_mr->sig->sig_status_checked = false;
3601 static int set_psv_wr(struct ib_sig_domain *domain,
3602 u32 psv_idx, void **seg, int *size)
3604 struct mlx5_seg_set_psv *psv_seg = *seg;
3606 memset(psv_seg, 0, sizeof(*psv_seg));
3607 psv_seg->psv_num = cpu_to_be32(psv_idx);
3608 switch (domain->sig_type) {
3609 case IB_SIG_TYPE_NONE:
3611 case IB_SIG_TYPE_T10_DIF:
3612 psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3613 domain->sig.dif.app_tag);
3614 psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
3617 pr_err("Bad signature type given.\n");
3621 *seg += sizeof(*psv_seg);
3622 *size += sizeof(*psv_seg) / 16;
3627 static int set_reg_wr(struct mlx5_ib_qp *qp,
3628 struct ib_reg_wr *wr,
3629 void **seg, int *size)
3631 struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3632 struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3634 if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3635 mlx5_ib_warn(to_mdev(qp->ibqp.device),
3636 "Invalid IB_SEND_INLINE send flag\n");
3640 set_reg_umr_seg(*seg, mr);
3641 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3642 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3643 if (unlikely((*seg == qp->sq.qend)))
3644 *seg = mlx5_get_send_wqe(qp, 0);
3646 set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3647 *seg += sizeof(struct mlx5_mkey_seg);
3648 *size += sizeof(struct mlx5_mkey_seg) / 16;
3649 if (unlikely((*seg == qp->sq.qend)))
3650 *seg = mlx5_get_send_wqe(qp, 0);
3652 set_reg_data_seg(*seg, mr, pd);
3653 *seg += sizeof(struct mlx5_wqe_data_seg);
3654 *size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3659 static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
3661 set_linv_umr_seg(*seg);
3662 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3663 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3664 if (unlikely((*seg == qp->sq.qend)))
3665 *seg = mlx5_get_send_wqe(qp, 0);
3666 set_linv_mkey_seg(*seg);
3667 *seg += sizeof(struct mlx5_mkey_seg);
3668 *size += sizeof(struct mlx5_mkey_seg) / 16;
3669 if (unlikely((*seg == qp->sq.qend)))
3670 *seg = mlx5_get_send_wqe(qp, 0);
3673 static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3679 pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3680 for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3681 if ((i & 0xf) == 0) {
3682 void *buf = mlx5_get_send_wqe(qp, tidx);
3683 tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3687 pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3688 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3689 be32_to_cpu(p[j + 3]));
3693 static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3694 unsigned bytecnt, struct mlx5_ib_qp *qp)
3696 while (bytecnt > 0) {
3697 __iowrite64_copy(dst++, src++, 8);
3698 __iowrite64_copy(dst++, src++, 8);
3699 __iowrite64_copy(dst++, src++, 8);
3700 __iowrite64_copy(dst++, src++, 8);
3701 __iowrite64_copy(dst++, src++, 8);
3702 __iowrite64_copy(dst++, src++, 8);
3703 __iowrite64_copy(dst++, src++, 8);
3704 __iowrite64_copy(dst++, src++, 8);
3706 if (unlikely(src == qp->sq.qend))
3707 src = mlx5_get_send_wqe(qp, 0);
3711 static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3713 if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3714 wr->send_flags & IB_SEND_FENCE))
3715 return MLX5_FENCE_MODE_STRONG_ORDERING;
3717 if (unlikely(fence)) {
3718 if (wr->send_flags & IB_SEND_FENCE)
3719 return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3722 } else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3723 return MLX5_FENCE_MODE_FENCE;
3729 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3730 struct mlx5_wqe_ctrl_seg **ctrl,
3731 struct ib_send_wr *wr, unsigned *idx,
3732 int *size, int nreq)
3734 if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3737 *idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3738 *seg = mlx5_get_send_wqe(qp, *idx);
3740 *(uint32_t *)(*seg + 8) = 0;
3741 (*ctrl)->imm = send_ieth(wr);
3742 (*ctrl)->fm_ce_se = qp->sq_signal_bits |
3743 (wr->send_flags & IB_SEND_SIGNALED ?
3744 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3745 (wr->send_flags & IB_SEND_SOLICITED ?
3746 MLX5_WQE_CTRL_SOLICITED : 0);
3748 *seg += sizeof(**ctrl);
3749 *size = sizeof(**ctrl) / 16;
3754 static void finish_wqe(struct mlx5_ib_qp *qp,
3755 struct mlx5_wqe_ctrl_seg *ctrl,
3756 u8 size, unsigned idx, u64 wr_id,
3757 int nreq, u8 fence, u8 next_fence,
3762 ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3763 mlx5_opcode | ((u32)opmod << 24));
3764 ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
3765 ctrl->fm_ce_se |= fence;
3766 qp->fm_cache = next_fence;
3767 if (unlikely(qp->wq_sig))
3768 ctrl->signature = wq_sig(ctrl);
3770 qp->sq.wrid[idx] = wr_id;
3771 qp->sq.w_list[idx].opcode = mlx5_opcode;
3772 qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3773 qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3774 qp->sq.w_list[idx].next = qp->sq.cur_post;
3778 int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3779 struct ib_send_wr **bad_wr)
3781 struct mlx5_wqe_ctrl_seg *ctrl = NULL; /* compiler warning */
3782 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
3783 struct mlx5_core_dev *mdev = dev->mdev;
3784 struct mlx5_ib_qp *qp;
3785 struct mlx5_ib_mr *mr;
3786 struct mlx5_wqe_data_seg *dpseg;
3787 struct mlx5_wqe_xrc_seg *xrc;
3789 int uninitialized_var(size);
3791 unsigned long flags;
3802 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3803 return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3809 spin_lock_irqsave(&qp->sq.lock, flags);
3811 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3818 for (nreq = 0; wr; nreq++, wr = wr->next) {
3819 if (unlikely(wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
3820 mlx5_ib_warn(dev, "\n");
3826 fence = qp->fm_cache;
3827 num_sge = wr->num_sge;
3828 if (unlikely(num_sge > qp->sq.max_gs)) {
3829 mlx5_ib_warn(dev, "\n");
3835 err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3837 mlx5_ib_warn(dev, "\n");
3843 switch (ibqp->qp_type) {
3844 case IB_QPT_XRC_INI:
3846 seg += sizeof(*xrc);
3847 size += sizeof(*xrc) / 16;
3850 switch (wr->opcode) {
3851 case IB_WR_RDMA_READ:
3852 case IB_WR_RDMA_WRITE:
3853 case IB_WR_RDMA_WRITE_WITH_IMM:
3854 set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3856 seg += sizeof(struct mlx5_wqe_raddr_seg);
3857 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3860 case IB_WR_ATOMIC_CMP_AND_SWP:
3861 case IB_WR_ATOMIC_FETCH_AND_ADD:
3862 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3863 mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3868 case IB_WR_LOCAL_INV:
3869 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3870 qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3871 ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
3872 set_linv_wr(qp, &seg, &size);
3877 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3878 qp->sq.wr_data[idx] = IB_WR_REG_MR;
3879 ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3880 err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3888 case IB_WR_REG_SIG_MR:
3889 qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
3890 mr = to_mmr(sig_handover_wr(wr)->sig_mr);
3892 ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3893 err = set_sig_umr_wr(wr, qp, &seg, &size);
3895 mlx5_ib_warn(dev, "\n");
3900 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3901 nreq, get_fence(fence, wr),
3902 next_fence, MLX5_OPCODE_UMR);
3904 * SET_PSV WQEs are not signaled and solicited
3907 wr->send_flags &= ~IB_SEND_SIGNALED;
3908 wr->send_flags |= IB_SEND_SOLICITED;
3909 err = begin_wqe(qp, &seg, &ctrl, wr,
3912 mlx5_ib_warn(dev, "\n");
3918 err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
3919 mr->sig->psv_memory.psv_idx, &seg,
3922 mlx5_ib_warn(dev, "\n");
3927 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3928 nreq, get_fence(fence, wr),
3929 next_fence, MLX5_OPCODE_SET_PSV);
3930 err = begin_wqe(qp, &seg, &ctrl, wr,
3933 mlx5_ib_warn(dev, "\n");
3939 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3940 err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
3941 mr->sig->psv_wire.psv_idx, &seg,
3944 mlx5_ib_warn(dev, "\n");
3949 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3950 nreq, get_fence(fence, wr),
3951 next_fence, MLX5_OPCODE_SET_PSV);
3961 switch (wr->opcode) {
3962 case IB_WR_RDMA_WRITE:
3963 case IB_WR_RDMA_WRITE_WITH_IMM:
3964 set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3966 seg += sizeof(struct mlx5_wqe_raddr_seg);
3967 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3976 case MLX5_IB_QPT_HW_GSI:
3977 set_datagram_seg(seg, wr);
3978 seg += sizeof(struct mlx5_wqe_datagram_seg);
3979 size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3980 if (unlikely((seg == qend)))
3981 seg = mlx5_get_send_wqe(qp, 0);
3984 set_datagram_seg(seg, wr);
3985 seg += sizeof(struct mlx5_wqe_datagram_seg);
3986 size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3988 if (unlikely((seg == qend)))
3989 seg = mlx5_get_send_wqe(qp, 0);
3991 /* handle qp that supports ud offload */
3992 if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
3993 struct mlx5_wqe_eth_pad *pad;
3996 memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
3997 seg += sizeof(struct mlx5_wqe_eth_pad);
3998 size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4000 seg = set_eth_seg(seg, wr, qend, qp, &size);
4002 if (unlikely((seg == qend)))
4003 seg = mlx5_get_send_wqe(qp, 0);
4006 case MLX5_IB_QPT_REG_UMR:
4007 if (wr->opcode != MLX5_IB_WR_UMR) {
4009 mlx5_ib_warn(dev, "bad opcode\n");
4012 qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
4013 ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
4014 set_reg_umr_segment(seg, wr);
4015 seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4016 size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4017 if (unlikely((seg == qend)))
4018 seg = mlx5_get_send_wqe(qp, 0);
4019 set_reg_mkey_segment(seg, wr);
4020 seg += sizeof(struct mlx5_mkey_seg);
4021 size += sizeof(struct mlx5_mkey_seg) / 16;
4022 if (unlikely((seg == qend)))
4023 seg = mlx5_get_send_wqe(qp, 0);
4030 if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4031 int uninitialized_var(sz);
4033 err = set_data_inl_seg(qp, wr, seg, &sz);
4034 if (unlikely(err)) {
4035 mlx5_ib_warn(dev, "\n");
4043 for (i = 0; i < num_sge; i++) {
4044 if (unlikely(dpseg == qend)) {
4045 seg = mlx5_get_send_wqe(qp, 0);
4048 if (likely(wr->sg_list[i].length)) {
4049 set_data_ptr_seg(dpseg, wr->sg_list + i);
4050 size += sizeof(struct mlx5_wqe_data_seg) / 16;
4056 finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4057 get_fence(fence, wr), next_fence,
4058 mlx5_ib_opcode[wr->opcode]);
4061 dump_wqe(qp, idx, size);
4066 qp->sq.head += nreq;
4068 /* Make sure that descriptors are written before
4069 * updating doorbell record and ringing the doorbell
4073 qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4075 /* Make sure doorbell record is visible to the HCA before
4076 * we hit doorbell */
4080 spin_lock(&bf->lock);
4082 __acquire(&bf->lock);
4085 if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
4086 mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
4089 mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
4090 MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4091 /* Make sure doorbells don't leak out of SQ spinlock
4092 * and reach the HCA out of order.
4096 bf->offset ^= bf->buf_size;
4098 spin_unlock(&bf->lock);
4100 __release(&bf->lock);
4103 spin_unlock_irqrestore(&qp->sq.lock, flags);
4108 static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4110 sig->signature = calc_sig(sig, size);
4113 int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
4114 struct ib_recv_wr **bad_wr)
4116 struct mlx5_ib_qp *qp = to_mqp(ibqp);
4117 struct mlx5_wqe_data_seg *scat;
4118 struct mlx5_rwqe_sig *sig;
4119 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4120 struct mlx5_core_dev *mdev = dev->mdev;
4121 unsigned long flags;
4127 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4128 return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4130 spin_lock_irqsave(&qp->rq.lock, flags);
4132 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4139 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4141 for (nreq = 0; wr; nreq++, wr = wr->next) {
4142 if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4148 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4154 scat = get_recv_wqe(qp, ind);
4158 for (i = 0; i < wr->num_sge; i++)
4159 set_data_ptr_seg(scat + i, wr->sg_list + i);
4161 if (i < qp->rq.max_gs) {
4162 scat[i].byte_count = 0;
4163 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
4168 sig = (struct mlx5_rwqe_sig *)scat;
4169 set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4172 qp->rq.wrid[ind] = wr->wr_id;
4174 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4179 qp->rq.head += nreq;
4181 /* Make sure that descriptors are written before
4186 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4189 spin_unlock_irqrestore(&qp->rq.lock, flags);
4194 static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4196 switch (mlx5_state) {
4197 case MLX5_QP_STATE_RST: return IB_QPS_RESET;
4198 case MLX5_QP_STATE_INIT: return IB_QPS_INIT;
4199 case MLX5_QP_STATE_RTR: return IB_QPS_RTR;
4200 case MLX5_QP_STATE_RTS: return IB_QPS_RTS;
4201 case MLX5_QP_STATE_SQ_DRAINING:
4202 case MLX5_QP_STATE_SQD: return IB_QPS_SQD;
4203 case MLX5_QP_STATE_SQER: return IB_QPS_SQE;
4204 case MLX5_QP_STATE_ERR: return IB_QPS_ERR;
4209 static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4211 switch (mlx5_mig_state) {
4212 case MLX5_QP_PM_ARMED: return IB_MIG_ARMED;
4213 case MLX5_QP_PM_REARM: return IB_MIG_REARM;
4214 case MLX5_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
4219 static int to_ib_qp_access_flags(int mlx5_flags)
4223 if (mlx5_flags & MLX5_QP_BIT_RRE)
4224 ib_flags |= IB_ACCESS_REMOTE_READ;
4225 if (mlx5_flags & MLX5_QP_BIT_RWE)
4226 ib_flags |= IB_ACCESS_REMOTE_WRITE;
4227 if (mlx5_flags & MLX5_QP_BIT_RAE)
4228 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4233 static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4234 struct mlx5_qp_path *path)
4236 struct mlx5_core_dev *dev = ibdev->mdev;
4238 memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4239 ib_ah_attr->port_num = path->port;
4241 if (ib_ah_attr->port_num == 0 ||
4242 ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
4245 ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
4247 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
4248 ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4249 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
4250 ib_ah_attr->ah_flags = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4251 if (ib_ah_attr->ah_flags) {
4252 ib_ah_attr->grh.sgid_index = path->mgid_index;
4253 ib_ah_attr->grh.hop_limit = path->hop_limit;
4254 ib_ah_attr->grh.traffic_class =
4255 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4256 ib_ah_attr->grh.flow_label =
4257 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4258 memcpy(ib_ah_attr->grh.dgid.raw,
4259 path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4263 static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4264 struct mlx5_ib_sq *sq,
4272 inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4273 out = mlx5_vzalloc(inlen);
4277 err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4281 sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4282 *sq_state = MLX5_GET(sqc, sqc, state);
4283 sq->state = *sq_state;
4290 static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4291 struct mlx5_ib_rq *rq,
4299 inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4300 out = mlx5_vzalloc(inlen);
4304 err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4308 rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4309 *rq_state = MLX5_GET(rqc, rqc, state);
4310 rq->state = *rq_state;
4317 static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4318 struct mlx5_ib_qp *qp, u8 *qp_state)
4320 static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4321 [MLX5_RQC_STATE_RST] = {
4322 [MLX5_SQC_STATE_RST] = IB_QPS_RESET,
4323 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE_BAD,
4324 [MLX5_SQC_STATE_ERR] = MLX5_QP_STATE_BAD,
4325 [MLX5_SQ_STATE_NA] = IB_QPS_RESET,
4327 [MLX5_RQC_STATE_RDY] = {
4328 [MLX5_SQC_STATE_RST] = MLX5_QP_STATE_BAD,
4329 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE,
4330 [MLX5_SQC_STATE_ERR] = IB_QPS_SQE,
4331 [MLX5_SQ_STATE_NA] = MLX5_QP_STATE,
4333 [MLX5_RQC_STATE_ERR] = {
4334 [MLX5_SQC_STATE_RST] = MLX5_QP_STATE_BAD,
4335 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE_BAD,
4336 [MLX5_SQC_STATE_ERR] = IB_QPS_ERR,
4337 [MLX5_SQ_STATE_NA] = IB_QPS_ERR,
4339 [MLX5_RQ_STATE_NA] = {
4340 [MLX5_SQC_STATE_RST] = IB_QPS_RESET,
4341 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE,
4342 [MLX5_SQC_STATE_ERR] = MLX5_QP_STATE,
4343 [MLX5_SQ_STATE_NA] = MLX5_QP_STATE_BAD,
4347 *qp_state = sqrq_trans[rq_state][sq_state];
4349 if (*qp_state == MLX5_QP_STATE_BAD) {
4350 WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4351 qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4352 qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4356 if (*qp_state == MLX5_QP_STATE)
4357 *qp_state = qp->state;
4362 static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4363 struct mlx5_ib_qp *qp,
4364 u8 *raw_packet_qp_state)
4366 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4367 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4368 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4370 u8 sq_state = MLX5_SQ_STATE_NA;
4371 u8 rq_state = MLX5_RQ_STATE_NA;
4373 if (qp->sq.wqe_cnt) {
4374 err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4379 if (qp->rq.wqe_cnt) {
4380 err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4385 return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4386 raw_packet_qp_state);
4389 static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4390 struct ib_qp_attr *qp_attr)
4392 int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
4393 struct mlx5_qp_context *context;
4398 outb = kzalloc(outlen, GFP_KERNEL);
4402 err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
4407 /* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4408 context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4410 mlx5_state = be32_to_cpu(context->flags) >> 28;
4412 qp->state = to_ib_qp_state(mlx5_state);
4413 qp_attr->path_mtu = context->mtu_msgmax >> 5;
4414 qp_attr->path_mig_state =
4415 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4416 qp_attr->qkey = be32_to_cpu(context->qkey);
4417 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4418 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
4419 qp_attr->dest_qp_num = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4420 qp_attr->qp_access_flags =
4421 to_ib_qp_access_flags(be32_to_cpu(context->params2));
4423 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4424 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4425 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
4426 qp_attr->alt_pkey_index =
4427 be16_to_cpu(context->alt_path.pkey_index);
4428 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
4431 qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
4432 qp_attr->port_num = context->pri_path.port;
4434 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4435 qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4437 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4439 qp_attr->max_dest_rd_atomic =
4440 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4441 qp_attr->min_rnr_timer =
4442 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4443 qp_attr->timeout = context->pri_path.ackto_lt >> 3;
4444 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
4445 qp_attr->rnr_retry = (be32_to_cpu(context->params1) >> 13) & 0x7;
4446 qp_attr->alt_timeout = context->alt_path.ackto_lt >> 3;
4453 int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4454 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4456 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4457 struct mlx5_ib_qp *qp = to_mqp(ibqp);
4459 u8 raw_packet_qp_state;
4461 if (ibqp->rwq_ind_tbl)
4464 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4465 return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4468 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4470 * Wait for any outstanding page faults, in case the user frees memory
4471 * based upon this query's result.
4473 flush_workqueue(mlx5_ib_page_fault_wq);
4476 mutex_lock(&qp->mutex);
4478 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4479 err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4482 qp->state = raw_packet_qp_state;
4483 qp_attr->port_num = 1;
4485 err = query_qp_attr(dev, qp, qp_attr);
4490 qp_attr->qp_state = qp->state;
4491 qp_attr->cur_qp_state = qp_attr->qp_state;
4492 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
4493 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
4495 if (!ibqp->uobject) {
4496 qp_attr->cap.max_send_wr = qp->sq.max_post;
4497 qp_attr->cap.max_send_sge = qp->sq.max_gs;
4498 qp_init_attr->qp_context = ibqp->qp_context;
4500 qp_attr->cap.max_send_wr = 0;
4501 qp_attr->cap.max_send_sge = 0;
4504 qp_init_attr->qp_type = ibqp->qp_type;
4505 qp_init_attr->recv_cq = ibqp->recv_cq;
4506 qp_init_attr->send_cq = ibqp->send_cq;
4507 qp_init_attr->srq = ibqp->srq;
4508 qp_attr->cap.max_inline_data = qp->max_inline_data;
4510 qp_init_attr->cap = qp_attr->cap;
4512 qp_init_attr->create_flags = 0;
4513 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4514 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4516 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4517 qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4518 if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4519 qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4520 if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4521 qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
4522 if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4523 qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
4525 qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4526 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4529 mutex_unlock(&qp->mutex);
4533 struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4534 struct ib_ucontext *context,
4535 struct ib_udata *udata)
4537 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4538 struct mlx5_ib_xrcd *xrcd;
4541 if (!MLX5_CAP_GEN(dev->mdev, xrc))
4542 return ERR_PTR(-ENOSYS);
4544 xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4546 return ERR_PTR(-ENOMEM);
4548 err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
4551 return ERR_PTR(-ENOMEM);
4554 return &xrcd->ibxrcd;
4557 int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4559 struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4560 u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4563 err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
4565 mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4574 static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4576 struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4577 struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4578 struct ib_event event;
4580 if (rwq->ibwq.event_handler) {
4581 event.device = rwq->ibwq.device;
4582 event.element.wq = &rwq->ibwq;
4584 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4585 event.event = IB_EVENT_WQ_FATAL;
4588 mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4592 rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4596 static int create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4597 struct ib_wq_init_attr *init_attr)
4599 struct mlx5_ib_dev *dev;
4607 dev = to_mdev(pd->device);
4609 inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4610 in = mlx5_vzalloc(inlen);
4614 rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4615 MLX5_SET(rqc, rqc, mem_rq_type,
4616 MLX5_RQC_RQ_TYPE_MEMORY_RQ_INLINE);
4617 MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4618 MLX5_SET(rqc, rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4619 MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
4620 MLX5_SET(rqc, rqc, flush_in_error_en, 1);
4621 wq = MLX5_ADDR_OF(rqc, rqc, wq);
4622 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4623 MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4624 MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4625 MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4626 MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4627 MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4628 MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4629 MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4630 MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4631 rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4632 mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
4633 err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
4638 static int set_user_rq_size(struct mlx5_ib_dev *dev,
4639 struct ib_wq_init_attr *wq_init_attr,
4640 struct mlx5_ib_create_wq *ucmd,
4641 struct mlx5_ib_rwq *rwq)
4643 /* Sanity check RQ size before proceeding */
4644 if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4647 if (!ucmd->rq_wqe_count)
4650 rwq->wqe_count = ucmd->rq_wqe_count;
4651 rwq->wqe_shift = ucmd->rq_wqe_shift;
4652 rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4653 rwq->log_rq_stride = rwq->wqe_shift;
4654 rwq->log_rq_size = ilog2(rwq->wqe_count);
4658 static int prepare_user_rq(struct ib_pd *pd,
4659 struct ib_wq_init_attr *init_attr,
4660 struct ib_udata *udata,
4661 struct mlx5_ib_rwq *rwq)
4663 struct mlx5_ib_dev *dev = to_mdev(pd->device);
4664 struct mlx5_ib_create_wq ucmd = {};
4666 size_t required_cmd_sz;
4668 required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4669 if (udata->inlen < required_cmd_sz) {
4670 mlx5_ib_dbg(dev, "invalid inlen\n");
4674 if (udata->inlen > sizeof(ucmd) &&
4675 !ib_is_udata_cleared(udata, sizeof(ucmd),
4676 udata->inlen - sizeof(ucmd))) {
4677 mlx5_ib_dbg(dev, "inlen is not supported\n");
4681 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4682 mlx5_ib_dbg(dev, "copy failed\n");
4686 if (ucmd.comp_mask) {
4687 mlx5_ib_dbg(dev, "invalid comp mask\n");
4691 if (ucmd.reserved) {
4692 mlx5_ib_dbg(dev, "invalid reserved\n");
4696 err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4698 mlx5_ib_dbg(dev, "err %d\n", err);
4702 err = create_user_rq(dev, pd, rwq, &ucmd);
4704 mlx5_ib_dbg(dev, "err %d\n", err);
4709 rwq->user_index = ucmd.user_index;
4713 struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4714 struct ib_wq_init_attr *init_attr,
4715 struct ib_udata *udata)
4717 struct mlx5_ib_dev *dev;
4718 struct mlx5_ib_rwq *rwq;
4719 struct mlx5_ib_create_wq_resp resp = {};
4720 size_t min_resp_len;
4724 return ERR_PTR(-ENOSYS);
4726 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4727 if (udata->outlen && udata->outlen < min_resp_len)
4728 return ERR_PTR(-EINVAL);
4730 dev = to_mdev(pd->device);
4731 switch (init_attr->wq_type) {
4733 rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4735 return ERR_PTR(-ENOMEM);
4736 err = prepare_user_rq(pd, init_attr, udata, rwq);
4739 err = create_rq(rwq, pd, init_attr);
4744 mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4745 init_attr->wq_type);
4746 return ERR_PTR(-EINVAL);
4749 rwq->ibwq.wq_num = rwq->core_qp.qpn;
4750 rwq->ibwq.state = IB_WQS_RESET;
4751 if (udata->outlen) {
4752 resp.response_length = offsetof(typeof(resp), response_length) +
4753 sizeof(resp.response_length);
4754 err = ib_copy_to_udata(udata, &resp, resp.response_length);
4759 rwq->core_qp.event = mlx5_ib_wq_event;
4760 rwq->ibwq.event_handler = init_attr->event_handler;
4764 mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4766 destroy_user_rq(pd, rwq);
4769 return ERR_PTR(err);
4772 int mlx5_ib_destroy_wq(struct ib_wq *wq)
4774 struct mlx5_ib_dev *dev = to_mdev(wq->device);
4775 struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4777 mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
4778 destroy_user_rq(wq->pd, rwq);
4784 struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4785 struct ib_rwq_ind_table_init_attr *init_attr,
4786 struct ib_udata *udata)
4788 struct mlx5_ib_dev *dev = to_mdev(device);
4789 struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4790 int sz = 1 << init_attr->log_ind_tbl_size;
4791 struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4792 size_t min_resp_len;
4799 if (udata->inlen > 0 &&
4800 !ib_is_udata_cleared(udata, 0,
4802 return ERR_PTR(-EOPNOTSUPP);
4804 if (init_attr->log_ind_tbl_size >
4805 MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
4806 mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
4807 init_attr->log_ind_tbl_size,
4808 MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
4809 return ERR_PTR(-EINVAL);
4812 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4813 if (udata->outlen && udata->outlen < min_resp_len)
4814 return ERR_PTR(-EINVAL);
4816 rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4818 return ERR_PTR(-ENOMEM);
4820 inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4821 in = mlx5_vzalloc(inlen);
4827 rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4829 MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4830 MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4832 for (i = 0; i < sz; i++)
4833 MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4835 err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4841 rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4842 if (udata->outlen) {
4843 resp.response_length = offsetof(typeof(resp), response_length) +
4844 sizeof(resp.response_length);
4845 err = ib_copy_to_udata(udata, &resp, resp.response_length);
4850 return &rwq_ind_tbl->ib_rwq_ind_tbl;
4853 mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4856 return ERR_PTR(err);
4859 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4861 struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4862 struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4864 mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4870 int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4871 u32 wq_attr_mask, struct ib_udata *udata)
4873 struct mlx5_ib_dev *dev = to_mdev(wq->device);
4874 struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4875 struct mlx5_ib_modify_wq ucmd = {};
4876 size_t required_cmd_sz;
4884 required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4885 if (udata->inlen < required_cmd_sz)
4888 if (udata->inlen > sizeof(ucmd) &&
4889 !ib_is_udata_cleared(udata, sizeof(ucmd),
4890 udata->inlen - sizeof(ucmd)))
4893 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4896 if (ucmd.comp_mask || ucmd.reserved)
4899 inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4900 in = mlx5_vzalloc(inlen);
4904 rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4906 MLX5_SET(modify_rq_in, in, rqn, rwq->core_qp.qpn);
4907 curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4908 wq_attr->curr_wq_state : wq->state;
4909 wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4910 wq_attr->wq_state : curr_wq_state;
4911 if (curr_wq_state == IB_WQS_ERR)
4912 curr_wq_state = MLX5_RQC_STATE_ERR;
4913 if (wq_state == IB_WQS_ERR)
4914 wq_state = MLX5_RQC_STATE_ERR;
4915 MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4916 MLX5_SET(rqc, rqc, state, wq_state);
4918 err = mlx5_core_modify_rq(dev->mdev, in, inlen);
4921 rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;