]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blame - drivers/infiniband/hw/mlx5/qp.c
projects / mirror_ubuntu-hirsute-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
IB/mlx5: Fix fatal error dispatching
[mirror_ubuntu-hirsute-kernel.git] / drivers / infiniband / hw / mlx5 / qp.c
CommitLineData
e126ba97 1/*
6cf0a15f 2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
e126ba97
EC		 3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <linux/module.h>
34#include <rdma/ib_umem.h>
2811ba51 35#include <rdma/ib_cache.h>
cfb5e088 36#include <rdma/ib_user_verbs.h>
e126ba97 37#include "mlx5_ib.h"
e126ba97
EC		 38
39/* not supported currently */
40static int wq_signature;
41
42enum {
43 MLX5_IB_ACK_REQ_FREQ = 8,
44};
45
46enum {
47 MLX5_IB_DEFAULT_SCHED_QUEUE = 0x83,
48 MLX5_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
49 MLX5_IB_LINK_TYPE_IB = 0,
50 MLX5_IB_LINK_TYPE_ETH = 1
51};
52
53enum {
54 MLX5_IB_SQ_STRIDE = 6,
55 MLX5_IB_CACHE_LINE_SIZE = 64,
56};
57
58static const u32 mlx5_ib_opcode[] = {
59 [IB_WR_SEND] = MLX5_OPCODE_SEND,
f0313965 60 [IB_WR_LSO] = MLX5_OPCODE_LSO,
e126ba97
EC		 61 [IB_WR_SEND_WITH_IMM] = MLX5_OPCODE_SEND_IMM,
62 [IB_WR_RDMA_WRITE] = MLX5_OPCODE_RDMA_WRITE,
63 [IB_WR_RDMA_WRITE_WITH_IMM] = MLX5_OPCODE_RDMA_WRITE_IMM,
64 [IB_WR_RDMA_READ] = MLX5_OPCODE_RDMA_READ,
65 [IB_WR_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_CS,
66 [IB_WR_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_FA,
67 [IB_WR_SEND_WITH_INV] = MLX5_OPCODE_SEND_INVAL,
68 [IB_WR_LOCAL_INV] = MLX5_OPCODE_UMR,
8a187ee5 69 [IB_WR_REG_MR] = MLX5_OPCODE_UMR,
e126ba97
EC		 70 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = MLX5_OPCODE_ATOMIC_MASKED_CS,
71 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = MLX5_OPCODE_ATOMIC_MASKED_FA,
72 [MLX5_IB_WR_UMR] = MLX5_OPCODE_UMR,
73};
74
f0313965
ES		 75struct mlx5_wqe_eth_pad {
76 u8 rsvd0[16];
77};
e126ba97 78
eb49ab0c
AV		 79enum raw_qp_set_mask_map {
80 MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID = 1UL << 0,
81};
82
0680efa2
AV		 83struct mlx5_modify_raw_qp_param {
84 u16 operation;
eb49ab0c
AV		 85
86 u32 set_mask; /* raw_qp_set_mask_map */
87 u8 rq_q_ctr_id;
0680efa2
AV		 88};
89
89ea94a7
MG		 90static void get_cqs(enum ib_qp_type qp_type,
91 struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
92 struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
93
e126ba97
EC		 94static int is_qp0(enum ib_qp_type qp_type)
95{
96 return qp_type == IB_QPT_SMI;
97}
98
e126ba97
EC		 99static int is_sqp(enum ib_qp_type qp_type)
100{
101 return is_qp0(qp_type) || is_qp1(qp_type);
102}
103
104static void *get_wqe(struct mlx5_ib_qp *qp, int offset)
105{
106 return mlx5_buf_offset(&qp->buf, offset);
107}
108
109static void *get_recv_wqe(struct mlx5_ib_qp *qp, int n)
110{
111 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
112}
113
114void *mlx5_get_send_wqe(struct mlx5_ib_qp *qp, int n)
115{
116 return get_wqe(qp, qp->sq.offset + (n << MLX5_IB_SQ_STRIDE));
117}
118
c1395a2a
HE		 119/**
120 * mlx5_ib_read_user_wqe() - Copy a user-space WQE to kernel space.
121 *
122 * @qp: QP to copy from.
123 * @send: copy from the send queue when non-zero, use the receive queue
124 * otherwise.
125 * @wqe_index: index to start copying from. For send work queues, the
126 * wqe_index is in units of MLX5_SEND_WQE_BB.
127 * For receive work queue, it is the number of work queue
128 * element in the queue.
129 * @buffer: destination buffer.
130 * @length: maximum number of bytes to copy.
131 *
132 * Copies at least a single WQE, but may copy more data.
133 *
134 * Return: the number of bytes copied, or an error code.
135 */
136int mlx5_ib_read_user_wqe(struct mlx5_ib_qp *qp, int send, int wqe_index,
19098df2 137 void *buffer, u32 length,
138 struct mlx5_ib_qp_base *base)
c1395a2a
HE		 139{
140 struct ib_device *ibdev = qp->ibqp.device;
141 struct mlx5_ib_dev *dev = to_mdev(ibdev);
142 struct mlx5_ib_wq *wq = send ? &qp->sq : &qp->rq;
143 size_t offset;
144 size_t wq_end;
19098df2 145 struct ib_umem *umem = base->ubuffer.umem;
c1395a2a
HE		 146 u32 first_copy_length;
147 int wqe_length;
148 int ret;
149
150 if (wq->wqe_cnt == 0) {
151 mlx5_ib_dbg(dev, "mlx5_ib_read_user_wqe for a QP with wqe_cnt == 0. qp_type: 0x%x\n",
152 qp->ibqp.qp_type);
153 return -EINVAL;
154 }
155
156 offset = wq->offset + ((wqe_index % wq->wqe_cnt) << wq->wqe_shift);
157 wq_end = wq->offset + (wq->wqe_cnt << wq->wqe_shift);
158
159 if (send && length < sizeof(struct mlx5_wqe_ctrl_seg))
160 return -EINVAL;
161
162 if (offset > umem->length ||
163 (send && offset + sizeof(struct mlx5_wqe_ctrl_seg) > umem->length))
164 return -EINVAL;
165
166 first_copy_length = min_t(u32, offset + length, wq_end) - offset;
167 ret = ib_umem_copy_from(buffer, umem, offset, first_copy_length);
168 if (ret)
169 return ret;
170
171 if (send) {
172 struct mlx5_wqe_ctrl_seg *ctrl = buffer;
173 int ds = be32_to_cpu(ctrl->qpn_ds) & MLX5_WQE_CTRL_DS_MASK;
174
175 wqe_length = ds * MLX5_WQE_DS_UNITS;
176 } else {
177 wqe_length = 1 << wq->wqe_shift;
178 }
179
180 if (wqe_length <= first_copy_length)
181 return first_copy_length;
182
183 ret = ib_umem_copy_from(buffer + first_copy_length, umem, wq->offset,
184 wqe_length - first_copy_length);
185 if (ret)
186 return ret;
187
188 return wqe_length;
189}
190
e126ba97
EC		 191static void mlx5_ib_qp_event(struct mlx5_core_qp *qp, int type)
192{
193 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
194 struct ib_event event;
195
19098df2 196 if (type == MLX5_EVENT_TYPE_PATH_MIG) {
197 /* This event is only valid for trans_qps */
198 to_mibqp(qp)->port = to_mibqp(qp)->trans_qp.alt_port;
199 }
e126ba97
EC		 200
201 if (ibqp->event_handler) {
202 event.device = ibqp->device;
203 event.element.qp = ibqp;
204 switch (type) {
205 case MLX5_EVENT_TYPE_PATH_MIG:
206 event.event = IB_EVENT_PATH_MIG;
207 break;
208 case MLX5_EVENT_TYPE_COMM_EST:
209 event.event = IB_EVENT_COMM_EST;
210 break;
211 case MLX5_EVENT_TYPE_SQ_DRAINED:
212 event.event = IB_EVENT_SQ_DRAINED;
213 break;
214 case MLX5_EVENT_TYPE_SRQ_LAST_WQE:
215 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
216 break;
217 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
218 event.event = IB_EVENT_QP_FATAL;
219 break;
220 case MLX5_EVENT_TYPE_PATH_MIG_FAILED:
221 event.event = IB_EVENT_PATH_MIG_ERR;
222 break;
223 case MLX5_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
224 event.event = IB_EVENT_QP_REQ_ERR;
225 break;
226 case MLX5_EVENT_TYPE_WQ_ACCESS_ERROR:
227 event.event = IB_EVENT_QP_ACCESS_ERR;
228 break;
229 default:
230 pr_warn("mlx5_ib: Unexpected event type %d on QP %06x\n", type, qp->qpn);
231 return;
232 }
233
234 ibqp->event_handler(&event, ibqp->qp_context);
235 }
236}
237
238static int set_rq_size(struct mlx5_ib_dev *dev, struct ib_qp_cap *cap,
239 int has_rq, struct mlx5_ib_qp *qp, struct mlx5_ib_create_qp *ucmd)
240{
241 int wqe_size;
242 int wq_size;
243
244 /* Sanity check RQ size before proceeding */
938fe83c 245 if (cap->max_recv_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz)))
e126ba97
EC		 246 return -EINVAL;
247
248 if (!has_rq) {
249 qp->rq.max_gs = 0;
250 qp->rq.wqe_cnt = 0;
251 qp->rq.wqe_shift = 0;
0540d814
NO		 252 cap->max_recv_wr = 0;
253 cap->max_recv_sge = 0;
e126ba97
EC		 254 } else {
255 if (ucmd) {
256 qp->rq.wqe_cnt = ucmd->rq_wqe_count;
257 qp->rq.wqe_shift = ucmd->rq_wqe_shift;
258 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
259 qp->rq.max_post = qp->rq.wqe_cnt;
260 } else {
261 wqe_size = qp->wq_sig ? sizeof(struct mlx5_wqe_signature_seg) : 0;
262 wqe_size += cap->max_recv_sge * sizeof(struct mlx5_wqe_data_seg);
263 wqe_size = roundup_pow_of_two(wqe_size);
264 wq_size = roundup_pow_of_two(cap->max_recv_wr) * wqe_size;
265 wq_size = max_t(int, wq_size, MLX5_SEND_WQE_BB);
266 qp->rq.wqe_cnt = wq_size / wqe_size;
938fe83c 267 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_rq)) {
e126ba97
EC		 268 mlx5_ib_dbg(dev, "wqe_size %d, max %d\n",
269 wqe_size,
938fe83c
SM		 270 MLX5_CAP_GEN(dev->mdev,
271 max_wqe_sz_rq));
e126ba97
EC		 272 return -EINVAL;
273 }
274 qp->rq.wqe_shift = ilog2(wqe_size);
275 qp->rq.max_gs = (1 << qp->rq.wqe_shift) / sizeof(struct mlx5_wqe_data_seg) - qp->wq_sig;
276 qp->rq.max_post = qp->rq.wqe_cnt;
277 }
278 }
279
280 return 0;
281}
282
f0313965 283static int sq_overhead(struct ib_qp_init_attr *attr)
e126ba97 284{
618af384 285 int size = 0;
e126ba97 286
f0313965 287 switch (attr->qp_type) {
e126ba97 288 case IB_QPT_XRC_INI:
b125a54b 289 size += sizeof(struct mlx5_wqe_xrc_seg);
e126ba97
EC		 290 /* fall through */
291 case IB_QPT_RC:
292 size += sizeof(struct mlx5_wqe_ctrl_seg) +
75c1657e
LR		 293 max(sizeof(struct mlx5_wqe_atomic_seg) +
294 sizeof(struct mlx5_wqe_raddr_seg),
295 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
296 sizeof(struct mlx5_mkey_seg));
e126ba97
EC		 297 break;
298
b125a54b
EC		 299 case IB_QPT_XRC_TGT:
300 return 0;
301
e126ba97 302 case IB_QPT_UC:
b125a54b 303 size += sizeof(struct mlx5_wqe_ctrl_seg) +
75c1657e
LR		 304 max(sizeof(struct mlx5_wqe_raddr_seg),
305 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
306 sizeof(struct mlx5_mkey_seg));
e126ba97
EC		 307 break;
308
309 case IB_QPT_UD:
f0313965
ES		 310 if (attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
311 size += sizeof(struct mlx5_wqe_eth_pad) +
312 sizeof(struct mlx5_wqe_eth_seg);
313 /* fall through */
e126ba97 314 case IB_QPT_SMI:
d16e91da 315 case MLX5_IB_QPT_HW_GSI:
b125a54b 316 size += sizeof(struct mlx5_wqe_ctrl_seg) +
e126ba97
EC		 317 sizeof(struct mlx5_wqe_datagram_seg);
318 break;
319
320 case MLX5_IB_QPT_REG_UMR:
b125a54b 321 size += sizeof(struct mlx5_wqe_ctrl_seg) +
e126ba97
EC		 322 sizeof(struct mlx5_wqe_umr_ctrl_seg) +
323 sizeof(struct mlx5_mkey_seg);
324 break;
325
326 default:
327 return -EINVAL;
328 }
329
330 return size;
331}
332
333static int calc_send_wqe(struct ib_qp_init_attr *attr)
334{
335 int inl_size = 0;
336 int size;
337
f0313965 338 size = sq_overhead(attr);
e126ba97
EC		 339 if (size < 0)
340 return size;
341
342 if (attr->cap.max_inline_data) {
343 inl_size = size + sizeof(struct mlx5_wqe_inline_seg) +
344 attr->cap.max_inline_data;
345 }
346
347 size += attr->cap.max_send_sge * sizeof(struct mlx5_wqe_data_seg);
e1e66cc2
SG		 348 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN &&
349 ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB) < MLX5_SIG_WQE_SIZE)
350 return MLX5_SIG_WQE_SIZE;
351 else
352 return ALIGN(max_t(int, inl_size, size), MLX5_SEND_WQE_BB);
e126ba97
EC		 353}
354
355static int calc_sq_size(struct mlx5_ib_dev *dev, struct ib_qp_init_attr *attr,
356 struct mlx5_ib_qp *qp)
357{
358 int wqe_size;
359 int wq_size;
360
361 if (!attr->cap.max_send_wr)
362 return 0;
363
364 wqe_size = calc_send_wqe(attr);
365 mlx5_ib_dbg(dev, "wqe_size %d\n", wqe_size);
366 if (wqe_size < 0)
367 return wqe_size;
368
938fe83c 369 if (wqe_size > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
b125a54b 370 mlx5_ib_dbg(dev, "wqe_size(%d) > max_sq_desc_sz(%d)\n",
938fe83c 371 wqe_size, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
e126ba97
EC		 372 return -EINVAL;
373 }
374
f0313965
ES		 375 qp->max_inline_data = wqe_size - sq_overhead(attr) -
376 sizeof(struct mlx5_wqe_inline_seg);
e126ba97
EC		 377 attr->cap.max_inline_data = qp->max_inline_data;
378
e1e66cc2
SG		 379 if (attr->create_flags & IB_QP_CREATE_SIGNATURE_EN)
380 qp->signature_en = true;
381
e126ba97
EC		 382 wq_size = roundup_pow_of_two(attr->cap.max_send_wr * wqe_size);
383 qp->sq.wqe_cnt = wq_size / MLX5_SEND_WQE_BB;
938fe83c 384 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
b125a54b 385 mlx5_ib_dbg(dev, "wqe count(%d) exceeds limits(%d)\n",
938fe83c
SM		 386 qp->sq.wqe_cnt,
387 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
b125a54b
EC		 388 return -ENOMEM;
389 }
e126ba97
EC		 390 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
391 qp->sq.max_gs = attr->cap.max_send_sge;
b125a54b
EC		 392 qp->sq.max_post = wq_size / wqe_size;
393 attr->cap.max_send_wr = qp->sq.max_post;
e126ba97
EC		 394
395 return wq_size;
396}
397
398static int set_user_buf_size(struct mlx5_ib_dev *dev,
399 struct mlx5_ib_qp *qp,
19098df2 400 struct mlx5_ib_create_qp *ucmd,
0fb2ed66 401 struct mlx5_ib_qp_base *base,
402 struct ib_qp_init_attr *attr)
e126ba97
EC		 403{
404 int desc_sz = 1 << qp->sq.wqe_shift;
405
938fe83c 406 if (desc_sz > MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq)) {
e126ba97 407 mlx5_ib_warn(dev, "desc_sz %d, max_sq_desc_sz %d\n",
938fe83c 408 desc_sz, MLX5_CAP_GEN(dev->mdev, max_wqe_sz_sq));
e126ba97
EC		 409 return -EINVAL;
410 }
411
412 if (ucmd->sq_wqe_count && ((1 << ilog2(ucmd->sq_wqe_count)) != ucmd->sq_wqe_count)) {
413 mlx5_ib_warn(dev, "sq_wqe_count %d, sq_wqe_count %d\n",
414 ucmd->sq_wqe_count, ucmd->sq_wqe_count);
415 return -EINVAL;
416 }
417
418 qp->sq.wqe_cnt = ucmd->sq_wqe_count;
419
938fe83c 420 if (qp->sq.wqe_cnt > (1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz))) {
e126ba97 421 mlx5_ib_warn(dev, "wqe_cnt %d, max_wqes %d\n",
938fe83c
SM		 422 qp->sq.wqe_cnt,
423 1 << MLX5_CAP_GEN(dev->mdev, log_max_qp_sz));
e126ba97
EC		 424 return -EINVAL;
425 }
426
0fb2ed66 427 if (attr->qp_type == IB_QPT_RAW_PACKET) {
428 base->ubuffer.buf_size = qp->rq.wqe_cnt << qp->rq.wqe_shift;
429 qp->raw_packet_qp.sq.ubuffer.buf_size = qp->sq.wqe_cnt << 6;
430 } else {
431 base->ubuffer.buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
432 (qp->sq.wqe_cnt << 6);
433 }
e126ba97
EC		 434
435 return 0;
436}
437
438static int qp_has_rq(struct ib_qp_init_attr *attr)
439{
440 if (attr->qp_type == IB_QPT_XRC_INI ||
441 attr->qp_type == IB_QPT_XRC_TGT || attr->srq ||
442 attr->qp_type == MLX5_IB_QPT_REG_UMR ||
443 !attr->cap.max_recv_wr)
444 return 0;
445
446 return 1;
447}
448
c1be5232
EC		 449static int first_med_uuar(void)
450{
451 return 1;
452}
453
454static int next_uuar(int n)
455{
456 n++;
457
458 while (((n % 4) & 2))
459 n++;
460
461 return n;
462}
463
464static int num_med_uuar(struct mlx5_uuar_info *uuari)
465{
466 int n;
467
468 n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
469 uuari->num_low_latency_uuars - 1;
470
471 return n >= 0 ? n : 0;
472}
473
474static int max_uuari(struct mlx5_uuar_info *uuari)
475{
476 return uuari->num_uars * 4;
477}
478
479static int first_hi_uuar(struct mlx5_uuar_info *uuari)
480{
481 int med;
482 int i;
483 int t;
484
485 med = num_med_uuar(uuari);
486 for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
487 t++;
488 if (t == med)
489 return next_uuar(i);
490 }
491
492 return 0;
493}
494
e126ba97
EC		 495static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
496{
e126ba97
EC		 497 int i;
498
c1be5232 499 for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
e126ba97
EC		 500 if (!test_bit(i, uuari->bitmap)) {
501 set_bit(i, uuari->bitmap);
502 uuari->count[i]++;
503 return i;
504 }
505 }
506
507 return -ENOMEM;
508}
509
510static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
511{
c1be5232 512 int minidx = first_med_uuar();
e126ba97
EC		 513 int i;
514
c1be5232 515 for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
e126ba97
EC		 516 if (uuari->count[i] < uuari->count[minidx])
517 minidx = i;
518 }
519
520 uuari->count[minidx]++;
521 return minidx;
522}
523
524static int alloc_uuar(struct mlx5_uuar_info *uuari,
525 enum mlx5_ib_latency_class lat)
526{
527 int uuarn = -EINVAL;
528
529 mutex_lock(&uuari->lock);
530 switch (lat) {
531 case MLX5_IB_LATENCY_CLASS_LOW:
532 uuarn = 0;
533 uuari->count[uuarn]++;
534 break;
535
536 case MLX5_IB_LATENCY_CLASS_MEDIUM:
78c0f98c
EC		 537 if (uuari->ver < 2)
538 uuarn = -ENOMEM;
539 else
540 uuarn = alloc_med_class_uuar(uuari);
e126ba97
EC		 541 break;
542
543 case MLX5_IB_LATENCY_CLASS_HIGH:
78c0f98c
EC		 544 if (uuari->ver < 2)
545 uuarn = -ENOMEM;
546 else
547 uuarn = alloc_high_class_uuar(uuari);
e126ba97
EC		 548 break;
549
550 case MLX5_IB_LATENCY_CLASS_FAST_PATH:
551 uuarn = 2;
552 break;
553 }
554 mutex_unlock(&uuari->lock);
555
556 return uuarn;
557}
558
559static void free_med_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
560{
561 clear_bit(uuarn, uuari->bitmap);
562 --uuari->count[uuarn];
563}
564
565static void free_high_class_uuar(struct mlx5_uuar_info *uuari, int uuarn)
566{
567 clear_bit(uuarn, uuari->bitmap);
568 --uuari->count[uuarn];
569}
570
571static void free_uuar(struct mlx5_uuar_info *uuari, int uuarn)
572{
573 int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
574 int high_uuar = nuuars - uuari->num_low_latency_uuars;
575
576 mutex_lock(&uuari->lock);
577 if (uuarn == 0) {
578 --uuari->count[uuarn];
579 goto out;
580 }
581
582 if (uuarn < high_uuar) {
583 free_med_class_uuar(uuari, uuarn);
584 goto out;
585 }
586
587 free_high_class_uuar(uuari, uuarn);
588
589out:
590 mutex_unlock(&uuari->lock);
591}
592
593static enum mlx5_qp_state to_mlx5_state(enum ib_qp_state state)
594{
595 switch (state) {
596 case IB_QPS_RESET: return MLX5_QP_STATE_RST;
597 case IB_QPS_INIT: return MLX5_QP_STATE_INIT;
598 case IB_QPS_RTR: return MLX5_QP_STATE_RTR;
599 case IB_QPS_RTS: return MLX5_QP_STATE_RTS;
600 case IB_QPS_SQD: return MLX5_QP_STATE_SQD;
601 case IB_QPS_SQE: return MLX5_QP_STATE_SQER;
602 case IB_QPS_ERR: return MLX5_QP_STATE_ERR;
603 default: return -1;
604 }
605}
606
607static int to_mlx5_st(enum ib_qp_type type)
608{
609 switch (type) {
610 case IB_QPT_RC: return MLX5_QP_ST_RC;
611 case IB_QPT_UC: return MLX5_QP_ST_UC;
612 case IB_QPT_UD: return MLX5_QP_ST_UD;
613 case MLX5_IB_QPT_REG_UMR: return MLX5_QP_ST_REG_UMR;
614 case IB_QPT_XRC_INI:
615 case IB_QPT_XRC_TGT: return MLX5_QP_ST_XRC;
616 case IB_QPT_SMI: return MLX5_QP_ST_QP0;
d16e91da 617 case MLX5_IB_QPT_HW_GSI: return MLX5_QP_ST_QP1;
e126ba97 618 case IB_QPT_RAW_IPV6: return MLX5_QP_ST_RAW_IPV6;
e126ba97 619 case IB_QPT_RAW_PACKET:
0fb2ed66 620 case IB_QPT_RAW_ETHERTYPE: return MLX5_QP_ST_RAW_ETHERTYPE;
e126ba97
EC		 621 case IB_QPT_MAX:
622 default: return -EINVAL;
623 }
624}
625
89ea94a7
MG		 626static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
627 struct mlx5_ib_cq *recv_cq);
628static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
629 struct mlx5_ib_cq *recv_cq);
630
e126ba97
EC		 631static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
632{
633 return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
634}
635
19098df2 636static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev,
637 struct ib_pd *pd,
638 unsigned long addr, size_t size,
639 struct ib_umem **umem,
640 int *npages, int *page_shift, int *ncont,
641 u32 *offset)
642{
643 int err;
644
645 *umem = ib_umem_get(pd->uobject->context, addr, size, 0, 0);
646 if (IS_ERR(*umem)) {
647 mlx5_ib_dbg(dev, "umem_get failed\n");
648 return PTR_ERR(*umem);
649 }
650
651 mlx5_ib_cont_pages(*umem, addr, npages, page_shift, ncont, NULL);
652
653 err = mlx5_ib_get_buf_offset(addr, *page_shift, offset);
654 if (err) {
655 mlx5_ib_warn(dev, "bad offset\n");
656 goto err_umem;
657 }
658
659 mlx5_ib_dbg(dev, "addr 0x%lx, size %zu, npages %d, page_shift %d, ncont %d, offset %d\n",
660 addr, size, *npages, *page_shift, *ncont, *offset);
661
662 return 0;
663
664err_umem:
665 ib_umem_release(*umem);
666 *umem = NULL;
667
668 return err;
669}
670
79b20a6c
YH		 671static void destroy_user_rq(struct ib_pd *pd, struct mlx5_ib_rwq *rwq)
672{
673 struct mlx5_ib_ucontext *context;
674
675 context = to_mucontext(pd->uobject->context);
676 mlx5_ib_db_unmap_user(context, &rwq->db);
677 if (rwq->umem)
678 ib_umem_release(rwq->umem);
679}
680
681static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
682 struct mlx5_ib_rwq *rwq,
683 struct mlx5_ib_create_wq *ucmd)
684{
685 struct mlx5_ib_ucontext *context;
686 int page_shift = 0;
687 int npages;
688 u32 offset = 0;
689 int ncont = 0;
690 int err;
691
692 if (!ucmd->buf_addr)
693 return -EINVAL;
694
695 context = to_mucontext(pd->uobject->context);
696 rwq->umem = ib_umem_get(pd->uobject->context, ucmd->buf_addr,
697 rwq->buf_size, 0, 0);
698 if (IS_ERR(rwq->umem)) {
699 mlx5_ib_dbg(dev, "umem_get failed\n");
700 err = PTR_ERR(rwq->umem);
701 return err;
702 }
703
704 mlx5_ib_cont_pages(rwq->umem, ucmd->buf_addr, &npages, &page_shift,
705 &ncont, NULL);
706 err = mlx5_ib_get_buf_offset(ucmd->buf_addr, page_shift,
707 &rwq->rq_page_offset);
708 if (err) {
709 mlx5_ib_warn(dev, "bad offset\n");
710 goto err_umem;
711 }
712
713 rwq->rq_num_pas = ncont;
714 rwq->page_shift = page_shift;
715 rwq->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
716 rwq->wq_sig = !!(ucmd->flags & MLX5_WQ_FLAG_SIGNATURE);
717
718 mlx5_ib_dbg(dev, "addr 0x%llx, size %zd, npages %d, page_shift %d, ncont %d, offset %d\n",
719 (unsigned long long)ucmd->buf_addr, rwq->buf_size,
720 npages, page_shift, ncont, offset);
721
722 err = mlx5_ib_db_map_user(context, ucmd->db_addr, &rwq->db);
723 if (err) {
724 mlx5_ib_dbg(dev, "map failed\n");
725 goto err_umem;
726 }
727
728 rwq->create_type = MLX5_WQ_USER;
729 return 0;
730
731err_umem:
732 ib_umem_release(rwq->umem);
733 return err;
734}
735
e126ba97
EC		 736static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
737 struct mlx5_ib_qp *qp, struct ib_udata *udata,
0fb2ed66 738 struct ib_qp_init_attr *attr,
09a7d9ec 739 u32 **in,
19098df2 740 struct mlx5_ib_create_qp_resp *resp, int *inlen,
741 struct mlx5_ib_qp_base *base)
e126ba97
EC		 742{
743 struct mlx5_ib_ucontext *context;
744 struct mlx5_ib_create_qp ucmd;
19098df2 745 struct mlx5_ib_ubuffer *ubuffer = &base->ubuffer;
9e9c47d0 746 int page_shift = 0;
e126ba97
EC		 747 int uar_index;
748 int npages;
9e9c47d0 749 u32 offset = 0;
e126ba97 750 int uuarn;
9e9c47d0 751 int ncont = 0;
09a7d9ec
SM		 752 __be64 *pas;
753 void *qpc;
e126ba97
EC		 754 int err;
755
756 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
757 if (err) {
758 mlx5_ib_dbg(dev, "copy failed\n");
759 return err;
760 }
761
762 context = to_mucontext(pd->uobject->context);
763 /*
764 * TBD: should come from the verbs when we have the API
765 */
051f2630
LR		 766 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
767 /* In CROSS_CHANNEL CQ and QP must use the same UAR */
768 uuarn = MLX5_CROSS_CHANNEL_UUAR;
769 else {
770 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
e126ba97 771 if (uuarn < 0) {
051f2630
LR		 772 mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
773 mlx5_ib_dbg(dev, "reverting to medium latency\n");
774 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
c1be5232 775 if (uuarn < 0) {
051f2630
LR		 776 mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
777 mlx5_ib_dbg(dev, "reverting to high latency\n");
778 uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
779 if (uuarn < 0) {
780 mlx5_ib_warn(dev, "uuar allocation failed\n");
781 return uuarn;
782 }
c1be5232 783 }
e126ba97
EC		 784 }
785 }
786
787 uar_index = uuarn_to_uar_index(&context->uuari, uuarn);
788 mlx5_ib_dbg(dev, "uuarn 0x%x, uar_index 0x%x\n", uuarn, uar_index);
789
48fea837
HE		 790 qp->rq.offset = 0;
791 qp->sq.wqe_shift = ilog2(MLX5_SEND_WQE_BB);
792 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
793
0fb2ed66 794 err = set_user_buf_size(dev, qp, &ucmd, base, attr);
e126ba97
EC		 795 if (err)
796 goto err_uuar;
797
19098df2 798 if (ucmd.buf_addr && ubuffer->buf_size) {
799 ubuffer->buf_addr = ucmd.buf_addr;
800 err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr,
801 ubuffer->buf_size,
802 &ubuffer->umem, &npages, &page_shift,
803 &ncont, &offset);
804 if (err)
9e9c47d0 805 goto err_uuar;
9e9c47d0 806 } else {
19098df2 807 ubuffer->umem = NULL;
e126ba97 808 }
e126ba97 809
09a7d9ec
SM		 810 *inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
811 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * ncont;
e126ba97
EC		 812 *in = mlx5_vzalloc(*inlen);
813 if (!*in) {
814 err = -ENOMEM;
815 goto err_umem;
816 }
09a7d9ec
SM		 817
818 pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas);
19098df2 819 if (ubuffer->umem)
09a7d9ec
SM		 820 mlx5_ib_populate_pas(dev, ubuffer->umem, page_shift, pas, 0);
821
822 qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
823
824 MLX5_SET(qpc, qpc, log_page_size, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
825 MLX5_SET(qpc, qpc, page_offset, offset);
e126ba97 826
09a7d9ec 827 MLX5_SET(qpc, qpc, uar_page, uar_index);
e126ba97
EC		 828 resp->uuar_index = uuarn;
829 qp->uuarn = uuarn;
830
831 err = mlx5_ib_db_map_user(context, ucmd.db_addr, &qp->db);
832 if (err) {
833 mlx5_ib_dbg(dev, "map failed\n");
834 goto err_free;
835 }
836
837 err = ib_copy_to_udata(udata, resp, sizeof(*resp));
838 if (err) {
839 mlx5_ib_dbg(dev, "copy failed\n");
840 goto err_unmap;
841 }
842 qp->create_type = MLX5_QP_USER;
843
844 return 0;
845
846err_unmap:
847 mlx5_ib_db_unmap_user(context, &qp->db);
848
849err_free:
479163f4 850 kvfree(*in);
e126ba97
EC		 851
852err_umem:
19098df2 853 if (ubuffer->umem)
854 ib_umem_release(ubuffer->umem);
e126ba97
EC		 855
856err_uuar:
857 free_uuar(&context->uuari, uuarn);
858 return err;
859}
860
19098df2 861static void destroy_qp_user(struct ib_pd *pd, struct mlx5_ib_qp *qp,
862 struct mlx5_ib_qp_base *base)
e126ba97
EC		 863{
864 struct mlx5_ib_ucontext *context;
865
866 context = to_mucontext(pd->uobject->context);
867 mlx5_ib_db_unmap_user(context, &qp->db);
19098df2 868 if (base->ubuffer.umem)
869 ib_umem_release(base->ubuffer.umem);
e126ba97
EC		 870 free_uuar(&context->uuari, qp->uuarn);
871}
872
873static int create_kernel_qp(struct mlx5_ib_dev *dev,
874 struct ib_qp_init_attr *init_attr,
875 struct mlx5_ib_qp *qp,
09a7d9ec 876 u32 **in, int *inlen,
19098df2 877 struct mlx5_ib_qp_base *base)
e126ba97
EC		 878{
879 enum mlx5_ib_latency_class lc = MLX5_IB_LATENCY_CLASS_LOW;
880 struct mlx5_uuar_info *uuari;
881 int uar_index;
09a7d9ec 882 void *qpc;
e126ba97
EC		 883 int uuarn;
884 int err;
885
9603b61d 886 uuari = &dev->mdev->priv.uuari;
f0313965
ES		 887 if (init_attr->create_flags & ~(IB_QP_CREATE_SIGNATURE_EN |
888 IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK |
b11a4f9c
HE		 889 IB_QP_CREATE_IPOIB_UD_LSO |
890 mlx5_ib_create_qp_sqpn_qp1()))
1a4c3a3d 891 return -EINVAL;
e126ba97
EC		 892
893 if (init_attr->qp_type == MLX5_IB_QPT_REG_UMR)
894 lc = MLX5_IB_LATENCY_CLASS_FAST_PATH;
895
896 uuarn = alloc_uuar(uuari, lc);
897 if (uuarn < 0) {
898 mlx5_ib_dbg(dev, "\n");
899 return -ENOMEM;
900 }
901
902 qp->bf = &uuari->bfs[uuarn];
903 uar_index = qp->bf->uar->index;
904
905 err = calc_sq_size(dev, init_attr, qp);
906 if (err < 0) {
907 mlx5_ib_dbg(dev, "err %d\n", err);
908 goto err_uuar;
909 }
910
911 qp->rq.offset = 0;
912 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
19098df2 913 base->ubuffer.buf_size = err + (qp->rq.wqe_cnt << qp->rq.wqe_shift);
e126ba97 914
19098df2 915 err = mlx5_buf_alloc(dev->mdev, base->ubuffer.buf_size, &qp->buf);
e126ba97
EC		 916 if (err) {
917 mlx5_ib_dbg(dev, "err %d\n", err);
918 goto err_uuar;
919 }
920
921 qp->sq.qend = mlx5_get_send_wqe(qp, qp->sq.wqe_cnt);
09a7d9ec
SM		 922 *inlen = MLX5_ST_SZ_BYTES(create_qp_in) +
923 MLX5_FLD_SZ_BYTES(create_qp_in, pas[0]) * qp->buf.npages;
e126ba97
EC		 924 *in = mlx5_vzalloc(*inlen);
925 if (!*in) {
926 err = -ENOMEM;
927 goto err_buf;
928 }
09a7d9ec
SM		 929
930 qpc = MLX5_ADDR_OF(create_qp_in, *in, qpc);
931 MLX5_SET(qpc, qpc, uar_page, uar_index);
932 MLX5_SET(qpc, qpc, log_page_size, qp->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT);
933
e126ba97 934 /* Set "fast registration enabled" for all kernel QPs */
09a7d9ec
SM		 935 MLX5_SET(qpc, qpc, fre, 1);
936 MLX5_SET(qpc, qpc, rlky, 1);
e126ba97 937
b11a4f9c 938 if (init_attr->create_flags & mlx5_ib_create_qp_sqpn_qp1()) {
09a7d9ec 939 MLX5_SET(qpc, qpc, deth_sqpn, 1);
b11a4f9c
HE		 940 qp->flags |= MLX5_IB_QP_SQPN_QP1;
941 }
942
09a7d9ec
SM		 943 mlx5_fill_page_array(&qp->buf,
944 (__be64 *)MLX5_ADDR_OF(create_qp_in, *in, pas));
e126ba97 945
9603b61d 946 err = mlx5_db_alloc(dev->mdev, &qp->db);
e126ba97
EC		 947 if (err) {
948 mlx5_ib_dbg(dev, "err %d\n", err);
949 goto err_free;
950 }
951
e126ba97
EC		 952 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
953 qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
954 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
955 qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
956 qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
957
958 if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
959 !qp->sq.w_list || !qp->sq.wqe_head) {
960 err = -ENOMEM;
961 goto err_wrid;
962 }
963 qp->create_type = MLX5_QP_KERNEL;
964
965 return 0;
966
967err_wrid:
9603b61d 968 mlx5_db_free(dev->mdev, &qp->db);
e126ba97
EC		 969 kfree(qp->sq.wqe_head);
970 kfree(qp->sq.w_list);
971 kfree(qp->sq.wrid);
972 kfree(qp->sq.wr_data);
973 kfree(qp->rq.wrid);
974
975err_free:
479163f4 976 kvfree(*in);
e126ba97
EC		 977
978err_buf:
9603b61d 979 mlx5_buf_free(dev->mdev, &qp->buf);
e126ba97
EC		 980
981err_uuar:
9603b61d 982 free_uuar(&dev->mdev->priv.uuari, uuarn);
e126ba97
EC		 983 return err;
984}
985
986static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
987{
9603b61d 988 mlx5_db_free(dev->mdev, &qp->db);
e126ba97
EC		 989 kfree(qp->sq.wqe_head);
990 kfree(qp->sq.w_list);
991 kfree(qp->sq.wrid);
992 kfree(qp->sq.wr_data);
993 kfree(qp->rq.wrid);
9603b61d
JM		 994 mlx5_buf_free(dev->mdev, &qp->buf);
995 free_uuar(&dev->mdev->priv.uuari, qp->bf->uuarn);
e126ba97
EC		 996}
997
09a7d9ec 998static u32 get_rx_type(struct mlx5_ib_qp *qp, struct ib_qp_init_attr *attr)
e126ba97
EC		 999{
1000 if (attr->srq || (attr->qp_type == IB_QPT_XRC_TGT) ||
1001 (attr->qp_type == IB_QPT_XRC_INI))
09a7d9ec 1002 return MLX5_SRQ_RQ;
e126ba97 1003 else if (!qp->has_rq)
09a7d9ec 1004 return MLX5_ZERO_LEN_RQ;
e126ba97 1005 else
09a7d9ec 1006 return MLX5_NON_ZERO_RQ;
e126ba97
EC		 1007}
1008
1009static int is_connected(enum ib_qp_type qp_type)
1010{
1011 if (qp_type == IB_QPT_RC || qp_type == IB_QPT_UC)
1012 return 1;
1013
1014 return 0;
1015}
1016
0fb2ed66 1017static int create_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1018 struct mlx5_ib_sq *sq, u32 tdn)
1019{
c4f287c4 1020 u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
0fb2ed66 1021 void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1022
0fb2ed66 1023 MLX5_SET(tisc, tisc, transport_domain, tdn);
0fb2ed66 1024 return mlx5_core_create_tis(dev->mdev, in, sizeof(in), &sq->tisn);
1025}
1026
1027static void destroy_raw_packet_qp_tis(struct mlx5_ib_dev *dev,
1028 struct mlx5_ib_sq *sq)
1029{
1030 mlx5_core_destroy_tis(dev->mdev, sq->tisn);
1031}
1032
1033static int create_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1034 struct mlx5_ib_sq *sq, void *qpin,
1035 struct ib_pd *pd)
1036{
1037 struct mlx5_ib_ubuffer *ubuffer = &sq->ubuffer;
1038 __be64 *pas;
1039 void *in;
1040 void *sqc;
1041 void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1042 void *wq;
1043 int inlen;
1044 int err;
1045 int page_shift = 0;
1046 int npages;
1047 int ncont = 0;
1048 u32 offset = 0;
1049
1050 err = mlx5_ib_umem_get(dev, pd, ubuffer->buf_addr, ubuffer->buf_size,
1051 &sq->ubuffer.umem, &npages, &page_shift,
1052 &ncont, &offset);
1053 if (err)
1054 return err;
1055
1056 inlen = MLX5_ST_SZ_BYTES(create_sq_in) + sizeof(u64) * ncont;
1057 in = mlx5_vzalloc(inlen);
1058 if (!in) {
1059 err = -ENOMEM;
1060 goto err_umem;
1061 }
1062
1063 sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1064 MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1065 MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1066 MLX5_SET(sqc, sqc, user_index, MLX5_GET(qpc, qpc, user_index));
1067 MLX5_SET(sqc, sqc, cqn, MLX5_GET(qpc, qpc, cqn_snd));
1068 MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1069 MLX5_SET(sqc, sqc, tis_num_0, sq->tisn);
1070
1071 wq = MLX5_ADDR_OF(sqc, sqc, wq);
1072 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1073 MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1074 MLX5_SET(wq, wq, uar_page, MLX5_GET(qpc, qpc, uar_page));
1075 MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1076 MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1077 MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_sq_size));
1078 MLX5_SET(wq, wq, log_wq_pg_sz, page_shift - MLX5_ADAPTER_PAGE_SHIFT);
1079 MLX5_SET(wq, wq, page_offset, offset);
1080
1081 pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1082 mlx5_ib_populate_pas(dev, sq->ubuffer.umem, page_shift, pas, 0);
1083
1084 err = mlx5_core_create_sq_tracked(dev->mdev, in, inlen, &sq->base.mqp);
1085
1086 kvfree(in);
1087
1088 if (err)
1089 goto err_umem;
1090
1091 return 0;
1092
1093err_umem:
1094 ib_umem_release(sq->ubuffer.umem);
1095 sq->ubuffer.umem = NULL;
1096
1097 return err;
1098}
1099
1100static void destroy_raw_packet_qp_sq(struct mlx5_ib_dev *dev,
1101 struct mlx5_ib_sq *sq)
1102{
1103 mlx5_core_destroy_sq_tracked(dev->mdev, &sq->base.mqp);
1104 ib_umem_release(sq->ubuffer.umem);
1105}
1106
1107static int get_rq_pas_size(void *qpc)
1108{
1109 u32 log_page_size = MLX5_GET(qpc, qpc, log_page_size) + 12;
1110 u32 log_rq_stride = MLX5_GET(qpc, qpc, log_rq_stride);
1111 u32 log_rq_size = MLX5_GET(qpc, qpc, log_rq_size);
1112 u32 page_offset = MLX5_GET(qpc, qpc, page_offset);
1113 u32 po_quanta = 1 << (log_page_size - 6);
1114 u32 rq_sz = 1 << (log_rq_size + 4 + log_rq_stride);
1115 u32 page_size = 1 << log_page_size;
1116 u32 rq_sz_po = rq_sz + (page_offset * po_quanta);
1117 u32 rq_num_pas = (rq_sz_po + page_size - 1) / page_size;
1118
1119 return rq_num_pas * sizeof(u64);
1120}
1121
1122static int create_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1123 struct mlx5_ib_rq *rq, void *qpin)
1124{
358e42ea 1125 struct mlx5_ib_qp *mqp = rq->base.container_mibqp;
0fb2ed66 1126 __be64 *pas;
1127 __be64 *qp_pas;
1128 void *in;
1129 void *rqc;
1130 void *wq;
1131 void *qpc = MLX5_ADDR_OF(create_qp_in, qpin, qpc);
1132 int inlen;
1133 int err;
1134 u32 rq_pas_size = get_rq_pas_size(qpc);
1135
1136 inlen = MLX5_ST_SZ_BYTES(create_rq_in) + rq_pas_size;
1137 in = mlx5_vzalloc(inlen);
1138 if (!in)
1139 return -ENOMEM;
1140
1141 rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1142 MLX5_SET(rqc, rqc, vsd, 1);
1143 MLX5_SET(rqc, rqc, mem_rq_type, MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
1144 MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1145 MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1146 MLX5_SET(rqc, rqc, user_index, MLX5_GET(qpc, qpc, user_index));
1147 MLX5_SET(rqc, rqc, cqn, MLX5_GET(qpc, qpc, cqn_rcv));
1148
358e42ea
MD		 1149 if (mqp->flags & MLX5_IB_QP_CAP_SCATTER_FCS)
1150 MLX5_SET(rqc, rqc, scatter_fcs, 1);
1151
0fb2ed66 1152 wq = MLX5_ADDR_OF(rqc, rqc, wq);
1153 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1154 MLX5_SET(wq, wq, end_padding_mode,
01581fb8 1155 MLX5_GET(qpc, qpc, end_padding_mode));
0fb2ed66 1156 MLX5_SET(wq, wq, page_offset, MLX5_GET(qpc, qpc, page_offset));
1157 MLX5_SET(wq, wq, pd, MLX5_GET(qpc, qpc, pd));
1158 MLX5_SET64(wq, wq, dbr_addr, MLX5_GET64(qpc, qpc, dbr_addr));
1159 MLX5_SET(wq, wq, log_wq_stride, MLX5_GET(qpc, qpc, log_rq_stride) + 4);
1160 MLX5_SET(wq, wq, log_wq_pg_sz, MLX5_GET(qpc, qpc, log_page_size));
1161 MLX5_SET(wq, wq, log_wq_sz, MLX5_GET(qpc, qpc, log_rq_size));
1162
1163 pas = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
1164 qp_pas = (__be64 *)MLX5_ADDR_OF(create_qp_in, qpin, pas);
1165 memcpy(pas, qp_pas, rq_pas_size);
1166
1167 err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rq->base.mqp);
1168
1169 kvfree(in);
1170
1171 return err;
1172}
1173
1174static void destroy_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
1175 struct mlx5_ib_rq *rq)
1176{
1177 mlx5_core_destroy_rq_tracked(dev->mdev, &rq->base.mqp);
1178}
1179
1180static int create_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1181 struct mlx5_ib_rq *rq, u32 tdn)
1182{
1183 u32 *in;
1184 void *tirc;
1185 int inlen;
1186 int err;
1187
1188 inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1189 in = mlx5_vzalloc(inlen);
1190 if (!in)
1191 return -ENOMEM;
1192
1193 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1194 MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_DIRECT);
1195 MLX5_SET(tirc, tirc, inline_rqn, rq->base.mqp.qpn);
1196 MLX5_SET(tirc, tirc, transport_domain, tdn);
1197
1198 err = mlx5_core_create_tir(dev->mdev, in, inlen, &rq->tirn);
1199
1200 kvfree(in);
1201
1202 return err;
1203}
1204
1205static void destroy_raw_packet_qp_tir(struct mlx5_ib_dev *dev,
1206 struct mlx5_ib_rq *rq)
1207{
1208 mlx5_core_destroy_tir(dev->mdev, rq->tirn);
1209}
1210
1211static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
09a7d9ec 1212 u32 *in,
0fb2ed66 1213 struct ib_pd *pd)
1214{
1215 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1216 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1217 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1218 struct ib_uobject *uobj = pd->uobject;
1219 struct ib_ucontext *ucontext = uobj->context;
1220 struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1221 int err;
1222 u32 tdn = mucontext->tdn;
1223
1224 if (qp->sq.wqe_cnt) {
1225 err = create_raw_packet_qp_tis(dev, sq, tdn);
1226 if (err)
1227 return err;
1228
1229 err = create_raw_packet_qp_sq(dev, sq, in, pd);
1230 if (err)
1231 goto err_destroy_tis;
1232
1233 sq->base.container_mibqp = qp;
1234 }
1235
1236 if (qp->rq.wqe_cnt) {
358e42ea
MD		 1237 rq->base.container_mibqp = qp;
1238
0fb2ed66 1239 err = create_raw_packet_qp_rq(dev, rq, in);
1240 if (err)
1241 goto err_destroy_sq;
1242
0fb2ed66 1243
1244 err = create_raw_packet_qp_tir(dev, rq, tdn);
1245 if (err)
1246 goto err_destroy_rq;
1247 }
1248
1249 qp->trans_qp.base.mqp.qpn = qp->sq.wqe_cnt ? sq->base.mqp.qpn :
1250 rq->base.mqp.qpn;
1251
1252 return 0;
1253
1254err_destroy_rq:
1255 destroy_raw_packet_qp_rq(dev, rq);
1256err_destroy_sq:
1257 if (!qp->sq.wqe_cnt)
1258 return err;
1259 destroy_raw_packet_qp_sq(dev, sq);
1260err_destroy_tis:
1261 destroy_raw_packet_qp_tis(dev, sq);
1262
1263 return err;
1264}
1265
1266static void destroy_raw_packet_qp(struct mlx5_ib_dev *dev,
1267 struct mlx5_ib_qp *qp)
1268{
1269 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
1270 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1271 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1272
1273 if (qp->rq.wqe_cnt) {
1274 destroy_raw_packet_qp_tir(dev, rq);
1275 destroy_raw_packet_qp_rq(dev, rq);
1276 }
1277
1278 if (qp->sq.wqe_cnt) {
1279 destroy_raw_packet_qp_sq(dev, sq);
1280 destroy_raw_packet_qp_tis(dev, sq);
1281 }
1282}
1283
1284static void raw_packet_qp_copy_info(struct mlx5_ib_qp *qp,
1285 struct mlx5_ib_raw_packet_qp *raw_packet_qp)
1286{
1287 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
1288 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
1289
1290 sq->sq = &qp->sq;
1291 rq->rq = &qp->rq;
1292 sq->doorbell = &qp->db;
1293 rq->doorbell = &qp->db;
1294}
1295
28d61370
YH		 1296static void destroy_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1297{
1298 mlx5_core_destroy_tir(dev->mdev, qp->rss_qp.tirn);
1299}
1300
1301static int create_rss_raw_qp_tir(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
1302 struct ib_pd *pd,
1303 struct ib_qp_init_attr *init_attr,
1304 struct ib_udata *udata)
1305{
1306 struct ib_uobject *uobj = pd->uobject;
1307 struct ib_ucontext *ucontext = uobj->context;
1308 struct mlx5_ib_ucontext *mucontext = to_mucontext(ucontext);
1309 struct mlx5_ib_create_qp_resp resp = {};
1310 int inlen;
1311 int err;
1312 u32 *in;
1313 void *tirc;
1314 void *hfso;
1315 u32 selected_fields = 0;
1316 size_t min_resp_len;
1317 u32 tdn = mucontext->tdn;
1318 struct mlx5_ib_create_qp_rss ucmd = {};
1319 size_t required_cmd_sz;
1320
1321 if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1322 return -EOPNOTSUPP;
1323
1324 if (init_attr->create_flags || init_attr->send_cq)
1325 return -EINVAL;
1326
1327 min_resp_len = offsetof(typeof(resp), uuar_index) + sizeof(resp.uuar_index);
1328 if (udata->outlen < min_resp_len)
1329 return -EINVAL;
1330
1331 required_cmd_sz = offsetof(typeof(ucmd), reserved1) + sizeof(ucmd.reserved1);
1332 if (udata->inlen < required_cmd_sz) {
1333 mlx5_ib_dbg(dev, "invalid inlen\n");
1334 return -EINVAL;
1335 }
1336
1337 if (udata->inlen > sizeof(ucmd) &&
1338 !ib_is_udata_cleared(udata, sizeof(ucmd),
1339 udata->inlen - sizeof(ucmd))) {
1340 mlx5_ib_dbg(dev, "inlen is not supported\n");
1341 return -EOPNOTSUPP;
1342 }
1343
1344 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
1345 mlx5_ib_dbg(dev, "copy failed\n");
1346 return -EFAULT;
1347 }
1348
1349 if (ucmd.comp_mask) {
1350 mlx5_ib_dbg(dev, "invalid comp mask\n");
1351 return -EOPNOTSUPP;
1352 }
1353
1354 if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved)) || ucmd.reserved1) {
1355 mlx5_ib_dbg(dev, "invalid reserved\n");
1356 return -EOPNOTSUPP;
1357 }
1358
1359 err = ib_copy_to_udata(udata, &resp, min_resp_len);
1360 if (err) {
1361 mlx5_ib_dbg(dev, "copy failed\n");
1362 return -EINVAL;
1363 }
1364
1365 inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1366 in = mlx5_vzalloc(inlen);
1367 if (!in)
1368 return -ENOMEM;
1369
1370 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1371 MLX5_SET(tirc, tirc, disp_type,
1372 MLX5_TIRC_DISP_TYPE_INDIRECT);
1373 MLX5_SET(tirc, tirc, indirect_table,
1374 init_attr->rwq_ind_tbl->ind_tbl_num);
1375 MLX5_SET(tirc, tirc, transport_domain, tdn);
1376
1377 hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1378 switch (ucmd.rx_hash_function) {
1379 case MLX5_RX_HASH_FUNC_TOEPLITZ:
1380 {
1381 void *rss_key = MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
1382 size_t len = MLX5_FLD_SZ_BYTES(tirc, rx_hash_toeplitz_key);
1383
1384 if (len != ucmd.rx_key_len) {
1385 err = -EINVAL;
1386 goto err;
1387 }
1388
1389 MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_TOEPLITZ);
1390 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1391 memcpy(rss_key, ucmd.rx_hash_key, len);
1392 break;
1393 }
1394 default:
1395 err = -EOPNOTSUPP;
1396 goto err;
1397 }
1398
1399 if (!ucmd.rx_hash_fields_mask) {
1400 /* special case when this TIR serves as steering entry without hashing */
1401 if (!init_attr->rwq_ind_tbl->log_ind_tbl_size)
1402 goto create_tir;
1403 err = -EINVAL;
1404 goto err;
1405 }
1406
1407 if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1408 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4)) &&
1409 ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1410 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))) {
1411 err = -EINVAL;
1412 goto err;
1413 }
1414
1415 /* If none of IPV4 & IPV6 SRC/DST was set - this bit field is ignored */
1416 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1417 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4))
1418 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1419 MLX5_L3_PROT_TYPE_IPV4);
1420 else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6) ||
1421 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1422 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1423 MLX5_L3_PROT_TYPE_IPV6);
1424
1425 if (((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1426 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP)) &&
1427 ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1428 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))) {
1429 err = -EINVAL;
1430 goto err;
1431 }
1432
1433 /* If none of TCP & UDP SRC/DST was set - this bit field is ignored */
1434 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1435 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP))
1436 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1437 MLX5_L4_PROT_TYPE_TCP);
1438 else if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP) ||
1439 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1440 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1441 MLX5_L4_PROT_TYPE_UDP);
1442
1443 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV4) ||
1444 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_IPV6))
1445 selected_fields |= MLX5_HASH_FIELD_SEL_SRC_IP;
1446
1447 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV4) ||
1448 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_IPV6))
1449 selected_fields |= MLX5_HASH_FIELD_SEL_DST_IP;
1450
1451 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_TCP) ||
1452 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_SRC_PORT_UDP))
1453 selected_fields |= MLX5_HASH_FIELD_SEL_L4_SPORT;
1454
1455 if ((ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_TCP) ||
1456 (ucmd.rx_hash_fields_mask & MLX5_RX_HASH_DST_PORT_UDP))
1457 selected_fields |= MLX5_HASH_FIELD_SEL_L4_DPORT;
1458
1459 MLX5_SET(rx_hash_field_select, hfso, selected_fields, selected_fields);
1460
1461create_tir:
1462 err = mlx5_core_create_tir(dev->mdev, in, inlen, &qp->rss_qp.tirn);
1463
1464 if (err)
1465 goto err;
1466
1467 kvfree(in);
1468 /* qpn is reserved for that QP */
1469 qp->trans_qp.base.mqp.qpn = 0;
d9f88e5a 1470 qp->flags |= MLX5_IB_QP_RSS;
28d61370
YH		 1471 return 0;
1472
1473err:
1474 kvfree(in);
1475 return err;
1476}
1477
e126ba97
EC		 1478static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
1479 struct ib_qp_init_attr *init_attr,
1480 struct ib_udata *udata, struct mlx5_ib_qp *qp)
1481{
1482 struct mlx5_ib_resources *devr = &dev->devr;
09a7d9ec 1483 int inlen = MLX5_ST_SZ_BYTES(create_qp_in);
938fe83c 1484 struct mlx5_core_dev *mdev = dev->mdev;
e126ba97 1485 struct mlx5_ib_create_qp_resp resp;
89ea94a7
MG		 1486 struct mlx5_ib_cq *send_cq;
1487 struct mlx5_ib_cq *recv_cq;
1488 unsigned long flags;
cfb5e088 1489 u32 uidx = MLX5_IB_DEFAULT_UIDX;
09a7d9ec
SM		 1490 struct mlx5_ib_create_qp ucmd;
1491 struct mlx5_ib_qp_base *base;
cfb5e088 1492 void *qpc;
09a7d9ec
SM		 1493 u32 *in;
1494 int err;
e126ba97 1495
0fb2ed66 1496 base = init_attr->qp_type == IB_QPT_RAW_PACKET ?
1497 &qp->raw_packet_qp.rq.base :
1498 &qp->trans_qp.base;
1499
1500 if (init_attr->qp_type != IB_QPT_RAW_PACKET)
1501 mlx5_ib_odp_create_qp(qp);
6aec21f6 1502
e126ba97
EC		 1503 mutex_init(&qp->mutex);
1504 spin_lock_init(&qp->sq.lock);
1505 spin_lock_init(&qp->rq.lock);
1506
28d61370
YH		 1507 if (init_attr->rwq_ind_tbl) {
1508 if (!udata)
1509 return -ENOSYS;
1510
1511 err = create_rss_raw_qp_tir(dev, qp, pd, init_attr, udata);
1512 return err;
1513 }
1514
f360d88a 1515 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {
938fe83c 1516 if (!MLX5_CAP_GEN(mdev, block_lb_mc)) {
f360d88a
EC		 1517 mlx5_ib_dbg(dev, "block multicast loopback isn't supported\n");
1518 return -EINVAL;
1519 } else {
1520 qp->flags |= MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1521 }
1522 }
1523
051f2630
LR		 1524 if (init_attr->create_flags &
1525 (IB_QP_CREATE_CROSS_CHANNEL |
1526 IB_QP_CREATE_MANAGED_SEND |
1527 IB_QP_CREATE_MANAGED_RECV)) {
1528 if (!MLX5_CAP_GEN(mdev, cd)) {
1529 mlx5_ib_dbg(dev, "cross-channel isn't supported\n");
1530 return -EINVAL;
1531 }
1532 if (init_attr->create_flags & IB_QP_CREATE_CROSS_CHANNEL)
1533 qp->flags |= MLX5_IB_QP_CROSS_CHANNEL;
1534 if (init_attr->create_flags & IB_QP_CREATE_MANAGED_SEND)
1535 qp->flags |= MLX5_IB_QP_MANAGED_SEND;
1536 if (init_attr->create_flags & IB_QP_CREATE_MANAGED_RECV)
1537 qp->flags |= MLX5_IB_QP_MANAGED_RECV;
1538 }
f0313965
ES		 1539
1540 if (init_attr->qp_type == IB_QPT_UD &&
1541 (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO))
1542 if (!MLX5_CAP_GEN(mdev, ipoib_basic_offloads)) {
1543 mlx5_ib_dbg(dev, "ipoib UD lso qp isn't supported\n");
1544 return -EOPNOTSUPP;
1545 }
1546
358e42ea
MD		 1547 if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) {
1548 if (init_attr->qp_type != IB_QPT_RAW_PACKET) {
1549 mlx5_ib_dbg(dev, "Scatter FCS is supported only for Raw Packet QPs");
1550 return -EOPNOTSUPP;
1551 }
1552 if (!MLX5_CAP_GEN(dev->mdev, eth_net_offloads) ||
1553 !MLX5_CAP_ETH(dev->mdev, scatter_fcs)) {
1554 mlx5_ib_dbg(dev, "Scatter FCS isn't supported\n");
1555 return -EOPNOTSUPP;
1556 }
1557 qp->flags |= MLX5_IB_QP_CAP_SCATTER_FCS;
1558 }
1559
e126ba97
EC		 1560 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
1561 qp->sq_signal_bits = MLX5_WQE_CTRL_CQ_UPDATE;
1562
1563 if (pd && pd->uobject) {
1564 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
1565 mlx5_ib_dbg(dev, "copy failed\n");
1566 return -EFAULT;
1567 }
1568
cfb5e088
HA		 1569 err = get_qp_user_index(to_mucontext(pd->uobject->context),
1570 &ucmd, udata->inlen, &uidx);
1571 if (err)
1572 return err;
1573
e126ba97
EC		 1574 qp->wq_sig = !!(ucmd.flags & MLX5_QP_FLAG_SIGNATURE);
1575 qp->scat_cqe = !!(ucmd.flags & MLX5_QP_FLAG_SCATTER_CQE);
1576 } else {
1577 qp->wq_sig = !!wq_signature;
1578 }
1579
1580 qp->has_rq = qp_has_rq(init_attr);
1581 err = set_rq_size(dev, &init_attr->cap, qp->has_rq,
1582 qp, (pd && pd->uobject) ? &ucmd : NULL);
1583 if (err) {
1584 mlx5_ib_dbg(dev, "err %d\n", err);
1585 return err;
1586 }
1587
1588 if (pd) {
1589 if (pd->uobject) {
938fe83c
SM		 1590 __u32 max_wqes =
1591 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
e126ba97
EC		 1592 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d)\n", ucmd.sq_wqe_count);
1593 if (ucmd.rq_wqe_shift != qp->rq.wqe_shift ||
1594 ucmd.rq_wqe_count != qp->rq.wqe_cnt) {
1595 mlx5_ib_dbg(dev, "invalid rq params\n");
1596 return -EINVAL;
1597 }
938fe83c 1598 if (ucmd.sq_wqe_count > max_wqes) {
e126ba97 1599 mlx5_ib_dbg(dev, "requested sq_wqe_count (%d) > max allowed (%d)\n",
938fe83c 1600 ucmd.sq_wqe_count, max_wqes);
e126ba97
EC		 1601 return -EINVAL;
1602 }
b11a4f9c
HE		 1603 if (init_attr->create_flags &
1604 mlx5_ib_create_qp_sqpn_qp1()) {
1605 mlx5_ib_dbg(dev, "user-space is not allowed to create UD QPs spoofing as QP1\n");
1606 return -EINVAL;
1607 }
0fb2ed66 1608 err = create_user_qp(dev, pd, qp, udata, init_attr, &in,
1609 &resp, &inlen, base);
e126ba97
EC		 1610 if (err)
1611 mlx5_ib_dbg(dev, "err %d\n", err);
1612 } else {
19098df2 1613 err = create_kernel_qp(dev, init_attr, qp, &in, &inlen,
1614 base);
e126ba97
EC		 1615 if (err)
1616 mlx5_ib_dbg(dev, "err %d\n", err);
e126ba97
EC		 1617 }
1618
1619 if (err)
1620 return err;
1621 } else {
09a7d9ec 1622 in = mlx5_vzalloc(inlen);
e126ba97
EC		 1623 if (!in)
1624 return -ENOMEM;
1625
1626 qp->create_type = MLX5_QP_EMPTY;
1627 }
1628
1629 if (is_sqp(init_attr->qp_type))
1630 qp->port = init_attr->port_num;
1631
09a7d9ec
SM		 1632 qpc = MLX5_ADDR_OF(create_qp_in, in, qpc);
1633
1634 MLX5_SET(qpc, qpc, st, to_mlx5_st(init_attr->qp_type));
1635 MLX5_SET(qpc, qpc, pm_state, MLX5_QP_PM_MIGRATED);
e126ba97
EC		 1636
1637 if (init_attr->qp_type != MLX5_IB_QPT_REG_UMR)
09a7d9ec 1638 MLX5_SET(qpc, qpc, pd, to_mpd(pd ? pd : devr->p0)->pdn);
e126ba97 1639 else
09a7d9ec
SM		 1640 MLX5_SET(qpc, qpc, latency_sensitive, 1);
1641
e126ba97
EC		 1642
1643 if (qp->wq_sig)
09a7d9ec 1644 MLX5_SET(qpc, qpc, wq_signature, 1);
e126ba97 1645
f360d88a 1646 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
09a7d9ec 1647 MLX5_SET(qpc, qpc, block_lb_mc, 1);
f360d88a 1648
051f2630 1649 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
09a7d9ec 1650 MLX5_SET(qpc, qpc, cd_master, 1);
051f2630 1651 if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
09a7d9ec 1652 MLX5_SET(qpc, qpc, cd_slave_send, 1);
051f2630 1653 if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
09a7d9ec 1654 MLX5_SET(qpc, qpc, cd_slave_receive, 1);
051f2630 1655
e126ba97
EC		 1656 if (qp->scat_cqe && is_connected(init_attr->qp_type)) {
1657 int rcqe_sz;
1658 int scqe_sz;
1659
1660 rcqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->recv_cq);
1661 scqe_sz = mlx5_ib_get_cqe_size(dev, init_attr->send_cq);
1662
1663 if (rcqe_sz == 128)
09a7d9ec 1664 MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA64_CQE);
e126ba97 1665 else
09a7d9ec 1666 MLX5_SET(qpc, qpc, cs_res, MLX5_RES_SCAT_DATA32_CQE);
e126ba97
EC		 1667
1668 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) {
1669 if (scqe_sz == 128)
09a7d9ec 1670 MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA64_CQE);
e126ba97 1671 else
09a7d9ec 1672 MLX5_SET(qpc, qpc, cs_req, MLX5_REQ_SCAT_DATA32_CQE);
e126ba97
EC		 1673 }
1674 }
1675
1676 if (qp->rq.wqe_cnt) {
09a7d9ec
SM		 1677 MLX5_SET(qpc, qpc, log_rq_stride, qp->rq.wqe_shift - 4);
1678 MLX5_SET(qpc, qpc, log_rq_size, ilog2(qp->rq.wqe_cnt));
e126ba97
EC		 1679 }
1680
09a7d9ec 1681 MLX5_SET(qpc, qpc, rq_type, get_rx_type(qp, init_attr));
e126ba97
EC		 1682
1683 if (qp->sq.wqe_cnt)
09a7d9ec 1684 MLX5_SET(qpc, qpc, log_sq_size, ilog2(qp->sq.wqe_cnt));
e126ba97 1685 else
09a7d9ec 1686 MLX5_SET(qpc, qpc, no_sq, 1);
e126ba97
EC		 1687
1688 /* Set default resources */
1689 switch (init_attr->qp_type) {
1690 case IB_QPT_XRC_TGT:
09a7d9ec
SM		 1691 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1692 MLX5_SET(qpc, qpc, cqn_snd, to_mcq(devr->c0)->mcq.cqn);
1693 MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
1694 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(init_attr->xrcd)->xrcdn);
e126ba97
EC		 1695 break;
1696 case IB_QPT_XRC_INI:
09a7d9ec
SM		 1697 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(devr->c0)->mcq.cqn);
1698 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1699 MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s0)->msrq.srqn);
e126ba97
EC		 1700 break;
1701 default:
1702 if (init_attr->srq) {
09a7d9ec
SM		 1703 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x0)->xrcdn);
1704 MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(init_attr->srq)->msrq.srqn);
e126ba97 1705 } else {
09a7d9ec
SM		 1706 MLX5_SET(qpc, qpc, xrcd, to_mxrcd(devr->x1)->xrcdn);
1707 MLX5_SET(qpc, qpc, srqn_rmpn_xrqn, to_msrq(devr->s1)->msrq.srqn);
e126ba97
EC		 1708 }
1709 }
1710
1711 if (init_attr->send_cq)
09a7d9ec 1712 MLX5_SET(qpc, qpc, cqn_snd, to_mcq(init_attr->send_cq)->mcq.cqn);
e126ba97
EC		 1713
1714 if (init_attr->recv_cq)
09a7d9ec 1715 MLX5_SET(qpc, qpc, cqn_rcv, to_mcq(init_attr->recv_cq)->mcq.cqn);
e126ba97 1716
09a7d9ec 1717 MLX5_SET64(qpc, qpc, dbr_addr, qp->db.dma);
e126ba97 1718
09a7d9ec
SM		 1719 /* 0xffffff means we ask to work with cqe version 0 */
1720 if (MLX5_CAP_GEN(mdev, cqe_version) == MLX5_CQE_VERSION_V1)
cfb5e088 1721 MLX5_SET(qpc, qpc, user_index, uidx);
09a7d9ec 1722
f0313965
ES		 1723 /* we use IB_QP_CREATE_IPOIB_UD_LSO to indicates ipoib qp */
1724 if (init_attr->qp_type == IB_QPT_UD &&
1725 (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)) {
f0313965
ES		 1726 MLX5_SET(qpc, qpc, ulp_stateless_offload_mode, 1);
1727 qp->flags |= MLX5_IB_QP_LSO;
1728 }
cfb5e088 1729
0fb2ed66 1730 if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1731 qp->raw_packet_qp.sq.ubuffer.buf_addr = ucmd.sq_buf_addr;
1732 raw_packet_qp_copy_info(qp, &qp->raw_packet_qp);
1733 err = create_raw_packet_qp(dev, qp, in, pd);
1734 } else {
1735 err = mlx5_core_create_qp(dev->mdev, &base->mqp, in, inlen);
1736 }
1737
e126ba97
EC		 1738 if (err) {
1739 mlx5_ib_dbg(dev, "create qp failed\n");
1740 goto err_create;
1741 }
1742
479163f4 1743 kvfree(in);
e126ba97 1744
19098df2 1745 base->container_mibqp = qp;
1746 base->mqp.event = mlx5_ib_qp_event;
e126ba97 1747
89ea94a7
MG		 1748 get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
1749 &send_cq, &recv_cq);
1750 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1751 mlx5_ib_lock_cqs(send_cq, recv_cq);
1752 /* Maintain device to QPs access, needed for further handling via reset
1753 * flow
1754 */
1755 list_add_tail(&qp->qps_list, &dev->qp_list);
1756 /* Maintain CQ to QPs access, needed for further handling via reset flow
1757 */
1758 if (send_cq)
1759 list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
1760 if (recv_cq)
1761 list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
1762 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1763 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1764
e126ba97
EC		 1765 return 0;
1766
1767err_create:
1768 if (qp->create_type == MLX5_QP_USER)
19098df2 1769 destroy_qp_user(pd, qp, base);
e126ba97
EC		 1770 else if (qp->create_type == MLX5_QP_KERNEL)
1771 destroy_qp_kernel(dev, qp);
1772
479163f4 1773 kvfree(in);
e126ba97
EC		 1774 return err;
1775}
1776
1777static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1778 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
1779{
1780 if (send_cq) {
1781 if (recv_cq) {
1782 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
89ea94a7 1783 spin_lock(&send_cq->lock);
e126ba97
EC		 1784 spin_lock_nested(&recv_cq->lock,
1785 SINGLE_DEPTH_NESTING);
1786 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
89ea94a7 1787 spin_lock(&send_cq->lock);
e126ba97
EC		 1788 __acquire(&recv_cq->lock);
1789 } else {
89ea94a7 1790 spin_lock(&recv_cq->lock);
e126ba97
EC		 1791 spin_lock_nested(&send_cq->lock,
1792 SINGLE_DEPTH_NESTING);
1793 }
1794 } else {
89ea94a7 1795 spin_lock(&send_cq->lock);
6a4f139a 1796 __acquire(&recv_cq->lock);
e126ba97
EC		 1797 }
1798 } else if (recv_cq) {
89ea94a7 1799 spin_lock(&recv_cq->lock);
6a4f139a
EC		 1800 __acquire(&send_cq->lock);
1801 } else {
1802 __acquire(&send_cq->lock);
1803 __acquire(&recv_cq->lock);
e126ba97
EC		 1804 }
1805}
1806
1807static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv_cq)
1808 __releases(&send_cq->lock) __releases(&recv_cq->lock)
1809{
1810 if (send_cq) {
1811 if (recv_cq) {
1812 if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
1813 spin_unlock(&recv_cq->lock);
89ea94a7 1814 spin_unlock(&send_cq->lock);
e126ba97
EC		 1815 } else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
1816 __release(&recv_cq->lock);
89ea94a7 1817 spin_unlock(&send_cq->lock);
e126ba97
EC		 1818 } else {
1819 spin_unlock(&send_cq->lock);
89ea94a7 1820 spin_unlock(&recv_cq->lock);
e126ba97
EC		 1821 }
1822 } else {
6a4f139a 1823 __release(&recv_cq->lock);
89ea94a7 1824 spin_unlock(&send_cq->lock);
e126ba97
EC		 1825 }
1826 } else if (recv_cq) {
6a4f139a 1827 __release(&send_cq->lock);
89ea94a7 1828 spin_unlock(&recv_cq->lock);
6a4f139a
EC		 1829 } else {
1830 __release(&recv_cq->lock);
1831 __release(&send_cq->lock);
e126ba97
EC		 1832 }
1833}
1834
1835static struct mlx5_ib_pd *get_pd(struct mlx5_ib_qp *qp)
1836{
1837 return to_mpd(qp->ibqp.pd);
1838}
1839
89ea94a7
MG		 1840static void get_cqs(enum ib_qp_type qp_type,
1841 struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
e126ba97
EC		 1842 struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
1843{
89ea94a7 1844 switch (qp_type) {
e126ba97
EC		 1845 case IB_QPT_XRC_TGT:
1846 *send_cq = NULL;
1847 *recv_cq = NULL;
1848 break;
1849 case MLX5_IB_QPT_REG_UMR:
1850 case IB_QPT_XRC_INI:
89ea94a7 1851 *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
e126ba97
EC		 1852 *recv_cq = NULL;
1853 break;
1854
1855 case IB_QPT_SMI:
d16e91da 1856 case MLX5_IB_QPT_HW_GSI:
e126ba97
EC		 1857 case IB_QPT_RC:
1858 case IB_QPT_UC:
1859 case IB_QPT_UD:
1860 case IB_QPT_RAW_IPV6:
1861 case IB_QPT_RAW_ETHERTYPE:
0fb2ed66 1862 case IB_QPT_RAW_PACKET:
89ea94a7
MG		 1863 *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
1864 *recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
e126ba97
EC		 1865 break;
1866
e126ba97
EC		 1867 case IB_QPT_MAX:
1868 default:
1869 *send_cq = NULL;
1870 *recv_cq = NULL;
1871 break;
1872 }
1873}
1874
ad5f8e96 1875static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
13eab21f
AH		 1876 const struct mlx5_modify_raw_qp_param *raw_qp_param,
1877 u8 lag_tx_affinity);
ad5f8e96 1878
e126ba97
EC		 1879static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
1880{
1881 struct mlx5_ib_cq *send_cq, *recv_cq;
19098df2 1882 struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
89ea94a7 1883 unsigned long flags;
e126ba97
EC		 1884 int err;
1885
28d61370
YH		 1886 if (qp->ibqp.rwq_ind_tbl) {
1887 destroy_rss_raw_qp_tir(dev, qp);
1888 return;
1889 }
1890
0fb2ed66 1891 base = qp->ibqp.qp_type == IB_QPT_RAW_PACKET ?
1892 &qp->raw_packet_qp.rq.base :
1893 &qp->trans_qp.base;
1894
6aec21f6 1895 if (qp->state != IB_QPS_RESET) {
ad5f8e96 1896 if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
1897 mlx5_ib_qp_disable_pagefaults(qp);
1898 err = mlx5_core_qp_modify(dev->mdev,
1a412fb1
SM		 1899 MLX5_CMD_OP_2RST_QP, 0,
1900 NULL, &base->mqp);
ad5f8e96 1901 } else {
0680efa2
AV		 1902 struct mlx5_modify_raw_qp_param raw_qp_param = {
1903 .operation = MLX5_CMD_OP_2RST_QP
1904 };
1905
13eab21f 1906 err = modify_raw_packet_qp(dev, qp, &raw_qp_param, 0);
ad5f8e96 1907 }
1908 if (err)
427c1e7b 1909 mlx5_ib_warn(dev, "mlx5_ib: modify QP 0x%06x to RESET failed\n",
19098df2 1910 base->mqp.qpn);
6aec21f6 1911 }
e126ba97 1912
89ea94a7
MG		 1913 get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
1914 &send_cq, &recv_cq);
1915
1916 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1917 mlx5_ib_lock_cqs(send_cq, recv_cq);
1918 /* del from lists under both locks above to protect reset flow paths */
1919 list_del(&qp->qps_list);
1920 if (send_cq)
1921 list_del(&qp->cq_send_list);
1922
1923 if (recv_cq)
1924 list_del(&qp->cq_recv_list);
e126ba97
EC		 1925
1926 if (qp->create_type == MLX5_QP_KERNEL) {
19098df2 1927 __mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
e126ba97
EC		 1928 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
1929 if (send_cq != recv_cq)
19098df2 1930 __mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
1931 NULL);
e126ba97 1932 }
89ea94a7
MG		 1933 mlx5_ib_unlock_cqs(send_cq, recv_cq);
1934 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
e126ba97 1935
0fb2ed66 1936 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1937 destroy_raw_packet_qp(dev, qp);
1938 } else {
1939 err = mlx5_core_destroy_qp(dev->mdev, &base->mqp);
1940 if (err)
1941 mlx5_ib_warn(dev, "failed to destroy QP 0x%x\n",
1942 base->mqp.qpn);
1943 }
e126ba97 1944
e126ba97
EC		 1945 if (qp->create_type == MLX5_QP_KERNEL)
1946 destroy_qp_kernel(dev, qp);
1947 else if (qp->create_type == MLX5_QP_USER)
19098df2 1948 destroy_qp_user(&get_pd(qp)->ibpd, qp, base);
e126ba97
EC		 1949}
1950
1951static const char *ib_qp_type_str(enum ib_qp_type type)
1952{
1953 switch (type) {
1954 case IB_QPT_SMI:
1955 return "IB_QPT_SMI";
1956 case IB_QPT_GSI:
1957 return "IB_QPT_GSI";
1958 case IB_QPT_RC:
1959 return "IB_QPT_RC";
1960 case IB_QPT_UC:
1961 return "IB_QPT_UC";
1962 case IB_QPT_UD:
1963 return "IB_QPT_UD";
1964 case IB_QPT_RAW_IPV6:
1965 return "IB_QPT_RAW_IPV6";
1966 case IB_QPT_RAW_ETHERTYPE:
1967 return "IB_QPT_RAW_ETHERTYPE";
1968 case IB_QPT_XRC_INI:
1969 return "IB_QPT_XRC_INI";
1970 case IB_QPT_XRC_TGT:
1971 return "IB_QPT_XRC_TGT";
1972 case IB_QPT_RAW_PACKET:
1973 return "IB_QPT_RAW_PACKET";
1974 case MLX5_IB_QPT_REG_UMR:
1975 return "MLX5_IB_QPT_REG_UMR";
1976 case IB_QPT_MAX:
1977 default:
1978 return "Invalid QP type";
1979 }
1980}
1981
1982struct ib_qp *mlx5_ib_create_qp(struct ib_pd *pd,
1983 struct ib_qp_init_attr *init_attr,
1984 struct ib_udata *udata)
1985{
1986 struct mlx5_ib_dev *dev;
1987 struct mlx5_ib_qp *qp;
1988 u16 xrcdn = 0;
1989 int err;
1990
1991 if (pd) {
1992 dev = to_mdev(pd->device);
0fb2ed66 1993
1994 if (init_attr->qp_type == IB_QPT_RAW_PACKET) {
1995 if (!pd->uobject) {
1996 mlx5_ib_dbg(dev, "Raw Packet QP is not supported for kernel consumers\n");
1997 return ERR_PTR(-EINVAL);
1998 } else if (!to_mucontext(pd->uobject->context)->cqe_version) {
1999 mlx5_ib_dbg(dev, "Raw Packet QP is only supported for CQE version > 0\n");
2000 return ERR_PTR(-EINVAL);
2001 }
2002 }
09f16cf5
MD		 2003 } else {
2004 /* being cautious here */
2005 if (init_attr->qp_type != IB_QPT_XRC_TGT &&
2006 init_attr->qp_type != MLX5_IB_QPT_REG_UMR) {
2007 pr_warn("%s: no PD for transport %s\n", __func__,
2008 ib_qp_type_str(init_attr->qp_type));
2009 return ERR_PTR(-EINVAL);
2010 }
2011 dev = to_mdev(to_mxrcd(init_attr->xrcd)->ibxrcd.device);
e126ba97
EC		 2012 }
2013
2014 switch (init_attr->qp_type) {
2015 case IB_QPT_XRC_TGT:
2016 case IB_QPT_XRC_INI:
938fe83c 2017 if (!MLX5_CAP_GEN(dev->mdev, xrc)) {
e126ba97
EC		 2018 mlx5_ib_dbg(dev, "XRC not supported\n");
2019 return ERR_PTR(-ENOSYS);
2020 }
2021 init_attr->recv_cq = NULL;
2022 if (init_attr->qp_type == IB_QPT_XRC_TGT) {
2023 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
2024 init_attr->send_cq = NULL;
2025 }
2026
2027 /* fall through */
0fb2ed66 2028 case IB_QPT_RAW_PACKET:
e126ba97
EC		 2029 case IB_QPT_RC:
2030 case IB_QPT_UC:
2031 case IB_QPT_UD:
2032 case IB_QPT_SMI:
d16e91da 2033 case MLX5_IB_QPT_HW_GSI:
e126ba97
EC		 2034 case MLX5_IB_QPT_REG_UMR:
2035 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2036 if (!qp)
2037 return ERR_PTR(-ENOMEM);
2038
2039 err = create_qp_common(dev, pd, init_attr, udata, qp);
2040 if (err) {
2041 mlx5_ib_dbg(dev, "create_qp_common failed\n");
2042 kfree(qp);
2043 return ERR_PTR(err);
2044 }
2045
2046 if (is_qp0(init_attr->qp_type))
2047 qp->ibqp.qp_num = 0;
2048 else if (is_qp1(init_attr->qp_type))
2049 qp->ibqp.qp_num = 1;
2050 else
19098df2 2051 qp->ibqp.qp_num = qp->trans_qp.base.mqp.qpn;
e126ba97
EC		 2052
2053 mlx5_ib_dbg(dev, "ib qpnum 0x%x, mlx qpn 0x%x, rcqn 0x%x, scqn 0x%x\n",
19098df2 2054 qp->ibqp.qp_num, qp->trans_qp.base.mqp.qpn,
2055 to_mcq(init_attr->recv_cq)->mcq.cqn,
e126ba97
EC		 2056 to_mcq(init_attr->send_cq)->mcq.cqn);
2057
19098df2 2058 qp->trans_qp.xrcdn = xrcdn;
e126ba97
EC		 2059
2060 break;
2061
d16e91da
HE		 2062 case IB_QPT_GSI:
2063 return mlx5_ib_gsi_create_qp(pd, init_attr);
2064
e126ba97
EC		 2065 case IB_QPT_RAW_IPV6:
2066 case IB_QPT_RAW_ETHERTYPE:
e126ba97
EC		 2067 case IB_QPT_MAX:
2068 default:
2069 mlx5_ib_dbg(dev, "unsupported qp type %d\n",
2070 init_attr->qp_type);
2071 /* Don't support raw QPs */
2072 return ERR_PTR(-EINVAL);
2073 }
2074
2075 return &qp->ibqp;
2076}
2077
2078int mlx5_ib_destroy_qp(struct ib_qp *qp)
2079{
2080 struct mlx5_ib_dev *dev = to_mdev(qp->device);
2081 struct mlx5_ib_qp *mqp = to_mqp(qp);
2082
d16e91da
HE		 2083 if (unlikely(qp->qp_type == IB_QPT_GSI))
2084 return mlx5_ib_gsi_destroy_qp(qp);
2085
e126ba97
EC		 2086 destroy_qp_common(dev, mqp);
2087
2088 kfree(mqp);
2089
2090 return 0;
2091}
2092
2093static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
2094 int attr_mask)
2095{
2096 u32 hw_access_flags = 0;
2097 u8 dest_rd_atomic;
2098 u32 access_flags;
2099
2100 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2101 dest_rd_atomic = attr->max_dest_rd_atomic;
2102 else
19098df2 2103 dest_rd_atomic = qp->trans_qp.resp_depth;
e126ba97
EC		 2104
2105 if (attr_mask & IB_QP_ACCESS_FLAGS)
2106 access_flags = attr->qp_access_flags;
2107 else
19098df2 2108 access_flags = qp->trans_qp.atomic_rd_en;
e126ba97
EC		 2109
2110 if (!dest_rd_atomic)
2111 access_flags &= IB_ACCESS_REMOTE_WRITE;
2112
2113 if (access_flags & IB_ACCESS_REMOTE_READ)
2114 hw_access_flags |= MLX5_QP_BIT_RRE;
2115 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
2116 hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
2117 if (access_flags & IB_ACCESS_REMOTE_WRITE)
2118 hw_access_flags |= MLX5_QP_BIT_RWE;
2119
2120 return cpu_to_be32(hw_access_flags);
2121}
2122
2123enum {
2124 MLX5_PATH_FLAG_FL = 1 << 0,
2125 MLX5_PATH_FLAG_FREE_AR = 1 << 1,
2126 MLX5_PATH_FLAG_COUNTER = 1 << 2,
2127};
2128
2129static int ib_rate_to_mlx5(struct mlx5_ib_dev *dev, u8 rate)
2130{
2131 if (rate == IB_RATE_PORT_CURRENT) {
2132 return 0;
2133 } else if (rate < IB_RATE_2_5_GBPS || rate > IB_RATE_300_GBPS) {
2134 return -EINVAL;
2135 } else {
2136 while (rate != IB_RATE_2_5_GBPS &&
2137 !(1 << (rate + MLX5_STAT_RATE_OFFSET) &
938fe83c 2138 MLX5_CAP_GEN(dev->mdev, stat_rate_support)))
e126ba97
EC		 2139 --rate;
2140 }
2141
2142 return rate + MLX5_STAT_RATE_OFFSET;
2143}
2144
75850d0b 2145static int modify_raw_packet_eth_prio(struct mlx5_core_dev *dev,
2146 struct mlx5_ib_sq *sq, u8 sl)
2147{
2148 void *in;
2149 void *tisc;
2150 int inlen;
2151 int err;
2152
2153 inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2154 in = mlx5_vzalloc(inlen);
2155 if (!in)
2156 return -ENOMEM;
2157
2158 MLX5_SET(modify_tis_in, in, bitmask.prio, 1);
2159
2160 tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2161 MLX5_SET(tisc, tisc, prio, ((sl & 0x7) << 1));
2162
2163 err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2164
2165 kvfree(in);
2166
2167 return err;
2168}
2169
13eab21f
AH		 2170static int modify_raw_packet_tx_affinity(struct mlx5_core_dev *dev,
2171 struct mlx5_ib_sq *sq, u8 tx_affinity)
2172{
2173 void *in;
2174 void *tisc;
2175 int inlen;
2176 int err;
2177
2178 inlen = MLX5_ST_SZ_BYTES(modify_tis_in);
2179 in = mlx5_vzalloc(inlen);
2180 if (!in)
2181 return -ENOMEM;
2182
2183 MLX5_SET(modify_tis_in, in, bitmask.lag_tx_port_affinity, 1);
2184
2185 tisc = MLX5_ADDR_OF(modify_tis_in, in, ctx);
2186 MLX5_SET(tisc, tisc, lag_tx_port_affinity, tx_affinity);
2187
2188 err = mlx5_core_modify_tis(dev, sq->tisn, in, inlen);
2189
2190 kvfree(in);
2191
2192 return err;
2193}
2194
75850d0b 2195static int mlx5_set_path(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
2196 const struct ib_ah_attr *ah,
e126ba97 2197 struct mlx5_qp_path *path, u8 port, int attr_mask,
f879ee8d
AS		 2198 u32 path_flags, const struct ib_qp_attr *attr,
2199 bool alt)
e126ba97 2200{
2811ba51 2201 enum rdma_link_layer ll = rdma_port_get_link_layer(&dev->ib_dev, port);
e126ba97
EC		 2202 int err;
2203
e126ba97 2204 if (attr_mask & IB_QP_PKEY_INDEX)
f879ee8d
AS		 2205 path->pkey_index = cpu_to_be16(alt ? attr->alt_pkey_index :
2206 attr->pkey_index);
e126ba97 2207
e126ba97 2208 if (ah->ah_flags & IB_AH_GRH) {
938fe83c
SM		 2209 if (ah->grh.sgid_index >=
2210 dev->mdev->port_caps[port - 1].gid_table_len) {
f4f01b54 2211 pr_err("sgid_index (%u) too large. max is %d\n",
938fe83c
SM		 2212 ah->grh.sgid_index,
2213 dev->mdev->port_caps[port - 1].gid_table_len);
f83b4263
EC		 2214 return -EINVAL;
2215 }
2811ba51
AS		 2216 }
2217
2218 if (ll == IB_LINK_LAYER_ETHERNET) {
2219 if (!(ah->ah_flags & IB_AH_GRH))
2220 return -EINVAL;
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 } else {
d3ae2bde
NO		 2226 path->fl_free_ar = (path_flags & MLX5_PATH_FLAG_FL) ? 0x80 : 0;
2227 path->fl_free_ar |=
2228 (path_flags & MLX5_PATH_FLAG_FREE_AR) ? 0x40 : 0;
2811ba51
AS		 2229 path->rlid = cpu_to_be16(ah->dlid);
2230 path->grh_mlid = ah->src_path_bits & 0x7f;
2231 if (ah->ah_flags & IB_AH_GRH)
2232 path->grh_mlid |= 1 << 7;
2233 path->dci_cfi_prio_sl = ah->sl & 0xf;
2234 }
2235
2236 if (ah->ah_flags & IB_AH_GRH) {
e126ba97
EC		 2237 path->mgid_index = ah->grh.sgid_index;
2238 path->hop_limit = ah->grh.hop_limit;
2239 path->tclass_flowlabel =
2240 cpu_to_be32((ah->grh.traffic_class << 20) |
2241 (ah->grh.flow_label));
2242 memcpy(path->rgid, ah->grh.dgid.raw, 16);
2243 }
2244
2245 err = ib_rate_to_mlx5(dev, ah->static_rate);
2246 if (err < 0)
2247 return err;
2248 path->static_rate = err;
2249 path->port = port;
2250
e126ba97 2251 if (attr_mask & IB_QP_TIMEOUT)
f879ee8d 2252 path->ackto_lt = (alt ? attr->alt_timeout : attr->timeout) << 3;
e126ba97 2253
75850d0b 2254 if ((qp->ibqp.qp_type == IB_QPT_RAW_PACKET) && qp->sq.wqe_cnt)
2255 return modify_raw_packet_eth_prio(dev->mdev,
2256 &qp->raw_packet_qp.sq,
2257 ah->sl & 0xf);
2258
e126ba97
EC		 2259 return 0;
2260}
2261
2262static enum mlx5_qp_optpar opt_mask[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE][MLX5_QP_ST_MAX] = {
2263 [MLX5_QP_STATE_INIT] = {
2264 [MLX5_QP_STATE_INIT] = {
2265 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
2266 MLX5_QP_OPTPAR_RAE |
2267 MLX5_QP_OPTPAR_RWE |
2268 MLX5_QP_OPTPAR_PKEY_INDEX |
2269 MLX5_QP_OPTPAR_PRI_PORT,
2270 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
2271 MLX5_QP_OPTPAR_PKEY_INDEX |
2272 MLX5_QP_OPTPAR_PRI_PORT,
2273 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
2274 MLX5_QP_OPTPAR_Q_KEY |
2275 MLX5_QP_OPTPAR_PRI_PORT,
2276 },
2277 [MLX5_QP_STATE_RTR] = {
2278 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2279 MLX5_QP_OPTPAR_RRE |
2280 MLX5_QP_OPTPAR_RAE |
2281 MLX5_QP_OPTPAR_RWE |
2282 MLX5_QP_OPTPAR_PKEY_INDEX,
2283 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2284 MLX5_QP_OPTPAR_RWE |
2285 MLX5_QP_OPTPAR_PKEY_INDEX,
2286 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_PKEY_INDEX |
2287 MLX5_QP_OPTPAR_Q_KEY,
2288 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_PKEY_INDEX |
2289 MLX5_QP_OPTPAR_Q_KEY,
a4774e90
EC		 2290 [MLX5_QP_ST_XRC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2291 MLX5_QP_OPTPAR_RRE |
2292 MLX5_QP_OPTPAR_RAE |
2293 MLX5_QP_OPTPAR_RWE |
2294 MLX5_QP_OPTPAR_PKEY_INDEX,
e126ba97
EC		 2295 },
2296 },
2297 [MLX5_QP_STATE_RTR] = {
2298 [MLX5_QP_STATE_RTS] = {
2299 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2300 MLX5_QP_OPTPAR_RRE |
2301 MLX5_QP_OPTPAR_RAE |
2302 MLX5_QP_OPTPAR_RWE |
2303 MLX5_QP_OPTPAR_PM_STATE |
2304 MLX5_QP_OPTPAR_RNR_TIMEOUT,
2305 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_ALT_ADDR_PATH |
2306 MLX5_QP_OPTPAR_RWE |
2307 MLX5_QP_OPTPAR_PM_STATE,
2308 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2309 },
2310 },
2311 [MLX5_QP_STATE_RTS] = {
2312 [MLX5_QP_STATE_RTS] = {
2313 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RRE |
2314 MLX5_QP_OPTPAR_RAE |
2315 MLX5_QP_OPTPAR_RWE |
2316 MLX5_QP_OPTPAR_RNR_TIMEOUT |
c2a3431e
EC		 2317 MLX5_QP_OPTPAR_PM_STATE |
2318 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
e126ba97 2319 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE |
c2a3431e
EC		 2320 MLX5_QP_OPTPAR_PM_STATE |
2321 MLX5_QP_OPTPAR_ALT_ADDR_PATH,
e126ba97
EC		 2322 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY |
2323 MLX5_QP_OPTPAR_SRQN |
2324 MLX5_QP_OPTPAR_CQN_RCV,
2325 },
2326 },
2327 [MLX5_QP_STATE_SQER] = {
2328 [MLX5_QP_STATE_RTS] = {
2329 [MLX5_QP_ST_UD] = MLX5_QP_OPTPAR_Q_KEY,
2330 [MLX5_QP_ST_MLX] = MLX5_QP_OPTPAR_Q_KEY,
75959f56 2331 [MLX5_QP_ST_UC] = MLX5_QP_OPTPAR_RWE,
a4774e90
EC		 2332 [MLX5_QP_ST_RC] = MLX5_QP_OPTPAR_RNR_TIMEOUT |
2333 MLX5_QP_OPTPAR_RWE |
2334 MLX5_QP_OPTPAR_RAE |
2335 MLX5_QP_OPTPAR_RRE,
e126ba97
EC		 2336 },
2337 },
2338};
2339
2340static int ib_nr_to_mlx5_nr(int ib_mask)
2341{
2342 switch (ib_mask) {
2343 case IB_QP_STATE:
2344 return 0;
2345 case IB_QP_CUR_STATE:
2346 return 0;
2347 case IB_QP_EN_SQD_ASYNC_NOTIFY:
2348 return 0;
2349 case IB_QP_ACCESS_FLAGS:
2350 return MLX5_QP_OPTPAR_RWE | MLX5_QP_OPTPAR_RRE |
2351 MLX5_QP_OPTPAR_RAE;
2352 case IB_QP_PKEY_INDEX:
2353 return MLX5_QP_OPTPAR_PKEY_INDEX;
2354 case IB_QP_PORT:
2355 return MLX5_QP_OPTPAR_PRI_PORT;
2356 case IB_QP_QKEY:
2357 return MLX5_QP_OPTPAR_Q_KEY;
2358 case IB_QP_AV:
2359 return MLX5_QP_OPTPAR_PRIMARY_ADDR_PATH |
2360 MLX5_QP_OPTPAR_PRI_PORT;
2361 case IB_QP_PATH_MTU:
2362 return 0;
2363 case IB_QP_TIMEOUT:
2364 return MLX5_QP_OPTPAR_ACK_TIMEOUT;
2365 case IB_QP_RETRY_CNT:
2366 return MLX5_QP_OPTPAR_RETRY_COUNT;
2367 case IB_QP_RNR_RETRY:
2368 return MLX5_QP_OPTPAR_RNR_RETRY;
2369 case IB_QP_RQ_PSN:
2370 return 0;
2371 case IB_QP_MAX_QP_RD_ATOMIC:
2372 return MLX5_QP_OPTPAR_SRA_MAX;
2373 case IB_QP_ALT_PATH:
2374 return MLX5_QP_OPTPAR_ALT_ADDR_PATH;
2375 case IB_QP_MIN_RNR_TIMER:
2376 return MLX5_QP_OPTPAR_RNR_TIMEOUT;
2377 case IB_QP_SQ_PSN:
2378 return 0;
2379 case IB_QP_MAX_DEST_RD_ATOMIC:
2380 return MLX5_QP_OPTPAR_RRA_MAX | MLX5_QP_OPTPAR_RWE |
2381 MLX5_QP_OPTPAR_RRE | MLX5_QP_OPTPAR_RAE;
2382 case IB_QP_PATH_MIG_STATE:
2383 return MLX5_QP_OPTPAR_PM_STATE;
2384 case IB_QP_CAP:
2385 return 0;
2386 case IB_QP_DEST_QPN:
2387 return 0;
2388 }
2389 return 0;
2390}
2391
2392static int ib_mask_to_mlx5_opt(int ib_mask)
2393{
2394 int result = 0;
2395 int i;
2396
2397 for (i = 0; i < 8 * sizeof(int); i++) {
2398 if ((1 << i) & ib_mask)
2399 result |= ib_nr_to_mlx5_nr(1 << i);
2400 }
2401
2402 return result;
2403}
2404
eb49ab0c
AV		 2405static int modify_raw_packet_qp_rq(struct mlx5_ib_dev *dev,
2406 struct mlx5_ib_rq *rq, int new_state,
2407 const struct mlx5_modify_raw_qp_param *raw_qp_param)
ad5f8e96 2408{
2409 void *in;
2410 void *rqc;
2411 int inlen;
2412 int err;
2413
2414 inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
2415 in = mlx5_vzalloc(inlen);
2416 if (!in)
2417 return -ENOMEM;
2418
2419 MLX5_SET(modify_rq_in, in, rq_state, rq->state);
2420
2421 rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
2422 MLX5_SET(rqc, rqc, state, new_state);
2423
eb49ab0c
AV		 2424 if (raw_qp_param->set_mask & MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID) {
2425 if (MLX5_CAP_GEN(dev->mdev, modify_rq_counter_set_id)) {
2426 MLX5_SET64(modify_rq_in, in, modify_bitmask,
2427 MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_MODIFY_RQ_COUNTER_SET_ID);
2428 MLX5_SET(rqc, rqc, counter_set_id, raw_qp_param->rq_q_ctr_id);
2429 } else
2430 pr_info_once("%s: RAW PACKET QP counters are not supported on current FW\n",
2431 dev->ib_dev.name);
2432 }
2433
2434 err = mlx5_core_modify_rq(dev->mdev, rq->base.mqp.qpn, in, inlen);
ad5f8e96 2435 if (err)
2436 goto out;
2437
2438 rq->state = new_state;
2439
2440out:
2441 kvfree(in);
2442 return err;
2443}
2444
2445static int modify_raw_packet_qp_sq(struct mlx5_core_dev *dev,
2446 struct mlx5_ib_sq *sq, int new_state)
2447{
2448 void *in;
2449 void *sqc;
2450 int inlen;
2451 int err;
2452
2453 inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
2454 in = mlx5_vzalloc(inlen);
2455 if (!in)
2456 return -ENOMEM;
2457
2458 MLX5_SET(modify_sq_in, in, sq_state, sq->state);
2459
2460 sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
2461 MLX5_SET(sqc, sqc, state, new_state);
2462
2463 err = mlx5_core_modify_sq(dev, sq->base.mqp.qpn, in, inlen);
2464 if (err)
2465 goto out;
2466
2467 sq->state = new_state;
2468
2469out:
2470 kvfree(in);
2471 return err;
2472}
2473
2474static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
13eab21f
AH		 2475 const struct mlx5_modify_raw_qp_param *raw_qp_param,
2476 u8 tx_affinity)
ad5f8e96 2477{
2478 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
2479 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
2480 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
2481 int rq_state;
2482 int sq_state;
2483 int err;
2484
0680efa2 2485 switch (raw_qp_param->operation) {
ad5f8e96 2486 case MLX5_CMD_OP_RST2INIT_QP:
2487 rq_state = MLX5_RQC_STATE_RDY;
2488 sq_state = MLX5_SQC_STATE_RDY;
2489 break;
2490 case MLX5_CMD_OP_2ERR_QP:
2491 rq_state = MLX5_RQC_STATE_ERR;
2492 sq_state = MLX5_SQC_STATE_ERR;
2493 break;
2494 case MLX5_CMD_OP_2RST_QP:
2495 rq_state = MLX5_RQC_STATE_RST;
2496 sq_state = MLX5_SQC_STATE_RST;
2497 break;
2498 case MLX5_CMD_OP_INIT2INIT_QP:
2499 case MLX5_CMD_OP_INIT2RTR_QP:
2500 case MLX5_CMD_OP_RTR2RTS_QP:
2501 case MLX5_CMD_OP_RTS2RTS_QP:
eb49ab0c
AV		 2502 if (raw_qp_param->set_mask)
2503 return -EINVAL;
2504 else
2505 return 0;
ad5f8e96 2506 default:
2507 WARN_ON(1);
2508 return -EINVAL;
2509 }
2510
2511 if (qp->rq.wqe_cnt) {
eb49ab0c 2512 err = modify_raw_packet_qp_rq(dev, rq, rq_state, raw_qp_param);
ad5f8e96 2513 if (err)
2514 return err;
2515 }
2516
13eab21f
AH		 2517 if (qp->sq.wqe_cnt) {
2518 if (tx_affinity) {
2519 err = modify_raw_packet_tx_affinity(dev->mdev, sq,
2520 tx_affinity);
2521 if (err)
2522 return err;
2523 }
2524
ad5f8e96 2525 return modify_raw_packet_qp_sq(dev->mdev, sq, sq_state);
13eab21f 2526 }
ad5f8e96 2527
2528 return 0;
2529}
2530
e126ba97
EC		 2531static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
2532 const struct ib_qp_attr *attr, int attr_mask,
2533 enum ib_qp_state cur_state, enum ib_qp_state new_state)
2534{
427c1e7b 2535 static const u16 optab[MLX5_QP_NUM_STATE][MLX5_QP_NUM_STATE] = {
2536 [MLX5_QP_STATE_RST] = {
2537 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2538 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2539 [MLX5_QP_STATE_INIT] = MLX5_CMD_OP_RST2INIT_QP,
2540 },
2541 [MLX5_QP_STATE_INIT] = {
2542 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2543 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2544 [MLX5_QP_STATE_INIT] = MLX5_CMD_OP_INIT2INIT_QP,
2545 [MLX5_QP_STATE_RTR] = MLX5_CMD_OP_INIT2RTR_QP,
2546 },
2547 [MLX5_QP_STATE_RTR] = {
2548 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2549 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2550 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_RTR2RTS_QP,
2551 },
2552 [MLX5_QP_STATE_RTS] = {
2553 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2554 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2555 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_RTS2RTS_QP,
2556 },
2557 [MLX5_QP_STATE_SQD] = {
2558 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2559 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2560 },
2561 [MLX5_QP_STATE_SQER] = {
2562 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2563 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2564 [MLX5_QP_STATE_RTS] = MLX5_CMD_OP_SQERR2RTS_QP,
2565 },
2566 [MLX5_QP_STATE_ERR] = {
2567 [MLX5_QP_STATE_RST] = MLX5_CMD_OP_2RST_QP,
2568 [MLX5_QP_STATE_ERR] = MLX5_CMD_OP_2ERR_QP,
2569 }
2570 };
2571
e126ba97
EC		 2572 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2573 struct mlx5_ib_qp *qp = to_mqp(ibqp);
19098df2 2574 struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
e126ba97
EC		 2575 struct mlx5_ib_cq *send_cq, *recv_cq;
2576 struct mlx5_qp_context *context;
e126ba97 2577 struct mlx5_ib_pd *pd;
eb49ab0c 2578 struct mlx5_ib_port *mibport = NULL;
e126ba97
EC		 2579 enum mlx5_qp_state mlx5_cur, mlx5_new;
2580 enum mlx5_qp_optpar optpar;
2581 int sqd_event;
2582 int mlx5_st;
2583 int err;
427c1e7b 2584 u16 op;
13eab21f 2585 u8 tx_affinity = 0;
e126ba97 2586
1a412fb1
SM		 2587 context = kzalloc(sizeof(*context), GFP_KERNEL);
2588 if (!context)
e126ba97
EC		 2589 return -ENOMEM;
2590
e126ba97 2591 err = to_mlx5_st(ibqp->qp_type);
158abf86
HE		 2592 if (err < 0) {
2593 mlx5_ib_dbg(dev, "unsupported qp type %d\n", ibqp->qp_type);
e126ba97 2594 goto out;
158abf86 2595 }
e126ba97
EC		 2596
2597 context->flags = cpu_to_be32(err << 16);
2598
2599 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) {
2600 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2601 } else {
2602 switch (attr->path_mig_state) {
2603 case IB_MIG_MIGRATED:
2604 context->flags |= cpu_to_be32(MLX5_QP_PM_MIGRATED << 11);
2605 break;
2606 case IB_MIG_REARM:
2607 context->flags |= cpu_to_be32(MLX5_QP_PM_REARM << 11);
2608 break;
2609 case IB_MIG_ARMED:
2610 context->flags |= cpu_to_be32(MLX5_QP_PM_ARMED << 11);
2611 break;
2612 }
2613 }
2614
13eab21f
AH		 2615 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2616 if ((ibqp->qp_type == IB_QPT_RC) ||
2617 (ibqp->qp_type == IB_QPT_UD &&
2618 !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
2619 (ibqp->qp_type == IB_QPT_UC) ||
2620 (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2621 (ibqp->qp_type == IB_QPT_XRC_INI) ||
2622 (ibqp->qp_type == IB_QPT_XRC_TGT)) {
2623 if (mlx5_lag_is_active(dev->mdev)) {
2624 tx_affinity = (unsigned int)atomic_add_return(1,
2625 &dev->roce.next_port) %
2626 MLX5_MAX_PORTS + 1;
2627 context->flags |= cpu_to_be32(tx_affinity << 24);
2628 }
2629 }
2630 }
2631
d16e91da 2632 if (is_sqp(ibqp->qp_type)) {
e126ba97
EC		 2633 context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
2634 } else if (ibqp->qp_type == IB_QPT_UD ||
2635 ibqp->qp_type == MLX5_IB_QPT_REG_UMR) {
2636 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
2637 } else if (attr_mask & IB_QP_PATH_MTU) {
2638 if (attr->path_mtu < IB_MTU_256 ||
2639 attr->path_mtu > IB_MTU_4096) {
2640 mlx5_ib_warn(dev, "invalid mtu %d\n", attr->path_mtu);
2641 err = -EINVAL;
2642 goto out;
2643 }
938fe83c
SM		 2644 context->mtu_msgmax = (attr->path_mtu << 5) |
2645 (u8)MLX5_CAP_GEN(dev->mdev, log_max_msg);
e126ba97
EC		 2646 }
2647
2648 if (attr_mask & IB_QP_DEST_QPN)
2649 context->log_pg_sz_remote_qpn = cpu_to_be32(attr->dest_qp_num);
2650
2651 if (attr_mask & IB_QP_PKEY_INDEX)
d3ae2bde 2652 context->pri_path.pkey_index = cpu_to_be16(attr->pkey_index);
e126ba97
EC		 2653
2654 /* todo implement counter_index functionality */
2655
2656 if (is_sqp(ibqp->qp_type))
2657 context->pri_path.port = qp->port;
2658
2659 if (attr_mask & IB_QP_PORT)
2660 context->pri_path.port = attr->port_num;
2661
2662 if (attr_mask & IB_QP_AV) {
75850d0b 2663 err = mlx5_set_path(dev, qp, &attr->ah_attr, &context->pri_path,
e126ba97 2664 attr_mask & IB_QP_PORT ? attr->port_num : qp->port,
f879ee8d 2665 attr_mask, 0, attr, false);
e126ba97
EC		 2666 if (err)
2667 goto out;
2668 }
2669
2670 if (attr_mask & IB_QP_TIMEOUT)
2671 context->pri_path.ackto_lt |= attr->timeout << 3;
2672
2673 if (attr_mask & IB_QP_ALT_PATH) {
75850d0b 2674 err = mlx5_set_path(dev, qp, &attr->alt_ah_attr,
2675 &context->alt_path,
f879ee8d
AS		 2676 attr->alt_port_num,
2677 attr_mask | IB_QP_PKEY_INDEX | IB_QP_TIMEOUT,
2678 0, attr, true);
e126ba97
EC		 2679 if (err)
2680 goto out;
2681 }
2682
2683 pd = get_pd(qp);
89ea94a7
MG		 2684 get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
2685 &send_cq, &recv_cq);
e126ba97
EC		 2686
2687 context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
2688 context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
2689 context->cqn_recv = recv_cq ? cpu_to_be32(recv_cq->mcq.cqn) : 0;
2690 context->params1 = cpu_to_be32(MLX5_IB_ACK_REQ_FREQ << 28);
2691
2692 if (attr_mask & IB_QP_RNR_RETRY)
2693 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
2694
2695 if (attr_mask & IB_QP_RETRY_CNT)
2696 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
2697
2698 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2699 if (attr->max_rd_atomic)
2700 context->params1 |=
2701 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
2702 }
2703
2704 if (attr_mask & IB_QP_SQ_PSN)
2705 context->next_send_psn = cpu_to_be32(attr->sq_psn);
2706
2707 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2708 if (attr->max_dest_rd_atomic)
2709 context->params2 |=
2710 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
2711 }
2712
2713 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
2714 context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
2715
2716 if (attr_mask & IB_QP_MIN_RNR_TIMER)
2717 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
2718
2719 if (attr_mask & IB_QP_RQ_PSN)
2720 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
2721
2722 if (attr_mask & IB_QP_QKEY)
2723 context->qkey = cpu_to_be32(attr->qkey);
2724
2725 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2726 context->db_rec_addr = cpu_to_be64(qp->db.dma);
2727
2728 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
2729 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
2730 sqd_event = 1;
2731 else
2732 sqd_event = 0;
2733
0837e86a
MB		 2734 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2735 u8 port_num = (attr_mask & IB_QP_PORT ? attr->port_num :
2736 qp->port) - 1;
eb49ab0c 2737 mibport = &dev->port[port_num];
0837e86a 2738 context->qp_counter_set_usr_page |=
321a9e3e 2739 cpu_to_be32((u32)(mibport->q_cnt_id) << 24);
0837e86a
MB		 2740 }
2741
e126ba97
EC		 2742 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
2743 context->sq_crq_size |= cpu_to_be16(1 << 4);
2744
b11a4f9c
HE		 2745 if (qp->flags & MLX5_IB_QP_SQPN_QP1)
2746 context->deth_sqpn = cpu_to_be32(1);
e126ba97
EC		 2747
2748 mlx5_cur = to_mlx5_state(cur_state);
2749 mlx5_new = to_mlx5_state(new_state);
2750 mlx5_st = to_mlx5_st(ibqp->qp_type);
07c9113f 2751 if (mlx5_st < 0)
e126ba97
EC		 2752 goto out;
2753
6aec21f6
HE		 2754 /* If moving to a reset or error state, we must disable page faults on
2755 * this QP and flush all current page faults. Otherwise a stale page
2756 * fault may attempt to work on this QP after it is reset and moved
2757 * again to RTS, and may cause the driver and the device to get out of
2758 * sync. */
2759 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
ad5f8e96 2760 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
2761 (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
6aec21f6
HE		 2762 mlx5_ib_qp_disable_pagefaults(qp);
2763
427c1e7b 2764 if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
2765 !optab[mlx5_cur][mlx5_new])
2766 goto out;
2767
2768 op = optab[mlx5_cur][mlx5_new];
e126ba97
EC		 2769 optpar = ib_mask_to_mlx5_opt(attr_mask);
2770 optpar &= opt_mask[mlx5_cur][mlx5_new][mlx5_st];
ad5f8e96 2771
0680efa2
AV		 2772 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
2773 struct mlx5_modify_raw_qp_param raw_qp_param = {};
2774
2775 raw_qp_param.operation = op;
eb49ab0c
AV		 2776 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
2777 raw_qp_param.rq_q_ctr_id = mibport->q_cnt_id;
2778 raw_qp_param.set_mask |= MLX5_RAW_QP_MOD_SET_RQ_Q_CTR_ID;
2779 }
13eab21f 2780 err = modify_raw_packet_qp(dev, qp, &raw_qp_param, tx_affinity);
0680efa2 2781 } else {
1a412fb1 2782 err = mlx5_core_qp_modify(dev->mdev, op, optpar, context,
ad5f8e96 2783 &base->mqp);
0680efa2
AV		 2784 }
2785
e126ba97
EC		 2786 if (err)
2787 goto out;
2788
ad5f8e96 2789 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
2790 (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
6aec21f6
HE		 2791 mlx5_ib_qp_enable_pagefaults(qp);
2792
e126ba97
EC		 2793 qp->state = new_state;
2794
2795 if (attr_mask & IB_QP_ACCESS_FLAGS)
19098df2 2796 qp->trans_qp.atomic_rd_en = attr->qp_access_flags;
e126ba97 2797 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
19098df2 2798 qp->trans_qp.resp_depth = attr->max_dest_rd_atomic;
e126ba97
EC		 2799 if (attr_mask & IB_QP_PORT)
2800 qp->port = attr->port_num;
2801 if (attr_mask & IB_QP_ALT_PATH)
19098df2 2802 qp->trans_qp.alt_port = attr->alt_port_num;
e126ba97
EC		 2803
2804 /*
2805 * If we moved a kernel QP to RESET, clean up all old CQ
2806 * entries and reinitialize the QP.
2807 */
2808 if (new_state == IB_QPS_RESET && !ibqp->uobject) {
19098df2 2809 mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
e126ba97
EC		 2810 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2811 if (send_cq != recv_cq)
19098df2 2812 mlx5_ib_cq_clean(send_cq, base->mqp.qpn, NULL);
e126ba97
EC		 2813
2814 qp->rq.head = 0;
2815 qp->rq.tail = 0;
2816 qp->sq.head = 0;
2817 qp->sq.tail = 0;
2818 qp->sq.cur_post = 0;
2819 qp->sq.last_poll = 0;
2820 qp->db.db[MLX5_RCV_DBR] = 0;
2821 qp->db.db[MLX5_SND_DBR] = 0;
2822 }
2823
2824out:
1a412fb1 2825 kfree(context);
e126ba97
EC		 2826 return err;
2827}
2828
2829int mlx5_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2830 int attr_mask, struct ib_udata *udata)
2831{
2832 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
2833 struct mlx5_ib_qp *qp = to_mqp(ibqp);
d16e91da 2834 enum ib_qp_type qp_type;
e126ba97
EC		 2835 enum ib_qp_state cur_state, new_state;
2836 int err = -EINVAL;
2837 int port;
2811ba51 2838 enum rdma_link_layer ll = IB_LINK_LAYER_UNSPECIFIED;
e126ba97 2839
28d61370
YH		 2840 if (ibqp->rwq_ind_tbl)
2841 return -ENOSYS;
2842
d16e91da
HE		 2843 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
2844 return mlx5_ib_gsi_modify_qp(ibqp, attr, attr_mask);
2845
2846 qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ?
2847 IB_QPT_GSI : ibqp->qp_type;
2848
e126ba97
EC		 2849 mutex_lock(&qp->mutex);
2850
2851 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2852 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2853
2811ba51
AS		 2854 if (!(cur_state == new_state && cur_state == IB_QPS_RESET)) {
2855 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2856 ll = dev->ib_dev.get_link_layer(&dev->ib_dev, port);
2857 }
2858
d16e91da
HE		 2859 if (qp_type != MLX5_IB_QPT_REG_UMR &&
2860 !ib_modify_qp_is_ok(cur_state, new_state, qp_type, attr_mask, ll)) {
158abf86
HE		 2861 mlx5_ib_dbg(dev, "invalid QP state transition from %d to %d, qp_type %d, attr_mask 0x%x\n",
2862 cur_state, new_state, ibqp->qp_type, attr_mask);
e126ba97 2863 goto out;
158abf86 2864 }
e126ba97
EC		 2865
2866 if ((attr_mask & IB_QP_PORT) &&
938fe83c 2867 (attr->port_num == 0 ||
158abf86
HE		 2868 attr->port_num > MLX5_CAP_GEN(dev->mdev, num_ports))) {
2869 mlx5_ib_dbg(dev, "invalid port number %d. number of ports is %d\n",
2870 attr->port_num, dev->num_ports);
e126ba97 2871 goto out;
158abf86 2872 }
e126ba97
EC		 2873
2874 if (attr_mask & IB_QP_PKEY_INDEX) {
2875 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
938fe83c 2876 if (attr->pkey_index >=
158abf86
HE		 2877 dev->mdev->port_caps[port - 1].pkey_table_len) {
2878 mlx5_ib_dbg(dev, "invalid pkey index %d\n",
2879 attr->pkey_index);
e126ba97 2880 goto out;
158abf86 2881 }
e126ba97
EC		 2882 }
2883
2884 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
938fe83c 2885 attr->max_rd_atomic >
158abf86
HE		 2886 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_res_qp))) {
2887 mlx5_ib_dbg(dev, "invalid max_rd_atomic value %d\n",
2888 attr->max_rd_atomic);
e126ba97 2889 goto out;
158abf86 2890 }
e126ba97
EC		 2891
2892 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
938fe83c 2893 attr->max_dest_rd_atomic >
158abf86
HE		 2894 (1 << MLX5_CAP_GEN(dev->mdev, log_max_ra_req_qp))) {
2895 mlx5_ib_dbg(dev, "invalid max_dest_rd_atomic value %d\n",
2896 attr->max_dest_rd_atomic);
e126ba97 2897 goto out;
158abf86 2898 }
e126ba97
EC		 2899
2900 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2901 err = 0;
2902 goto out;
2903 }
2904
2905 err = __mlx5_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2906
2907out:
2908 mutex_unlock(&qp->mutex);
2909 return err;
2910}
2911
2912static int mlx5_wq_overflow(struct mlx5_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2913{
2914 struct mlx5_ib_cq *cq;
2915 unsigned cur;
2916
2917 cur = wq->head - wq->tail;
2918 if (likely(cur + nreq < wq->max_post))
2919 return 0;
2920
2921 cq = to_mcq(ib_cq);
2922 spin_lock(&cq->lock);
2923 cur = wq->head - wq->tail;
2924 spin_unlock(&cq->lock);
2925
2926 return cur + nreq >= wq->max_post;
2927}
2928
2929static __always_inline void set_raddr_seg(struct mlx5_wqe_raddr_seg *rseg,
2930 u64 remote_addr, u32 rkey)
2931{
2932 rseg->raddr = cpu_to_be64(remote_addr);
2933 rseg->rkey = cpu_to_be32(rkey);
2934 rseg->reserved = 0;
2935}
2936
f0313965
ES		 2937static void *set_eth_seg(struct mlx5_wqe_eth_seg *eseg,
2938 struct ib_send_wr *wr, void *qend,
2939 struct mlx5_ib_qp *qp, int *size)
2940{
2941 void *seg = eseg;
2942
2943 memset(eseg, 0, sizeof(struct mlx5_wqe_eth_seg));
2944
2945 if (wr->send_flags & IB_SEND_IP_CSUM)
2946 eseg->cs_flags = MLX5_ETH_WQE_L3_CSUM |
2947 MLX5_ETH_WQE_L4_CSUM;
2948
2949 seg += sizeof(struct mlx5_wqe_eth_seg);
2950 *size += sizeof(struct mlx5_wqe_eth_seg) / 16;
2951
2952 if (wr->opcode == IB_WR_LSO) {
2953 struct ib_ud_wr *ud_wr = container_of(wr, struct ib_ud_wr, wr);
2954 int size_of_inl_hdr_start = sizeof(eseg->inline_hdr_start);
2955 u64 left, leftlen, copysz;
2956 void *pdata = ud_wr->header;
2957
2958 left = ud_wr->hlen;
2959 eseg->mss = cpu_to_be16(ud_wr->mss);
2960 eseg->inline_hdr_sz = cpu_to_be16(left);
2961
2962 /*
2963 * check if there is space till the end of queue, if yes,
2964 * copy all in one shot, otherwise copy till the end of queue,
2965 * rollback and than the copy the left
2966 */
2967 leftlen = qend - (void *)eseg->inline_hdr_start;
2968 copysz = min_t(u64, leftlen, left);
2969
2970 memcpy(seg - size_of_inl_hdr_start, pdata, copysz);
2971
2972 if (likely(copysz > size_of_inl_hdr_start)) {
2973 seg += ALIGN(copysz - size_of_inl_hdr_start, 16);
2974 *size += ALIGN(copysz - size_of_inl_hdr_start, 16) / 16;
2975 }
2976
2977 if (unlikely(copysz < left)) { /* the last wqe in the queue */
2978 seg = mlx5_get_send_wqe(qp, 0);
2979 left -= copysz;
2980 pdata += copysz;
2981 memcpy(seg, pdata, left);
2982 seg += ALIGN(left, 16);
2983 *size += ALIGN(left, 16) / 16;
2984 }
2985 }
2986
2987 return seg;
2988}
2989
e126ba97
EC		 2990static void set_datagram_seg(struct mlx5_wqe_datagram_seg *dseg,
2991 struct ib_send_wr *wr)
2992{
e622f2f4
CH		 2993 memcpy(&dseg->av, &to_mah(ud_wr(wr)->ah)->av, sizeof(struct mlx5_av));
2994 dseg->av.dqp_dct = cpu_to_be32(ud_wr(wr)->remote_qpn | MLX5_EXTENDED_UD_AV);
2995 dseg->av.key.qkey.qkey = cpu_to_be32(ud_wr(wr)->remote_qkey);
e126ba97
EC		 2996}
2997
2998static void set_data_ptr_seg(struct mlx5_wqe_data_seg *dseg, struct ib_sge *sg)
2999{
3000 dseg->byte_count = cpu_to_be32(sg->length);
3001 dseg->lkey = cpu_to_be32(sg->lkey);
3002 dseg->addr = cpu_to_be64(sg->addr);
3003}
3004
3005static __be16 get_klm_octo(int npages)
3006{
3007 return cpu_to_be16(ALIGN(npages, 8) / 2);
3008}
3009
3010static __be64 frwr_mkey_mask(void)
3011{
3012 u64 result;
3013
3014 result = MLX5_MKEY_MASK_LEN |
3015 MLX5_MKEY_MASK_PAGE_SIZE |
3016 MLX5_MKEY_MASK_START_ADDR |
3017 MLX5_MKEY_MASK_EN_RINVAL |
3018 MLX5_MKEY_MASK_KEY |
3019 MLX5_MKEY_MASK_LR |
3020 MLX5_MKEY_MASK_LW |
3021 MLX5_MKEY_MASK_RR |
3022 MLX5_MKEY_MASK_RW |
3023 MLX5_MKEY_MASK_A |
3024 MLX5_MKEY_MASK_SMALL_FENCE |
3025 MLX5_MKEY_MASK_FREE;
3026
3027 return cpu_to_be64(result);
3028}
3029
e6631814
SG		 3030static __be64 sig_mkey_mask(void)
3031{
3032 u64 result;
3033
3034 result = MLX5_MKEY_MASK_LEN |
3035 MLX5_MKEY_MASK_PAGE_SIZE |
3036 MLX5_MKEY_MASK_START_ADDR |
d5436ba0 3037 MLX5_MKEY_MASK_EN_SIGERR |
e6631814
SG		 3038 MLX5_MKEY_MASK_EN_RINVAL |
3039 MLX5_MKEY_MASK_KEY |
3040 MLX5_MKEY_MASK_LR |
3041 MLX5_MKEY_MASK_LW |
3042 MLX5_MKEY_MASK_RR |
3043 MLX5_MKEY_MASK_RW |
3044 MLX5_MKEY_MASK_SMALL_FENCE |
3045 MLX5_MKEY_MASK_FREE |
3046 MLX5_MKEY_MASK_BSF_EN;
3047
3048 return cpu_to_be64(result);
3049}
3050
8a187ee5
SG		 3051static void set_reg_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr,
3052 struct mlx5_ib_mr *mr)
3053{
3054 int ndescs = mr->ndescs;
3055
3056 memset(umr, 0, sizeof(*umr));
b005d316 3057
ec22eb53 3058 if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
b005d316
SG		 3059 /* KLMs take twice the size of MTTs */
3060 ndescs *= 2;
3061
8a187ee5
SG		 3062 umr->flags = MLX5_UMR_CHECK_NOT_FREE;
3063 umr->klm_octowords = get_klm_octo(ndescs);
3064 umr->mkey_mask = frwr_mkey_mask();
3065}
3066
dd01e66a 3067static void set_linv_umr_seg(struct mlx5_wqe_umr_ctrl_seg *umr)
e126ba97
EC		 3068{
3069 memset(umr, 0, sizeof(*umr));
dd01e66a
SG		 3070 umr->mkey_mask = cpu_to_be64(MLX5_MKEY_MASK_FREE);
3071 umr->flags = 1 << 7;
e126ba97
EC		 3072}
3073
968e78dd
HE		 3074static __be64 get_umr_reg_mr_mask(void)
3075{
3076 u64 result;
3077
3078 result = MLX5_MKEY_MASK_LEN |
3079 MLX5_MKEY_MASK_PAGE_SIZE |
3080 MLX5_MKEY_MASK_START_ADDR |
3081 MLX5_MKEY_MASK_PD |
3082 MLX5_MKEY_MASK_LR |
3083 MLX5_MKEY_MASK_LW |
3084 MLX5_MKEY_MASK_KEY |
3085 MLX5_MKEY_MASK_RR |
3086 MLX5_MKEY_MASK_RW |
3087 MLX5_MKEY_MASK_A |
3088 MLX5_MKEY_MASK_FREE;
3089
3090 return cpu_to_be64(result);
3091}
3092
3093static __be64 get_umr_unreg_mr_mask(void)
3094{
3095 u64 result;
3096
3097 result = MLX5_MKEY_MASK_FREE;
3098
3099 return cpu_to_be64(result);
3100}
3101
3102static __be64 get_umr_update_mtt_mask(void)
3103{
3104 u64 result;
3105
3106 result = MLX5_MKEY_MASK_FREE;
3107
3108 return cpu_to_be64(result);
3109}
3110
56e11d62
NO		 3111static __be64 get_umr_update_translation_mask(void)
3112{
3113 u64 result;
3114
3115 result = MLX5_MKEY_MASK_LEN |
3116 MLX5_MKEY_MASK_PAGE_SIZE |
3117 MLX5_MKEY_MASK_START_ADDR |
3118 MLX5_MKEY_MASK_KEY |
3119 MLX5_MKEY_MASK_FREE;
3120
3121 return cpu_to_be64(result);
3122}
3123
3124static __be64 get_umr_update_access_mask(void)
3125{
3126 u64 result;
3127
3128 result = MLX5_MKEY_MASK_LW |
3129 MLX5_MKEY_MASK_RR |
3130 MLX5_MKEY_MASK_RW |
3131 MLX5_MKEY_MASK_A |
3132 MLX5_MKEY_MASK_KEY |
3133 MLX5_MKEY_MASK_FREE;
3134
3135 return cpu_to_be64(result);
3136}
3137
3138static __be64 get_umr_update_pd_mask(void)
3139{
3140 u64 result;
3141
3142 result = MLX5_MKEY_MASK_PD |
3143 MLX5_MKEY_MASK_KEY |
3144 MLX5_MKEY_MASK_FREE;
3145
3146 return cpu_to_be64(result);
3147}
3148
e126ba97
EC		 3149static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
3150 struct ib_send_wr *wr)
3151{
e622f2f4 3152 struct mlx5_umr_wr *umrwr = umr_wr(wr);
e126ba97
EC		 3153
3154 memset(umr, 0, sizeof(*umr));
3155
968e78dd
HE		 3156 if (wr->send_flags & MLX5_IB_SEND_UMR_FAIL_IF_FREE)
3157 umr->flags = MLX5_UMR_CHECK_FREE; /* fail if free */
3158 else
3159 umr->flags = MLX5_UMR_CHECK_NOT_FREE; /* fail if not free */
3160
e126ba97 3161 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UNREG)) {
e126ba97 3162 umr->klm_octowords = get_klm_octo(umrwr->npages);
968e78dd
HE		 3163 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT) {
3164 umr->mkey_mask = get_umr_update_mtt_mask();
3165 umr->bsf_octowords = get_klm_octo(umrwr->target.offset);
3166 umr->flags |= MLX5_UMR_TRANSLATION_OFFSET_EN;
968e78dd 3167 }
56e11d62
NO		 3168 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_TRANSLATION)
3169 umr->mkey_mask |= get_umr_update_translation_mask();
3170 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_ACCESS)
3171 umr->mkey_mask |= get_umr_update_access_mask();
3172 if (wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_PD)
3173 umr->mkey_mask |= get_umr_update_pd_mask();
3174 if (!umr->mkey_mask)
3175 umr->mkey_mask = get_umr_reg_mr_mask();
e126ba97 3176 } else {
968e78dd 3177 umr->mkey_mask = get_umr_unreg_mr_mask();
e126ba97
EC		 3178 }
3179
3180 if (!wr->num_sge)
968e78dd 3181 umr->flags |= MLX5_UMR_INLINE;
e126ba97
EC		 3182}
3183
3184static u8 get_umr_flags(int acc)
3185{
3186 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX5_PERM_ATOMIC : 0) |
3187 (acc & IB_ACCESS_REMOTE_WRITE ? MLX5_PERM_REMOTE_WRITE : 0) |
3188 (acc & IB_ACCESS_REMOTE_READ ? MLX5_PERM_REMOTE_READ : 0) |
3189 (acc & IB_ACCESS_LOCAL_WRITE ? MLX5_PERM_LOCAL_WRITE : 0) |
2ac45934 3190 MLX5_PERM_LOCAL_READ | MLX5_PERM_UMR_EN;
e126ba97
EC		 3191}
3192
8a187ee5
SG		 3193static void set_reg_mkey_seg(struct mlx5_mkey_seg *seg,
3194 struct mlx5_ib_mr *mr,
3195 u32 key, int access)
3196{
3197 int ndescs = ALIGN(mr->ndescs, 8) >> 1;
3198
3199 memset(seg, 0, sizeof(*seg));
b005d316 3200
ec22eb53 3201 if (mr->access_mode == MLX5_MKC_ACCESS_MODE_MTT)
b005d316 3202 seg->log2_page_size = ilog2(mr->ibmr.page_size);
ec22eb53 3203 else if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
b005d316
SG		 3204 /* KLMs take twice the size of MTTs */
3205 ndescs *= 2;
3206
3207 seg->flags = get_umr_flags(access) | mr->access_mode;
8a187ee5
SG		 3208 seg->qpn_mkey7_0 = cpu_to_be32((key & 0xff) | 0xffffff00);
3209 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL);
3210 seg->start_addr = cpu_to_be64(mr->ibmr.iova);
3211 seg->len = cpu_to_be64(mr->ibmr.length);
3212 seg->xlt_oct_size = cpu_to_be32(ndescs);
8a187ee5
SG		 3213}
3214
dd01e66a 3215static void set_linv_mkey_seg(struct mlx5_mkey_seg *seg)
e126ba97
EC		 3216{
3217 memset(seg, 0, sizeof(*seg));
dd01e66a 3218 seg->status = MLX5_MKEY_STATUS_FREE;
e126ba97
EC		 3219}
3220
3221static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *wr)
3222{
e622f2f4 3223 struct mlx5_umr_wr *umrwr = umr_wr(wr);
968e78dd 3224
e126ba97
EC		 3225 memset(seg, 0, sizeof(*seg));
3226 if (wr->send_flags & MLX5_IB_SEND_UMR_UNREG) {
968e78dd 3227 seg->status = MLX5_MKEY_STATUS_FREE;
e126ba97
EC		 3228 return;
3229 }
3230
968e78dd
HE		 3231 seg->flags = convert_access(umrwr->access_flags);
3232 if (!(wr->send_flags & MLX5_IB_SEND_UMR_UPDATE_MTT)) {
56e11d62
NO		 3233 if (umrwr->pd)
3234 seg->flags_pd = cpu_to_be32(to_mpd(umrwr->pd)->pdn);
968e78dd
HE		 3235 seg->start_addr = cpu_to_be64(umrwr->target.virt_addr);
3236 }
3237 seg->len = cpu_to_be64(umrwr->length);
3238 seg->log2_page_size = umrwr->page_shift;
746b5583 3239 seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
968e78dd 3240 mlx5_mkey_variant(umrwr->mkey));
e126ba97
EC		 3241}
3242
8a187ee5
SG		 3243static void set_reg_data_seg(struct mlx5_wqe_data_seg *dseg,
3244 struct mlx5_ib_mr *mr,
3245 struct mlx5_ib_pd *pd)
3246{
3247 int bcount = mr->desc_size * mr->ndescs;
3248
3249 dseg->addr = cpu_to_be64(mr->desc_map);
3250 dseg->byte_count = cpu_to_be32(ALIGN(bcount, 64));
3251 dseg->lkey = cpu_to_be32(pd->ibpd.local_dma_lkey);
3252}
3253
e126ba97
EC		 3254static __be32 send_ieth(struct ib_send_wr *wr)
3255{
3256 switch (wr->opcode) {
3257 case IB_WR_SEND_WITH_IMM:
3258 case IB_WR_RDMA_WRITE_WITH_IMM:
3259 return wr->ex.imm_data;
3260
3261 case IB_WR_SEND_WITH_INV:
3262 return cpu_to_be32(wr->ex.invalidate_rkey);
3263
3264 default:
3265 return 0;
3266 }
3267}
3268
3269static u8 calc_sig(void *wqe, int size)
3270{
3271 u8 *p = wqe;
3272 u8 res = 0;
3273 int i;
3274
3275 for (i = 0; i < size; i++)
3276 res ^= p[i];
3277
3278 return ~res;
3279}
3280
3281static u8 wq_sig(void *wqe)
3282{
3283 return calc_sig(wqe, (*((u8 *)wqe + 8) & 0x3f) << 4);
3284}
3285
3286static int set_data_inl_seg(struct mlx5_ib_qp *qp, struct ib_send_wr *wr,
3287 void *wqe, int *sz)
3288{
3289 struct mlx5_wqe_inline_seg *seg;
3290 void *qend = qp->sq.qend;
3291 void *addr;
3292 int inl = 0;
3293 int copy;
3294 int len;
3295 int i;
3296
3297 seg = wqe;
3298 wqe += sizeof(*seg);
3299 for (i = 0; i < wr->num_sge; i++) {
3300 addr = (void *)(unsigned long)(wr->sg_list[i].addr);
3301 len = wr->sg_list[i].length;
3302 inl += len;
3303
3304 if (unlikely(inl > qp->max_inline_data))
3305 return -ENOMEM;
3306
3307 if (unlikely(wqe + len > qend)) {
3308 copy = qend - wqe;
3309 memcpy(wqe, addr, copy);
3310 addr += copy;
3311 len -= copy;
3312 wqe = mlx5_get_send_wqe(qp, 0);
3313 }
3314 memcpy(wqe, addr, len);
3315 wqe += len;
3316 }
3317
3318 seg->byte_count = cpu_to_be32(inl | MLX5_INLINE_SEG);
3319
3320 *sz = ALIGN(inl + sizeof(seg->byte_count), 16) / 16;
3321
3322 return 0;
3323}
3324
e6631814
SG		 3325static u16 prot_field_size(enum ib_signature_type type)
3326{
3327 switch (type) {
3328 case IB_SIG_TYPE_T10_DIF:
3329 return MLX5_DIF_SIZE;
3330 default:
3331 return 0;
3332 }
3333}
3334
3335static u8 bs_selector(int block_size)
3336{
3337 switch (block_size) {
3338 case 512: return 0x1;
3339 case 520: return 0x2;
3340 case 4096: return 0x3;
3341 case 4160: return 0x4;
3342 case 1073741824: return 0x5;
3343 default: return 0;
3344 }
3345}
3346
78eda2bb
SG		 3347static void mlx5_fill_inl_bsf(struct ib_sig_domain *domain,
3348 struct mlx5_bsf_inl *inl)
e6631814 3349{
142537f4
SG		 3350 /* Valid inline section and allow BSF refresh */
3351 inl->vld_refresh = cpu_to_be16(MLX5_BSF_INL_VALID |
3352 MLX5_BSF_REFRESH_DIF);
3353 inl->dif_apptag = cpu_to_be16(domain->sig.dif.app_tag);
3354 inl->dif_reftag = cpu_to_be32(domain->sig.dif.ref_tag);
78eda2bb
SG		 3355 /* repeating block */
3356 inl->rp_inv_seed = MLX5_BSF_REPEAT_BLOCK;
3357 inl->sig_type = domain->sig.dif.bg_type == IB_T10DIF_CRC ?
3358 MLX5_DIF_CRC : MLX5_DIF_IPCS;
e6631814 3359
78eda2bb
SG		 3360 if (domain->sig.dif.ref_remap)
3361 inl->dif_inc_ref_guard_check |= MLX5_BSF_INC_REFTAG;
e6631814 3362
78eda2bb
SG		 3363 if (domain->sig.dif.app_escape) {
3364 if (domain->sig.dif.ref_escape)
3365 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPREF_ESCAPE;
3366 else
3367 inl->dif_inc_ref_guard_check |= MLX5_BSF_APPTAG_ESCAPE;
e6631814
SG		 3368 }
3369
78eda2bb
SG		 3370 inl->dif_app_bitmask_check =
3371 cpu_to_be16(domain->sig.dif.apptag_check_mask);
e6631814
SG		 3372}
3373
3374static int mlx5_set_bsf(struct ib_mr *sig_mr,
3375 struct ib_sig_attrs *sig_attrs,
3376 struct mlx5_bsf *bsf, u32 data_size)
3377{
3378 struct mlx5_core_sig_ctx *msig = to_mmr(sig_mr)->sig;
3379 struct mlx5_bsf_basic *basic = &bsf->basic;
3380 struct ib_sig_domain *mem = &sig_attrs->mem;
3381 struct ib_sig_domain *wire = &sig_attrs->wire;
e6631814 3382
c7f44fbd 3383 memset(bsf, 0, sizeof(*bsf));
78eda2bb
SG		 3384
3385 /* Basic + Extended + Inline */
3386 basic->bsf_size_sbs = 1 << 7;
3387 /* Input domain check byte mask */
3388 basic->check_byte_mask = sig_attrs->check_mask;
3389 basic->raw_data_size = cpu_to_be32(data_size);
3390
3391 /* Memory domain */
e6631814 3392 switch (sig_attrs->mem.sig_type) {
78eda2bb
SG		 3393 case IB_SIG_TYPE_NONE:
3394 break;
e6631814 3395 case IB_SIG_TYPE_T10_DIF:
78eda2bb
SG		 3396 basic->mem.bs_selector = bs_selector(mem->sig.dif.pi_interval);
3397 basic->m_bfs_psv = cpu_to_be32(msig->psv_memory.psv_idx);
3398 mlx5_fill_inl_bsf(mem, &bsf->m_inl);
3399 break;
3400 default:
3401 return -EINVAL;
3402 }
e6631814 3403
78eda2bb
SG		 3404 /* Wire domain */
3405 switch (sig_attrs->wire.sig_type) {
3406 case IB_SIG_TYPE_NONE:
3407 break;
3408 case IB_SIG_TYPE_T10_DIF:
e6631814 3409 if (mem->sig.dif.pi_interval == wire->sig.dif.pi_interval &&
78eda2bb 3410 mem->sig_type == wire->sig_type) {
e6631814 3411 /* Same block structure */
142537f4 3412 basic->bsf_size_sbs |= 1 << 4;
e6631814 3413 if (mem->sig.dif.bg_type == wire->sig.dif.bg_type)
fd22f78c 3414 basic->wire.copy_byte_mask |= MLX5_CPY_GRD_MASK;
c7f44fbd 3415 if (mem->sig.dif.app_tag == wire->sig.dif.app_tag)
fd22f78c 3416 basic->wire.copy_byte_mask |= MLX5_CPY_APP_MASK;
c7f44fbd 3417 if (mem->sig.dif.ref_tag == wire->sig.dif.ref_tag)
fd22f78c 3418 basic->wire.copy_byte_mask |= MLX5_CPY_REF_MASK;
e6631814
SG		 3419 } else
3420 basic->wire.bs_selector = bs_selector(wire->sig.dif.pi_interval);
3421
142537f4 3422 basic->w_bfs_psv = cpu_to_be32(msig->psv_wire.psv_idx);
78eda2bb 3423 mlx5_fill_inl_bsf(wire, &bsf->w_inl);
e6631814 3424 break;
e6631814
SG		 3425 default:
3426 return -EINVAL;
3427 }
3428
3429 return 0;
3430}
3431
e622f2f4
CH		 3432static int set_sig_data_segment(struct ib_sig_handover_wr *wr,
3433 struct mlx5_ib_qp *qp, void **seg, int *size)
e6631814 3434{
e622f2f4
CH		 3435 struct ib_sig_attrs *sig_attrs = wr->sig_attrs;
3436 struct ib_mr *sig_mr = wr->sig_mr;
e6631814 3437 struct mlx5_bsf *bsf;
e622f2f4
CH		 3438 u32 data_len = wr->wr.sg_list->length;
3439 u32 data_key = wr->wr.sg_list->lkey;
3440 u64 data_va = wr->wr.sg_list->addr;
e6631814
SG		 3441 int ret;
3442 int wqe_size;
3443
e622f2f4
CH		 3444 if (!wr->prot ||
3445 (data_key == wr->prot->lkey &&
3446 data_va == wr->prot->addr &&
3447 data_len == wr->prot->length)) {
e6631814
SG		 3448 /**
3449 * Source domain doesn't contain signature information
5c273b16 3450 * or data and protection are interleaved in memory.
e6631814
SG		 3451 * So need construct:
3452 * ------------------
3453 * | data_klm |
3454 * ------------------
3455 * | BSF |
3456 * ------------------
3457 **/
3458 struct mlx5_klm *data_klm = *seg;
3459
3460 data_klm->bcount = cpu_to_be32(data_len);
3461 data_klm->key = cpu_to_be32(data_key);
3462 data_klm->va = cpu_to_be64(data_va);
3463 wqe_size = ALIGN(sizeof(*data_klm), 64);
3464 } else {
3465 /**
3466 * Source domain contains signature information
3467 * So need construct a strided block format:
3468 * ---------------------------
3469 * | stride_block_ctrl |
3470 * ---------------------------
3471 * | data_klm |
3472 * ---------------------------
3473 * | prot_klm |
3474 * ---------------------------
3475 * | BSF |
3476 * ---------------------------
3477 **/
3478 struct mlx5_stride_block_ctrl_seg *sblock_ctrl;
3479 struct mlx5_stride_block_entry *data_sentry;
3480 struct mlx5_stride_block_entry *prot_sentry;
e622f2f4
CH		 3481 u32 prot_key = wr->prot->lkey;
3482 u64 prot_va = wr->prot->addr;
e6631814
SG		 3483 u16 block_size = sig_attrs->mem.sig.dif.pi_interval;
3484 int prot_size;
3485
3486 sblock_ctrl = *seg;
3487 data_sentry = (void *)sblock_ctrl + sizeof(*sblock_ctrl);
3488 prot_sentry = (void *)data_sentry + sizeof(*data_sentry);
3489
3490 prot_size = prot_field_size(sig_attrs->mem.sig_type);
3491 if (!prot_size) {
3492 pr_err("Bad block size given: %u\n", block_size);
3493 return -EINVAL;
3494 }
3495 sblock_ctrl->bcount_per_cycle = cpu_to_be32(block_size +
3496 prot_size);
3497 sblock_ctrl->op = cpu_to_be32(MLX5_STRIDE_BLOCK_OP);
3498 sblock_ctrl->repeat_count = cpu_to_be32(data_len / block_size);
3499 sblock_ctrl->num_entries = cpu_to_be16(2);
3500
3501 data_sentry->bcount = cpu_to_be16(block_size);
3502 data_sentry->key = cpu_to_be32(data_key);
3503 data_sentry->va = cpu_to_be64(data_va);
5c273b16
SG		 3504 data_sentry->stride = cpu_to_be16(block_size);
3505
e6631814
SG		 3506 prot_sentry->bcount = cpu_to_be16(prot_size);
3507 prot_sentry->key = cpu_to_be32(prot_key);
5c273b16
SG		 3508 prot_sentry->va = cpu_to_be64(prot_va);
3509 prot_sentry->stride = cpu_to_be16(prot_size);
e6631814 3510
e6631814
SG		 3511 wqe_size = ALIGN(sizeof(*sblock_ctrl) + sizeof(*data_sentry) +
3512 sizeof(*prot_sentry), 64);
3513 }
3514
3515 *seg += wqe_size;
3516 *size += wqe_size / 16;
3517 if (unlikely((*seg == qp->sq.qend)))
3518 *seg = mlx5_get_send_wqe(qp, 0);
3519
3520 bsf = *seg;
3521 ret = mlx5_set_bsf(sig_mr, sig_attrs, bsf, data_len);
3522 if (ret)
3523 return -EINVAL;
3524
3525 *seg += sizeof(*bsf);
3526 *size += sizeof(*bsf) / 16;
3527 if (unlikely((*seg == qp->sq.qend)))
3528 *seg = mlx5_get_send_wqe(qp, 0);
3529
3530 return 0;
3531}
3532
3533static void set_sig_mkey_segment(struct mlx5_mkey_seg *seg,
e622f2f4 3534 struct ib_sig_handover_wr *wr, u32 nelements,
e6631814
SG		 3535 u32 length, u32 pdn)
3536{
e622f2f4 3537 struct ib_mr *sig_mr = wr->sig_mr;
e6631814 3538 u32 sig_key = sig_mr->rkey;
d5436ba0 3539 u8 sigerr = to_mmr(sig_mr)->sig->sigerr_count & 1;
e6631814
SG		 3540
3541 memset(seg, 0, sizeof(*seg));
3542
e622f2f4 3543 seg->flags = get_umr_flags(wr->access_flags) |
ec22eb53 3544 MLX5_MKC_ACCESS_MODE_KLMS;
e6631814 3545 seg->qpn_mkey7_0 = cpu_to_be32((sig_key & 0xff) | 0xffffff00);
d5436ba0 3546 seg->flags_pd = cpu_to_be32(MLX5_MKEY_REMOTE_INVAL | sigerr << 26 |
e6631814
SG		 3547 MLX5_MKEY_BSF_EN | pdn);
3548 seg->len = cpu_to_be64(length);
3549 seg->xlt_oct_size = cpu_to_be32(be16_to_cpu(get_klm_octo(nelements)));
3550 seg->bsfs_octo_size = cpu_to_be32(MLX5_MKEY_BSF_OCTO_SIZE);
3551}
3552
3553static void set_sig_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
e622f2f4 3554 u32 nelements)
e6631814
SG		 3555{
3556 memset(umr, 0, sizeof(*umr));
3557
3558 umr->flags = MLX5_FLAGS_INLINE | MLX5_FLAGS_CHECK_FREE;
3559 umr->klm_octowords = get_klm_octo(nelements);
3560 umr->bsf_octowords = cpu_to_be16(MLX5_MKEY_BSF_OCTO_SIZE);
3561 umr->mkey_mask = sig_mkey_mask();
3562}
3563
3564
e622f2f4 3565static int set_sig_umr_wr(struct ib_send_wr *send_wr, struct mlx5_ib_qp *qp,
e6631814
SG		 3566 void **seg, int *size)
3567{
e622f2f4
CH		 3568 struct ib_sig_handover_wr *wr = sig_handover_wr(send_wr);
3569 struct mlx5_ib_mr *sig_mr = to_mmr(wr->sig_mr);
e6631814
SG		 3570 u32 pdn = get_pd(qp)->pdn;
3571 u32 klm_oct_size;
3572 int region_len, ret;
3573
e622f2f4
CH		 3574 if (unlikely(wr->wr.num_sge != 1) ||
3575 unlikely(wr->access_flags & IB_ACCESS_REMOTE_ATOMIC) ||
d5436ba0
SG		 3576 unlikely(!sig_mr->sig) || unlikely(!qp->signature_en) ||
3577 unlikely(!sig_mr->sig->sig_status_checked))
e6631814
SG		 3578 return -EINVAL;
3579
3580 /* length of the protected region, data + protection */
e622f2f4
CH		 3581 region_len = wr->wr.sg_list->length;
3582 if (wr->prot &&
3583 (wr->prot->lkey != wr->wr.sg_list->lkey ||
3584 wr->prot->addr != wr->wr.sg_list->addr ||
3585 wr->prot->length != wr->wr.sg_list->length))
3586 region_len += wr->prot->length;
e6631814
SG		 3587
3588 /**
3589 * KLM octoword size - if protection was provided
3590 * then we use strided block format (3 octowords),
3591 * else we use single KLM (1 octoword)
3592 **/
e622f2f4 3593 klm_oct_size = wr->prot ? 3 : 1;
e6631814 3594
e622f2f4 3595 set_sig_umr_segment(*seg, klm_oct_size);
e6631814
SG		 3596 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3597 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3598 if (unlikely((*seg == qp->sq.qend)))
3599 *seg = mlx5_get_send_wqe(qp, 0);
3600
3601 set_sig_mkey_segment(*seg, wr, klm_oct_size, region_len, pdn);
3602 *seg += sizeof(struct mlx5_mkey_seg);
3603 *size += sizeof(struct mlx5_mkey_seg) / 16;
3604 if (unlikely((*seg == qp->sq.qend)))
3605 *seg = mlx5_get_send_wqe(qp, 0);
3606
3607 ret = set_sig_data_segment(wr, qp, seg, size);
3608 if (ret)
3609 return ret;
3610
d5436ba0 3611 sig_mr->sig->sig_status_checked = false;
e6631814
SG		 3612 return 0;
3613}
3614
3615static int set_psv_wr(struct ib_sig_domain *domain,
3616 u32 psv_idx, void **seg, int *size)
3617{
3618 struct mlx5_seg_set_psv *psv_seg = *seg;
3619
3620 memset(psv_seg, 0, sizeof(*psv_seg));
3621 psv_seg->psv_num = cpu_to_be32(psv_idx);
3622 switch (domain->sig_type) {
78eda2bb
SG		 3623 case IB_SIG_TYPE_NONE:
3624 break;
e6631814
SG		 3625 case IB_SIG_TYPE_T10_DIF:
3626 psv_seg->transient_sig = cpu_to_be32(domain->sig.dif.bg << 16 |
3627 domain->sig.dif.app_tag);
3628 psv_seg->ref_tag = cpu_to_be32(domain->sig.dif.ref_tag);
e6631814 3629 break;
e6631814
SG		 3630 default:
3631 pr_err("Bad signature type given.\n");
3632 return 1;
3633 }
3634
78eda2bb
SG		 3635 *seg += sizeof(*psv_seg);
3636 *size += sizeof(*psv_seg) / 16;
3637
e6631814
SG		 3638 return 0;
3639}
3640
8a187ee5
SG		 3641static int set_reg_wr(struct mlx5_ib_qp *qp,
3642 struct ib_reg_wr *wr,
3643 void **seg, int *size)
3644{
3645 struct mlx5_ib_mr *mr = to_mmr(wr->mr);
3646 struct mlx5_ib_pd *pd = to_mpd(qp->ibqp.pd);
3647
3648 if (unlikely(wr->wr.send_flags & IB_SEND_INLINE)) {
3649 mlx5_ib_warn(to_mdev(qp->ibqp.device),
3650 "Invalid IB_SEND_INLINE send flag\n");
3651 return -EINVAL;
3652 }
3653
3654 set_reg_umr_seg(*seg, mr);
3655 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3656 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3657 if (unlikely((*seg == qp->sq.qend)))
3658 *seg = mlx5_get_send_wqe(qp, 0);
3659
3660 set_reg_mkey_seg(*seg, mr, wr->key, wr->access);
3661 *seg += sizeof(struct mlx5_mkey_seg);
3662 *size += sizeof(struct mlx5_mkey_seg) / 16;
3663 if (unlikely((*seg == qp->sq.qend)))
3664 *seg = mlx5_get_send_wqe(qp, 0);
3665
3666 set_reg_data_seg(*seg, mr, pd);
3667 *seg += sizeof(struct mlx5_wqe_data_seg);
3668 *size += (sizeof(struct mlx5_wqe_data_seg) / 16);
3669
3670 return 0;
3671}
3672
dd01e66a 3673static void set_linv_wr(struct mlx5_ib_qp *qp, void **seg, int *size)
e126ba97 3674{
dd01e66a 3675 set_linv_umr_seg(*seg);
e126ba97
EC		 3676 *seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
3677 *size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
3678 if (unlikely((*seg == qp->sq.qend)))
3679 *seg = mlx5_get_send_wqe(qp, 0);
dd01e66a 3680 set_linv_mkey_seg(*seg);
e126ba97
EC		 3681 *seg += sizeof(struct mlx5_mkey_seg);
3682 *size += sizeof(struct mlx5_mkey_seg) / 16;
3683 if (unlikely((*seg == qp->sq.qend)))
3684 *seg = mlx5_get_send_wqe(qp, 0);
e126ba97
EC		 3685}
3686
3687static void dump_wqe(struct mlx5_ib_qp *qp, int idx, int size_16)
3688{
3689 __be32 *p = NULL;
3690 int tidx = idx;
3691 int i, j;
3692
3693 pr_debug("dump wqe at %p\n", mlx5_get_send_wqe(qp, tidx));
3694 for (i = 0, j = 0; i < size_16 * 4; i += 4, j += 4) {
3695 if ((i & 0xf) == 0) {
3696 void *buf = mlx5_get_send_wqe(qp, tidx);
3697 tidx = (tidx + 1) & (qp->sq.wqe_cnt - 1);
3698 p = buf;
3699 j = 0;
3700 }
3701 pr_debug("%08x %08x %08x %08x\n", be32_to_cpu(p[j]),
3702 be32_to_cpu(p[j + 1]), be32_to_cpu(p[j + 2]),
3703 be32_to_cpu(p[j + 3]));
3704 }
3705}
3706
3707static void mlx5_bf_copy(u64 __iomem *dst, u64 *src,
3708 unsigned bytecnt, struct mlx5_ib_qp *qp)
3709{
3710 while (bytecnt > 0) {
3711 __iowrite64_copy(dst++, src++, 8);
3712 __iowrite64_copy(dst++, src++, 8);
3713 __iowrite64_copy(dst++, src++, 8);
3714 __iowrite64_copy(dst++, src++, 8);
3715 __iowrite64_copy(dst++, src++, 8);
3716 __iowrite64_copy(dst++, src++, 8);
3717 __iowrite64_copy(dst++, src++, 8);
3718 __iowrite64_copy(dst++, src++, 8);
3719 bytecnt -= 64;
3720 if (unlikely(src == qp->sq.qend))
3721 src = mlx5_get_send_wqe(qp, 0);
3722 }
3723}
3724
3725static u8 get_fence(u8 fence, struct ib_send_wr *wr)
3726{
3727 if (unlikely(wr->opcode == IB_WR_LOCAL_INV &&
3728 wr->send_flags & IB_SEND_FENCE))
3729 return MLX5_FENCE_MODE_STRONG_ORDERING;
3730
3731 if (unlikely(fence)) {
3732 if (wr->send_flags & IB_SEND_FENCE)
3733 return MLX5_FENCE_MODE_SMALL_AND_FENCE;
3734 else
3735 return fence;
c9b25495
EC		 3736 } else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
3737 return MLX5_FENCE_MODE_FENCE;
e126ba97 3738 }
c9b25495
EC		 3739
3740 return 0;
e126ba97
EC		 3741}
3742
6e5eadac
SG		 3743static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
3744 struct mlx5_wqe_ctrl_seg **ctrl,
6a4f139a 3745 struct ib_send_wr *wr, unsigned *idx,
6e5eadac
SG		 3746 int *size, int nreq)
3747{
b2a232d2
LR		 3748 if (unlikely(mlx5_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)))
3749 return -ENOMEM;
6e5eadac
SG		 3750
3751 *idx = qp->sq.cur_post & (qp->sq.wqe_cnt - 1);
3752 *seg = mlx5_get_send_wqe(qp, *idx);
3753 *ctrl = *seg;
3754 *(uint32_t *)(*seg + 8) = 0;
3755 (*ctrl)->imm = send_ieth(wr);
3756 (*ctrl)->fm_ce_se = qp->sq_signal_bits |
3757 (wr->send_flags & IB_SEND_SIGNALED ?
3758 MLX5_WQE_CTRL_CQ_UPDATE : 0) |
3759 (wr->send_flags & IB_SEND_SOLICITED ?
3760 MLX5_WQE_CTRL_SOLICITED : 0);
3761
3762 *seg += sizeof(**ctrl);
3763 *size = sizeof(**ctrl) / 16;
3764
b2a232d2 3765 return 0;
6e5eadac
SG		 3766}
3767
3768static void finish_wqe(struct mlx5_ib_qp *qp,
3769 struct mlx5_wqe_ctrl_seg *ctrl,
3770 u8 size, unsigned idx, u64 wr_id,
3771 int nreq, u8 fence, u8 next_fence,
3772 u32 mlx5_opcode)
3773{
3774 u8 opmod = 0;
3775
3776 ctrl->opmod_idx_opcode = cpu_to_be32(((u32)(qp->sq.cur_post) << 8) |
3777 mlx5_opcode | ((u32)opmod << 24));
19098df2 3778 ctrl->qpn_ds = cpu_to_be32(size | (qp->trans_qp.base.mqp.qpn << 8));
6e5eadac
SG		 3779 ctrl->fm_ce_se |= fence;
3780 qp->fm_cache = next_fence;
3781 if (unlikely(qp->wq_sig))
3782 ctrl->signature = wq_sig(ctrl);
3783
3784 qp->sq.wrid[idx] = wr_id;
3785 qp->sq.w_list[idx].opcode = mlx5_opcode;
3786 qp->sq.wqe_head[idx] = qp->sq.head + nreq;
3787 qp->sq.cur_post += DIV_ROUND_UP(size * 16, MLX5_SEND_WQE_BB);
3788 qp->sq.w_list[idx].next = qp->sq.cur_post;
3789}
3790
3791
e126ba97
EC		 3792int mlx5_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
3793 struct ib_send_wr **bad_wr)
3794{
3795 struct mlx5_wqe_ctrl_seg *ctrl = NULL; /* compiler warning */
3796 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
89ea94a7 3797 struct mlx5_core_dev *mdev = dev->mdev;
d16e91da 3798 struct mlx5_ib_qp *qp;
e6631814 3799 struct mlx5_ib_mr *mr;
e126ba97
EC		 3800 struct mlx5_wqe_data_seg *dpseg;
3801 struct mlx5_wqe_xrc_seg *xrc;
d16e91da 3802 struct mlx5_bf *bf;
e126ba97 3803 int uninitialized_var(size);
d16e91da 3804 void *qend;
e126ba97 3805 unsigned long flags;
e126ba97
EC		 3806 unsigned idx;
3807 int err = 0;
3808 int inl = 0;
3809 int num_sge;
3810 void *seg;
3811 int nreq;
3812 int i;
3813 u8 next_fence = 0;
e126ba97
EC		 3814 u8 fence;
3815
d16e91da
HE		 3816 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
3817 return mlx5_ib_gsi_post_send(ibqp, wr, bad_wr);
3818
3819 qp = to_mqp(ibqp);
3820 bf = qp->bf;
3821 qend = qp->sq.qend;
3822
e126ba97
EC		 3823 spin_lock_irqsave(&qp->sq.lock, flags);
3824
89ea94a7
MG		 3825 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
3826 err = -EIO;
3827 *bad_wr = wr;
3828 nreq = 0;
3829 goto out;
3830 }
3831
e126ba97 3832 for (nreq = 0; wr; nreq++, wr = wr->next) {
a8f731eb 3833 if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
e126ba97
EC		 3834 mlx5_ib_warn(dev, "\n");
3835 err = -EINVAL;
3836 *bad_wr = wr;
3837 goto out;
3838 }
3839
6e5eadac
SG		 3840 fence = qp->fm_cache;
3841 num_sge = wr->num_sge;
3842 if (unlikely(num_sge > qp->sq.max_gs)) {
e126ba97 3843 mlx5_ib_warn(dev, "\n");
24be409b 3844 err = -EINVAL;
e126ba97
EC		 3845 *bad_wr = wr;
3846 goto out;
3847 }
3848
6e5eadac
SG		 3849 err = begin_wqe(qp, &seg, &ctrl, wr, &idx, &size, nreq);
3850 if (err) {
e126ba97
EC		 3851 mlx5_ib_warn(dev, "\n");
3852 err = -ENOMEM;
3853 *bad_wr = wr;
3854 goto out;
3855 }
3856
e126ba97
EC		 3857 switch (ibqp->qp_type) {
3858 case IB_QPT_XRC_INI:
3859 xrc = seg;
e126ba97
EC		 3860 seg += sizeof(*xrc);
3861 size += sizeof(*xrc) / 16;
3862 /* fall through */
3863 case IB_QPT_RC:
3864 switch (wr->opcode) {
3865 case IB_WR_RDMA_READ:
3866 case IB_WR_RDMA_WRITE:
3867 case IB_WR_RDMA_WRITE_WITH_IMM:
e622f2f4
CH		 3868 set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3869 rdma_wr(wr)->rkey);
f241e749 3870 seg += sizeof(struct mlx5_wqe_raddr_seg);
e126ba97
EC		 3871 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3872 break;
3873
3874 case IB_WR_ATOMIC_CMP_AND_SWP:
3875 case IB_WR_ATOMIC_FETCH_AND_ADD:
e126ba97 3876 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
81bea28f
EC		 3877 mlx5_ib_warn(dev, "Atomic operations are not supported yet\n");
3878 err = -ENOSYS;
3879 *bad_wr = wr;
3880 goto out;
e126ba97
EC		 3881
3882 case IB_WR_LOCAL_INV:
3883 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3884 qp->sq.wr_data[idx] = IB_WR_LOCAL_INV;
3885 ctrl->imm = cpu_to_be32(wr->ex.invalidate_rkey);
dd01e66a 3886 set_linv_wr(qp, &seg, &size);
e126ba97
EC		 3887 num_sge = 0;
3888 break;
3889
8a187ee5
SG		 3890 case IB_WR_REG_MR:
3891 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
3892 qp->sq.wr_data[idx] = IB_WR_REG_MR;
3893 ctrl->imm = cpu_to_be32(reg_wr(wr)->key);
3894 err = set_reg_wr(qp, reg_wr(wr), &seg, &size);
3895 if (err) {
3896 *bad_wr = wr;
3897 goto out;
3898 }
3899 num_sge = 0;
3900 break;
3901
e6631814
SG		 3902 case IB_WR_REG_SIG_MR:
3903 qp->sq.wr_data[idx] = IB_WR_REG_SIG_MR;
e622f2f4 3904 mr = to_mmr(sig_handover_wr(wr)->sig_mr);
e6631814
SG		 3905
3906 ctrl->imm = cpu_to_be32(mr->ibmr.rkey);
3907 err = set_sig_umr_wr(wr, qp, &seg, &size);
3908 if (err) {
3909 mlx5_ib_warn(dev, "\n");
3910 *bad_wr = wr;
3911 goto out;
3912 }
3913
3914 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3915 nreq, get_fence(fence, wr),
3916 next_fence, MLX5_OPCODE_UMR);
3917 /*
3918 * SET_PSV WQEs are not signaled and solicited
3919 * on error
3920 */
3921 wr->send_flags &= ~IB_SEND_SIGNALED;
3922 wr->send_flags |= IB_SEND_SOLICITED;
3923 err = begin_wqe(qp, &seg, &ctrl, wr,
3924 &idx, &size, nreq);
3925 if (err) {
3926 mlx5_ib_warn(dev, "\n");
3927 err = -ENOMEM;
3928 *bad_wr = wr;
3929 goto out;
3930 }
3931
e622f2f4 3932 err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->mem,
e6631814
SG		 3933 mr->sig->psv_memory.psv_idx, &seg,
3934 &size);
3935 if (err) {
3936 mlx5_ib_warn(dev, "\n");
3937 *bad_wr = wr;
3938 goto out;
3939 }
3940
3941 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3942 nreq, get_fence(fence, wr),
3943 next_fence, MLX5_OPCODE_SET_PSV);
3944 err = begin_wqe(qp, &seg, &ctrl, wr,
3945 &idx, &size, nreq);
3946 if (err) {
3947 mlx5_ib_warn(dev, "\n");
3948 err = -ENOMEM;
3949 *bad_wr = wr;
3950 goto out;
3951 }
3952
3953 next_fence = MLX5_FENCE_MODE_INITIATOR_SMALL;
e622f2f4 3954 err = set_psv_wr(&sig_handover_wr(wr)->sig_attrs->wire,
e6631814
SG		 3955 mr->sig->psv_wire.psv_idx, &seg,
3956 &size);
3957 if (err) {
3958 mlx5_ib_warn(dev, "\n");
3959 *bad_wr = wr;
3960 goto out;
3961 }
3962
3963 finish_wqe(qp, ctrl, size, idx, wr->wr_id,
3964 nreq, get_fence(fence, wr),
3965 next_fence, MLX5_OPCODE_SET_PSV);
3966 num_sge = 0;
3967 goto skip_psv;
3968
e126ba97
EC		 3969 default:
3970 break;
3971 }
3972 break;
3973
3974 case IB_QPT_UC:
3975 switch (wr->opcode) {
3976 case IB_WR_RDMA_WRITE:
3977 case IB_WR_RDMA_WRITE_WITH_IMM:
e622f2f4
CH		 3978 set_raddr_seg(seg, rdma_wr(wr)->remote_addr,
3979 rdma_wr(wr)->rkey);
e126ba97
EC		 3980 seg += sizeof(struct mlx5_wqe_raddr_seg);
3981 size += sizeof(struct mlx5_wqe_raddr_seg) / 16;
3982 break;
3983
3984 default:
3985 break;
3986 }
3987 break;
3988
e126ba97 3989 case IB_QPT_SMI:
d16e91da 3990 case MLX5_IB_QPT_HW_GSI:
e126ba97 3991 set_datagram_seg(seg, wr);
f241e749 3992 seg += sizeof(struct mlx5_wqe_datagram_seg);
e126ba97
EC		 3993 size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
3994 if (unlikely((seg == qend)))
3995 seg = mlx5_get_send_wqe(qp, 0);
3996 break;
f0313965
ES		 3997 case IB_QPT_UD:
3998 set_datagram_seg(seg, wr);
3999 seg += sizeof(struct mlx5_wqe_datagram_seg);
4000 size += sizeof(struct mlx5_wqe_datagram_seg) / 16;
4001
4002 if (unlikely((seg == qend)))
4003 seg = mlx5_get_send_wqe(qp, 0);
4004
4005 /* handle qp that supports ud offload */
4006 if (qp->flags & IB_QP_CREATE_IPOIB_UD_LSO) {
4007 struct mlx5_wqe_eth_pad *pad;
e126ba97 4008
f0313965
ES		 4009 pad = seg;
4010 memset(pad, 0, sizeof(struct mlx5_wqe_eth_pad));
4011 seg += sizeof(struct mlx5_wqe_eth_pad);
4012 size += sizeof(struct mlx5_wqe_eth_pad) / 16;
4013
4014 seg = set_eth_seg(seg, wr, qend, qp, &size);
4015
4016 if (unlikely((seg == qend)))
4017 seg = mlx5_get_send_wqe(qp, 0);
4018 }
4019 break;
e126ba97
EC		 4020 case MLX5_IB_QPT_REG_UMR:
4021 if (wr->opcode != MLX5_IB_WR_UMR) {
4022 err = -EINVAL;
4023 mlx5_ib_warn(dev, "bad opcode\n");
4024 goto out;
4025 }
4026 qp->sq.wr_data[idx] = MLX5_IB_WR_UMR;
e622f2f4 4027 ctrl->imm = cpu_to_be32(umr_wr(wr)->mkey);
e126ba97
EC		 4028 set_reg_umr_segment(seg, wr);
4029 seg += sizeof(struct mlx5_wqe_umr_ctrl_seg);
4030 size += sizeof(struct mlx5_wqe_umr_ctrl_seg) / 16;
4031 if (unlikely((seg == qend)))
4032 seg = mlx5_get_send_wqe(qp, 0);
4033 set_reg_mkey_segment(seg, wr);
4034 seg += sizeof(struct mlx5_mkey_seg);
4035 size += sizeof(struct mlx5_mkey_seg) / 16;
4036 if (unlikely((seg == qend)))
4037 seg = mlx5_get_send_wqe(qp, 0);
4038 break;
4039
4040 default:
4041 break;
4042 }
4043
4044 if (wr->send_flags & IB_SEND_INLINE && num_sge) {
4045 int uninitialized_var(sz);
4046
4047 err = set_data_inl_seg(qp, wr, seg, &sz);
4048 if (unlikely(err)) {
4049 mlx5_ib_warn(dev, "\n");
4050 *bad_wr = wr;
4051 goto out;
4052 }
4053 inl = 1;
4054 size += sz;
4055 } else {
4056 dpseg = seg;
4057 for (i = 0; i < num_sge; i++) {
4058 if (unlikely(dpseg == qend)) {
4059 seg = mlx5_get_send_wqe(qp, 0);
4060 dpseg = seg;
4061 }
4062 if (likely(wr->sg_list[i].length)) {
4063 set_data_ptr_seg(dpseg, wr->sg_list + i);
4064 size += sizeof(struct mlx5_wqe_data_seg) / 16;
4065 dpseg++;
4066 }
4067 }
4068 }
4069
6e5eadac
SG		 4070 finish_wqe(qp, ctrl, size, idx, wr->wr_id, nreq,
4071 get_fence(fence, wr), next_fence,
4072 mlx5_ib_opcode[wr->opcode]);
e6631814 4073skip_psv:
e126ba97
EC		 4074 if (0)
4075 dump_wqe(qp, idx, size);
4076 }
4077
4078out:
4079 if (likely(nreq)) {
4080 qp->sq.head += nreq;
4081
4082 /* Make sure that descriptors are written before
4083 * updating doorbell record and ringing the doorbell
4084 */
4085 wmb();
4086
4087 qp->db.db[MLX5_SND_DBR] = cpu_to_be32(qp->sq.cur_post);
4088
ada388f7
EC		 4089 /* Make sure doorbell record is visible to the HCA before
4090 * we hit doorbell */
4091 wmb();
4092
e126ba97
EC		 4093 if (bf->need_lock)
4094 spin_lock(&bf->lock);
6a4f139a
EC		 4095 else
4096 __acquire(&bf->lock);
e126ba97
EC		 4097
4098 /* TBD enable WC */
4099 if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
4100 mlx5_bf_copy(bf->reg + bf->offset, (u64 *)ctrl, ALIGN(size * 16, 64), qp);
4101 /* wc_wmb(); */
4102 } else {
4103 mlx5_write64((__be32 *)ctrl, bf->regreg + bf->offset,
4104 MLX5_GET_DOORBELL_LOCK(&bf->lock32));
4105 /* Make sure doorbells don't leak out of SQ spinlock
4106 * and reach the HCA out of order.
4107 */
4108 mmiowb();
4109 }
4110 bf->offset ^= bf->buf_size;
4111 if (bf->need_lock)
4112 spin_unlock(&bf->lock);
6a4f139a
EC		 4113 else
4114 __release(&bf->lock);
e126ba97
EC		 4115 }
4116
4117 spin_unlock_irqrestore(&qp->sq.lock, flags);
4118
4119 return err;
4120}
4121
4122static void set_sig_seg(struct mlx5_rwqe_sig *sig, int size)
4123{
4124 sig->signature = calc_sig(sig, size);
4125}
4126
4127int mlx5_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
4128 struct ib_recv_wr **bad_wr)
4129{
4130 struct mlx5_ib_qp *qp = to_mqp(ibqp);
4131 struct mlx5_wqe_data_seg *scat;
4132 struct mlx5_rwqe_sig *sig;
89ea94a7
MG		 4133 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4134 struct mlx5_core_dev *mdev = dev->mdev;
e126ba97
EC		 4135 unsigned long flags;
4136 int err = 0;
4137 int nreq;
4138 int ind;
4139 int i;
4140
d16e91da
HE		 4141 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4142 return mlx5_ib_gsi_post_recv(ibqp, wr, bad_wr);
4143
e126ba97
EC		 4144 spin_lock_irqsave(&qp->rq.lock, flags);
4145
89ea94a7
MG		 4146 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
4147 err = -EIO;
4148 *bad_wr = wr;
4149 nreq = 0;
4150 goto out;
4151 }
4152
e126ba97
EC		 4153 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
4154
4155 for (nreq = 0; wr; nreq++, wr = wr->next) {
4156 if (mlx5_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
4157 err = -ENOMEM;
4158 *bad_wr = wr;
4159 goto out;
4160 }
4161
4162 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
4163 err = -EINVAL;
4164 *bad_wr = wr;
4165 goto out;
4166 }
4167
4168 scat = get_recv_wqe(qp, ind);
4169 if (qp->wq_sig)
4170 scat++;
4171
4172 for (i = 0; i < wr->num_sge; i++)
4173 set_data_ptr_seg(scat + i, wr->sg_list + i);
4174
4175 if (i < qp->rq.max_gs) {
4176 scat[i].byte_count = 0;
4177 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
4178 scat[i].addr = 0;
4179 }
4180
4181 if (qp->wq_sig) {
4182 sig = (struct mlx5_rwqe_sig *)scat;
4183 set_sig_seg(sig, (qp->rq.max_gs + 1) << 2);
4184 }
4185
4186 qp->rq.wrid[ind] = wr->wr_id;
4187
4188 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
4189 }
4190
4191out:
4192 if (likely(nreq)) {
4193 qp->rq.head += nreq;
4194
4195 /* Make sure that descriptors are written before
4196 * doorbell record.
4197 */
4198 wmb();
4199
4200 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
4201 }
4202
4203 spin_unlock_irqrestore(&qp->rq.lock, flags);
4204
4205 return err;
4206}
4207
4208static inline enum ib_qp_state to_ib_qp_state(enum mlx5_qp_state mlx5_state)
4209{
4210 switch (mlx5_state) {
4211 case MLX5_QP_STATE_RST: return IB_QPS_RESET;
4212 case MLX5_QP_STATE_INIT: return IB_QPS_INIT;
4213 case MLX5_QP_STATE_RTR: return IB_QPS_RTR;
4214 case MLX5_QP_STATE_RTS: return IB_QPS_RTS;
4215 case MLX5_QP_STATE_SQ_DRAINING:
4216 case MLX5_QP_STATE_SQD: return IB_QPS_SQD;
4217 case MLX5_QP_STATE_SQER: return IB_QPS_SQE;
4218 case MLX5_QP_STATE_ERR: return IB_QPS_ERR;
4219 default: return -1;
4220 }
4221}
4222
4223static inline enum ib_mig_state to_ib_mig_state(int mlx5_mig_state)
4224{
4225 switch (mlx5_mig_state) {
4226 case MLX5_QP_PM_ARMED: return IB_MIG_ARMED;
4227 case MLX5_QP_PM_REARM: return IB_MIG_REARM;
4228 case MLX5_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
4229 default: return -1;
4230 }
4231}
4232
4233static int to_ib_qp_access_flags(int mlx5_flags)
4234{
4235 int ib_flags = 0;
4236
4237 if (mlx5_flags & MLX5_QP_BIT_RRE)
4238 ib_flags |= IB_ACCESS_REMOTE_READ;
4239 if (mlx5_flags & MLX5_QP_BIT_RWE)
4240 ib_flags |= IB_ACCESS_REMOTE_WRITE;
4241 if (mlx5_flags & MLX5_QP_BIT_RAE)
4242 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
4243
4244 return ib_flags;
4245}
4246
4247static void to_ib_ah_attr(struct mlx5_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
4248 struct mlx5_qp_path *path)
4249{
9603b61d 4250 struct mlx5_core_dev *dev = ibdev->mdev;
e126ba97
EC		 4251
4252 memset(ib_ah_attr, 0, sizeof(*ib_ah_attr));
4253 ib_ah_attr->port_num = path->port;
4254
c7a08ac7 4255 if (ib_ah_attr->port_num == 0 ||
938fe83c 4256 ib_ah_attr->port_num > MLX5_CAP_GEN(dev, num_ports))
e126ba97
EC		 4257 return;
4258
2811ba51 4259 ib_ah_attr->sl = path->dci_cfi_prio_sl & 0xf;
e126ba97
EC		 4260
4261 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
4262 ib_ah_attr->src_path_bits = path->grh_mlid & 0x7f;
4263 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
4264 ib_ah_attr->ah_flags = (path->grh_mlid & (1 << 7)) ? IB_AH_GRH : 0;
4265 if (ib_ah_attr->ah_flags) {
4266 ib_ah_attr->grh.sgid_index = path->mgid_index;
4267 ib_ah_attr->grh.hop_limit = path->hop_limit;
4268 ib_ah_attr->grh.traffic_class =
4269 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
4270 ib_ah_attr->grh.flow_label =
4271 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
4272 memcpy(ib_ah_attr->grh.dgid.raw,
4273 path->rgid, sizeof(ib_ah_attr->grh.dgid.raw));
4274 }
4275}
4276
6d2f89df 4277static int query_raw_packet_qp_sq_state(struct mlx5_ib_dev *dev,
4278 struct mlx5_ib_sq *sq,
4279 u8 *sq_state)
4280{
4281 void *out;
4282 void *sqc;
4283 int inlen;
4284 int err;
4285
4286 inlen = MLX5_ST_SZ_BYTES(query_sq_out);
4287 out = mlx5_vzalloc(inlen);
4288 if (!out)
4289 return -ENOMEM;
4290
4291 err = mlx5_core_query_sq(dev->mdev, sq->base.mqp.qpn, out);
4292 if (err)
4293 goto out;
4294
4295 sqc = MLX5_ADDR_OF(query_sq_out, out, sq_context);
4296 *sq_state = MLX5_GET(sqc, sqc, state);
4297 sq->state = *sq_state;
4298
4299out:
4300 kvfree(out);
4301 return err;
4302}
4303
4304static int query_raw_packet_qp_rq_state(struct mlx5_ib_dev *dev,
4305 struct mlx5_ib_rq *rq,
4306 u8 *rq_state)
4307{
4308 void *out;
4309 void *rqc;
4310 int inlen;
4311 int err;
4312
4313 inlen = MLX5_ST_SZ_BYTES(query_rq_out);
4314 out = mlx5_vzalloc(inlen);
4315 if (!out)
4316 return -ENOMEM;
4317
4318 err = mlx5_core_query_rq(dev->mdev, rq->base.mqp.qpn, out);
4319 if (err)
4320 goto out;
4321
4322 rqc = MLX5_ADDR_OF(query_rq_out, out, rq_context);
4323 *rq_state = MLX5_GET(rqc, rqc, state);
4324 rq->state = *rq_state;
4325
4326out:
4327 kvfree(out);
4328 return err;
4329}
4330
4331static int sqrq_state_to_qp_state(u8 sq_state, u8 rq_state,
4332 struct mlx5_ib_qp *qp, u8 *qp_state)
4333{
4334 static const u8 sqrq_trans[MLX5_RQ_NUM_STATE][MLX5_SQ_NUM_STATE] = {
4335 [MLX5_RQC_STATE_RST] = {
4336 [MLX5_SQC_STATE_RST] = IB_QPS_RESET,
4337 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE_BAD,
4338 [MLX5_SQC_STATE_ERR] = MLX5_QP_STATE_BAD,
4339 [MLX5_SQ_STATE_NA] = IB_QPS_RESET,
4340 },
4341 [MLX5_RQC_STATE_RDY] = {
4342 [MLX5_SQC_STATE_RST] = MLX5_QP_STATE_BAD,
4343 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE,
4344 [MLX5_SQC_STATE_ERR] = IB_QPS_SQE,
4345 [MLX5_SQ_STATE_NA] = MLX5_QP_STATE,
4346 },
4347 [MLX5_RQC_STATE_ERR] = {
4348 [MLX5_SQC_STATE_RST] = MLX5_QP_STATE_BAD,
4349 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE_BAD,
4350 [MLX5_SQC_STATE_ERR] = IB_QPS_ERR,
4351 [MLX5_SQ_STATE_NA] = IB_QPS_ERR,
4352 },
4353 [MLX5_RQ_STATE_NA] = {
4354 [MLX5_SQC_STATE_RST] = IB_QPS_RESET,
4355 [MLX5_SQC_STATE_RDY] = MLX5_QP_STATE,
4356 [MLX5_SQC_STATE_ERR] = MLX5_QP_STATE,
4357 [MLX5_SQ_STATE_NA] = MLX5_QP_STATE_BAD,
4358 },
4359 };
4360
4361 *qp_state = sqrq_trans[rq_state][sq_state];
4362
4363 if (*qp_state == MLX5_QP_STATE_BAD) {
4364 WARN(1, "Buggy Raw Packet QP state, SQ 0x%x state: 0x%x, RQ 0x%x state: 0x%x",
4365 qp->raw_packet_qp.sq.base.mqp.qpn, sq_state,
4366 qp->raw_packet_qp.rq.base.mqp.qpn, rq_state);
4367 return -EINVAL;
4368 }
4369
4370 if (*qp_state == MLX5_QP_STATE)
4371 *qp_state = qp->state;
4372
4373 return 0;
4374}
4375
4376static int query_raw_packet_qp_state(struct mlx5_ib_dev *dev,
4377 struct mlx5_ib_qp *qp,
4378 u8 *raw_packet_qp_state)
4379{
4380 struct mlx5_ib_raw_packet_qp *raw_packet_qp = &qp->raw_packet_qp;
4381 struct mlx5_ib_sq *sq = &raw_packet_qp->sq;
4382 struct mlx5_ib_rq *rq = &raw_packet_qp->rq;
4383 int err;
4384 u8 sq_state = MLX5_SQ_STATE_NA;
4385 u8 rq_state = MLX5_RQ_STATE_NA;
4386
4387 if (qp->sq.wqe_cnt) {
4388 err = query_raw_packet_qp_sq_state(dev, sq, &sq_state);
4389 if (err)
4390 return err;
4391 }
4392
4393 if (qp->rq.wqe_cnt) {
4394 err = query_raw_packet_qp_rq_state(dev, rq, &rq_state);
4395 if (err)
4396 return err;
4397 }
4398
4399 return sqrq_state_to_qp_state(sq_state, rq_state, qp,
4400 raw_packet_qp_state);
4401}
4402
4403static int query_qp_attr(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
4404 struct ib_qp_attr *qp_attr)
e126ba97 4405{
09a7d9ec 4406 int outlen = MLX5_ST_SZ_BYTES(query_qp_out);
e126ba97
EC		 4407 struct mlx5_qp_context *context;
4408 int mlx5_state;
09a7d9ec 4409 u32 *outb;
e126ba97
EC		 4410 int err = 0;
4411
09a7d9ec 4412 outb = kzalloc(outlen, GFP_KERNEL);
6d2f89df 4413 if (!outb)
4414 return -ENOMEM;
4415
19098df2 4416 err = mlx5_core_qp_query(dev->mdev, &qp->trans_qp.base.mqp, outb,
09a7d9ec 4417 outlen);
e126ba97 4418 if (err)
6d2f89df 4419 goto out;
e126ba97 4420
09a7d9ec
SM		 4421 /* FIXME: use MLX5_GET rather than mlx5_qp_context manual struct */
4422 context = (struct mlx5_qp_context *)MLX5_ADDR_OF(query_qp_out, outb, qpc);
4423
e126ba97
EC		 4424 mlx5_state = be32_to_cpu(context->flags) >> 28;
4425
4426 qp->state = to_ib_qp_state(mlx5_state);
e126ba97
EC		 4427 qp_attr->path_mtu = context->mtu_msgmax >> 5;
4428 qp_attr->path_mig_state =
4429 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3);
4430 qp_attr->qkey = be32_to_cpu(context->qkey);
4431 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff;
4432 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff;
4433 qp_attr->dest_qp_num = be32_to_cpu(context->log_pg_sz_remote_qpn) & 0xffffff;
4434 qp_attr->qp_access_flags =
4435 to_ib_qp_access_flags(be32_to_cpu(context->params2));
4436
4437 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
4438 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path);
4439 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path);
d3ae2bde
NO		 4440 qp_attr->alt_pkey_index =
4441 be16_to_cpu(context->alt_path.pkey_index);
e126ba97
EC		 4442 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
4443 }
4444
d3ae2bde 4445 qp_attr->pkey_index = be16_to_cpu(context->pri_path.pkey_index);
e126ba97
EC		 4446 qp_attr->port_num = context->pri_path.port;
4447
4448 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
4449 qp_attr->sq_draining = mlx5_state == MLX5_QP_STATE_SQ_DRAINING;
4450
4451 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7);
4452
4453 qp_attr->max_dest_rd_atomic =
4454 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7);
4455 qp_attr->min_rnr_timer =
4456 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f;
4457 qp_attr->timeout = context->pri_path.ackto_lt >> 3;
4458 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7;
4459 qp_attr->rnr_retry = (be32_to_cpu(context->params1) >> 13) & 0x7;
4460 qp_attr->alt_timeout = context->alt_path.ackto_lt >> 3;
6d2f89df 4461
4462out:
4463 kfree(outb);
4464 return err;
4465}
4466
4467int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
4468 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
4469{
4470 struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
4471 struct mlx5_ib_qp *qp = to_mqp(ibqp);
4472 int err = 0;
4473 u8 raw_packet_qp_state;
4474
28d61370
YH		 4475 if (ibqp->rwq_ind_tbl)
4476 return -ENOSYS;
4477
d16e91da
HE		 4478 if (unlikely(ibqp->qp_type == IB_QPT_GSI))
4479 return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
4480 qp_init_attr);
4481
6d2f89df 4482#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
4483 /*
4484 * Wait for any outstanding page faults, in case the user frees memory
4485 * based upon this query's result.
4486 */
4487 flush_workqueue(mlx5_ib_page_fault_wq);
4488#endif
4489
4490 mutex_lock(&qp->mutex);
4491
4492 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
4493 err = query_raw_packet_qp_state(dev, qp, &raw_packet_qp_state);
4494 if (err)
4495 goto out;
4496 qp->state = raw_packet_qp_state;
4497 qp_attr->port_num = 1;
4498 } else {
4499 err = query_qp_attr(dev, qp, qp_attr);
4500 if (err)
4501 goto out;
4502 }
4503
4504 qp_attr->qp_state = qp->state;
e126ba97
EC		 4505 qp_attr->cur_qp_state = qp_attr->qp_state;
4506 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
4507 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
4508
4509 if (!ibqp->uobject) {
0540d814 4510 qp_attr->cap.max_send_wr = qp->sq.max_post;
e126ba97 4511 qp_attr->cap.max_send_sge = qp->sq.max_gs;
0540d814 4512 qp_init_attr->qp_context = ibqp->qp_context;
e126ba97
EC		 4513 } else {
4514 qp_attr->cap.max_send_wr = 0;
4515 qp_attr->cap.max_send_sge = 0;
4516 }
4517
0540d814
NO		 4518 qp_init_attr->qp_type = ibqp->qp_type;
4519 qp_init_attr->recv_cq = ibqp->recv_cq;
4520 qp_init_attr->send_cq = ibqp->send_cq;
4521 qp_init_attr->srq = ibqp->srq;
4522 qp_attr->cap.max_inline_data = qp->max_inline_data;
e126ba97
EC		 4523
4524 qp_init_attr->cap = qp_attr->cap;
4525
4526 qp_init_attr->create_flags = 0;
4527 if (qp->flags & MLX5_IB_QP_BLOCK_MULTICAST_LOOPBACK)
4528 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
4529
051f2630
LR		 4530 if (qp->flags & MLX5_IB_QP_CROSS_CHANNEL)
4531 qp_init_attr->create_flags |= IB_QP_CREATE_CROSS_CHANNEL;
4532 if (qp->flags & MLX5_IB_QP_MANAGED_SEND)
4533 qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_SEND;
4534 if (qp->flags & MLX5_IB_QP_MANAGED_RECV)
4535 qp_init_attr->create_flags |= IB_QP_CREATE_MANAGED_RECV;
b11a4f9c
HE		 4536 if (qp->flags & MLX5_IB_QP_SQPN_QP1)
4537 qp_init_attr->create_flags |= mlx5_ib_create_qp_sqpn_qp1();
051f2630 4538
e126ba97
EC		 4539 qp_init_attr->sq_sig_type = qp->sq_signal_bits & MLX5_WQE_CTRL_CQ_UPDATE ?
4540 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
4541
e126ba97
EC		 4542out:
4543 mutex_unlock(&qp->mutex);
4544 return err;
4545}
4546
4547struct ib_xrcd *mlx5_ib_alloc_xrcd(struct ib_device *ibdev,
4548 struct ib_ucontext *context,
4549 struct ib_udata *udata)
4550{
4551 struct mlx5_ib_dev *dev = to_mdev(ibdev);
4552 struct mlx5_ib_xrcd *xrcd;
4553 int err;
4554
938fe83c 4555 if (!MLX5_CAP_GEN(dev->mdev, xrc))
e126ba97
EC		 4556 return ERR_PTR(-ENOSYS);
4557
4558 xrcd = kmalloc(sizeof(*xrcd), GFP_KERNEL);
4559 if (!xrcd)
4560 return ERR_PTR(-ENOMEM);
4561
9603b61d 4562 err = mlx5_core_xrcd_alloc(dev->mdev, &xrcd->xrcdn);
e126ba97
EC		 4563 if (err) {
4564 kfree(xrcd);
4565 return ERR_PTR(-ENOMEM);
4566 }
4567
4568 return &xrcd->ibxrcd;
4569}
4570
4571int mlx5_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
4572{
4573 struct mlx5_ib_dev *dev = to_mdev(xrcd->device);
4574 u32 xrcdn = to_mxrcd(xrcd)->xrcdn;
4575 int err;
4576
9603b61d 4577 err = mlx5_core_xrcd_dealloc(dev->mdev, xrcdn);
e126ba97
EC		 4578 if (err) {
4579 mlx5_ib_warn(dev, "failed to dealloc xrcdn 0x%x\n", xrcdn);
4580 return err;
4581 }
4582
4583 kfree(xrcd);
4584
4585 return 0;
4586}
79b20a6c 4587
350d0e4c
YH		 4588static void mlx5_ib_wq_event(struct mlx5_core_qp *core_qp, int type)
4589{
4590 struct mlx5_ib_rwq *rwq = to_mibrwq(core_qp);
4591 struct mlx5_ib_dev *dev = to_mdev(rwq->ibwq.device);
4592 struct ib_event event;
4593
4594 if (rwq->ibwq.event_handler) {
4595 event.device = rwq->ibwq.device;
4596 event.element.wq = &rwq->ibwq;
4597 switch (type) {
4598 case MLX5_EVENT_TYPE_WQ_CATAS_ERROR:
4599 event.event = IB_EVENT_WQ_FATAL;
4600 break;
4601 default:
4602 mlx5_ib_warn(dev, "Unexpected event type %d on WQ %06x\n", type, core_qp->qpn);
4603 return;
4604 }
4605
4606 rwq->ibwq.event_handler(&event, rwq->ibwq.wq_context);
4607 }
4608}
4609
79b20a6c
YH		 4610static int create_rq(struct mlx5_ib_rwq *rwq, struct ib_pd *pd,
4611 struct ib_wq_init_attr *init_attr)
4612{
4613 struct mlx5_ib_dev *dev;
4614 __be64 *rq_pas0;
4615 void *in;
4616 void *rqc;
4617 void *wq;
4618 int inlen;
4619 int err;
4620
4621 dev = to_mdev(pd->device);
4622
4623 inlen = MLX5_ST_SZ_BYTES(create_rq_in) + sizeof(u64) * rwq->rq_num_pas;
4624 in = mlx5_vzalloc(inlen);
4625 if (!in)
4626 return -ENOMEM;
4627
4628 rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
4629 MLX5_SET(rqc, rqc, mem_rq_type,
4630 MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE);
4631 MLX5_SET(rqc, rqc, user_index, rwq->user_index);
4632 MLX5_SET(rqc, rqc, cqn, to_mcq(init_attr->cq)->mcq.cqn);
4633 MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
4634 MLX5_SET(rqc, rqc, flush_in_error_en, 1);
4635 wq = MLX5_ADDR_OF(rqc, rqc, wq);
4636 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
4637 MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
4638 MLX5_SET(wq, wq, log_wq_stride, rwq->log_rq_stride);
4639 MLX5_SET(wq, wq, log_wq_sz, rwq->log_rq_size);
4640 MLX5_SET(wq, wq, pd, to_mpd(pd)->pdn);
4641 MLX5_SET(wq, wq, page_offset, rwq->rq_page_offset);
4642 MLX5_SET(wq, wq, log_wq_pg_sz, rwq->log_page_size);
4643 MLX5_SET(wq, wq, wq_signature, rwq->wq_sig);
4644 MLX5_SET64(wq, wq, dbr_addr, rwq->db.dma);
4645 rq_pas0 = (__be64 *)MLX5_ADDR_OF(wq, wq, pas);
4646 mlx5_ib_populate_pas(dev, rwq->umem, rwq->page_shift, rq_pas0, 0);
350d0e4c 4647 err = mlx5_core_create_rq_tracked(dev->mdev, in, inlen, &rwq->core_qp);
79b20a6c
YH		 4648 kvfree(in);
4649 return err;
4650}
4651
4652static int set_user_rq_size(struct mlx5_ib_dev *dev,
4653 struct ib_wq_init_attr *wq_init_attr,
4654 struct mlx5_ib_create_wq *ucmd,
4655 struct mlx5_ib_rwq *rwq)
4656{
4657 /* Sanity check RQ size before proceeding */
4658 if (wq_init_attr->max_wr > (1 << MLX5_CAP_GEN(dev->mdev, log_max_wq_sz)))
4659 return -EINVAL;
4660
4661 if (!ucmd->rq_wqe_count)
4662 return -EINVAL;
4663
4664 rwq->wqe_count = ucmd->rq_wqe_count;
4665 rwq->wqe_shift = ucmd->rq_wqe_shift;
4666 rwq->buf_size = (rwq->wqe_count << rwq->wqe_shift);
4667 rwq->log_rq_stride = rwq->wqe_shift;
4668 rwq->log_rq_size = ilog2(rwq->wqe_count);
4669 return 0;
4670}
4671
4672static int prepare_user_rq(struct ib_pd *pd,
4673 struct ib_wq_init_attr *init_attr,
4674 struct ib_udata *udata,
4675 struct mlx5_ib_rwq *rwq)
4676{
4677 struct mlx5_ib_dev *dev = to_mdev(pd->device);
4678 struct mlx5_ib_create_wq ucmd = {};
4679 int err;
4680 size_t required_cmd_sz;
4681
4682 required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4683 if (udata->inlen < required_cmd_sz) {
4684 mlx5_ib_dbg(dev, "invalid inlen\n");
4685 return -EINVAL;
4686 }
4687
4688 if (udata->inlen > sizeof(ucmd) &&
4689 !ib_is_udata_cleared(udata, sizeof(ucmd),
4690 udata->inlen - sizeof(ucmd))) {
4691 mlx5_ib_dbg(dev, "inlen is not supported\n");
4692 return -EOPNOTSUPP;
4693 }
4694
4695 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) {
4696 mlx5_ib_dbg(dev, "copy failed\n");
4697 return -EFAULT;
4698 }
4699
4700 if (ucmd.comp_mask) {
4701 mlx5_ib_dbg(dev, "invalid comp mask\n");
4702 return -EOPNOTSUPP;
4703 }
4704
4705 if (ucmd.reserved) {
4706 mlx5_ib_dbg(dev, "invalid reserved\n");
4707 return -EOPNOTSUPP;
4708 }
4709
4710 err = set_user_rq_size(dev, init_attr, &ucmd, rwq);
4711 if (err) {
4712 mlx5_ib_dbg(dev, "err %d\n", err);
4713 return err;
4714 }
4715
4716 err = create_user_rq(dev, pd, rwq, &ucmd);
4717 if (err) {
4718 mlx5_ib_dbg(dev, "err %d\n", err);
4719 if (err)
4720 return err;
4721 }
4722
4723 rwq->user_index = ucmd.user_index;
4724 return 0;
4725}
4726
4727struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
4728 struct ib_wq_init_attr *init_attr,
4729 struct ib_udata *udata)
4730{
4731 struct mlx5_ib_dev *dev;
4732 struct mlx5_ib_rwq *rwq;
4733 struct mlx5_ib_create_wq_resp resp = {};
4734 size_t min_resp_len;
4735 int err;
4736
4737 if (!udata)
4738 return ERR_PTR(-ENOSYS);
4739
4740 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4741 if (udata->outlen && udata->outlen < min_resp_len)
4742 return ERR_PTR(-EINVAL);
4743
4744 dev = to_mdev(pd->device);
4745 switch (init_attr->wq_type) {
4746 case IB_WQT_RQ:
4747 rwq = kzalloc(sizeof(*rwq), GFP_KERNEL);
4748 if (!rwq)
4749 return ERR_PTR(-ENOMEM);
4750 err = prepare_user_rq(pd, init_attr, udata, rwq);
4751 if (err)
4752 goto err;
4753 err = create_rq(rwq, pd, init_attr);
4754 if (err)
4755 goto err_user_rq;
4756 break;
4757 default:
4758 mlx5_ib_dbg(dev, "unsupported wq type %d\n",
4759 init_attr->wq_type);
4760 return ERR_PTR(-EINVAL);
4761 }
4762
350d0e4c 4763 rwq->ibwq.wq_num = rwq->core_qp.qpn;
79b20a6c
YH		 4764 rwq->ibwq.state = IB_WQS_RESET;
4765 if (udata->outlen) {
4766 resp.response_length = offsetof(typeof(resp), response_length) +
4767 sizeof(resp.response_length);
4768 err = ib_copy_to_udata(udata, &resp, resp.response_length);
4769 if (err)
4770 goto err_copy;
4771 }
4772
350d0e4c
YH		 4773 rwq->core_qp.event = mlx5_ib_wq_event;
4774 rwq->ibwq.event_handler = init_attr->event_handler;
79b20a6c
YH		 4775 return &rwq->ibwq;
4776
4777err_copy:
350d0e4c 4778 mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
79b20a6c
YH		 4779err_user_rq:
4780 destroy_user_rq(pd, rwq);
4781err:
4782 kfree(rwq);
4783 return ERR_PTR(err);
4784}
4785
4786int mlx5_ib_destroy_wq(struct ib_wq *wq)
4787{
4788 struct mlx5_ib_dev *dev = to_mdev(wq->device);
4789 struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4790
350d0e4c 4791 mlx5_core_destroy_rq_tracked(dev->mdev, &rwq->core_qp);
79b20a6c
YH		 4792 destroy_user_rq(wq->pd, rwq);
4793 kfree(rwq);
4794
4795 return 0;
4796}
4797
c5f90929
YH		 4798struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
4799 struct ib_rwq_ind_table_init_attr *init_attr,
4800 struct ib_udata *udata)
4801{
4802 struct mlx5_ib_dev *dev = to_mdev(device);
4803 struct mlx5_ib_rwq_ind_table *rwq_ind_tbl;
4804 int sz = 1 << init_attr->log_ind_tbl_size;
4805 struct mlx5_ib_create_rwq_ind_tbl_resp resp = {};
4806 size_t min_resp_len;
4807 int inlen;
4808 int err;
4809 int i;
4810 u32 *in;
4811 void *rqtc;
4812
4813 if (udata->inlen > 0 &&
4814 !ib_is_udata_cleared(udata, 0,
4815 udata->inlen))
4816 return ERR_PTR(-EOPNOTSUPP);
4817
efd7f400
MG		 4818 if (init_attr->log_ind_tbl_size >
4819 MLX5_CAP_GEN(dev->mdev, log_max_rqt_size)) {
4820 mlx5_ib_dbg(dev, "log_ind_tbl_size = %d is bigger than supported = %d\n",
4821 init_attr->log_ind_tbl_size,
4822 MLX5_CAP_GEN(dev->mdev, log_max_rqt_size));
4823 return ERR_PTR(-EINVAL);
4824 }
4825
c5f90929
YH		 4826 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved);
4827 if (udata->outlen && udata->outlen < min_resp_len)
4828 return ERR_PTR(-EINVAL);
4829
4830 rwq_ind_tbl = kzalloc(sizeof(*rwq_ind_tbl), GFP_KERNEL);
4831 if (!rwq_ind_tbl)
4832 return ERR_PTR(-ENOMEM);
4833
4834 inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
4835 in = mlx5_vzalloc(inlen);
4836 if (!in) {
4837 err = -ENOMEM;
4838 goto err;
4839 }
4840
4841 rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
4842
4843 MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
4844 MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
4845
4846 for (i = 0; i < sz; i++)
4847 MLX5_SET(rqtc, rqtc, rq_num[i], init_attr->ind_tbl[i]->wq_num);
4848
4849 err = mlx5_core_create_rqt(dev->mdev, in, inlen, &rwq_ind_tbl->rqtn);
4850 kvfree(in);
4851
4852 if (err)
4853 goto err;
4854
4855 rwq_ind_tbl->ib_rwq_ind_tbl.ind_tbl_num = rwq_ind_tbl->rqtn;
4856 if (udata->outlen) {
4857 resp.response_length = offsetof(typeof(resp), response_length) +
4858 sizeof(resp.response_length);
4859 err = ib_copy_to_udata(udata, &resp, resp.response_length);
4860 if (err)
4861 goto err_copy;
4862 }
4863
4864 return &rwq_ind_tbl->ib_rwq_ind_tbl;
4865
4866err_copy:
4867 mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4868err:
4869 kfree(rwq_ind_tbl);
4870 return ERR_PTR(err);
4871}
4872
4873int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *ib_rwq_ind_tbl)
4874{
4875 struct mlx5_ib_rwq_ind_table *rwq_ind_tbl = to_mrwq_ind_table(ib_rwq_ind_tbl);
4876 struct mlx5_ib_dev *dev = to_mdev(ib_rwq_ind_tbl->device);
4877
4878 mlx5_core_destroy_rqt(dev->mdev, rwq_ind_tbl->rqtn);
4879
4880 kfree(rwq_ind_tbl);
4881 return 0;
4882}
4883
79b20a6c
YH		 4884int mlx5_ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
4885 u32 wq_attr_mask, struct ib_udata *udata)
4886{
4887 struct mlx5_ib_dev *dev = to_mdev(wq->device);
4888 struct mlx5_ib_rwq *rwq = to_mrwq(wq);
4889 struct mlx5_ib_modify_wq ucmd = {};
4890 size_t required_cmd_sz;
4891 int curr_wq_state;
4892 int wq_state;
4893 int inlen;
4894 int err;
4895 void *rqc;
4896 void *in;
4897
4898 required_cmd_sz = offsetof(typeof(ucmd), reserved) + sizeof(ucmd.reserved);
4899 if (udata->inlen < required_cmd_sz)
4900 return -EINVAL;
4901
4902 if (udata->inlen > sizeof(ucmd) &&
4903 !ib_is_udata_cleared(udata, sizeof(ucmd),
4904 udata->inlen - sizeof(ucmd)))
4905 return -EOPNOTSUPP;
4906
4907 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen)))
4908 return -EFAULT;
4909
4910 if (ucmd.comp_mask || ucmd.reserved)
4911 return -EOPNOTSUPP;
4912
4913 inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
4914 in = mlx5_vzalloc(inlen);
4915 if (!in)
4916 return -ENOMEM;
4917
4918 rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
4919
4920 curr_wq_state = (wq_attr_mask & IB_WQ_CUR_STATE) ?
4921 wq_attr->curr_wq_state : wq->state;
4922 wq_state = (wq_attr_mask & IB_WQ_STATE) ?
4923 wq_attr->wq_state : curr_wq_state;
4924 if (curr_wq_state == IB_WQS_ERR)
4925 curr_wq_state = MLX5_RQC_STATE_ERR;
4926 if (wq_state == IB_WQS_ERR)
4927 wq_state = MLX5_RQC_STATE_ERR;
4928 MLX5_SET(modify_rq_in, in, rq_state, curr_wq_state);
4929 MLX5_SET(rqc, rqc, state, wq_state);
4930
350d0e4c 4931 err = mlx5_core_modify_rq(dev->mdev, rwq->core_qp.qpn, in, inlen);
79b20a6c
YH		 4932 kvfree(in);
4933 if (!err)
4934 rwq->ibwq.state = (wq_state == MLX5_RQC_STATE_ERR) ? IB_WQS_ERR : wq_state;
4935
4936 return err;
4937}
Ubuntu Hirsute Linux kernel mirror