1 /*
2  * Copyright (c) 2016-2017 Hisilicon Limited.
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/acpi.h>
34 #include <linux/etherdevice.h>
35 #include <linux/interrupt.h>
36 #include <linux/iopoll.h>
37 #include <linux/kernel.h>
38 #include <linux/types.h>
39 #include <net/addrconf.h>
40 #include <rdma/ib_addr.h>
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/uverbs_ioctl.h>
44 
45 #include "hnae3.h"
46 #include "hns_roce_common.h"
47 #include "hns_roce_device.h"
48 #include "hns_roce_cmd.h"
49 #include "hns_roce_hem.h"
50 #include "hns_roce_hw_v2.h"
51 
52 enum {
53 	CMD_RST_PRC_OTHERS,
54 	CMD_RST_PRC_SUCCESS,
55 	CMD_RST_PRC_EBUSY,
56 };
57 
58 enum ecc_resource_type {
59 	ECC_RESOURCE_QPC,
60 	ECC_RESOURCE_CQC,
61 	ECC_RESOURCE_MPT,
62 	ECC_RESOURCE_SRQC,
63 	ECC_RESOURCE_GMV,
64 	ECC_RESOURCE_QPC_TIMER,
65 	ECC_RESOURCE_CQC_TIMER,
66 	ECC_RESOURCE_SCCC,
67 	ECC_RESOURCE_COUNT,
68 };
69 
70 static const struct {
71 	const char *name;
72 	u8 read_bt0_op;
73 	u8 write_bt0_op;
74 } fmea_ram_res[] = {
75 	{ "ECC_RESOURCE_QPC",
76 	  HNS_ROCE_CMD_READ_QPC_BT0, HNS_ROCE_CMD_WRITE_QPC_BT0 },
77 	{ "ECC_RESOURCE_CQC",
78 	  HNS_ROCE_CMD_READ_CQC_BT0, HNS_ROCE_CMD_WRITE_CQC_BT0 },
79 	{ "ECC_RESOURCE_MPT",
80 	  HNS_ROCE_CMD_READ_MPT_BT0, HNS_ROCE_CMD_WRITE_MPT_BT0 },
81 	{ "ECC_RESOURCE_SRQC",
82 	  HNS_ROCE_CMD_READ_SRQC_BT0, HNS_ROCE_CMD_WRITE_SRQC_BT0 },
83 	/* ECC_RESOURCE_GMV is handled by cmdq, not mailbox */
84 	{ "ECC_RESOURCE_GMV",
85 	  0, 0 },
86 	{ "ECC_RESOURCE_QPC_TIMER",
87 	  HNS_ROCE_CMD_READ_QPC_TIMER_BT0, HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0 },
88 	{ "ECC_RESOURCE_CQC_TIMER",
89 	  HNS_ROCE_CMD_READ_CQC_TIMER_BT0, HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0 },
90 	{ "ECC_RESOURCE_SCCC",
91 	  HNS_ROCE_CMD_READ_SCCC_BT0, HNS_ROCE_CMD_WRITE_SCCC_BT0 },
92 };
93 
94 static inline void set_data_seg_v2(struct hns_roce_v2_wqe_data_seg *dseg,
95 				   struct ib_sge *sg)
96 {
97 	dseg->lkey = cpu_to_le32(sg->lkey);
98 	dseg->addr = cpu_to_le64(sg->addr);
99 	dseg->len  = cpu_to_le32(sg->length);
100 }
101 
102 /*
103  * mapped-value = 1 + real-value
104  * The hns wr opcode real value is start from 0, In order to distinguish between
105  * initialized and uninitialized map values, we plus 1 to the actual value when
106  * defining the mapping, so that the validity can be identified by checking the
107  * mapped value is greater than 0.
108  */
109 #define HR_OPC_MAP(ib_key, hr_key) \
110 		[IB_WR_ ## ib_key] = 1 + HNS_ROCE_V2_WQE_OP_ ## hr_key
111 
112 static const u32 hns_roce_op_code[] = {
113 	HR_OPC_MAP(RDMA_WRITE,			RDMA_WRITE),
114 	HR_OPC_MAP(RDMA_WRITE_WITH_IMM,		RDMA_WRITE_WITH_IMM),
115 	HR_OPC_MAP(SEND,			SEND),
116 	HR_OPC_MAP(SEND_WITH_IMM,		SEND_WITH_IMM),
117 	HR_OPC_MAP(RDMA_READ,			RDMA_READ),
118 	HR_OPC_MAP(ATOMIC_CMP_AND_SWP,		ATOM_CMP_AND_SWAP),
119 	HR_OPC_MAP(ATOMIC_FETCH_AND_ADD,	ATOM_FETCH_AND_ADD),
120 	HR_OPC_MAP(SEND_WITH_INV,		SEND_WITH_INV),
121 	HR_OPC_MAP(LOCAL_INV,			LOCAL_INV),
122 	HR_OPC_MAP(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
123 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
124 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
125 };
126 
127 static u32 to_hr_opcode(u32 ib_opcode)
128 {
129 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
130 		return HNS_ROCE_V2_WQE_OP_MASK;
131 
132 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
133 					     HNS_ROCE_V2_WQE_OP_MASK;
134 }
135 
136 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
137 			 const struct ib_reg_wr *wr)
138 {
139 	struct hns_roce_wqe_frmr_seg *fseg =
140 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
141 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
142 	u64 pbl_ba;
143 
144 	/* use ib_access_flags */
145 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
146 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
147 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
148 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
149 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
150 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
151 
152 	/* Data structure reuse may lead to confusion */
153 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
154 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
155 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
156 
157 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
158 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
159 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
160 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
161 
162 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
163 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
164 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
165 	hr_reg_clear(fseg, FRMR_BLK_MODE);
166 }
167 
168 static void set_atomic_seg(const struct ib_send_wr *wr,
169 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
170 			   unsigned int valid_num_sge)
171 {
172 	struct hns_roce_v2_wqe_data_seg *dseg =
173 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
174 	struct hns_roce_wqe_atomic_seg *aseg =
175 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
176 
177 	set_data_seg_v2(dseg, wr->sg_list);
178 
179 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
180 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
181 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
182 	} else {
183 		aseg->fetchadd_swap_data =
184 			cpu_to_le64(atomic_wr(wr)->compare_add);
185 		aseg->cmp_data = 0;
186 	}
187 
188 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
189 }
190 
191 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
192 				 const struct ib_send_wr *wr,
193 				 unsigned int *sge_idx, u32 msg_len)
194 {
195 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
196 	unsigned int dseg_len = sizeof(struct hns_roce_v2_wqe_data_seg);
197 	unsigned int ext_sge_sz = qp->sq.max_gs * dseg_len;
198 	unsigned int left_len_in_pg;
199 	unsigned int idx = *sge_idx;
200 	unsigned int i = 0;
201 	unsigned int len;
202 	void *addr;
203 	void *dseg;
204 
205 	if (msg_len > ext_sge_sz) {
206 		ibdev_err(ibdev,
207 			  "no enough extended sge space for inline data.\n");
208 		return -EINVAL;
209 	}
210 
211 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
212 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
213 	len = wr->sg_list[0].length;
214 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
215 
216 	/* When copying data to extended sge space, the left length in page may
217 	 * not long enough for current user's sge. So the data should be
218 	 * splited into several parts, one in the first page, and the others in
219 	 * the subsequent pages.
220 	 */
221 	while (1) {
222 		if (len <= left_len_in_pg) {
223 			memcpy(dseg, addr, len);
224 
225 			idx += len / dseg_len;
226 
227 			i++;
228 			if (i >= wr->num_sge)
229 				break;
230 
231 			left_len_in_pg -= len;
232 			len = wr->sg_list[i].length;
233 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
234 			dseg += len;
235 		} else {
236 			memcpy(dseg, addr, left_len_in_pg);
237 
238 			len -= left_len_in_pg;
239 			addr += left_len_in_pg;
240 			idx += left_len_in_pg / dseg_len;
241 			dseg = hns_roce_get_extend_sge(qp,
242 						idx & (qp->sge.sge_cnt - 1));
243 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
244 		}
245 	}
246 
247 	*sge_idx = idx;
248 
249 	return 0;
250 }
251 
252 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
253 			   unsigned int *sge_ind, unsigned int cnt)
254 {
255 	struct hns_roce_v2_wqe_data_seg *dseg;
256 	unsigned int idx = *sge_ind;
257 
258 	while (cnt > 0) {
259 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
260 		if (likely(sge->length)) {
261 			set_data_seg_v2(dseg, sge);
262 			idx++;
263 			cnt--;
264 		}
265 		sge++;
266 	}
267 
268 	*sge_ind = idx;
269 }
270 
271 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
272 {
273 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
274 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
275 
276 	if (len > qp->max_inline_data || len > mtu) {
277 		ibdev_err(&hr_dev->ib_dev,
278 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
279 			  len, qp->max_inline_data, mtu);
280 		return false;
281 	}
282 
283 	return true;
284 }
285 
286 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
287 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
288 		      unsigned int *sge_idx)
289 {
290 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
291 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
292 	struct ib_device *ibdev = &hr_dev->ib_dev;
293 	unsigned int curr_idx = *sge_idx;
294 	void *dseg = rc_sq_wqe;
295 	unsigned int i;
296 	int ret;
297 
298 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
299 		ibdev_err(ibdev, "invalid inline parameters!\n");
300 		return -EINVAL;
301 	}
302 
303 	if (!check_inl_data_len(qp, msg_len))
304 		return -EINVAL;
305 
306 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
307 
308 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
309 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
310 
311 		for (i = 0; i < wr->num_sge; i++) {
312 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
313 			       wr->sg_list[i].length);
314 			dseg += wr->sg_list[i].length;
315 		}
316 	} else {
317 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
318 
319 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
320 		if (ret)
321 			return ret;
322 
323 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
324 	}
325 
326 	*sge_idx = curr_idx;
327 
328 	return 0;
329 }
330 
331 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
332 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
333 			     unsigned int *sge_ind,
334 			     unsigned int valid_num_sge)
335 {
336 	struct hns_roce_v2_wqe_data_seg *dseg =
337 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
338 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
339 	int j = 0;
340 	int i;
341 
342 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
343 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
344 
345 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
346 		     !!(wr->send_flags & IB_SEND_INLINE));
347 	if (wr->send_flags & IB_SEND_INLINE)
348 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
349 
350 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
351 		for (i = 0; i < wr->num_sge; i++) {
352 			if (likely(wr->sg_list[i].length)) {
353 				set_data_seg_v2(dseg, wr->sg_list + i);
354 				dseg++;
355 			}
356 		}
357 	} else {
358 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
359 			if (likely(wr->sg_list[i].length)) {
360 				set_data_seg_v2(dseg, wr->sg_list + i);
361 				dseg++;
362 				j++;
363 			}
364 		}
365 
366 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
367 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
368 	}
369 
370 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
371 
372 	return 0;
373 }
374 
375 static int check_send_valid(struct hns_roce_dev *hr_dev,
376 			    struct hns_roce_qp *hr_qp)
377 {
378 	struct ib_device *ibdev = &hr_dev->ib_dev;
379 	struct ib_qp *ibqp = &hr_qp->ibqp;
380 
381 	if (unlikely(ibqp->qp_type != IB_QPT_RC &&
382 		     ibqp->qp_type != IB_QPT_GSI &&
383 		     ibqp->qp_type != IB_QPT_UD)) {
384 		ibdev_err(ibdev, "Not supported QP(0x%x)type!\n",
385 			  ibqp->qp_type);
386 		return -EOPNOTSUPP;
387 	} else if (unlikely(hr_qp->state == IB_QPS_RESET ||
388 		   hr_qp->state == IB_QPS_INIT ||
389 		   hr_qp->state == IB_QPS_RTR)) {
390 		ibdev_err(ibdev, "failed to post WQE, QP state %u!\n",
391 			  hr_qp->state);
392 		return -EINVAL;
393 	} else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)) {
394 		ibdev_err(ibdev, "failed to post WQE, dev state %d!\n",
395 			  hr_dev->state);
396 		return -EIO;
397 	}
398 
399 	return 0;
400 }
401 
402 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
403 				    unsigned int *sge_len)
404 {
405 	unsigned int valid_num = 0;
406 	unsigned int len = 0;
407 	int i;
408 
409 	for (i = 0; i < wr->num_sge; i++) {
410 		if (likely(wr->sg_list[i].length)) {
411 			len += wr->sg_list[i].length;
412 			valid_num++;
413 		}
414 	}
415 
416 	*sge_len = len;
417 	return valid_num;
418 }
419 
420 static __le32 get_immtdata(const struct ib_send_wr *wr)
421 {
422 	switch (wr->opcode) {
423 	case IB_WR_SEND_WITH_IMM:
424 	case IB_WR_RDMA_WRITE_WITH_IMM:
425 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
426 	default:
427 		return 0;
428 	}
429 }
430 
431 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
432 			 const struct ib_send_wr *wr)
433 {
434 	u32 ib_op = wr->opcode;
435 
436 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
437 		return -EINVAL;
438 
439 	ud_sq_wqe->immtdata = get_immtdata(wr);
440 
441 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
442 
443 	return 0;
444 }
445 
446 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
447 		      struct hns_roce_ah *ah)
448 {
449 	struct ib_device *ib_dev = ah->ibah.device;
450 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
451 
452 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
453 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
454 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
455 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
456 
457 	if (WARN_ON(ah->av.sl > MAX_SERVICE_LEVEL))
458 		return -EINVAL;
459 
460 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
461 
462 	ud_sq_wqe->sgid_index = ah->av.gid_index;
463 
464 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
465 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
466 
467 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
468 		return 0;
469 
470 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
471 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
472 
473 	return 0;
474 }
475 
476 static inline int set_ud_wqe(struct hns_roce_qp *qp,
477 			     const struct ib_send_wr *wr,
478 			     void *wqe, unsigned int *sge_idx,
479 			     unsigned int owner_bit)
480 {
481 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
482 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
483 	unsigned int curr_idx = *sge_idx;
484 	unsigned int valid_num_sge;
485 	u32 msg_len = 0;
486 	int ret;
487 
488 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
489 
490 	ret = set_ud_opcode(ud_sq_wqe, wr);
491 	if (WARN_ON(ret))
492 		return ret;
493 
494 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
495 
496 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
497 		     !!(wr->send_flags & IB_SEND_SIGNALED));
498 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
499 		     !!(wr->send_flags & IB_SEND_SOLICITED));
500 
501 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
502 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
503 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
504 		     curr_idx & (qp->sge.sge_cnt - 1));
505 
506 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
507 			  qp->qkey : ud_wr(wr)->remote_qkey);
508 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
509 
510 	ret = fill_ud_av(ud_sq_wqe, ah);
511 	if (ret)
512 		return ret;
513 
514 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
515 
516 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
517 
518 	/*
519 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
520 	 * including new WQEs waiting for the doorbell to update the PI again.
521 	 * Therefore, the owner bit of WQE MUST be updated after all fields
522 	 * and extSGEs have been written into DDR instead of cache.
523 	 */
524 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
525 		dma_wmb();
526 
527 	*sge_idx = curr_idx;
528 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
529 
530 	return 0;
531 }
532 
533 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
534 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
535 			 const struct ib_send_wr *wr)
536 {
537 	u32 ib_op = wr->opcode;
538 	int ret = 0;
539 
540 	rc_sq_wqe->immtdata = get_immtdata(wr);
541 
542 	switch (ib_op) {
543 	case IB_WR_RDMA_READ:
544 	case IB_WR_RDMA_WRITE:
545 	case IB_WR_RDMA_WRITE_WITH_IMM:
546 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
547 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
548 		break;
549 	case IB_WR_SEND:
550 	case IB_WR_SEND_WITH_IMM:
551 		break;
552 	case IB_WR_ATOMIC_CMP_AND_SWP:
553 	case IB_WR_ATOMIC_FETCH_AND_ADD:
554 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
555 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
556 		break;
557 	case IB_WR_REG_MR:
558 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
559 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
560 		else
561 			ret = -EOPNOTSUPP;
562 		break;
563 	case IB_WR_LOCAL_INV:
564 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_SO);
565 		fallthrough;
566 	case IB_WR_SEND_WITH_INV:
567 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
568 		break;
569 	default:
570 		ret = -EINVAL;
571 	}
572 
573 	if (unlikely(ret))
574 		return ret;
575 
576 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
577 
578 	return ret;
579 }
580 
581 static inline int set_rc_wqe(struct hns_roce_qp *qp,
582 			     const struct ib_send_wr *wr,
583 			     void *wqe, unsigned int *sge_idx,
584 			     unsigned int owner_bit)
585 {
586 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
587 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
588 	unsigned int curr_idx = *sge_idx;
589 	unsigned int valid_num_sge;
590 	u32 msg_len = 0;
591 	int ret;
592 
593 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
594 
595 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
596 
597 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
598 	if (WARN_ON(ret))
599 		return ret;
600 
601 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_FENCE,
602 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
603 
604 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
605 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
606 
607 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
608 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
609 
610 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
611 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
612 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
613 	else if (wr->opcode != IB_WR_REG_MR)
614 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
615 					&curr_idx, valid_num_sge);
616 
617 	/*
618 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
619 	 * including new WQEs waiting for the doorbell to update the PI again.
620 	 * Therefore, the owner bit of WQE MUST be updated after all fields
621 	 * and extSGEs have been written into DDR instead of cache.
622 	 */
623 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
624 		dma_wmb();
625 
626 	*sge_idx = curr_idx;
627 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
628 
629 	return ret;
630 }
631 
632 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
633 				struct hns_roce_qp *qp)
634 {
635 	if (unlikely(qp->state == IB_QPS_ERR)) {
636 		flush_cqe(hr_dev, qp);
637 	} else {
638 		struct hns_roce_v2_db sq_db = {};
639 
640 		hr_reg_write(&sq_db, DB_TAG, qp->doorbell_qpn);
641 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
642 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
643 		hr_reg_write(&sq_db, DB_SL, qp->sl);
644 
645 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
646 	}
647 }
648 
649 static inline void update_rq_db(struct hns_roce_dev *hr_dev,
650 				struct hns_roce_qp *qp)
651 {
652 	if (unlikely(qp->state == IB_QPS_ERR)) {
653 		flush_cqe(hr_dev, qp);
654 	} else {
655 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
656 			*qp->rdb.db_record =
657 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
658 		} else {
659 			struct hns_roce_v2_db rq_db = {};
660 
661 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
662 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
663 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
664 
665 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
666 					 qp->rq.db_reg);
667 		}
668 	}
669 }
670 
671 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
672 			      u64 __iomem *dest)
673 {
674 #define HNS_ROCE_WRITE_TIMES 8
675 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
676 	struct hnae3_handle *handle = priv->handle;
677 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
678 	int i;
679 
680 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
681 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
682 			writeq_relaxed(*(val + i), dest + i);
683 }
684 
685 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
686 		       void *wqe)
687 {
688 #define HNS_ROCE_SL_SHIFT 2
689 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
690 
691 	/* All kinds of DirectWQE have the same header field layout */
692 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
693 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
694 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
695 		     qp->sl >> HNS_ROCE_SL_SHIFT);
696 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
697 
698 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
699 }
700 
701 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
702 				 const struct ib_send_wr *wr,
703 				 const struct ib_send_wr **bad_wr)
704 {
705 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
706 	struct ib_device *ibdev = &hr_dev->ib_dev;
707 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
708 	unsigned long flags = 0;
709 	unsigned int owner_bit;
710 	unsigned int sge_idx;
711 	unsigned int wqe_idx;
712 	void *wqe = NULL;
713 	u32 nreq;
714 	int ret;
715 
716 	spin_lock_irqsave(&qp->sq.lock, flags);
717 
718 	ret = check_send_valid(hr_dev, qp);
719 	if (unlikely(ret)) {
720 		*bad_wr = wr;
721 		nreq = 0;
722 		goto out;
723 	}
724 
725 	sge_idx = qp->next_sge;
726 
727 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
728 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
729 			ret = -ENOMEM;
730 			*bad_wr = wr;
731 			goto out;
732 		}
733 
734 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
735 
736 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
737 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
738 				  wr->num_sge, qp->sq.max_gs);
739 			ret = -EINVAL;
740 			*bad_wr = wr;
741 			goto out;
742 		}
743 
744 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
745 		qp->sq.wrid[wqe_idx] = wr->wr_id;
746 		owner_bit =
747 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
748 
749 		/* Corresponding to the QP type, wqe process separately */
750 		if (ibqp->qp_type == IB_QPT_RC)
751 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
752 		else
753 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
754 
755 		if (unlikely(ret)) {
756 			*bad_wr = wr;
757 			goto out;
758 		}
759 	}
760 
761 out:
762 	if (likely(nreq)) {
763 		qp->sq.head += nreq;
764 		qp->next_sge = sge_idx;
765 
766 		if (nreq == 1 && (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
767 			write_dwqe(hr_dev, qp, wqe);
768 		else
769 			update_sq_db(hr_dev, qp);
770 	}
771 
772 	spin_unlock_irqrestore(&qp->sq.lock, flags);
773 
774 	return ret;
775 }
776 
777 static int check_recv_valid(struct hns_roce_dev *hr_dev,
778 			    struct hns_roce_qp *hr_qp)
779 {
780 	struct ib_device *ibdev = &hr_dev->ib_dev;
781 	struct ib_qp *ibqp = &hr_qp->ibqp;
782 
783 	if (unlikely(ibqp->qp_type != IB_QPT_RC &&
784 		     ibqp->qp_type != IB_QPT_GSI &&
785 		     ibqp->qp_type != IB_QPT_UD)) {
786 		ibdev_err(ibdev, "unsupported qp type, qp_type = %d.\n",
787 			  ibqp->qp_type);
788 		return -EOPNOTSUPP;
789 	}
790 
791 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
792 		return -EIO;
793 
794 	if (hr_qp->state == IB_QPS_RESET)
795 		return -EINVAL;
796 
797 	return 0;
798 }
799 
800 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
801 				 u32 max_sge, bool rsv)
802 {
803 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
804 	u32 i, cnt;
805 
806 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
807 		/* Skip zero-length sge */
808 		if (!wr->sg_list[i].length)
809 			continue;
810 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
811 		cnt++;
812 	}
813 
814 	/* Fill a reserved sge to make hw stop reading remaining segments */
815 	if (rsv) {
816 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
817 		dseg[cnt].addr = 0;
818 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
819 	} else {
820 		/* Clear remaining segments to make ROCEE ignore sges */
821 		if (cnt < max_sge)
822 			memset(dseg + cnt, 0,
823 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
824 	}
825 }
826 
827 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
828 			u32 wqe_idx, u32 max_sge)
829 {
830 	struct hns_roce_rinl_sge *sge_list;
831 	void *wqe = NULL;
832 	u32 i;
833 
834 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
835 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
836 
837 	/* rq support inline data */
838 	if (hr_qp->rq_inl_buf.wqe_cnt) {
839 		sge_list = hr_qp->rq_inl_buf.wqe_list[wqe_idx].sg_list;
840 		hr_qp->rq_inl_buf.wqe_list[wqe_idx].sge_cnt = (u32)wr->num_sge;
841 		for (i = 0; i < wr->num_sge; i++) {
842 			sge_list[i].addr = (void *)(u64)wr->sg_list[i].addr;
843 			sge_list[i].len = wr->sg_list[i].length;
844 		}
845 	}
846 }
847 
848 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
849 				 const struct ib_recv_wr *wr,
850 				 const struct ib_recv_wr **bad_wr)
851 {
852 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
853 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
854 	struct ib_device *ibdev = &hr_dev->ib_dev;
855 	u32 wqe_idx, nreq, max_sge;
856 	unsigned long flags;
857 	int ret;
858 
859 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
860 
861 	ret = check_recv_valid(hr_dev, hr_qp);
862 	if (unlikely(ret)) {
863 		*bad_wr = wr;
864 		nreq = 0;
865 		goto out;
866 	}
867 
868 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
869 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
870 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
871 						  hr_qp->ibqp.recv_cq))) {
872 			ret = -ENOMEM;
873 			*bad_wr = wr;
874 			goto out;
875 		}
876 
877 		if (unlikely(wr->num_sge > max_sge)) {
878 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
879 				  wr->num_sge, max_sge);
880 			ret = -EINVAL;
881 			*bad_wr = wr;
882 			goto out;
883 		}
884 
885 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
886 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
887 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
888 	}
889 
890 out:
891 	if (likely(nreq)) {
892 		hr_qp->rq.head += nreq;
893 
894 		update_rq_db(hr_dev, hr_qp);
895 	}
896 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
897 
898 	return ret;
899 }
900 
901 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
902 {
903 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
904 }
905 
906 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
907 {
908 	return hns_roce_buf_offset(idx_que->mtr.kmem,
909 				   n << idx_que->entry_shift);
910 }
911 
912 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
913 {
914 	/* always called with interrupts disabled. */
915 	spin_lock(&srq->lock);
916 
917 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
918 	srq->idx_que.tail++;
919 
920 	spin_unlock(&srq->lock);
921 }
922 
923 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
924 {
925 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
926 
927 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
928 }
929 
930 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
931 				const struct ib_recv_wr *wr)
932 {
933 	struct ib_device *ib_dev = srq->ibsrq.device;
934 
935 	if (unlikely(wr->num_sge > max_sge)) {
936 		ibdev_err(ib_dev,
937 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
938 			  wr->num_sge, max_sge);
939 		return -EINVAL;
940 	}
941 
942 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
943 		ibdev_err(ib_dev,
944 			  "failed to check srqwq status, srqwq is full.\n");
945 		return -ENOMEM;
946 	}
947 
948 	return 0;
949 }
950 
951 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
952 {
953 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
954 	u32 pos;
955 
956 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
957 	if (unlikely(pos == srq->wqe_cnt))
958 		return -ENOSPC;
959 
960 	bitmap_set(idx_que->bitmap, pos, 1);
961 	*wqe_idx = pos;
962 	return 0;
963 }
964 
965 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
966 {
967 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
968 	unsigned int head;
969 	__le32 *buf;
970 
971 	head = idx_que->head & (srq->wqe_cnt - 1);
972 
973 	buf = get_idx_buf(idx_que, head);
974 	*buf = cpu_to_le32(wqe_idx);
975 
976 	idx_que->head++;
977 }
978 
979 static void update_srq_db(struct hns_roce_v2_db *db, struct hns_roce_srq *srq)
980 {
981 	hr_reg_write(db, DB_TAG, srq->srqn);
982 	hr_reg_write(db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
983 	hr_reg_write(db, DB_PI, srq->idx_que.head);
984 }
985 
986 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
987 				     const struct ib_recv_wr *wr,
988 				     const struct ib_recv_wr **bad_wr)
989 {
990 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
991 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
992 	struct hns_roce_v2_db srq_db;
993 	unsigned long flags;
994 	int ret = 0;
995 	u32 max_sge;
996 	u32 wqe_idx;
997 	void *wqe;
998 	u32 nreq;
999 
1000 	spin_lock_irqsave(&srq->lock, flags);
1001 
1002 	max_sge = srq->max_gs - srq->rsv_sge;
1003 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
1004 		ret = check_post_srq_valid(srq, max_sge, wr);
1005 		if (ret) {
1006 			*bad_wr = wr;
1007 			break;
1008 		}
1009 
1010 		ret = get_srq_wqe_idx(srq, &wqe_idx);
1011 		if (unlikely(ret)) {
1012 			*bad_wr = wr;
1013 			break;
1014 		}
1015 
1016 		wqe = get_srq_wqe_buf(srq, wqe_idx);
1017 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
1018 		fill_wqe_idx(srq, wqe_idx);
1019 		srq->wrid[wqe_idx] = wr->wr_id;
1020 	}
1021 
1022 	if (likely(nreq)) {
1023 		update_srq_db(&srq_db, srq);
1024 
1025 		hns_roce_write64(hr_dev, (__le32 *)&srq_db, srq->db_reg);
1026 	}
1027 
1028 	spin_unlock_irqrestore(&srq->lock, flags);
1029 
1030 	return ret;
1031 }
1032 
1033 static u32 hns_roce_v2_cmd_hw_reseted(struct hns_roce_dev *hr_dev,
1034 				      unsigned long instance_stage,
1035 				      unsigned long reset_stage)
1036 {
1037 	/* When hardware reset has been completed once or more, we should stop
1038 	 * sending mailbox&cmq&doorbell to hardware. If now in .init_instance()
1039 	 * function, we should exit with error. If now at HNAE3_INIT_CLIENT
1040 	 * stage of soft reset process, we should exit with error, and then
1041 	 * HNAE3_INIT_CLIENT related process can rollback the operation like
1042 	 * notifing hardware to free resources, HNAE3_INIT_CLIENT related
1043 	 * process will exit with error to notify NIC driver to reschedule soft
1044 	 * reset process once again.
1045 	 */
1046 	hr_dev->is_reset = true;
1047 	hr_dev->dis_db = true;
1048 
1049 	if (reset_stage == HNS_ROCE_STATE_RST_INIT ||
1050 	    instance_stage == HNS_ROCE_STATE_INIT)
1051 		return CMD_RST_PRC_EBUSY;
1052 
1053 	return CMD_RST_PRC_SUCCESS;
1054 }
1055 
1056 static u32 hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
1057 					unsigned long instance_stage,
1058 					unsigned long reset_stage)
1059 {
1060 #define HW_RESET_TIMEOUT_US 1000000
1061 #define HW_RESET_SLEEP_US 1000
1062 
1063 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1064 	struct hnae3_handle *handle = priv->handle;
1065 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1066 	unsigned long val;
1067 	int ret;
1068 
1069 	/* When hardware reset is detected, we should stop sending mailbox&cmq&
1070 	 * doorbell to hardware. If now in .init_instance() function, we should
1071 	 * exit with error. If now at HNAE3_INIT_CLIENT stage of soft reset
1072 	 * process, we should exit with error, and then HNAE3_INIT_CLIENT
1073 	 * related process can rollback the operation like notifing hardware to
1074 	 * free resources, HNAE3_INIT_CLIENT related process will exit with
1075 	 * error to notify NIC driver to reschedule soft reset process once
1076 	 * again.
1077 	 */
1078 	hr_dev->dis_db = true;
1079 
1080 	ret = read_poll_timeout(ops->ae_dev_reset_cnt, val,
1081 				val > hr_dev->reset_cnt, HW_RESET_SLEEP_US,
1082 				HW_RESET_TIMEOUT_US, false, handle);
1083 	if (!ret)
1084 		hr_dev->is_reset = true;
1085 
1086 	if (!hr_dev->is_reset || reset_stage == HNS_ROCE_STATE_RST_INIT ||
1087 	    instance_stage == HNS_ROCE_STATE_INIT)
1088 		return CMD_RST_PRC_EBUSY;
1089 
1090 	return CMD_RST_PRC_SUCCESS;
1091 }
1092 
1093 static u32 hns_roce_v2_cmd_sw_resetting(struct hns_roce_dev *hr_dev)
1094 {
1095 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1096 	struct hnae3_handle *handle = priv->handle;
1097 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1098 
1099 	/* When software reset is detected at .init_instance() function, we
1100 	 * should stop sending mailbox&cmq&doorbell to hardware, and exit
1101 	 * with error.
1102 	 */
1103 	hr_dev->dis_db = true;
1104 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt)
1105 		hr_dev->is_reset = true;
1106 
1107 	return CMD_RST_PRC_EBUSY;
1108 }
1109 
1110 static u32 check_aedev_reset_status(struct hns_roce_dev *hr_dev,
1111 				    struct hnae3_handle *handle)
1112 {
1113 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1114 	unsigned long instance_stage; /* the current instance stage */
1115 	unsigned long reset_stage; /* the current reset stage */
1116 	unsigned long reset_cnt;
1117 	bool sw_resetting;
1118 	bool hw_resetting;
1119 
1120 	/* Get information about reset from NIC driver or RoCE driver itself,
1121 	 * the meaning of the following variables from NIC driver are described
1122 	 * as below:
1123 	 * reset_cnt -- The count value of completed hardware reset.
1124 	 * hw_resetting -- Whether hardware device is resetting now.
1125 	 * sw_resetting -- Whether NIC's software reset process is running now.
1126 	 */
1127 	instance_stage = handle->rinfo.instance_state;
1128 	reset_stage = handle->rinfo.reset_state;
1129 	reset_cnt = ops->ae_dev_reset_cnt(handle);
1130 	if (reset_cnt != hr_dev->reset_cnt)
1131 		return hns_roce_v2_cmd_hw_reseted(hr_dev, instance_stage,
1132 						  reset_stage);
1133 
1134 	hw_resetting = ops->get_cmdq_stat(handle);
1135 	if (hw_resetting)
1136 		return hns_roce_v2_cmd_hw_resetting(hr_dev, instance_stage,
1137 						    reset_stage);
1138 
1139 	sw_resetting = ops->ae_dev_resetting(handle);
1140 	if (sw_resetting && instance_stage == HNS_ROCE_STATE_INIT)
1141 		return hns_roce_v2_cmd_sw_resetting(hr_dev);
1142 
1143 	return CMD_RST_PRC_OTHERS;
1144 }
1145 
1146 static bool check_device_is_in_reset(struct hns_roce_dev *hr_dev)
1147 {
1148 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1149 	struct hnae3_handle *handle = priv->handle;
1150 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1151 
1152 	if (hr_dev->reset_cnt != ops->ae_dev_reset_cnt(handle))
1153 		return true;
1154 
1155 	if (ops->get_hw_reset_stat(handle))
1156 		return true;
1157 
1158 	if (ops->ae_dev_resetting(handle))
1159 		return true;
1160 
1161 	return false;
1162 }
1163 
1164 static bool v2_chk_mbox_is_avail(struct hns_roce_dev *hr_dev, bool *busy)
1165 {
1166 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1167 	u32 status;
1168 
1169 	if (hr_dev->is_reset)
1170 		status = CMD_RST_PRC_SUCCESS;
1171 	else
1172 		status = check_aedev_reset_status(hr_dev, priv->handle);
1173 
1174 	*busy = (status == CMD_RST_PRC_EBUSY);
1175 
1176 	return status == CMD_RST_PRC_OTHERS;
1177 }
1178 
1179 static int hns_roce_alloc_cmq_desc(struct hns_roce_dev *hr_dev,
1180 				   struct hns_roce_v2_cmq_ring *ring)
1181 {
1182 	int size = ring->desc_num * sizeof(struct hns_roce_cmq_desc);
1183 
1184 	ring->desc = dma_alloc_coherent(hr_dev->dev, size,
1185 					&ring->desc_dma_addr, GFP_KERNEL);
1186 	if (!ring->desc)
1187 		return -ENOMEM;
1188 
1189 	return 0;
1190 }
1191 
1192 static void hns_roce_free_cmq_desc(struct hns_roce_dev *hr_dev,
1193 				   struct hns_roce_v2_cmq_ring *ring)
1194 {
1195 	dma_free_coherent(hr_dev->dev,
1196 			  ring->desc_num * sizeof(struct hns_roce_cmq_desc),
1197 			  ring->desc, ring->desc_dma_addr);
1198 
1199 	ring->desc_dma_addr = 0;
1200 }
1201 
1202 static int init_csq(struct hns_roce_dev *hr_dev,
1203 		    struct hns_roce_v2_cmq_ring *csq)
1204 {
1205 	dma_addr_t dma;
1206 	int ret;
1207 
1208 	csq->desc_num = CMD_CSQ_DESC_NUM;
1209 	spin_lock_init(&csq->lock);
1210 	csq->flag = TYPE_CSQ;
1211 	csq->head = 0;
1212 
1213 	ret = hns_roce_alloc_cmq_desc(hr_dev, csq);
1214 	if (ret)
1215 		return ret;
1216 
1217 	dma = csq->desc_dma_addr;
1218 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_L_REG, lower_32_bits(dma));
1219 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_H_REG, upper_32_bits(dma));
1220 	roce_write(hr_dev, ROCEE_TX_CMQ_DEPTH_REG,
1221 		   (u32)csq->desc_num >> HNS_ROCE_CMQ_DESC_NUM_S);
1222 
1223 	/* Make sure to write CI first and then PI */
1224 	roce_write(hr_dev, ROCEE_TX_CMQ_CI_REG, 0);
1225 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, 0);
1226 
1227 	return 0;
1228 }
1229 
1230 static int hns_roce_v2_cmq_init(struct hns_roce_dev *hr_dev)
1231 {
1232 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1233 	int ret;
1234 
1235 	priv->cmq.tx_timeout = HNS_ROCE_CMQ_TX_TIMEOUT;
1236 
1237 	ret = init_csq(hr_dev, &priv->cmq.csq);
1238 	if (ret)
1239 		dev_err(hr_dev->dev, "failed to init CSQ, ret = %d.\n", ret);
1240 
1241 	return ret;
1242 }
1243 
1244 static void hns_roce_v2_cmq_exit(struct hns_roce_dev *hr_dev)
1245 {
1246 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1247 
1248 	hns_roce_free_cmq_desc(hr_dev, &priv->cmq.csq);
1249 }
1250 
1251 static void hns_roce_cmq_setup_basic_desc(struct hns_roce_cmq_desc *desc,
1252 					  enum hns_roce_opcode_type opcode,
1253 					  bool is_read)
1254 {
1255 	memset((void *)desc, 0, sizeof(struct hns_roce_cmq_desc));
1256 	desc->opcode = cpu_to_le16(opcode);
1257 	desc->flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1258 	if (is_read)
1259 		desc->flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_WR);
1260 	else
1261 		desc->flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1262 }
1263 
1264 static int hns_roce_cmq_csq_done(struct hns_roce_dev *hr_dev)
1265 {
1266 	u32 tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1267 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1268 
1269 	return tail == priv->cmq.csq.head;
1270 }
1271 
1272 static void update_cmdq_status(struct hns_roce_dev *hr_dev)
1273 {
1274 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1275 	struct hnae3_handle *handle = priv->handle;
1276 
1277 	if (handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
1278 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT)
1279 		hr_dev->cmd.state = HNS_ROCE_CMDQ_STATE_FATAL_ERR;
1280 }
1281 
1282 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1283 			       struct hns_roce_cmq_desc *desc, int num)
1284 {
1285 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1286 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1287 	u32 timeout = 0;
1288 	u16 desc_ret;
1289 	u32 tail;
1290 	int ret;
1291 	int i;
1292 
1293 	spin_lock_bh(&csq->lock);
1294 
1295 	tail = csq->head;
1296 
1297 	for (i = 0; i < num; i++) {
1298 		csq->desc[csq->head++] = desc[i];
1299 		if (csq->head == csq->desc_num)
1300 			csq->head = 0;
1301 	}
1302 
1303 	/* Write to hardware */
1304 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1305 
1306 	do {
1307 		if (hns_roce_cmq_csq_done(hr_dev))
1308 			break;
1309 		udelay(1);
1310 	} while (++timeout < priv->cmq.tx_timeout);
1311 
1312 	if (hns_roce_cmq_csq_done(hr_dev)) {
1313 		ret = 0;
1314 		for (i = 0; i < num; i++) {
1315 			/* check the result of hardware write back */
1316 			desc[i] = csq->desc[tail++];
1317 			if (tail == csq->desc_num)
1318 				tail = 0;
1319 
1320 			desc_ret = le16_to_cpu(desc[i].retval);
1321 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1322 				continue;
1323 
1324 			dev_err_ratelimited(hr_dev->dev,
1325 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1326 					    desc->opcode, desc_ret);
1327 			ret = -EIO;
1328 		}
1329 	} else {
1330 		/* FW/HW reset or incorrect number of desc */
1331 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1332 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1333 			 csq->head, tail);
1334 		csq->head = tail;
1335 
1336 		update_cmdq_status(hr_dev);
1337 
1338 		ret = -EAGAIN;
1339 	}
1340 
1341 	spin_unlock_bh(&csq->lock);
1342 
1343 	return ret;
1344 }
1345 
1346 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1347 			     struct hns_roce_cmq_desc *desc, int num)
1348 {
1349 	bool busy;
1350 	int ret;
1351 
1352 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1353 		return -EIO;
1354 
1355 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1356 		return busy ? -EBUSY : 0;
1357 
1358 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1359 	if (ret) {
1360 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1361 			return busy ? -EBUSY : 0;
1362 	}
1363 
1364 	return ret;
1365 }
1366 
1367 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1368 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1369 {
1370 	struct hns_roce_cmd_mailbox *mbox;
1371 	int ret;
1372 
1373 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1374 	if (IS_ERR(mbox))
1375 		return PTR_ERR(mbox);
1376 
1377 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1378 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1379 	return ret;
1380 }
1381 
1382 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1383 {
1384 	struct hns_roce_query_version *resp;
1385 	struct hns_roce_cmq_desc desc;
1386 	int ret;
1387 
1388 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1389 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1390 	if (ret)
1391 		return ret;
1392 
1393 	resp = (struct hns_roce_query_version *)desc.data;
1394 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1395 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1396 
1397 	return 0;
1398 }
1399 
1400 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1401 					struct hnae3_handle *handle)
1402 {
1403 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1404 	unsigned long end;
1405 
1406 	hr_dev->dis_db = true;
1407 
1408 	dev_warn(hr_dev->dev,
1409 		 "Func clear is pending, device in resetting state.\n");
1410 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1411 	while (end) {
1412 		if (!ops->get_hw_reset_stat(handle)) {
1413 			hr_dev->is_reset = true;
1414 			dev_info(hr_dev->dev,
1415 				 "Func clear success after reset.\n");
1416 			return;
1417 		}
1418 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1419 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1420 	}
1421 
1422 	dev_warn(hr_dev->dev, "Func clear failed.\n");
1423 }
1424 
1425 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1426 					struct hnae3_handle *handle)
1427 {
1428 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1429 	unsigned long end;
1430 
1431 	hr_dev->dis_db = true;
1432 
1433 	dev_warn(hr_dev->dev,
1434 		 "Func clear is pending, device in resetting state.\n");
1435 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1436 	while (end) {
1437 		if (ops->ae_dev_reset_cnt(handle) !=
1438 		    hr_dev->reset_cnt) {
1439 			hr_dev->is_reset = true;
1440 			dev_info(hr_dev->dev,
1441 				 "Func clear success after sw reset\n");
1442 			return;
1443 		}
1444 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1445 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1446 	}
1447 
1448 	dev_warn(hr_dev->dev, "Func clear failed because of unfinished sw reset\n");
1449 }
1450 
1451 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1452 				       int flag)
1453 {
1454 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1455 	struct hnae3_handle *handle = priv->handle;
1456 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1457 
1458 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1459 		hr_dev->dis_db = true;
1460 		hr_dev->is_reset = true;
1461 		dev_info(hr_dev->dev, "Func clear success after reset.\n");
1462 		return;
1463 	}
1464 
1465 	if (ops->get_hw_reset_stat(handle)) {
1466 		func_clr_hw_resetting_state(hr_dev, handle);
1467 		return;
1468 	}
1469 
1470 	if (ops->ae_dev_resetting(handle) &&
1471 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1472 		func_clr_sw_resetting_state(hr_dev, handle);
1473 		return;
1474 	}
1475 
1476 	if (retval && !flag)
1477 		dev_warn(hr_dev->dev,
1478 			 "Func clear read failed, ret = %d.\n", retval);
1479 
1480 	dev_warn(hr_dev->dev, "Func clear failed.\n");
1481 }
1482 
1483 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1484 {
1485 	bool fclr_write_fail_flag = false;
1486 	struct hns_roce_func_clear *resp;
1487 	struct hns_roce_cmq_desc desc;
1488 	unsigned long end;
1489 	int ret = 0;
1490 
1491 	if (check_device_is_in_reset(hr_dev))
1492 		goto out;
1493 
1494 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1495 	resp = (struct hns_roce_func_clear *)desc.data;
1496 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1497 
1498 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1499 	if (ret) {
1500 		fclr_write_fail_flag = true;
1501 		dev_err(hr_dev->dev, "Func clear write failed, ret = %d.\n",
1502 			 ret);
1503 		goto out;
1504 	}
1505 
1506 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1507 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1508 	while (end) {
1509 		if (check_device_is_in_reset(hr_dev))
1510 			goto out;
1511 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1512 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1513 
1514 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1515 					      true);
1516 
1517 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1518 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1519 		if (ret)
1520 			continue;
1521 
1522 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1523 			if (vf_id == 0)
1524 				hr_dev->is_reset = true;
1525 			return;
1526 		}
1527 	}
1528 
1529 out:
1530 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1531 }
1532 
1533 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1534 {
1535 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1536 	struct hns_roce_cmq_desc desc[2];
1537 	struct hns_roce_cmq_req *req_a;
1538 
1539 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1540 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1541 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1542 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1543 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1544 
1545 	return hns_roce_cmq_send(hr_dev, desc, 2);
1546 }
1547 
1548 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1549 {
1550 	int ret;
1551 	int i;
1552 
1553 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1554 		return;
1555 
1556 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1557 		__hns_roce_function_clear(hr_dev, i);
1558 
1559 		if (i == 0)
1560 			continue;
1561 
1562 		ret = hns_roce_free_vf_resource(hr_dev, i);
1563 		if (ret)
1564 			ibdev_err(&hr_dev->ib_dev,
1565 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1566 				  i, ret);
1567 	}
1568 }
1569 
1570 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1571 {
1572 	struct hns_roce_cmq_desc desc;
1573 	int ret;
1574 
1575 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1576 				      false);
1577 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1578 	if (ret)
1579 		ibdev_err(&hr_dev->ib_dev,
1580 			  "failed to clear extended doorbell info, ret = %d.\n",
1581 			  ret);
1582 
1583 	return ret;
1584 }
1585 
1586 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1587 {
1588 	struct hns_roce_query_fw_info *resp;
1589 	struct hns_roce_cmq_desc desc;
1590 	int ret;
1591 
1592 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1593 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1594 	if (ret)
1595 		return ret;
1596 
1597 	resp = (struct hns_roce_query_fw_info *)desc.data;
1598 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1599 
1600 	return 0;
1601 }
1602 
1603 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1604 {
1605 	struct hns_roce_cmq_desc desc;
1606 	int ret;
1607 
1608 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1609 		hr_dev->func_num = 1;
1610 		return 0;
1611 	}
1612 
1613 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1614 				      true);
1615 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1616 	if (ret) {
1617 		hr_dev->func_num = 1;
1618 		return ret;
1619 	}
1620 
1621 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1622 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1623 
1624 	return 0;
1625 }
1626 
1627 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1628 {
1629 	struct hns_roce_cmq_desc desc;
1630 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1631 	u32 clock_cycles_of_1us;
1632 
1633 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1634 				      false);
1635 
1636 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1637 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1638 	else
1639 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1640 
1641 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1642 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1643 
1644 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1645 }
1646 
1647 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1648 {
1649 	struct hns_roce_cmq_desc desc[2];
1650 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1651 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1652 	struct hns_roce_caps *caps = &hr_dev->caps;
1653 	enum hns_roce_opcode_type opcode;
1654 	u32 func_num;
1655 	int ret;
1656 
1657 	if (is_vf) {
1658 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1659 		func_num = 1;
1660 	} else {
1661 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1662 		func_num = hr_dev->func_num;
1663 	}
1664 
1665 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1666 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1667 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1668 
1669 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1670 	if (ret)
1671 		return ret;
1672 
1673 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1674 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1675 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1676 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1677 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1678 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1679 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1680 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1681 
1682 	if (is_vf) {
1683 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1684 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1685 					       func_num;
1686 	} else {
1687 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1688 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1689 					       func_num;
1690 	}
1691 
1692 	return 0;
1693 }
1694 
1695 static int load_ext_cfg_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1696 {
1697 	struct hns_roce_cmq_desc desc;
1698 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1699 	struct hns_roce_caps *caps = &hr_dev->caps;
1700 	u32 func_num, qp_num;
1701 	int ret;
1702 
1703 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_EXT_CFG, true);
1704 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1705 	if (ret)
1706 		return ret;
1707 
1708 	func_num = is_vf ? 1 : max_t(u32, 1, hr_dev->func_num);
1709 	qp_num = hr_reg_read(req, EXT_CFG_QP_PI_NUM) / func_num;
1710 	caps->num_pi_qps = round_down(qp_num, HNS_ROCE_QP_BANK_NUM);
1711 
1712 	qp_num = hr_reg_read(req, EXT_CFG_QP_NUM) / func_num;
1713 	caps->num_qps = round_down(qp_num, HNS_ROCE_QP_BANK_NUM);
1714 
1715 	return 0;
1716 }
1717 
1718 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1719 {
1720 	struct hns_roce_cmq_desc desc;
1721 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1722 	struct hns_roce_caps *caps = &hr_dev->caps;
1723 	int ret;
1724 
1725 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1726 				      true);
1727 
1728 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1729 	if (ret)
1730 		return ret;
1731 
1732 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1733 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1734 
1735 	return 0;
1736 }
1737 
1738 static int query_func_resource_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1739 {
1740 	struct device *dev = hr_dev->dev;
1741 	int ret;
1742 
1743 	ret = load_func_res_caps(hr_dev, is_vf);
1744 	if (ret) {
1745 		dev_err(dev, "failed to load res caps, ret = %d (%s).\n", ret,
1746 			is_vf ? "vf" : "pf");
1747 		return ret;
1748 	}
1749 
1750 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1751 		ret = load_ext_cfg_caps(hr_dev, is_vf);
1752 		if (ret)
1753 			dev_err(dev, "failed to load ext cfg, ret = %d (%s).\n",
1754 				ret, is_vf ? "vf" : "pf");
1755 	}
1756 
1757 	return ret;
1758 }
1759 
1760 static int hns_roce_query_pf_resource(struct hns_roce_dev *hr_dev)
1761 {
1762 	struct device *dev = hr_dev->dev;
1763 	int ret;
1764 
1765 	ret = query_func_resource_caps(hr_dev, false);
1766 	if (ret)
1767 		return ret;
1768 
1769 	ret = load_pf_timer_res_caps(hr_dev);
1770 	if (ret)
1771 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1772 			ret);
1773 
1774 	return ret;
1775 }
1776 
1777 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1778 {
1779 	return query_func_resource_caps(hr_dev, true);
1780 }
1781 
1782 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1783 					  u32 vf_id)
1784 {
1785 	struct hns_roce_vf_switch *swt;
1786 	struct hns_roce_cmq_desc desc;
1787 	int ret;
1788 
1789 	swt = (struct hns_roce_vf_switch *)desc.data;
1790 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1791 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1792 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1793 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1794 	if (ret)
1795 		return ret;
1796 
1797 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1798 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1799 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1800 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1801 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1802 
1803 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1804 }
1805 
1806 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1807 {
1808 	u32 vf_id;
1809 	int ret;
1810 
1811 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1812 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1813 		if (ret)
1814 			return ret;
1815 	}
1816 	return 0;
1817 }
1818 
1819 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1820 {
1821 	struct hns_roce_cmq_desc desc[2];
1822 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1823 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1824 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1825 	struct hns_roce_caps *caps = &hr_dev->caps;
1826 
1827 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1828 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1829 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1830 
1831 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1832 
1833 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1834 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1835 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1836 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1837 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1838 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1839 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1840 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1841 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1842 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1843 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1844 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1845 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1846 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1847 
1848 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1849 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1850 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1851 			     vf_id * caps->gmv_bt_num);
1852 	} else {
1853 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1854 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1855 			     vf_id * caps->sgid_bt_num);
1856 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1857 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1858 			     vf_id * caps->smac_bt_num);
1859 	}
1860 
1861 	return hns_roce_cmq_send(hr_dev, desc, 2);
1862 }
1863 
1864 static int config_vf_ext_resource(struct hns_roce_dev *hr_dev, u32 vf_id)
1865 {
1866 	struct hns_roce_cmq_desc desc;
1867 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1868 	struct hns_roce_caps *caps = &hr_dev->caps;
1869 
1870 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_EXT_CFG, false);
1871 
1872 	hr_reg_write(req, EXT_CFG_VF_ID, vf_id);
1873 
1874 	hr_reg_write(req, EXT_CFG_QP_PI_NUM, caps->num_pi_qps);
1875 	hr_reg_write(req, EXT_CFG_QP_PI_IDX, vf_id * caps->num_pi_qps);
1876 	hr_reg_write(req, EXT_CFG_QP_NUM, caps->num_qps);
1877 	hr_reg_write(req, EXT_CFG_QP_IDX, vf_id * caps->num_qps);
1878 
1879 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1880 }
1881 
1882 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1883 {
1884 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1885 	u32 vf_id;
1886 	int ret;
1887 
1888 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1889 		ret = config_vf_hem_resource(hr_dev, vf_id);
1890 		if (ret) {
1891 			dev_err(hr_dev->dev,
1892 				"failed to config vf-%u hem res, ret = %d.\n",
1893 				vf_id, ret);
1894 			return ret;
1895 		}
1896 
1897 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1898 			ret = config_vf_ext_resource(hr_dev, vf_id);
1899 			if (ret) {
1900 				dev_err(hr_dev->dev,
1901 					"failed to config vf-%u ext res, ret = %d.\n",
1902 					vf_id, ret);
1903 				return ret;
1904 			}
1905 		}
1906 	}
1907 
1908 	return 0;
1909 }
1910 
1911 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1912 {
1913 	struct hns_roce_cmq_desc desc;
1914 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1915 	struct hns_roce_caps *caps = &hr_dev->caps;
1916 
1917 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1918 
1919 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1920 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1921 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1922 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1923 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1924 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1925 
1926 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1927 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1928 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1929 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1930 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1931 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1932 
1933 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1934 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1935 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1936 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1937 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1938 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1939 
1940 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1941 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1942 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1943 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1944 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1945 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1946 
1947 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1948 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1949 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1950 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1951 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1952 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1953 
1954 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1955 }
1956 
1957 /* Use default caps when hns_roce_query_pf_caps() failed or init VF profile */
1958 static void set_default_caps(struct hns_roce_dev *hr_dev)
1959 {
1960 	struct hns_roce_caps *caps = &hr_dev->caps;
1961 
1962 	caps->num_qps		= HNS_ROCE_V2_MAX_QP_NUM;
1963 	caps->max_wqes		= HNS_ROCE_V2_MAX_WQE_NUM;
1964 	caps->num_cqs		= HNS_ROCE_V2_MAX_CQ_NUM;
1965 	caps->num_srqs		= HNS_ROCE_V2_MAX_SRQ_NUM;
1966 	caps->min_cqes		= HNS_ROCE_MIN_CQE_NUM;
1967 	caps->max_cqes		= HNS_ROCE_V2_MAX_CQE_NUM;
1968 	caps->max_sq_sg		= HNS_ROCE_V2_MAX_SQ_SGE_NUM;
1969 	caps->max_extend_sg	= HNS_ROCE_V2_MAX_EXTEND_SGE_NUM;
1970 	caps->max_rq_sg		= HNS_ROCE_V2_MAX_RQ_SGE_NUM;
1971 
1972 	caps->num_uars		= HNS_ROCE_V2_UAR_NUM;
1973 	caps->phy_num_uars	= HNS_ROCE_V2_PHY_UAR_NUM;
1974 	caps->num_aeq_vectors	= HNS_ROCE_V2_AEQE_VEC_NUM;
1975 	caps->num_other_vectors = HNS_ROCE_V2_ABNORMAL_VEC_NUM;
1976 	caps->num_comp_vectors	= 0;
1977 
1978 	caps->num_mtpts		= HNS_ROCE_V2_MAX_MTPT_NUM;
1979 	caps->num_pds		= HNS_ROCE_V2_MAX_PD_NUM;
1980 	caps->num_qpc_timer	= HNS_ROCE_V2_MAX_QPC_TIMER_NUM;
1981 	caps->cqc_timer_bt_num	= HNS_ROCE_V2_MAX_CQC_TIMER_BT_NUM;
1982 
1983 	caps->max_qp_init_rdma	= HNS_ROCE_V2_MAX_QP_INIT_RDMA;
1984 	caps->max_qp_dest_rdma	= HNS_ROCE_V2_MAX_QP_DEST_RDMA;
1985 	caps->max_sq_desc_sz	= HNS_ROCE_V2_MAX_SQ_DESC_SZ;
1986 	caps->max_rq_desc_sz	= HNS_ROCE_V2_MAX_RQ_DESC_SZ;
1987 	caps->max_srq_desc_sz	= HNS_ROCE_V2_MAX_SRQ_DESC_SZ;
1988 	caps->irrl_entry_sz	= HNS_ROCE_V2_IRRL_ENTRY_SZ;
1989 	caps->trrl_entry_sz	= HNS_ROCE_V2_EXT_ATOMIC_TRRL_ENTRY_SZ;
1990 	caps->cqc_entry_sz	= HNS_ROCE_V2_CQC_ENTRY_SZ;
1991 	caps->srqc_entry_sz	= HNS_ROCE_V2_SRQC_ENTRY_SZ;
1992 	caps->mtpt_entry_sz	= HNS_ROCE_V2_MTPT_ENTRY_SZ;
1993 	caps->idx_entry_sz	= HNS_ROCE_V2_IDX_ENTRY_SZ;
1994 	caps->page_size_cap	= HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
1995 	caps->reserved_lkey	= 0;
1996 	caps->reserved_pds	= 0;
1997 	caps->reserved_mrws	= 1;
1998 	caps->reserved_uars	= 0;
1999 	caps->reserved_cqs	= 0;
2000 	caps->reserved_srqs	= 0;
2001 	caps->reserved_qps	= HNS_ROCE_V2_RSV_QPS;
2002 
2003 	caps->qpc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
2004 	caps->srqc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
2005 	caps->cqc_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
2006 	caps->mpt_hop_num	= HNS_ROCE_CONTEXT_HOP_NUM;
2007 	caps->sccc_hop_num	= HNS_ROCE_SCCC_HOP_NUM;
2008 
2009 	caps->mtt_hop_num	= HNS_ROCE_MTT_HOP_NUM;
2010 	caps->wqe_sq_hop_num	= HNS_ROCE_SQWQE_HOP_NUM;
2011 	caps->wqe_sge_hop_num	= HNS_ROCE_EXT_SGE_HOP_NUM;
2012 	caps->wqe_rq_hop_num	= HNS_ROCE_RQWQE_HOP_NUM;
2013 	caps->cqe_hop_num	= HNS_ROCE_CQE_HOP_NUM;
2014 	caps->srqwqe_hop_num	= HNS_ROCE_SRQWQE_HOP_NUM;
2015 	caps->idx_hop_num	= HNS_ROCE_IDX_HOP_NUM;
2016 	caps->chunk_sz          = HNS_ROCE_V2_TABLE_CHUNK_SIZE;
2017 
2018 	caps->flags		= HNS_ROCE_CAP_FLAG_REREG_MR |
2019 				  HNS_ROCE_CAP_FLAG_ROCE_V1_V2 |
2020 				  HNS_ROCE_CAP_FLAG_CQ_RECORD_DB |
2021 				  HNS_ROCE_CAP_FLAG_QP_RECORD_DB;
2022 
2023 	caps->pkey_table_len[0] = 1;
2024 	caps->ceqe_depth	= HNS_ROCE_V2_COMP_EQE_NUM;
2025 	caps->aeqe_depth	= HNS_ROCE_V2_ASYNC_EQE_NUM;
2026 	caps->local_ca_ack_delay = 0;
2027 	caps->max_mtu = IB_MTU_4096;
2028 
2029 	caps->max_srq_wrs	= HNS_ROCE_V2_MAX_SRQ_WR;
2030 	caps->max_srq_sges	= HNS_ROCE_V2_MAX_SRQ_SGE;
2031 
2032 	caps->flags |= HNS_ROCE_CAP_FLAG_ATOMIC | HNS_ROCE_CAP_FLAG_MW |
2033 		       HNS_ROCE_CAP_FLAG_SRQ | HNS_ROCE_CAP_FLAG_FRMR |
2034 		       HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL | HNS_ROCE_CAP_FLAG_XRC;
2035 
2036 	caps->gid_table_len[0] = HNS_ROCE_V2_GID_INDEX_NUM;
2037 
2038 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2039 		caps->flags |= HNS_ROCE_CAP_FLAG_STASH |
2040 			       HNS_ROCE_CAP_FLAG_DIRECT_WQE;
2041 		caps->max_sq_inline = HNS_ROCE_V3_MAX_SQ_INLINE;
2042 	} else {
2043 		caps->max_sq_inline = HNS_ROCE_V2_MAX_SQ_INLINE;
2044 
2045 		/* The following configuration are only valid for HIP08 */
2046 		caps->qpc_sz = HNS_ROCE_V2_QPC_SZ;
2047 		caps->sccc_sz = HNS_ROCE_V2_SCCC_SZ;
2048 		caps->cqe_sz = HNS_ROCE_V2_CQE_SIZE;
2049 	}
2050 }
2051 
2052 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
2053 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
2054 {
2055 	u64 obj_per_chunk;
2056 	u64 bt_chunk_size = PAGE_SIZE;
2057 	u64 buf_chunk_size = PAGE_SIZE;
2058 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
2059 
2060 	*buf_page_size = 0;
2061 	*bt_page_size = 0;
2062 
2063 	switch (hop_num) {
2064 	case 3:
2065 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2066 				(bt_chunk_size / BA_BYTE_LEN) *
2067 				(bt_chunk_size / BA_BYTE_LEN) *
2068 				 obj_per_chunk_default;
2069 		break;
2070 	case 2:
2071 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2072 				(bt_chunk_size / BA_BYTE_LEN) *
2073 				 obj_per_chunk_default;
2074 		break;
2075 	case 1:
2076 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
2077 				obj_per_chunk_default;
2078 		break;
2079 	case HNS_ROCE_HOP_NUM_0:
2080 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
2081 		break;
2082 	default:
2083 		pr_err("table %u not support hop_num = %u!\n", hem_type,
2084 		       hop_num);
2085 		return;
2086 	}
2087 
2088 	if (hem_type >= HEM_TYPE_MTT)
2089 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2090 	else
2091 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
2092 }
2093 
2094 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
2095 {
2096 	struct hns_roce_caps *caps = &hr_dev->caps;
2097 
2098 	/* EQ */
2099 	caps->eqe_ba_pg_sz = 0;
2100 	caps->eqe_buf_pg_sz = 0;
2101 
2102 	/* Link Table */
2103 	caps->llm_buf_pg_sz = 0;
2104 
2105 	/* MR */
2106 	caps->mpt_ba_pg_sz = 0;
2107 	caps->mpt_buf_pg_sz = 0;
2108 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
2109 	caps->pbl_buf_pg_sz = 0;
2110 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
2111 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
2112 		   HEM_TYPE_MTPT);
2113 
2114 	/* QP */
2115 	caps->qpc_ba_pg_sz = 0;
2116 	caps->qpc_buf_pg_sz = 0;
2117 	caps->qpc_timer_ba_pg_sz = 0;
2118 	caps->qpc_timer_buf_pg_sz = 0;
2119 	caps->sccc_ba_pg_sz = 0;
2120 	caps->sccc_buf_pg_sz = 0;
2121 	caps->mtt_ba_pg_sz = 0;
2122 	caps->mtt_buf_pg_sz = 0;
2123 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2124 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2125 		   HEM_TYPE_QPC);
2126 
2127 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2128 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2129 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2130 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2131 
2132 	/* CQ */
2133 	caps->cqc_ba_pg_sz = 0;
2134 	caps->cqc_buf_pg_sz = 0;
2135 	caps->cqc_timer_ba_pg_sz = 0;
2136 	caps->cqc_timer_buf_pg_sz = 0;
2137 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2138 	caps->cqe_buf_pg_sz = 0;
2139 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2140 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2141 		   HEM_TYPE_CQC);
2142 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2143 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2144 
2145 	/* SRQ */
2146 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2147 		caps->srqc_ba_pg_sz = 0;
2148 		caps->srqc_buf_pg_sz = 0;
2149 		caps->srqwqe_ba_pg_sz = 0;
2150 		caps->srqwqe_buf_pg_sz = 0;
2151 		caps->idx_ba_pg_sz = 0;
2152 		caps->idx_buf_pg_sz = 0;
2153 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2154 			   caps->srqc_hop_num, caps->srqc_bt_num,
2155 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2156 			   HEM_TYPE_SRQC);
2157 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2158 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2159 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2160 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2161 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2162 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2163 	}
2164 
2165 	/* GMV */
2166 	caps->gmv_ba_pg_sz = 0;
2167 	caps->gmv_buf_pg_sz = 0;
2168 }
2169 
2170 /* Apply all loaded caps before setting to hardware */
2171 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2172 {
2173 	struct hns_roce_caps *caps = &hr_dev->caps;
2174 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2175 
2176 	/* The following configurations don't need to be got from firmware. */
2177 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2178 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2179 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2180 
2181 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2182 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2183 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2184 
2185 	caps->num_xrcds = HNS_ROCE_V2_MAX_XRCD_NUM;
2186 	caps->reserved_xrcds = HNS_ROCE_V2_RSV_XRCD_NUM;
2187 
2188 	caps->num_mtt_segs = HNS_ROCE_V2_MAX_MTT_SEGS;
2189 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2190 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2191 
2192 	if (!caps->num_comp_vectors)
2193 		caps->num_comp_vectors = min_t(u32, caps->eqc_bt_num - 1,
2194 				  (u32)priv->handle->rinfo.num_vectors - 2);
2195 
2196 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2197 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2198 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2199 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2200 
2201 		/* The following configurations will be overwritten */
2202 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2203 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2204 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2205 
2206 		/* The following configurations are not got from firmware */
2207 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2208 
2209 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2210 		caps->gid_table_len[0] = caps->gmv_bt_num *
2211 					(HNS_HW_PAGE_SIZE / caps->gmv_entry_sz);
2212 
2213 		caps->gmv_entry_num = caps->gmv_bt_num * (PAGE_SIZE /
2214 							  caps->gmv_entry_sz);
2215 	} else {
2216 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2217 
2218 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2219 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2220 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2221 		caps->gid_table_len[0] /= func_num;
2222 	}
2223 
2224 	if (hr_dev->is_vf) {
2225 		caps->default_aeq_arm_st = 0x3;
2226 		caps->default_ceq_arm_st = 0x3;
2227 		caps->default_ceq_max_cnt = 0x1;
2228 		caps->default_ceq_period = 0x10;
2229 		caps->default_aeq_max_cnt = 0x1;
2230 		caps->default_aeq_period = 0x10;
2231 	}
2232 
2233 	set_hem_page_size(hr_dev);
2234 }
2235 
2236 static int hns_roce_query_pf_caps(struct hns_roce_dev *hr_dev)
2237 {
2238 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2239 	struct hns_roce_caps *caps = &hr_dev->caps;
2240 	struct hns_roce_query_pf_caps_a *resp_a;
2241 	struct hns_roce_query_pf_caps_b *resp_b;
2242 	struct hns_roce_query_pf_caps_c *resp_c;
2243 	struct hns_roce_query_pf_caps_d *resp_d;
2244 	struct hns_roce_query_pf_caps_e *resp_e;
2245 	int ctx_hop_num;
2246 	int pbl_hop_num;
2247 	int ret;
2248 	int i;
2249 
2250 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2251 		hns_roce_cmq_setup_basic_desc(&desc[i],
2252 					      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM,
2253 					      true);
2254 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2255 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2256 		else
2257 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2258 	}
2259 
2260 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2261 	if (ret)
2262 		return ret;
2263 
2264 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2265 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2266 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2267 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2268 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2269 
2270 	caps->local_ca_ack_delay     = resp_a->local_ca_ack_delay;
2271 	caps->max_sq_sg		     = le16_to_cpu(resp_a->max_sq_sg);
2272 	caps->max_sq_inline	     = le16_to_cpu(resp_a->max_sq_inline);
2273 	caps->max_rq_sg		     = le16_to_cpu(resp_a->max_rq_sg);
2274 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2275 	caps->max_extend_sg	     = le32_to_cpu(resp_a->max_extend_sg);
2276 	caps->num_qpc_timer	     = le16_to_cpu(resp_a->num_qpc_timer);
2277 	caps->max_srq_sges	     = le16_to_cpu(resp_a->max_srq_sges);
2278 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2279 	caps->num_aeq_vectors	     = resp_a->num_aeq_vectors;
2280 	caps->num_other_vectors	     = resp_a->num_other_vectors;
2281 	caps->max_sq_desc_sz	     = resp_a->max_sq_desc_sz;
2282 	caps->max_rq_desc_sz	     = resp_a->max_rq_desc_sz;
2283 	caps->max_srq_desc_sz	     = resp_a->max_srq_desc_sz;
2284 	caps->cqe_sz		     = resp_a->cqe_sz;
2285 
2286 	caps->mtpt_entry_sz	     = resp_b->mtpt_entry_sz;
2287 	caps->irrl_entry_sz	     = resp_b->irrl_entry_sz;
2288 	caps->trrl_entry_sz	     = resp_b->trrl_entry_sz;
2289 	caps->cqc_entry_sz	     = resp_b->cqc_entry_sz;
2290 	caps->srqc_entry_sz	     = resp_b->srqc_entry_sz;
2291 	caps->idx_entry_sz	     = resp_b->idx_entry_sz;
2292 	caps->sccc_sz		     = resp_b->sccc_sz;
2293 	caps->max_mtu		     = resp_b->max_mtu;
2294 	caps->qpc_sz		     = le16_to_cpu(resp_b->qpc_sz);
2295 	caps->min_cqes		     = resp_b->min_cqes;
2296 	caps->min_wqes		     = resp_b->min_wqes;
2297 	caps->page_size_cap	     = le32_to_cpu(resp_b->page_size_cap);
2298 	caps->pkey_table_len[0]	     = resp_b->pkey_table_len;
2299 	caps->phy_num_uars	     = resp_b->phy_num_uars;
2300 	ctx_hop_num		     = resp_b->ctx_hop_num;
2301 	pbl_hop_num		     = resp_b->pbl_hop_num;
2302 
2303 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2304 
2305 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2306 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2307 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2308 
2309 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2310 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2311 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2312 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2313 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2314 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2315 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2316 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2317 
2318 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2319 	caps->cong_type = hr_reg_read(resp_d, PF_CAPS_D_CONG_TYPE);
2320 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2321 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2322 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2323 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2324 	caps->default_aeq_arm_st = hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2325 	caps->default_ceq_arm_st = hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2326 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2327 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2328 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2329 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2330 
2331 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2332 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2333 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2334 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2335 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2336 	caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2337 	caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2338 	caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2339 	caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2340 
2341 	caps->qpc_hop_num = ctx_hop_num;
2342 	caps->sccc_hop_num = ctx_hop_num;
2343 	caps->srqc_hop_num = ctx_hop_num;
2344 	caps->cqc_hop_num = ctx_hop_num;
2345 	caps->mpt_hop_num = ctx_hop_num;
2346 	caps->mtt_hop_num = pbl_hop_num;
2347 	caps->cqe_hop_num = pbl_hop_num;
2348 	caps->srqwqe_hop_num = pbl_hop_num;
2349 	caps->idx_hop_num = pbl_hop_num;
2350 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2351 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2352 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2353 
2354 	return 0;
2355 }
2356 
2357 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2358 {
2359 	struct hns_roce_cmq_desc desc;
2360 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2361 
2362 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2363 				      false);
2364 
2365 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2366 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2367 
2368 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2369 }
2370 
2371 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2372 {
2373 	struct hns_roce_caps *caps = &hr_dev->caps;
2374 	int ret;
2375 
2376 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2377 		return 0;
2378 
2379 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2380 				    caps->qpc_sz);
2381 	if (ret) {
2382 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2383 		return ret;
2384 	}
2385 
2386 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2387 				    caps->sccc_sz);
2388 	if (ret)
2389 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2390 
2391 	return ret;
2392 }
2393 
2394 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2395 {
2396 	struct device *dev = hr_dev->dev;
2397 	int ret;
2398 
2399 	hr_dev->func_num = 1;
2400 
2401 	set_default_caps(hr_dev);
2402 
2403 	ret = hns_roce_query_vf_resource(hr_dev);
2404 	if (ret) {
2405 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2406 		return ret;
2407 	}
2408 
2409 	apply_func_caps(hr_dev);
2410 
2411 	ret = hns_roce_v2_set_bt(hr_dev);
2412 	if (ret)
2413 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2414 
2415 	return ret;
2416 }
2417 
2418 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2419 {
2420 	struct device *dev = hr_dev->dev;
2421 	int ret;
2422 
2423 	ret = hns_roce_query_func_info(hr_dev);
2424 	if (ret) {
2425 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2426 		return ret;
2427 	}
2428 
2429 	ret = hns_roce_config_global_param(hr_dev);
2430 	if (ret) {
2431 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2432 		return ret;
2433 	}
2434 
2435 	ret = hns_roce_set_vf_switch_param(hr_dev);
2436 	if (ret) {
2437 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2438 		return ret;
2439 	}
2440 
2441 	ret = hns_roce_query_pf_caps(hr_dev);
2442 	if (ret)
2443 		set_default_caps(hr_dev);
2444 
2445 	ret = hns_roce_query_pf_resource(hr_dev);
2446 	if (ret) {
2447 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2448 		return ret;
2449 	}
2450 
2451 	apply_func_caps(hr_dev);
2452 
2453 	ret = hns_roce_alloc_vf_resource(hr_dev);
2454 	if (ret) {
2455 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2456 		return ret;
2457 	}
2458 
2459 	ret = hns_roce_v2_set_bt(hr_dev);
2460 	if (ret) {
2461 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2462 		return ret;
2463 	}
2464 
2465 	/* Configure the size of QPC, SCCC, etc. */
2466 	return hns_roce_config_entry_size(hr_dev);
2467 }
2468 
2469 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2470 {
2471 	struct device *dev = hr_dev->dev;
2472 	int ret;
2473 
2474 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2475 	if (ret) {
2476 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2477 		return ret;
2478 	}
2479 
2480 	ret = hns_roce_query_fw_ver(hr_dev);
2481 	if (ret) {
2482 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2483 		return ret;
2484 	}
2485 
2486 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2487 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2488 
2489 	if (hr_dev->is_vf)
2490 		return hns_roce_v2_vf_profile(hr_dev);
2491 	else
2492 		return hns_roce_v2_pf_profile(hr_dev);
2493 }
2494 
2495 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2496 {
2497 	u32 i, next_ptr, page_num;
2498 	__le64 *entry = cfg_buf;
2499 	dma_addr_t addr;
2500 	u64 val;
2501 
2502 	page_num = data_buf->npages;
2503 	for (i = 0; i < page_num; i++) {
2504 		addr = hns_roce_buf_page(data_buf, i);
2505 		if (i == (page_num - 1))
2506 			next_ptr = 0;
2507 		else
2508 			next_ptr = i + 1;
2509 
2510 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2511 		entry[i] = cpu_to_le64(val);
2512 	}
2513 }
2514 
2515 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2516 			     struct hns_roce_link_table *table)
2517 {
2518 	struct hns_roce_cmq_desc desc[2];
2519 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2520 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2521 	struct hns_roce_buf *buf = table->buf;
2522 	enum hns_roce_opcode_type opcode;
2523 	dma_addr_t addr;
2524 
2525 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2526 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2527 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2528 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2529 
2530 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2531 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2532 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2533 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2534 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2535 
2536 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2537 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2538 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2539 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2540 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2541 
2542 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2543 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2544 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2545 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2546 
2547 	return hns_roce_cmq_send(hr_dev, desc, 2);
2548 }
2549 
2550 static struct hns_roce_link_table *
2551 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2552 {
2553 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2554 	struct hns_roce_link_table *link_tbl;
2555 	u32 pg_shift, size, min_size;
2556 
2557 	link_tbl = &priv->ext_llm;
2558 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2559 	size = hr_dev->caps.num_qps * HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2560 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(hr_dev->caps.sl_num) << pg_shift;
2561 
2562 	/* Alloc data table */
2563 	size = max(size, min_size);
2564 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2565 	if (IS_ERR(link_tbl->buf))
2566 		return ERR_PTR(-ENOMEM);
2567 
2568 	/* Alloc config table */
2569 	size = link_tbl->buf->npages * sizeof(u64);
2570 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2571 						 &link_tbl->table.map,
2572 						 GFP_KERNEL);
2573 	if (!link_tbl->table.buf) {
2574 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2575 		return ERR_PTR(-ENOMEM);
2576 	}
2577 
2578 	return link_tbl;
2579 }
2580 
2581 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2582 				struct hns_roce_link_table *tbl)
2583 {
2584 	if (tbl->buf) {
2585 		u32 size = tbl->buf->npages * sizeof(u64);
2586 
2587 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2588 				  tbl->table.map);
2589 	}
2590 
2591 	hns_roce_buf_free(hr_dev, tbl->buf);
2592 }
2593 
2594 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2595 {
2596 	struct hns_roce_link_table *link_tbl;
2597 	int ret;
2598 
2599 	link_tbl = alloc_link_table_buf(hr_dev);
2600 	if (IS_ERR(link_tbl))
2601 		return -ENOMEM;
2602 
2603 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2604 		ret = -EINVAL;
2605 		goto err_alloc;
2606 	}
2607 
2608 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2609 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2610 	if (ret)
2611 		goto err_alloc;
2612 
2613 	return 0;
2614 
2615 err_alloc:
2616 	free_link_table_buf(hr_dev, link_tbl);
2617 	return ret;
2618 }
2619 
2620 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2621 {
2622 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2623 
2624 	free_link_table_buf(hr_dev, &priv->ext_llm);
2625 }
2626 
2627 static void free_dip_list(struct hns_roce_dev *hr_dev)
2628 {
2629 	struct hns_roce_dip *hr_dip;
2630 	struct hns_roce_dip *tmp;
2631 	unsigned long flags;
2632 
2633 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
2634 
2635 	list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) {
2636 		list_del(&hr_dip->node);
2637 		kfree(hr_dip);
2638 	}
2639 
2640 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
2641 }
2642 
2643 static void free_mr_exit(struct hns_roce_dev *hr_dev)
2644 {
2645 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2646 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2647 	int ret;
2648 	int i;
2649 
2650 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2651 		if (free_mr->rsv_qp[i]) {
2652 			ret = ib_destroy_qp(free_mr->rsv_qp[i]);
2653 			if (ret)
2654 				ibdev_err(&hr_dev->ib_dev,
2655 					  "failed to destroy qp in free mr.\n");
2656 
2657 			free_mr->rsv_qp[i] = NULL;
2658 		}
2659 	}
2660 
2661 	if (free_mr->rsv_cq) {
2662 		ib_destroy_cq(free_mr->rsv_cq);
2663 		free_mr->rsv_cq = NULL;
2664 	}
2665 
2666 	if (free_mr->rsv_pd) {
2667 		ib_dealloc_pd(free_mr->rsv_pd);
2668 		free_mr->rsv_pd = NULL;
2669 	}
2670 }
2671 
2672 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2673 {
2674 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2675 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2676 	struct ib_device *ibdev = &hr_dev->ib_dev;
2677 	struct ib_cq_init_attr cq_init_attr = {};
2678 	struct ib_qp_init_attr qp_init_attr = {};
2679 	struct ib_pd *pd;
2680 	struct ib_cq *cq;
2681 	struct ib_qp *qp;
2682 	int ret;
2683 	int i;
2684 
2685 	pd = ib_alloc_pd(ibdev, 0);
2686 	if (IS_ERR(pd)) {
2687 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2688 		return PTR_ERR(pd);
2689 	}
2690 	free_mr->rsv_pd = pd;
2691 
2692 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2693 	cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_init_attr);
2694 	if (IS_ERR(cq)) {
2695 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2696 		ret = PTR_ERR(cq);
2697 		goto create_failed;
2698 	}
2699 	free_mr->rsv_cq = cq;
2700 
2701 	qp_init_attr.qp_type = IB_QPT_RC;
2702 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2703 	qp_init_attr.send_cq = free_mr->rsv_cq;
2704 	qp_init_attr.recv_cq = free_mr->rsv_cq;
2705 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2706 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2707 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2708 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2709 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2710 
2711 		qp = ib_create_qp(free_mr->rsv_pd, &qp_init_attr);
2712 		if (IS_ERR(qp)) {
2713 			ibdev_err(ibdev, "failed to create qp for free mr.\n");
2714 			ret = PTR_ERR(qp);
2715 			goto create_failed;
2716 		}
2717 
2718 		free_mr->rsv_qp[i] = qp;
2719 	}
2720 
2721 	return 0;
2722 
2723 create_failed:
2724 	free_mr_exit(hr_dev);
2725 
2726 	return ret;
2727 }
2728 
2729 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2730 				 struct ib_qp_attr *attr, int sl_num)
2731 {
2732 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2733 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2734 	struct ib_device *ibdev = &hr_dev->ib_dev;
2735 	struct hns_roce_qp *hr_qp;
2736 	int loopback;
2737 	int mask;
2738 	int ret;
2739 
2740 	hr_qp = to_hr_qp(free_mr->rsv_qp[sl_num]);
2741 	hr_qp->free_mr_en = 1;
2742 
2743 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2744 	attr->qp_state = IB_QPS_INIT;
2745 	attr->port_num = 1;
2746 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2747 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2748 	if (ret) {
2749 		ibdev_err(ibdev, "failed to modify qp to init, ret = %d.\n",
2750 			  ret);
2751 		return ret;
2752 	}
2753 
2754 	loopback = hr_dev->loop_idc;
2755 	/* Set qpc lbi = 1 incidate loopback IO */
2756 	hr_dev->loop_idc = 1;
2757 
2758 	mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
2759 	       IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
2760 	attr->qp_state = IB_QPS_RTR;
2761 	attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2762 	attr->path_mtu = IB_MTU_256;
2763 	attr->dest_qp_num = hr_qp->qpn;
2764 	attr->rq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2765 
2766 	rdma_ah_set_sl(&attr->ah_attr, (u8)sl_num);
2767 
2768 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2769 	hr_dev->loop_idc = loopback;
2770 	if (ret) {
2771 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2772 			  ret);
2773 		return ret;
2774 	}
2775 
2776 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2777 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2778 	attr->qp_state = IB_QPS_RTS;
2779 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2780 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2781 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2782 	ret = ib_modify_qp(&hr_qp->ibqp, attr, mask);
2783 	if (ret)
2784 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2785 			  ret);
2786 
2787 	return ret;
2788 }
2789 
2790 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2791 {
2792 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2793 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2794 	struct ib_qp_attr attr = {};
2795 	int ret;
2796 	int i;
2797 
2798 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2799 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2800 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2801 
2802 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2803 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2804 		if (ret)
2805 			return ret;
2806 	}
2807 
2808 	return 0;
2809 }
2810 
2811 static int free_mr_init(struct hns_roce_dev *hr_dev)
2812 {
2813 	int ret;
2814 
2815 	ret = free_mr_alloc_res(hr_dev);
2816 	if (ret)
2817 		return ret;
2818 
2819 	ret = free_mr_modify_qp(hr_dev);
2820 	if (ret)
2821 		goto err_modify_qp;
2822 
2823 	return 0;
2824 
2825 err_modify_qp:
2826 	free_mr_exit(hr_dev);
2827 
2828 	return ret;
2829 }
2830 
2831 static int get_hem_table(struct hns_roce_dev *hr_dev)
2832 {
2833 	unsigned int qpc_count;
2834 	unsigned int cqc_count;
2835 	unsigned int gmv_count;
2836 	int ret;
2837 	int i;
2838 
2839 	/* Alloc memory for source address table buffer space chunk */
2840 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2841 	     gmv_count++) {
2842 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2843 		if (ret)
2844 			goto err_gmv_failed;
2845 	}
2846 
2847 	if (hr_dev->is_vf)
2848 		return 0;
2849 
2850 	/* Alloc memory for QPC Timer buffer space chunk */
2851 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2852 	     qpc_count++) {
2853 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2854 					 qpc_count);
2855 		if (ret) {
2856 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2857 			goto err_qpc_timer_failed;
2858 		}
2859 	}
2860 
2861 	/* Alloc memory for CQC Timer buffer space chunk */
2862 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2863 	     cqc_count++) {
2864 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2865 					 cqc_count);
2866 		if (ret) {
2867 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2868 			goto err_cqc_timer_failed;
2869 		}
2870 	}
2871 
2872 	return 0;
2873 
2874 err_cqc_timer_failed:
2875 	for (i = 0; i < cqc_count; i++)
2876 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2877 
2878 err_qpc_timer_failed:
2879 	for (i = 0; i < qpc_count; i++)
2880 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2881 
2882 err_gmv_failed:
2883 	for (i = 0; i < gmv_count; i++)
2884 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2885 
2886 	return ret;
2887 }
2888 
2889 static void put_hem_table(struct hns_roce_dev *hr_dev)
2890 {
2891 	int i;
2892 
2893 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2894 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2895 
2896 	if (hr_dev->is_vf)
2897 		return;
2898 
2899 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2900 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2901 
2902 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2903 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2904 }
2905 
2906 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2907 {
2908 	int ret;
2909 
2910 	/* The hns ROCEE requires the extdb info to be cleared before using */
2911 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2912 	if (ret)
2913 		return ret;
2914 
2915 	ret = get_hem_table(hr_dev);
2916 	if (ret)
2917 		return ret;
2918 
2919 	if (hr_dev->is_vf)
2920 		return 0;
2921 
2922 	ret = hns_roce_init_link_table(hr_dev);
2923 	if (ret) {
2924 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2925 		goto err_llm_init_failed;
2926 	}
2927 
2928 	return 0;
2929 
2930 err_llm_init_failed:
2931 	put_hem_table(hr_dev);
2932 
2933 	return ret;
2934 }
2935 
2936 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2937 {
2938 	hns_roce_function_clear(hr_dev);
2939 
2940 	if (!hr_dev->is_vf)
2941 		hns_roce_free_link_table(hr_dev);
2942 
2943 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2944 		free_dip_list(hr_dev);
2945 }
2946 
2947 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2948 			      struct hns_roce_mbox_msg *mbox_msg)
2949 {
2950 	struct hns_roce_cmq_desc desc;
2951 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2952 
2953 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2954 
2955 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2956 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2957 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2958 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2959 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
2960 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
2961 					 mbox_msg->token);
2962 
2963 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2964 }
2965 
2966 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
2967 				 u8 *complete_status)
2968 {
2969 	struct hns_roce_mbox_status *mb_st;
2970 	struct hns_roce_cmq_desc desc;
2971 	unsigned long end;
2972 	int ret = -EBUSY;
2973 	u32 status;
2974 	bool busy;
2975 
2976 	mb_st = (struct hns_roce_mbox_status *)desc.data;
2977 	end = msecs_to_jiffies(timeout) + jiffies;
2978 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
2979 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
2980 			return -EIO;
2981 
2982 		status = 0;
2983 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
2984 					      true);
2985 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
2986 		if (!ret) {
2987 			status = le32_to_cpu(mb_st->mb_status_hw_run);
2988 			/* No pending message exists in ROCEE mbox. */
2989 			if (!(status & MB_ST_HW_RUN_M))
2990 				break;
2991 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
2992 			break;
2993 		}
2994 
2995 		if (time_after(jiffies, end)) {
2996 			dev_err_ratelimited(hr_dev->dev,
2997 					    "failed to wait mbox status 0x%x\n",
2998 					    status);
2999 			return -ETIMEDOUT;
3000 		}
3001 
3002 		cond_resched();
3003 		ret = -EBUSY;
3004 	}
3005 
3006 	if (!ret) {
3007 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
3008 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3009 		/* Ignore all errors if the mbox is unavailable. */
3010 		ret = 0;
3011 		*complete_status = MB_ST_COMPLETE_M;
3012 	}
3013 
3014 	return ret;
3015 }
3016 
3017 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
3018 			struct hns_roce_mbox_msg *mbox_msg)
3019 {
3020 	u8 status = 0;
3021 	int ret;
3022 
3023 	/* Waiting for the mbox to be idle */
3024 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
3025 				    &status);
3026 	if (unlikely(ret)) {
3027 		dev_err_ratelimited(hr_dev->dev,
3028 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
3029 				    status, ret);
3030 		return ret;
3031 	}
3032 
3033 	/* Post new message to mbox */
3034 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
3035 	if (ret)
3036 		dev_err_ratelimited(hr_dev->dev,
3037 				    "failed to post mailbox, ret = %d.\n", ret);
3038 
3039 	return ret;
3040 }
3041 
3042 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3043 {
3044 	u8 status = 0;
3045 	int ret;
3046 
3047 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3048 				    &status);
3049 	if (!ret) {
3050 		if (status != MB_ST_COMPLETE_SUCC)
3051 			return -EBUSY;
3052 	} else {
3053 		dev_err_ratelimited(hr_dev->dev,
3054 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3055 				    status, ret);
3056 	}
3057 
3058 	return ret;
3059 }
3060 
3061 static void copy_gid(void *dest, const union ib_gid *gid)
3062 {
3063 #define GID_SIZE 4
3064 	const union ib_gid *src = gid;
3065 	__le32 (*p)[GID_SIZE] = dest;
3066 	int i;
3067 
3068 	if (!gid)
3069 		src = &zgid;
3070 
3071 	for (i = 0; i < GID_SIZE; i++)
3072 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3073 }
3074 
3075 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3076 			     int gid_index, const union ib_gid *gid,
3077 			     enum hns_roce_sgid_type sgid_type)
3078 {
3079 	struct hns_roce_cmq_desc desc;
3080 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3081 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3082 
3083 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3084 
3085 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3086 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3087 
3088 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3089 
3090 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3091 }
3092 
3093 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3094 			    int gid_index, const union ib_gid *gid,
3095 			    enum hns_roce_sgid_type sgid_type,
3096 			    const struct ib_gid_attr *attr)
3097 {
3098 	struct hns_roce_cmq_desc desc[2];
3099 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3100 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3101 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3102 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3103 
3104 	u16 vlan_id = VLAN_CFI_MASK;
3105 	u8 mac[ETH_ALEN] = {};
3106 	int ret;
3107 
3108 	if (gid) {
3109 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3110 		if (ret)
3111 			return ret;
3112 	}
3113 
3114 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3115 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3116 
3117 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3118 
3119 	copy_gid(&tb_a->vf_sgid_l, gid);
3120 
3121 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3122 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3123 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3124 
3125 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3126 
3127 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3128 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3129 
3130 	return hns_roce_cmq_send(hr_dev, desc, 2);
3131 }
3132 
3133 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3134 			       const union ib_gid *gid,
3135 			       const struct ib_gid_attr *attr)
3136 {
3137 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3138 	int ret;
3139 
3140 	if (gid) {
3141 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3142 			if (ipv6_addr_v4mapped((void *)gid))
3143 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3144 			else
3145 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3146 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3147 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3148 		}
3149 	}
3150 
3151 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3152 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3153 	else
3154 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3155 
3156 	if (ret)
3157 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3158 			  ret);
3159 
3160 	return ret;
3161 }
3162 
3163 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3164 			       const u8 *addr)
3165 {
3166 	struct hns_roce_cmq_desc desc;
3167 	struct hns_roce_cfg_smac_tb *smac_tb =
3168 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3169 	u16 reg_smac_h;
3170 	u32 reg_smac_l;
3171 
3172 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3173 
3174 	reg_smac_l = *(u32 *)(&addr[0]);
3175 	reg_smac_h = *(u16 *)(&addr[4]);
3176 
3177 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3178 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3179 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3180 
3181 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3182 }
3183 
3184 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3185 			struct hns_roce_v2_mpt_entry *mpt_entry,
3186 			struct hns_roce_mr *mr)
3187 {
3188 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3189 	struct ib_device *ibdev = &hr_dev->ib_dev;
3190 	dma_addr_t pbl_ba;
3191 	int i, count;
3192 
3193 	count = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3194 				  ARRAY_SIZE(pages), &pbl_ba);
3195 	if (count < 1) {
3196 		ibdev_err(ibdev, "failed to find PBL mtr, count = %d.\n",
3197 			  count);
3198 		return -ENOBUFS;
3199 	}
3200 
3201 	/* Aligned to the hardware address access unit */
3202 	for (i = 0; i < count; i++)
3203 		pages[i] >>= 6;
3204 
3205 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3206 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> 3);
3207 	hr_reg_write(mpt_entry, MPT_PBL_BA_H, upper_32_bits(pbl_ba >> 3));
3208 
3209 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3210 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3211 
3212 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3213 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3214 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3215 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3216 
3217 	return 0;
3218 }
3219 
3220 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3221 				  void *mb_buf, struct hns_roce_mr *mr)
3222 {
3223 	struct hns_roce_v2_mpt_entry *mpt_entry;
3224 
3225 	mpt_entry = mb_buf;
3226 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3227 
3228 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3229 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3230 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3231 
3232 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3233 			  mr->access & IB_ACCESS_MW_BIND);
3234 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3235 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3236 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3237 			  mr->access & IB_ACCESS_REMOTE_READ);
3238 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3239 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3240 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3241 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3242 
3243 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3244 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3245 	mpt_entry->lkey = cpu_to_le32(mr->key);
3246 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3247 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3248 
3249 	if (mr->type != MR_TYPE_MR)
3250 		hr_reg_enable(mpt_entry, MPT_PA);
3251 
3252 	if (mr->type == MR_TYPE_DMA)
3253 		return 0;
3254 
3255 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3256 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3257 
3258 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3259 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3260 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3261 
3262 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3263 }
3264 
3265 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3266 					struct hns_roce_mr *mr, int flags,
3267 					void *mb_buf)
3268 {
3269 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3270 	u32 mr_access_flags = mr->access;
3271 	int ret = 0;
3272 
3273 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3274 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3275 
3276 	if (flags & IB_MR_REREG_ACCESS) {
3277 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3278 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3279 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3280 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3281 		hr_reg_write(mpt_entry, MPT_RR_EN,
3282 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3283 		hr_reg_write(mpt_entry, MPT_RW_EN,
3284 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3285 		hr_reg_write(mpt_entry, MPT_LW_EN,
3286 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3287 	}
3288 
3289 	if (flags & IB_MR_REREG_TRANS) {
3290 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3291 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3292 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3293 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3294 
3295 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3296 	}
3297 
3298 	return ret;
3299 }
3300 
3301 static int hns_roce_v2_frmr_write_mtpt(struct hns_roce_dev *hr_dev,
3302 				       void *mb_buf, struct hns_roce_mr *mr)
3303 {
3304 	struct ib_device *ibdev = &hr_dev->ib_dev;
3305 	struct hns_roce_v2_mpt_entry *mpt_entry;
3306 	dma_addr_t pbl_ba = 0;
3307 
3308 	mpt_entry = mb_buf;
3309 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3310 
3311 	if (hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, NULL, 0, &pbl_ba) < 0) {
3312 		ibdev_err(ibdev, "failed to find frmr mtr.\n");
3313 		return -ENOBUFS;
3314 	}
3315 
3316 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3317 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3318 
3319 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3320 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3321 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3322 
3323 	hr_reg_enable(mpt_entry, MPT_FRE);
3324 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3325 	hr_reg_enable(mpt_entry, MPT_BPD);
3326 	hr_reg_clear(mpt_entry, MPT_PA);
3327 
3328 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3329 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3330 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3331 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3332 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3333 
3334 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3335 
3336 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >> 3));
3337 	hr_reg_write(mpt_entry, MPT_PBL_BA_H, upper_32_bits(pbl_ba >> 3));
3338 
3339 	return 0;
3340 }
3341 
3342 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3343 {
3344 	struct hns_roce_v2_mpt_entry *mpt_entry;
3345 
3346 	mpt_entry = mb_buf;
3347 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3348 
3349 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3350 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3351 
3352 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3353 	hr_reg_enable(mpt_entry, MPT_L_INV_EN);
3354 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3355 
3356 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3357 	hr_reg_enable(mpt_entry, MPT_BPD);
3358 	hr_reg_clear(mpt_entry, MPT_PA);
3359 	hr_reg_write(mpt_entry, MPT_BQP,
3360 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3361 
3362 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3363 
3364 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3365 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3366 							     mw->pbl_hop_num);
3367 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3368 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3369 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3370 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3371 
3372 	return 0;
3373 }
3374 
3375 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3376 {
3377 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3378 	struct ib_device *ibdev = &hr_dev->ib_dev;
3379 	const struct ib_send_wr *bad_wr;
3380 	struct ib_rdma_wr rdma_wr = {};
3381 	struct ib_send_wr *send_wr;
3382 	int ret;
3383 
3384 	send_wr = &rdma_wr.wr;
3385 	send_wr->opcode = IB_WR_RDMA_WRITE;
3386 
3387 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3388 	if (ret) {
3389 		ibdev_err(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3390 			  ret);
3391 		return ret;
3392 	}
3393 
3394 	return 0;
3395 }
3396 
3397 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3398 			       struct ib_wc *wc);
3399 
3400 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3401 {
3402 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3403 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3404 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3405 	struct ib_device *ibdev = &hr_dev->ib_dev;
3406 	struct hns_roce_qp *hr_qp;
3407 	unsigned long end;
3408 	int cqe_cnt = 0;
3409 	int npolled;
3410 	int ret;
3411 	int i;
3412 
3413 	/*
3414 	 * If the device initialization is not complete or in the uninstall
3415 	 * process, then there is no need to execute free mr.
3416 	 */
3417 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3418 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3419 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3420 		return;
3421 
3422 	mutex_lock(&free_mr->mutex);
3423 
3424 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3425 		hr_qp = to_hr_qp(free_mr->rsv_qp[i]);
3426 
3427 		ret = free_mr_post_send_lp_wqe(hr_qp);
3428 		if (ret) {
3429 			ibdev_err(ibdev,
3430 				  "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3431 				  hr_qp->qpn, ret);
3432 			break;
3433 		}
3434 
3435 		cqe_cnt++;
3436 	}
3437 
3438 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3439 	while (cqe_cnt) {
3440 		npolled = hns_roce_v2_poll_cq(free_mr->rsv_cq, cqe_cnt, wc);
3441 		if (npolled < 0) {
3442 			ibdev_err(ibdev,
3443 				  "failed to poll cqe for free mr, remain %d cqe.\n",
3444 				  cqe_cnt);
3445 			goto out;
3446 		}
3447 
3448 		if (time_after(jiffies, end)) {
3449 			ibdev_err(ibdev,
3450 				  "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3451 				  cqe_cnt);
3452 			goto out;
3453 		}
3454 		cqe_cnt -= npolled;
3455 	}
3456 
3457 out:
3458 	mutex_unlock(&free_mr->mutex);
3459 }
3460 
3461 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3462 {
3463 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3464 		free_mr_send_cmd_to_hw(hr_dev);
3465 }
3466 
3467 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3468 {
3469 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3470 }
3471 
3472 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3473 {
3474 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3475 
3476 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3477 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3478 									 NULL;
3479 }
3480 
3481 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3482 				struct hns_roce_cq *hr_cq)
3483 {
3484 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3485 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3486 	} else {
3487 		struct hns_roce_v2_db cq_db = {};
3488 
3489 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3490 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3491 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3492 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3493 
3494 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3495 	}
3496 }
3497 
3498 static void __hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3499 				   struct hns_roce_srq *srq)
3500 {
3501 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3502 	struct hns_roce_v2_cqe *cqe, *dest;
3503 	u32 prod_index;
3504 	int nfreed = 0;
3505 	int wqe_index;
3506 	u8 owner_bit;
3507 
3508 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3509 	     ++prod_index) {
3510 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3511 			break;
3512 	}
3513 
3514 	/*
3515 	 * Now backwards through the CQ, removing CQ entries
3516 	 * that match our QP by overwriting them with next entries.
3517 	 */
3518 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3519 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3520 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3521 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3522 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3523 				hns_roce_free_srq_wqe(srq, wqe_index);
3524 			}
3525 			++nfreed;
3526 		} else if (nfreed) {
3527 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3528 					  hr_cq->ib_cq.cqe);
3529 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3530 			memcpy(dest, cqe, hr_cq->cqe_size);
3531 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3532 		}
3533 	}
3534 
3535 	if (nfreed) {
3536 		hr_cq->cons_index += nfreed;
3537 		update_cq_db(hr_dev, hr_cq);
3538 	}
3539 }
3540 
3541 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3542 				 struct hns_roce_srq *srq)
3543 {
3544 	spin_lock_irq(&hr_cq->lock);
3545 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3546 	spin_unlock_irq(&hr_cq->lock);
3547 }
3548 
3549 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3550 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3551 				  u64 *mtts, dma_addr_t dma_handle)
3552 {
3553 	struct hns_roce_v2_cq_context *cq_context;
3554 
3555 	cq_context = mb_buf;
3556 	memset(cq_context, 0, sizeof(*cq_context));
3557 
3558 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3559 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3560 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3561 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3562 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3563 
3564 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3565 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3566 
3567 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3568 		hr_reg_enable(cq_context, CQC_STASH);
3569 
3570 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3571 		     to_hr_hw_page_addr(mtts[0]));
3572 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3573 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3574 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3575 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3576 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3577 		     to_hr_hw_page_addr(mtts[1]));
3578 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3579 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3580 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3581 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3582 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3583 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3584 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> 3);
3585 	hr_reg_write(cq_context, CQC_CQE_BA_H, (dma_handle >> (32 + 3)));
3586 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3587 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3588 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3589 		     ((u32)hr_cq->db.dma) >> 1);
3590 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3591 		     hr_cq->db.dma >> 32);
3592 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3593 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3594 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3595 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3596 }
3597 
3598 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3599 				     enum ib_cq_notify_flags flags)
3600 {
3601 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3602 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3603 	struct hns_roce_v2_db cq_db = {};
3604 	u32 notify_flag;
3605 
3606 	/*
3607 	 * flags = 0, then notify_flag : next
3608 	 * flags = 1, then notify flag : solocited
3609 	 */
3610 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3611 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3612 
3613 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3614 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3615 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3616 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3617 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3618 
3619 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3620 
3621 	return 0;
3622 }
3623 
3624 static int hns_roce_handle_recv_inl_wqe(struct hns_roce_v2_cqe *cqe,
3625 					struct hns_roce_qp *qp,
3626 					struct ib_wc *wc)
3627 {
3628 	struct hns_roce_rinl_sge *sge_list;
3629 	u32 wr_num, wr_cnt, sge_num;
3630 	u32 sge_cnt, data_len, size;
3631 	void *wqe_buf;
3632 
3633 	wr_num = hr_reg_read(cqe, CQE_WQE_IDX);
3634 	wr_cnt = wr_num & (qp->rq.wqe_cnt - 1);
3635 
3636 	sge_list = qp->rq_inl_buf.wqe_list[wr_cnt].sg_list;
3637 	sge_num = qp->rq_inl_buf.wqe_list[wr_cnt].sge_cnt;
3638 	wqe_buf = hns_roce_get_recv_wqe(qp, wr_cnt);
3639 	data_len = wc->byte_len;
3640 
3641 	for (sge_cnt = 0; (sge_cnt < sge_num) && (data_len); sge_cnt++) {
3642 		size = min(sge_list[sge_cnt].len, data_len);
3643 		memcpy((void *)sge_list[sge_cnt].addr, wqe_buf, size);
3644 
3645 		data_len -= size;
3646 		wqe_buf += size;
3647 	}
3648 
3649 	if (unlikely(data_len)) {
3650 		wc->status = IB_WC_LOC_LEN_ERR;
3651 		return -EAGAIN;
3652 	}
3653 
3654 	return 0;
3655 }
3656 
3657 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3658 		   int num_entries, struct ib_wc *wc)
3659 {
3660 	unsigned int left;
3661 	int npolled = 0;
3662 
3663 	left = wq->head - wq->tail;
3664 	if (left == 0)
3665 		return 0;
3666 
3667 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3668 	while (npolled < left) {
3669 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3670 		wc->status = IB_WC_WR_FLUSH_ERR;
3671 		wc->vendor_err = 0;
3672 		wc->qp = &hr_qp->ibqp;
3673 
3674 		wq->tail++;
3675 		wc++;
3676 		npolled++;
3677 	}
3678 
3679 	return npolled;
3680 }
3681 
3682 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3683 				  struct ib_wc *wc)
3684 {
3685 	struct hns_roce_qp *hr_qp;
3686 	int npolled = 0;
3687 
3688 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3689 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3690 				   num_entries - npolled, wc + npolled);
3691 		if (npolled >= num_entries)
3692 			goto out;
3693 	}
3694 
3695 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3696 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3697 				   num_entries - npolled, wc + npolled);
3698 		if (npolled >= num_entries)
3699 			goto out;
3700 	}
3701 
3702 out:
3703 	return npolled;
3704 }
3705 
3706 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3707 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3708 			   struct ib_wc *wc)
3709 {
3710 	static const struct {
3711 		u32 cqe_status;
3712 		enum ib_wc_status wc_status;
3713 	} map[] = {
3714 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3715 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3716 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3717 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3718 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3719 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3720 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3721 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3722 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3723 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3724 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3725 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3726 		  IB_WC_RETRY_EXC_ERR },
3727 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3728 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3729 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3730 	};
3731 
3732 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3733 	int i;
3734 
3735 	wc->status = IB_WC_GENERAL_ERR;
3736 	for (i = 0; i < ARRAY_SIZE(map); i++)
3737 		if (cqe_status == map[i].cqe_status) {
3738 			wc->status = map[i].wc_status;
3739 			break;
3740 		}
3741 
3742 	if (likely(wc->status == IB_WC_SUCCESS ||
3743 		   wc->status == IB_WC_WR_FLUSH_ERR))
3744 		return;
3745 
3746 	ibdev_err(&hr_dev->ib_dev, "error cqe status 0x%x:\n", cqe_status);
3747 	print_hex_dump(KERN_ERR, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3748 		       cq->cqe_size, false);
3749 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3750 
3751 	/*
3752 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3753 	 * the standard protocol, the driver must ignore it and needn't to set
3754 	 * the QP to an error state.
3755 	 */
3756 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3757 		return;
3758 
3759 	flush_cqe(hr_dev, qp);
3760 }
3761 
3762 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3763 		      struct hns_roce_qp **cur_qp)
3764 {
3765 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3766 	struct hns_roce_qp *hr_qp = *cur_qp;
3767 	u32 qpn;
3768 
3769 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3770 
3771 	if (!hr_qp || qpn != hr_qp->qpn) {
3772 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3773 		if (unlikely(!hr_qp)) {
3774 			ibdev_err(&hr_dev->ib_dev,
3775 				  "CQ %06lx with entry for unknown QPN %06x\n",
3776 				  hr_cq->cqn, qpn);
3777 			return -EINVAL;
3778 		}
3779 		*cur_qp = hr_qp;
3780 	}
3781 
3782 	return 0;
3783 }
3784 
3785 /*
3786  * mapped-value = 1 + real-value
3787  * The ib wc opcode's real value is start from 0, In order to distinguish
3788  * between initialized and uninitialized map values, we plus 1 to the actual
3789  * value when defining the mapping, so that the validity can be identified by
3790  * checking whether the mapped value is greater than 0.
3791  */
3792 #define HR_WC_OP_MAP(hr_key, ib_key) \
3793 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3794 
3795 static const u32 wc_send_op_map[] = {
3796 	HR_WC_OP_MAP(SEND,			SEND),
3797 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3798 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3799 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3800 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3801 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3802 	HR_WC_OP_MAP(LOCAL_INV,			LOCAL_INV),
3803 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3804 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3805 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3806 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3807 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3808 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3809 };
3810 
3811 static int to_ib_wc_send_op(u32 hr_opcode)
3812 {
3813 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3814 		return -EINVAL;
3815 
3816 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3817 					   -EINVAL;
3818 }
3819 
3820 static const u32 wc_recv_op_map[] = {
3821 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3822 	HR_WC_OP_MAP(SEND,				RECV),
3823 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3824 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3825 };
3826 
3827 static int to_ib_wc_recv_op(u32 hr_opcode)
3828 {
3829 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3830 		return -EINVAL;
3831 
3832 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3833 					   -EINVAL;
3834 }
3835 
3836 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3837 {
3838 	u32 hr_opcode;
3839 	int ib_opcode;
3840 
3841 	wc->wc_flags = 0;
3842 
3843 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3844 	switch (hr_opcode) {
3845 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3846 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3847 		break;
3848 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3849 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3850 		wc->wc_flags |= IB_WC_WITH_IMM;
3851 		break;
3852 	case HNS_ROCE_V2_WQE_OP_LOCAL_INV:
3853 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3854 		break;
3855 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3856 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3857 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3858 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3859 		wc->byte_len  = 8;
3860 		break;
3861 	default:
3862 		break;
3863 	}
3864 
3865 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3866 	if (ib_opcode < 0)
3867 		wc->status = IB_WC_GENERAL_ERR;
3868 	else
3869 		wc->opcode = ib_opcode;
3870 }
3871 
3872 static inline bool is_rq_inl_enabled(struct ib_wc *wc, u32 hr_opcode,
3873 				     struct hns_roce_v2_cqe *cqe)
3874 {
3875 	return wc->qp->qp_type != IB_QPT_UD && wc->qp->qp_type != IB_QPT_GSI &&
3876 	       (hr_opcode == HNS_ROCE_V2_OPCODE_SEND ||
3877 		hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_IMM ||
3878 		hr_opcode == HNS_ROCE_V2_OPCODE_SEND_WITH_INV) &&
3879 	       hr_reg_read(cqe, CQE_RQ_INLINE);
3880 }
3881 
3882 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3883 {
3884 	struct hns_roce_qp *qp = to_hr_qp(wc->qp);
3885 	u32 hr_opcode;
3886 	int ib_opcode;
3887 	int ret;
3888 
3889 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3890 
3891 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3892 	switch (hr_opcode) {
3893 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3894 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3895 		wc->wc_flags = IB_WC_WITH_IMM;
3896 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3897 		break;
3898 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3899 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3900 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3901 		break;
3902 	default:
3903 		wc->wc_flags = 0;
3904 	}
3905 
3906 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3907 	if (ib_opcode < 0)
3908 		wc->status = IB_WC_GENERAL_ERR;
3909 	else
3910 		wc->opcode = ib_opcode;
3911 
3912 	if (is_rq_inl_enabled(wc, hr_opcode, cqe)) {
3913 		ret = hns_roce_handle_recv_inl_wqe(cqe, qp, wc);
3914 		if (unlikely(ret))
3915 			return ret;
3916 	}
3917 
3918 	wc->sl = hr_reg_read(cqe, CQE_SL);
3919 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3920 	wc->slid = 0;
3921 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3922 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3923 	wc->pkey_index = 0;
3924 
3925 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3926 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3927 		wc->wc_flags |= IB_WC_WITH_VLAN;
3928 	} else {
3929 		wc->vlan_id = 0xffff;
3930 	}
3931 
3932 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3933 
3934 	return 0;
3935 }
3936 
3937 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3938 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3939 {
3940 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3941 	struct hns_roce_qp *qp = *cur_qp;
3942 	struct hns_roce_srq *srq = NULL;
3943 	struct hns_roce_v2_cqe *cqe;
3944 	struct hns_roce_wq *wq;
3945 	int is_send;
3946 	u16 wqe_idx;
3947 	int ret;
3948 
3949 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3950 	if (!cqe)
3951 		return -EAGAIN;
3952 
3953 	++hr_cq->cons_index;
3954 	/* Memory barrier */
3955 	rmb();
3956 
3957 	ret = get_cur_qp(hr_cq, cqe, &qp);
3958 	if (ret)
3959 		return ret;
3960 
3961 	wc->qp = &qp->ibqp;
3962 	wc->vendor_err = 0;
3963 
3964 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3965 
3966 	is_send = !hr_reg_read(cqe, CQE_S_R);
3967 	if (is_send) {
3968 		wq = &qp->sq;
3969 
3970 		/* If sg_signal_bit is set, tail pointer will be updated to
3971 		 * the WQE corresponding to the current CQE.
3972 		 */
3973 		if (qp->sq_signal_bits)
3974 			wq->tail += (wqe_idx - (u16)wq->tail) &
3975 				    (wq->wqe_cnt - 1);
3976 
3977 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3978 		++wq->tail;
3979 
3980 		fill_send_wc(wc, cqe);
3981 	} else {
3982 		if (qp->ibqp.srq) {
3983 			srq = to_hr_srq(qp->ibqp.srq);
3984 			wc->wr_id = srq->wrid[wqe_idx];
3985 			hns_roce_free_srq_wqe(srq, wqe_idx);
3986 		} else {
3987 			wq = &qp->rq;
3988 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3989 			++wq->tail;
3990 		}
3991 
3992 		ret = fill_recv_wc(wc, cqe);
3993 	}
3994 
3995 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3996 	if (unlikely(wc->status != IB_WC_SUCCESS))
3997 		return 0;
3998 
3999 	return ret;
4000 }
4001 
4002 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
4003 			       struct ib_wc *wc)
4004 {
4005 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
4006 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
4007 	struct hns_roce_qp *cur_qp = NULL;
4008 	unsigned long flags;
4009 	int npolled;
4010 
4011 	spin_lock_irqsave(&hr_cq->lock, flags);
4012 
4013 	/*
4014 	 * When the device starts to reset, the state is RST_DOWN. At this time,
4015 	 * there may still be some valid CQEs in the hardware that are not
4016 	 * polled. Therefore, it is not allowed to switch to the software mode
4017 	 * immediately. When the state changes to UNINIT, CQE no longer exists
4018 	 * in the hardware, and then switch to software mode.
4019 	 */
4020 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
4021 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
4022 		goto out;
4023 	}
4024 
4025 	for (npolled = 0; npolled < num_entries; ++npolled) {
4026 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
4027 			break;
4028 	}
4029 
4030 	if (npolled)
4031 		update_cq_db(hr_dev, hr_cq);
4032 
4033 out:
4034 	spin_unlock_irqrestore(&hr_cq->lock, flags);
4035 
4036 	return npolled;
4037 }
4038 
4039 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
4040 			      u32 step_idx, u8 *mbox_cmd)
4041 {
4042 	u8 cmd;
4043 
4044 	switch (type) {
4045 	case HEM_TYPE_QPC:
4046 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
4047 		break;
4048 	case HEM_TYPE_MTPT:
4049 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
4050 		break;
4051 	case HEM_TYPE_CQC:
4052 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4053 		break;
4054 	case HEM_TYPE_SRQC:
4055 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4056 		break;
4057 	case HEM_TYPE_SCCC:
4058 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4059 		break;
4060 	case HEM_TYPE_QPC_TIMER:
4061 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4062 		break;
4063 	case HEM_TYPE_CQC_TIMER:
4064 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4065 		break;
4066 	default:
4067 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4068 		return -EINVAL;
4069 	}
4070 
4071 	*mbox_cmd = cmd + step_idx;
4072 
4073 	return 0;
4074 }
4075 
4076 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4077 			       dma_addr_t base_addr)
4078 {
4079 	struct hns_roce_cmq_desc desc;
4080 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4081 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4082 	u64 addr = to_hr_hw_page_addr(base_addr);
4083 
4084 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4085 
4086 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4087 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4088 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4089 
4090 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4091 }
4092 
4093 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4094 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4095 {
4096 	int ret;
4097 	u8 cmd;
4098 
4099 	if (unlikely(hem_type == HEM_TYPE_GMV))
4100 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4101 
4102 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4103 		return 0;
4104 
4105 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4106 	if (ret < 0)
4107 		return ret;
4108 
4109 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4110 }
4111 
4112 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4113 			       struct hns_roce_hem_table *table, int obj,
4114 			       u32 step_idx)
4115 {
4116 	struct hns_roce_hem_iter iter;
4117 	struct hns_roce_hem_mhop mhop;
4118 	struct hns_roce_hem *hem;
4119 	unsigned long mhop_obj = obj;
4120 	int i, j, k;
4121 	int ret = 0;
4122 	u64 hem_idx = 0;
4123 	u64 l1_idx = 0;
4124 	u64 bt_ba = 0;
4125 	u32 chunk_ba_num;
4126 	u32 hop_num;
4127 
4128 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4129 		return 0;
4130 
4131 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4132 	i = mhop.l0_idx;
4133 	j = mhop.l1_idx;
4134 	k = mhop.l2_idx;
4135 	hop_num = mhop.hop_num;
4136 	chunk_ba_num = mhop.bt_chunk_size / 8;
4137 
4138 	if (hop_num == 2) {
4139 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4140 			  k;
4141 		l1_idx = i * chunk_ba_num + j;
4142 	} else if (hop_num == 1) {
4143 		hem_idx = i * chunk_ba_num + j;
4144 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4145 		hem_idx = i;
4146 	}
4147 
4148 	if (table->type == HEM_TYPE_SCCC)
4149 		obj = mhop.l0_idx;
4150 
4151 	if (check_whether_last_step(hop_num, step_idx)) {
4152 		hem = table->hem[hem_idx];
4153 		for (hns_roce_hem_first(hem, &iter);
4154 		     !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
4155 			bt_ba = hns_roce_hem_addr(&iter);
4156 			ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type,
4157 					    step_idx);
4158 		}
4159 	} else {
4160 		if (step_idx == 0)
4161 			bt_ba = table->bt_l0_dma_addr[i];
4162 		else if (step_idx == 1 && hop_num == 2)
4163 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4164 
4165 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4166 	}
4167 
4168 	return ret;
4169 }
4170 
4171 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4172 				 struct hns_roce_hem_table *table,
4173 				 int tag, u32 step_idx)
4174 {
4175 	struct hns_roce_cmd_mailbox *mailbox;
4176 	struct device *dev = hr_dev->dev;
4177 	u8 cmd = 0xff;
4178 	int ret;
4179 
4180 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4181 		return 0;
4182 
4183 	switch (table->type) {
4184 	case HEM_TYPE_QPC:
4185 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4186 		break;
4187 	case HEM_TYPE_MTPT:
4188 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4189 		break;
4190 	case HEM_TYPE_CQC:
4191 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4192 		break;
4193 	case HEM_TYPE_SRQC:
4194 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4195 		break;
4196 	case HEM_TYPE_SCCC:
4197 	case HEM_TYPE_QPC_TIMER:
4198 	case HEM_TYPE_CQC_TIMER:
4199 	case HEM_TYPE_GMV:
4200 		return 0;
4201 	default:
4202 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4203 			 table->type);
4204 		return 0;
4205 	}
4206 
4207 	cmd += step_idx;
4208 
4209 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4210 	if (IS_ERR(mailbox))
4211 		return PTR_ERR(mailbox);
4212 
4213 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4214 
4215 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4216 	return ret;
4217 }
4218 
4219 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4220 				 struct hns_roce_v2_qp_context *context,
4221 				 struct hns_roce_v2_qp_context *qpc_mask,
4222 				 struct hns_roce_qp *hr_qp)
4223 {
4224 	struct hns_roce_cmd_mailbox *mailbox;
4225 	int qpc_size;
4226 	int ret;
4227 
4228 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4229 	if (IS_ERR(mailbox))
4230 		return PTR_ERR(mailbox);
4231 
4232 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4233 	qpc_size = hr_dev->caps.qpc_sz;
4234 	memcpy(mailbox->buf, context, qpc_size);
4235 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4236 
4237 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4238 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4239 
4240 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4241 
4242 	return ret;
4243 }
4244 
4245 static void set_access_flags(struct hns_roce_qp *hr_qp,
4246 			     struct hns_roce_v2_qp_context *context,
4247 			     struct hns_roce_v2_qp_context *qpc_mask,
4248 			     const struct ib_qp_attr *attr, int attr_mask)
4249 {
4250 	u8 dest_rd_atomic;
4251 	u32 access_flags;
4252 
4253 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4254 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4255 
4256 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4257 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4258 
4259 	if (!dest_rd_atomic)
4260 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4261 
4262 	hr_reg_write_bool(context, QPC_RRE,
4263 			  access_flags & IB_ACCESS_REMOTE_READ);
4264 	hr_reg_clear(qpc_mask, QPC_RRE);
4265 
4266 	hr_reg_write_bool(context, QPC_RWE,
4267 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4268 	hr_reg_clear(qpc_mask, QPC_RWE);
4269 
4270 	hr_reg_write_bool(context, QPC_ATE,
4271 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4272 	hr_reg_clear(qpc_mask, QPC_ATE);
4273 	hr_reg_write_bool(context, QPC_EXT_ATE,
4274 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4275 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4276 }
4277 
4278 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4279 			    struct hns_roce_v2_qp_context *context,
4280 			    struct hns_roce_v2_qp_context *qpc_mask)
4281 {
4282 	hr_reg_write(context, QPC_SGE_SHIFT,
4283 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4284 					     hr_qp->sge.sge_shift));
4285 
4286 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4287 
4288 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4289 }
4290 
4291 static inline int get_cqn(struct ib_cq *ib_cq)
4292 {
4293 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4294 }
4295 
4296 static inline int get_pdn(struct ib_pd *ib_pd)
4297 {
4298 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4299 }
4300 
4301 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4302 				    const struct ib_qp_attr *attr,
4303 				    int attr_mask,
4304 				    struct hns_roce_v2_qp_context *context,
4305 				    struct hns_roce_v2_qp_context *qpc_mask)
4306 {
4307 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4308 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4309 
4310 	/*
4311 	 * In v2 engine, software pass context and context mask to hardware
4312 	 * when modifying qp. If software need modify some fields in context,
4313 	 * we should set all bits of the relevant fields in context mask to
4314 	 * 0 at the same time, else set them to 0x1.
4315 	 */
4316 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4317 
4318 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4319 
4320 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4321 
4322 	set_qpc_wqe_cnt(hr_qp, context, qpc_mask);
4323 
4324 	/* No VLAN need to set 0xFFF */
4325 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4326 
4327 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4328 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4329 
4330 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4331 	}
4332 
4333 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4334 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4335 
4336 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4337 		hr_reg_enable(context, QPC_OWNER_MODE);
4338 
4339 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4340 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4341 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4342 		     upper_32_bits(hr_qp->rdb.dma));
4343 
4344 	if (ibqp->qp_type != IB_QPT_UD && ibqp->qp_type != IB_QPT_GSI)
4345 		hr_reg_write_bool(context, QPC_RQIE,
4346 			     hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_RQ_INLINE);
4347 
4348 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4349 
4350 	if (ibqp->srq) {
4351 		hr_reg_enable(context, QPC_SRQ_EN);
4352 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4353 	}
4354 
4355 	hr_reg_enable(context, QPC_FRE);
4356 
4357 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4358 
4359 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4360 		return;
4361 
4362 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4363 		hr_reg_enable(&context->ext, QPCEX_STASH);
4364 }
4365 
4366 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4367 				   const struct ib_qp_attr *attr, int attr_mask,
4368 				   struct hns_roce_v2_qp_context *context,
4369 				   struct hns_roce_v2_qp_context *qpc_mask)
4370 {
4371 	/*
4372 	 * In v2 engine, software pass context and context mask to hardware
4373 	 * when modifying qp. If software need modify some fields in context,
4374 	 * we should set all bits of the relevant fields in context mask to
4375 	 * 0 at the same time, else set them to 0x1.
4376 	 */
4377 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4378 	hr_reg_clear(qpc_mask, QPC_TST);
4379 
4380 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4381 	hr_reg_clear(qpc_mask, QPC_PD);
4382 
4383 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4384 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4385 
4386 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4387 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4388 
4389 	if (ibqp->srq) {
4390 		hr_reg_enable(context, QPC_SRQ_EN);
4391 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4392 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4393 		hr_reg_clear(qpc_mask, QPC_SRQN);
4394 	}
4395 }
4396 
4397 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4398 			    struct hns_roce_qp *hr_qp,
4399 			    struct hns_roce_v2_qp_context *context,
4400 			    struct hns_roce_v2_qp_context *qpc_mask)
4401 {
4402 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4403 	u64 wqe_sge_ba;
4404 	int count;
4405 
4406 	/* Search qp buf's mtts */
4407 	count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4408 				  MTT_MIN_COUNT, &wqe_sge_ba);
4409 	if (hr_qp->rq.wqe_cnt && count < 1) {
4410 		ibdev_err(&hr_dev->ib_dev,
4411 			  "failed to find RQ WQE, QPN = 0x%lx.\n", hr_qp->qpn);
4412 		return -EINVAL;
4413 	}
4414 
4415 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4416 	qpc_mask->wqe_sge_ba = 0;
4417 
4418 	/*
4419 	 * In v2 engine, software pass context and context mask to hardware
4420 	 * when modifying qp. If software need modify some fields in context,
4421 	 * we should set all bits of the relevant fields in context mask to
4422 	 * 0 at the same time, else set them to 0x1.
4423 	 */
4424 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4425 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4426 
4427 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4428 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4429 				      hr_qp->sq.wqe_cnt));
4430 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4431 
4432 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4433 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4434 				      hr_qp->sge.sge_cnt));
4435 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4436 
4437 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4438 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4439 				      hr_qp->rq.wqe_cnt));
4440 
4441 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4442 
4443 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4444 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4445 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4446 
4447 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4448 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4449 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4450 
4451 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4452 	qpc_mask->rq_cur_blk_addr = 0;
4453 
4454 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4455 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4456 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4457 
4458 	context->rq_nxt_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4459 	qpc_mask->rq_nxt_blk_addr = 0;
4460 
4461 	hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4462 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4463 	hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4464 
4465 	return 0;
4466 }
4467 
4468 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4469 			    struct hns_roce_qp *hr_qp,
4470 			    struct hns_roce_v2_qp_context *context,
4471 			    struct hns_roce_v2_qp_context *qpc_mask)
4472 {
4473 	struct ib_device *ibdev = &hr_dev->ib_dev;
4474 	u64 sge_cur_blk = 0;
4475 	u64 sq_cur_blk = 0;
4476 	int count;
4477 
4478 	/* search qp buf's mtts */
4479 	count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, 0, &sq_cur_blk, 1, NULL);
4480 	if (count < 1) {
4481 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ buf.\n",
4482 			  hr_qp->qpn);
4483 		return -EINVAL;
4484 	}
4485 	if (hr_qp->sge.sge_cnt > 0) {
4486 		count = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4487 					  hr_qp->sge.offset,
4488 					  &sge_cur_blk, 1, NULL);
4489 		if (count < 1) {
4490 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf.\n",
4491 				  hr_qp->qpn);
4492 			return -EINVAL;
4493 		}
4494 	}
4495 
4496 	/*
4497 	 * In v2 engine, software pass context and context mask to hardware
4498 	 * when modifying qp. If software need modify some fields in context,
4499 	 * we should set all bits of the relevant fields in context mask to
4500 	 * 0 at the same time, else set them to 0x1.
4501 	 */
4502 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4503 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4504 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4505 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4506 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4507 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4508 
4509 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4510 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4511 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4512 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4513 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4514 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4515 
4516 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4517 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4518 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4519 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4520 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4521 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4522 
4523 	return 0;
4524 }
4525 
4526 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4527 				  const struct ib_qp_attr *attr)
4528 {
4529 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4530 		return IB_MTU_4096;
4531 
4532 	return attr->path_mtu;
4533 }
4534 
4535 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4536 				 const struct ib_qp_attr *attr, int attr_mask,
4537 				 struct hns_roce_v2_qp_context *context,
4538 				 struct hns_roce_v2_qp_context *qpc_mask)
4539 {
4540 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4541 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4542 	struct ib_device *ibdev = &hr_dev->ib_dev;
4543 	dma_addr_t trrl_ba;
4544 	dma_addr_t irrl_ba;
4545 	enum ib_mtu ib_mtu;
4546 	const u8 *smac;
4547 	u8 lp_pktn_ini;
4548 	u64 *mtts;
4549 	u8 *dmac;
4550 	u32 port;
4551 	int mtu;
4552 	int ret;
4553 
4554 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4555 	if (ret) {
4556 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4557 		return ret;
4558 	}
4559 
4560 	/* Search IRRL's mtts */
4561 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4562 				   hr_qp->qpn, &irrl_ba);
4563 	if (!mtts) {
4564 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4565 		return -EINVAL;
4566 	}
4567 
4568 	/* Search TRRL's mtts */
4569 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4570 				   hr_qp->qpn, &trrl_ba);
4571 	if (!mtts) {
4572 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4573 		return -EINVAL;
4574 	}
4575 
4576 	if (attr_mask & IB_QP_ALT_PATH) {
4577 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4578 			  attr_mask);
4579 		return -EINVAL;
4580 	}
4581 
4582 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> 4);
4583 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4584 	context->trrl_ba = cpu_to_le32(trrl_ba >> (16 + 4));
4585 	qpc_mask->trrl_ba = 0;
4586 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> (32 + 16 + 4));
4587 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4588 
4589 	context->irrl_ba = cpu_to_le32(irrl_ba >> 6);
4590 	qpc_mask->irrl_ba = 0;
4591 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> (32 + 6));
4592 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4593 
4594 	hr_reg_enable(context, QPC_RMT_E2E);
4595 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4596 
4597 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4598 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4599 
4600 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4601 
4602 	smac = (const u8 *)hr_dev->dev_addr[port];
4603 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4604 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4605 	if (ether_addr_equal_unaligned(dmac, smac) ||
4606 	    hr_dev->loop_idc == 0x1) {
4607 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4608 		hr_reg_clear(qpc_mask, QPC_LBI);
4609 	}
4610 
4611 	if (attr_mask & IB_QP_DEST_QPN) {
4612 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4613 		hr_reg_clear(qpc_mask, QPC_DQPN);
4614 	}
4615 
4616 	memcpy(&(context->dmac), dmac, sizeof(u32));
4617 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4618 	qpc_mask->dmac = 0;
4619 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4620 
4621 	ib_mtu = get_mtu(ibqp, attr);
4622 	hr_qp->path_mtu = ib_mtu;
4623 
4624 	mtu = ib_mtu_enum_to_int(ib_mtu);
4625 	if (WARN_ON(mtu <= 0))
4626 		return -EINVAL;
4627 #define MAX_LP_MSG_LEN 16384
4628 	/* MTU * (2 ^ LP_PKTN_INI) shouldn't be bigger than 16KB */
4629 	lp_pktn_ini = ilog2(MAX_LP_MSG_LEN / mtu);
4630 	if (WARN_ON(lp_pktn_ini >= 0xF))
4631 		return -EINVAL;
4632 
4633 	if (attr_mask & IB_QP_PATH_MTU) {
4634 		hr_reg_write(context, QPC_MTU, ib_mtu);
4635 		hr_reg_clear(qpc_mask, QPC_MTU);
4636 	}
4637 
4638 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4639 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4640 
4641 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4642 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4643 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4644 
4645 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4646 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4647 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4648 
4649 	context->rq_rnr_timer = 0;
4650 	qpc_mask->rq_rnr_timer = 0;
4651 
4652 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4653 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4654 
4655 	/* rocee send 2^lp_sgen_ini segs every time */
4656 	hr_reg_write(context, QPC_LP_SGEN_INI, 3);
4657 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4658 
4659 	return 0;
4660 }
4661 
4662 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp,
4663 				const struct ib_qp_attr *attr, int attr_mask,
4664 				struct hns_roce_v2_qp_context *context,
4665 				struct hns_roce_v2_qp_context *qpc_mask)
4666 {
4667 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4668 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4669 	struct ib_device *ibdev = &hr_dev->ib_dev;
4670 	int ret;
4671 
4672 	/* Not support alternate path and path migration */
4673 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4674 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4675 		return -EINVAL;
4676 	}
4677 
4678 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4679 	if (ret) {
4680 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4681 		return ret;
4682 	}
4683 
4684 	/*
4685 	 * Set some fields in context to zero, Because the default values
4686 	 * of all fields in context are zero, we need not set them to 0 again.
4687 	 * but we should set the relevant fields of context mask to 0.
4688 	 */
4689 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4690 
4691 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4692 
4693 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4694 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4695 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4696 
4697 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4698 
4699 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4700 
4701 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4702 
4703 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4704 
4705 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4706 
4707 	return 0;
4708 }
4709 
4710 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4711 			   u32 *dip_idx)
4712 {
4713 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4714 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4715 	u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx;
4716 	u32 *head =  &hr_dev->qp_table.idx_table.head;
4717 	u32 *tail =  &hr_dev->qp_table.idx_table.tail;
4718 	struct hns_roce_dip *hr_dip;
4719 	unsigned long flags;
4720 	int ret = 0;
4721 
4722 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
4723 
4724 	spare_idx[*tail] = ibqp->qp_num;
4725 	*tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1);
4726 
4727 	list_for_each_entry(hr_dip, &hr_dev->dip_list, node) {
4728 		if (!memcmp(grh->dgid.raw, hr_dip->dgid, 16)) {
4729 			*dip_idx = hr_dip->dip_idx;
4730 			goto out;
4731 		}
4732 	}
4733 
4734 	/* If no dgid is found, a new dip and a mapping between dgid and
4735 	 * dip_idx will be created.
4736 	 */
4737 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC);
4738 	if (!hr_dip) {
4739 		ret = -ENOMEM;
4740 		goto out;
4741 	}
4742 
4743 	memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4744 	hr_dip->dip_idx = *dip_idx = spare_idx[*head];
4745 	*head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1);
4746 	list_add_tail(&hr_dip->node, &hr_dev->dip_list);
4747 
4748 out:
4749 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
4750 	return ret;
4751 }
4752 
4753 enum {
4754 	CONG_DCQCN,
4755 	CONG_WINDOW,
4756 };
4757 
4758 enum {
4759 	UNSUPPORT_CONG_LEVEL,
4760 	SUPPORT_CONG_LEVEL,
4761 };
4762 
4763 enum {
4764 	CONG_LDCP,
4765 	CONG_HC3,
4766 };
4767 
4768 enum {
4769 	DIP_INVALID,
4770 	DIP_VALID,
4771 };
4772 
4773 enum {
4774 	WND_LIMIT,
4775 	WND_UNLIMIT,
4776 };
4777 
4778 static int check_cong_type(struct ib_qp *ibqp,
4779 			   struct hns_roce_congestion_algorithm *cong_alg)
4780 {
4781 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4782 
4783 	/* different congestion types match different configurations */
4784 	switch (hr_dev->caps.cong_type) {
4785 	case CONG_TYPE_DCQCN:
4786 		cong_alg->alg_sel = CONG_DCQCN;
4787 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4788 		cong_alg->dip_vld = DIP_INVALID;
4789 		cong_alg->wnd_mode_sel = WND_LIMIT;
4790 		break;
4791 	case CONG_TYPE_LDCP:
4792 		cong_alg->alg_sel = CONG_WINDOW;
4793 		cong_alg->alg_sub_sel = CONG_LDCP;
4794 		cong_alg->dip_vld = DIP_INVALID;
4795 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4796 		break;
4797 	case CONG_TYPE_HC3:
4798 		cong_alg->alg_sel = CONG_WINDOW;
4799 		cong_alg->alg_sub_sel = CONG_HC3;
4800 		cong_alg->dip_vld = DIP_INVALID;
4801 		cong_alg->wnd_mode_sel = WND_LIMIT;
4802 		break;
4803 	case CONG_TYPE_DIP:
4804 		cong_alg->alg_sel = CONG_DCQCN;
4805 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4806 		cong_alg->dip_vld = DIP_VALID;
4807 		cong_alg->wnd_mode_sel = WND_LIMIT;
4808 		break;
4809 	default:
4810 		ibdev_err(&hr_dev->ib_dev,
4811 			  "error type(%u) for congestion selection.\n",
4812 			  hr_dev->caps.cong_type);
4813 		return -EINVAL;
4814 	}
4815 
4816 	return 0;
4817 }
4818 
4819 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4820 			   struct hns_roce_v2_qp_context *context,
4821 			   struct hns_roce_v2_qp_context *qpc_mask)
4822 {
4823 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4824 	struct hns_roce_congestion_algorithm cong_field;
4825 	struct ib_device *ibdev = ibqp->device;
4826 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4827 	u32 dip_idx = 0;
4828 	int ret;
4829 
4830 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4831 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4832 		return 0;
4833 
4834 	ret = check_cong_type(ibqp, &cong_field);
4835 	if (ret)
4836 		return ret;
4837 
4838 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4839 		     hr_dev->caps.cong_type * HNS_ROCE_CONG_SIZE);
4840 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4841 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4842 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4843 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4844 		     cong_field.alg_sub_sel);
4845 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4846 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4847 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4848 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4849 		     cong_field.wnd_mode_sel);
4850 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4851 
4852 	/* if dip is disabled, there is no need to set dip idx */
4853 	if (cong_field.dip_vld == 0)
4854 		return 0;
4855 
4856 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4857 	if (ret) {
4858 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4859 		return ret;
4860 	}
4861 
4862 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4863 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4864 
4865 	return 0;
4866 }
4867 
4868 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4869 				const struct ib_qp_attr *attr,
4870 				int attr_mask,
4871 				struct hns_roce_v2_qp_context *context,
4872 				struct hns_roce_v2_qp_context *qpc_mask)
4873 {
4874 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4875 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4876 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4877 	struct ib_device *ibdev = &hr_dev->ib_dev;
4878 	const struct ib_gid_attr *gid_attr = NULL;
4879 	int is_roce_protocol;
4880 	u16 vlan_id = 0xffff;
4881 	bool is_udp = false;
4882 	u8 ib_port;
4883 	u8 hr_port;
4884 	int ret;
4885 
4886 	/*
4887 	 * If free_mr_en of qp is set, it means that this qp comes from
4888 	 * free mr. This qp will perform the loopback operation.
4889 	 * In the loopback scenario, only sl needs to be set.
4890 	 */
4891 	if (hr_qp->free_mr_en) {
4892 		hr_reg_write(context, QPC_SL, rdma_ah_get_sl(&attr->ah_attr));
4893 		hr_reg_clear(qpc_mask, QPC_SL);
4894 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4895 		return 0;
4896 	}
4897 
4898 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4899 	hr_port = ib_port - 1;
4900 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4901 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4902 
4903 	if (is_roce_protocol) {
4904 		gid_attr = attr->ah_attr.grh.sgid_attr;
4905 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4906 		if (ret)
4907 			return ret;
4908 
4909 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4910 	}
4911 
4912 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4913 	if (vlan_id < VLAN_N_VID &&
4914 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4915 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4916 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
4917 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
4918 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
4919 	}
4920 
4921 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
4922 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
4923 
4924 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
4925 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
4926 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
4927 		return -EINVAL;
4928 	}
4929 
4930 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
4931 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
4932 		return -EINVAL;
4933 	}
4934 
4935 	hr_reg_write(context, QPC_UDPSPN,
4936 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
4937 						 attr->dest_qp_num) :
4938 				    0);
4939 
4940 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
4941 
4942 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
4943 
4944 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
4945 
4946 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
4947 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
4948 
4949 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
4950 	if (ret)
4951 		return ret;
4952 
4953 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
4954 	hr_reg_clear(qpc_mask, QPC_TC);
4955 
4956 	hr_reg_write(context, QPC_FL, grh->flow_label);
4957 	hr_reg_clear(qpc_mask, QPC_FL);
4958 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4959 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
4960 
4961 	hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4962 	if (unlikely(hr_qp->sl > MAX_SERVICE_LEVEL)) {
4963 		ibdev_err(ibdev,
4964 			  "failed to fill QPC, sl (%u) shouldn't be larger than %d.\n",
4965 			  hr_qp->sl, MAX_SERVICE_LEVEL);
4966 		return -EINVAL;
4967 	}
4968 
4969 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4970 	hr_reg_clear(qpc_mask, QPC_SL);
4971 
4972 	return 0;
4973 }
4974 
4975 static bool check_qp_state(enum ib_qp_state cur_state,
4976 			   enum ib_qp_state new_state)
4977 {
4978 	static const bool sm[][IB_QPS_ERR + 1] = {
4979 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
4980 				   [IB_QPS_INIT] = true },
4981 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
4982 				  [IB_QPS_INIT] = true,
4983 				  [IB_QPS_RTR] = true,
4984 				  [IB_QPS_ERR] = true },
4985 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
4986 				 [IB_QPS_RTS] = true,
4987 				 [IB_QPS_ERR] = true },
4988 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
4989 				 [IB_QPS_RTS] = true,
4990 				 [IB_QPS_ERR] = true },
4991 		[IB_QPS_SQD] = {},
4992 		[IB_QPS_SQE] = {},
4993 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
4994 				 [IB_QPS_ERR] = true }
4995 	};
4996 
4997 	return sm[cur_state][new_state];
4998 }
4999 
5000 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
5001 				      const struct ib_qp_attr *attr,
5002 				      int attr_mask,
5003 				      enum ib_qp_state cur_state,
5004 				      enum ib_qp_state new_state,
5005 				      struct hns_roce_v2_qp_context *context,
5006 				      struct hns_roce_v2_qp_context *qpc_mask)
5007 {
5008 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5009 	int ret = 0;
5010 
5011 	if (!check_qp_state(cur_state, new_state)) {
5012 		ibdev_err(&hr_dev->ib_dev, "Illegal state for QP!\n");
5013 		return -EINVAL;
5014 	}
5015 
5016 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
5017 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
5018 		modify_qp_reset_to_init(ibqp, attr, attr_mask, context,
5019 					qpc_mask);
5020 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
5021 		modify_qp_init_to_init(ibqp, attr, attr_mask, context,
5022 				       qpc_mask);
5023 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
5024 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
5025 					    qpc_mask);
5026 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
5027 		ret = modify_qp_rtr_to_rts(ibqp, attr, attr_mask, context,
5028 					   qpc_mask);
5029 	}
5030 
5031 	return ret;
5032 }
5033 
5034 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
5035 {
5036 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5037 #define QP_ACK_TIMEOUT_OFFSET 10
5038 #define QP_ACK_TIMEOUT_MAX 31
5039 
5040 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5041 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5042 			ibdev_warn(&hr_dev->ib_dev,
5043 				   "Local ACK timeout shall be 0 to 20.\n");
5044 			return false;
5045 		}
5046 		*timeout += QP_ACK_TIMEOUT_OFFSET;
5047 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5048 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5049 			ibdev_warn(&hr_dev->ib_dev,
5050 				   "Local ACK timeout shall be 0 to 31.\n");
5051 			return false;
5052 		}
5053 	}
5054 
5055 	return true;
5056 }
5057 
5058 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5059 				      const struct ib_qp_attr *attr,
5060 				      int attr_mask,
5061 				      struct hns_roce_v2_qp_context *context,
5062 				      struct hns_roce_v2_qp_context *qpc_mask)
5063 {
5064 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5065 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5066 	int ret = 0;
5067 	u8 timeout;
5068 
5069 	if (attr_mask & IB_QP_AV) {
5070 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5071 					   qpc_mask);
5072 		if (ret)
5073 			return ret;
5074 	}
5075 
5076 	if (attr_mask & IB_QP_TIMEOUT) {
5077 		timeout = attr->timeout;
5078 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5079 			hr_reg_write(context, QPC_AT, timeout);
5080 			hr_reg_clear(qpc_mask, QPC_AT);
5081 		}
5082 	}
5083 
5084 	if (attr_mask & IB_QP_RETRY_CNT) {
5085 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5086 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5087 
5088 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5089 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5090 	}
5091 
5092 	if (attr_mask & IB_QP_RNR_RETRY) {
5093 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5094 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5095 
5096 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5097 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5098 	}
5099 
5100 	if (attr_mask & IB_QP_SQ_PSN) {
5101 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5102 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5103 
5104 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5105 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5106 
5107 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5108 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5109 
5110 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5111 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5112 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5113 
5114 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5115 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5116 
5117 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5118 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5119 	}
5120 
5121 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5122 	     attr->max_dest_rd_atomic) {
5123 		hr_reg_write(context, QPC_RR_MAX,
5124 			     fls(attr->max_dest_rd_atomic - 1));
5125 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5126 	}
5127 
5128 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5129 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5130 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5131 	}
5132 
5133 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5134 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5135 
5136 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5137 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5138 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5139 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5140 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5141 	}
5142 
5143 	if (attr_mask & IB_QP_RQ_PSN) {
5144 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5145 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5146 
5147 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5148 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5149 	}
5150 
5151 	if (attr_mask & IB_QP_QKEY) {
5152 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5153 		qpc_mask->qkey_xrcd = 0;
5154 		hr_qp->qkey = attr->qkey;
5155 	}
5156 
5157 	return ret;
5158 }
5159 
5160 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5161 					  const struct ib_qp_attr *attr,
5162 					  int attr_mask)
5163 {
5164 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5165 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5166 
5167 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5168 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5169 
5170 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5171 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5172 	if (attr_mask & IB_QP_PORT) {
5173 		hr_qp->port = attr->port_num - 1;
5174 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5175 	}
5176 }
5177 
5178 static void clear_qp(struct hns_roce_qp *hr_qp)
5179 {
5180 	struct ib_qp *ibqp = &hr_qp->ibqp;
5181 
5182 	if (ibqp->send_cq)
5183 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5184 				     hr_qp->qpn, NULL);
5185 
5186 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5187 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5188 				     hr_qp->qpn, ibqp->srq ?
5189 				     to_hr_srq(ibqp->srq) : NULL);
5190 
5191 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5192 		*hr_qp->rdb.db_record = 0;
5193 
5194 	hr_qp->rq.head = 0;
5195 	hr_qp->rq.tail = 0;
5196 	hr_qp->sq.head = 0;
5197 	hr_qp->sq.tail = 0;
5198 	hr_qp->next_sge = 0;
5199 }
5200 
5201 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5202 				  struct hns_roce_v2_qp_context *context,
5203 				  struct hns_roce_v2_qp_context *qpc_mask)
5204 {
5205 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5206 	unsigned long sq_flag = 0;
5207 	unsigned long rq_flag = 0;
5208 
5209 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5210 		return;
5211 
5212 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5213 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5214 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5215 	hr_qp->state = IB_QPS_ERR;
5216 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5217 
5218 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5219 		return;
5220 
5221 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5222 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5223 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5224 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5225 }
5226 
5227 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5228 				 const struct ib_qp_attr *attr,
5229 				 int attr_mask, enum ib_qp_state cur_state,
5230 				 enum ib_qp_state new_state)
5231 {
5232 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5233 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5234 	struct hns_roce_v2_qp_context ctx[2];
5235 	struct hns_roce_v2_qp_context *context = ctx;
5236 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5237 	struct ib_device *ibdev = &hr_dev->ib_dev;
5238 	int ret;
5239 
5240 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5241 		return -EOPNOTSUPP;
5242 
5243 	/*
5244 	 * In v2 engine, software pass context and context mask to hardware
5245 	 * when modifying qp. If software need modify some fields in context,
5246 	 * we should set all bits of the relevant fields in context mask to
5247 	 * 0 at the same time, else set them to 0x1.
5248 	 */
5249 	memset(context, 0, hr_dev->caps.qpc_sz);
5250 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5251 
5252 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5253 					 new_state, context, qpc_mask);
5254 	if (ret)
5255 		goto out;
5256 
5257 	/* When QP state is err, SQ and RQ WQE should be flushed */
5258 	if (new_state == IB_QPS_ERR)
5259 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5260 
5261 	/* Configure the optional fields */
5262 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5263 					 qpc_mask);
5264 	if (ret)
5265 		goto out;
5266 
5267 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5268 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5269 			  ibqp->srq);
5270 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5271 
5272 	/* Every status migrate must change state */
5273 	hr_reg_write(context, QPC_QP_ST, new_state);
5274 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5275 
5276 	/* SW pass context to HW */
5277 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5278 	if (ret) {
5279 		ibdev_err(ibdev, "failed to modify QP, ret = %d.\n", ret);
5280 		goto out;
5281 	}
5282 
5283 	hr_qp->state = new_state;
5284 
5285 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5286 
5287 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5288 		clear_qp(hr_qp);
5289 
5290 out:
5291 	return ret;
5292 }
5293 
5294 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5295 {
5296 	static const enum ib_qp_state map[] = {
5297 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5298 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5299 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5300 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5301 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5302 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5303 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5304 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5305 	};
5306 
5307 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5308 }
5309 
5310 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev,
5311 				 struct hns_roce_qp *hr_qp,
5312 				 struct hns_roce_v2_qp_context *hr_context)
5313 {
5314 	struct hns_roce_cmd_mailbox *mailbox;
5315 	int ret;
5316 
5317 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5318 	if (IS_ERR(mailbox))
5319 		return PTR_ERR(mailbox);
5320 
5321 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5322 				hr_qp->qpn);
5323 	if (ret)
5324 		goto out;
5325 
5326 	memcpy(hr_context, mailbox->buf, hr_dev->caps.qpc_sz);
5327 
5328 out:
5329 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5330 	return ret;
5331 }
5332 
5333 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5334 				int qp_attr_mask,
5335 				struct ib_qp_init_attr *qp_init_attr)
5336 {
5337 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5338 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5339 	struct hns_roce_v2_qp_context context = {};
5340 	struct ib_device *ibdev = &hr_dev->ib_dev;
5341 	int tmp_qp_state;
5342 	int state;
5343 	int ret;
5344 
5345 	memset(qp_attr, 0, sizeof(*qp_attr));
5346 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5347 
5348 	mutex_lock(&hr_qp->mutex);
5349 
5350 	if (hr_qp->state == IB_QPS_RESET) {
5351 		qp_attr->qp_state = IB_QPS_RESET;
5352 		ret = 0;
5353 		goto done;
5354 	}
5355 
5356 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp, &context);
5357 	if (ret) {
5358 		ibdev_err(ibdev, "failed to query QPC, ret = %d.\n", ret);
5359 		ret = -EINVAL;
5360 		goto out;
5361 	}
5362 
5363 	state = hr_reg_read(&context, QPC_QP_ST);
5364 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5365 	if (tmp_qp_state == -1) {
5366 		ibdev_err(ibdev, "Illegal ib_qp_state\n");
5367 		ret = -EINVAL;
5368 		goto out;
5369 	}
5370 	hr_qp->state = (u8)tmp_qp_state;
5371 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5372 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5373 	qp_attr->path_mig_state = IB_MIG_ARMED;
5374 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5375 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5376 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5377 
5378 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5379 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5380 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5381 	qp_attr->qp_access_flags =
5382 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5383 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5384 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5385 
5386 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5387 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5388 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5389 		struct ib_global_route *grh =
5390 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5391 
5392 		rdma_ah_set_sl(&qp_attr->ah_attr,
5393 			       hr_reg_read(&context, QPC_SL));
5394 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5395 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5396 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5397 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5398 
5399 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5400 	}
5401 
5402 	qp_attr->port_num = hr_qp->port + 1;
5403 	qp_attr->sq_draining = 0;
5404 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5405 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5406 
5407 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5408 	qp_attr->timeout = (u8)hr_reg_read(&context, QPC_AT);
5409 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5410 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5411 
5412 done:
5413 	qp_attr->cur_qp_state = qp_attr->qp_state;
5414 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5415 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5416 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5417 
5418 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5419 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5420 
5421 	qp_init_attr->qp_context = ibqp->qp_context;
5422 	qp_init_attr->qp_type = ibqp->qp_type;
5423 	qp_init_attr->recv_cq = ibqp->recv_cq;
5424 	qp_init_attr->send_cq = ibqp->send_cq;
5425 	qp_init_attr->srq = ibqp->srq;
5426 	qp_init_attr->cap = qp_attr->cap;
5427 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5428 
5429 out:
5430 	mutex_unlock(&hr_qp->mutex);
5431 	return ret;
5432 }
5433 
5434 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5435 {
5436 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5437 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5438 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5439 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5440 		hr_qp->state != IB_QPS_RESET);
5441 }
5442 
5443 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5444 					 struct hns_roce_qp *hr_qp,
5445 					 struct ib_udata *udata)
5446 {
5447 	struct ib_device *ibdev = &hr_dev->ib_dev;
5448 	struct hns_roce_cq *send_cq, *recv_cq;
5449 	unsigned long flags;
5450 	int ret = 0;
5451 
5452 	if (modify_qp_is_ok(hr_qp)) {
5453 		/* Modify qp to reset before destroying qp */
5454 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5455 					    hr_qp->state, IB_QPS_RESET);
5456 		if (ret)
5457 			ibdev_err(ibdev,
5458 				  "failed to modify QP to RST, ret = %d.\n",
5459 				  ret);
5460 	}
5461 
5462 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5463 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5464 
5465 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5466 	hns_roce_lock_cqs(send_cq, recv_cq);
5467 
5468 	if (!udata) {
5469 		if (recv_cq)
5470 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5471 					       (hr_qp->ibqp.srq ?
5472 						to_hr_srq(hr_qp->ibqp.srq) :
5473 						NULL));
5474 
5475 		if (send_cq && send_cq != recv_cq)
5476 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5477 	}
5478 
5479 	hns_roce_qp_remove(hr_dev, hr_qp);
5480 
5481 	hns_roce_unlock_cqs(send_cq, recv_cq);
5482 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5483 
5484 	return ret;
5485 }
5486 
5487 static int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5488 {
5489 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5490 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5491 	int ret;
5492 
5493 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5494 	if (ret)
5495 		ibdev_err(&hr_dev->ib_dev,
5496 			  "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5497 			  hr_qp->qpn, ret);
5498 
5499 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5500 
5501 	return 0;
5502 }
5503 
5504 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5505 					    struct hns_roce_qp *hr_qp)
5506 {
5507 	struct ib_device *ibdev = &hr_dev->ib_dev;
5508 	struct hns_roce_sccc_clr_done *resp;
5509 	struct hns_roce_sccc_clr *clr;
5510 	struct hns_roce_cmq_desc desc;
5511 	int ret, i;
5512 
5513 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5514 		return 0;
5515 
5516 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5517 
5518 	/* set scc ctx clear done flag */
5519 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5520 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5521 	if (ret) {
5522 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5523 		goto out;
5524 	}
5525 
5526 	/* clear scc context */
5527 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5528 	clr = (struct hns_roce_sccc_clr *)desc.data;
5529 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5530 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5531 	if (ret) {
5532 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5533 		goto out;
5534 	}
5535 
5536 	/* query scc context clear is done or not */
5537 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5538 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5539 		hns_roce_cmq_setup_basic_desc(&desc,
5540 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5541 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5542 		if (ret) {
5543 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5544 				  ret);
5545 			goto out;
5546 		}
5547 
5548 		if (resp->clr_done)
5549 			goto out;
5550 
5551 		msleep(20);
5552 	}
5553 
5554 	ibdev_err(ibdev, "Query SCC clr done flag overtime.\n");
5555 	ret = -ETIMEDOUT;
5556 
5557 out:
5558 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5559 	return ret;
5560 }
5561 
5562 #define DMA_IDX_SHIFT 3
5563 #define DMA_WQE_SHIFT 3
5564 
5565 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5566 					      struct hns_roce_srq_context *ctx)
5567 {
5568 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5569 	struct ib_device *ibdev = srq->ibsrq.device;
5570 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5571 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5572 	dma_addr_t dma_handle_idx = 0;
5573 	int ret;
5574 
5575 	/* Get physical address of idx que buf */
5576 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5577 				ARRAY_SIZE(mtts_idx), &dma_handle_idx);
5578 	if (ret < 1) {
5579 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5580 			  ret);
5581 		return -ENOBUFS;
5582 	}
5583 
5584 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5585 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5586 
5587 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5588 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5589 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5590 
5591 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5592 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5593 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5594 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5595 
5596 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5597 		     to_hr_hw_page_addr(mtts_idx[0]));
5598 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5599 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5600 
5601 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5602 		     to_hr_hw_page_addr(mtts_idx[1]));
5603 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5604 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5605 
5606 	return 0;
5607 }
5608 
5609 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5610 {
5611 	struct ib_device *ibdev = srq->ibsrq.device;
5612 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5613 	struct hns_roce_srq_context *ctx = mb_buf;
5614 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5615 	dma_addr_t dma_handle_wqe = 0;
5616 	int ret;
5617 
5618 	memset(ctx, 0, sizeof(*ctx));
5619 
5620 	/* Get the physical address of srq buf */
5621 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5622 				ARRAY_SIZE(mtts_wqe), &dma_handle_wqe);
5623 	if (ret < 1) {
5624 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5625 			  ret);
5626 		return -ENOBUFS;
5627 	}
5628 
5629 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5630 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5631 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5632 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5633 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5634 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5635 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5636 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5637 	hr_reg_write(ctx, SRQC_RQWS,
5638 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5639 
5640 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5641 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5642 				      srq->wqe_cnt));
5643 
5644 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5645 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5646 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5647 
5648 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5649 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5650 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5651 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5652 
5653 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5654 }
5655 
5656 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5657 				  struct ib_srq_attr *srq_attr,
5658 				  enum ib_srq_attr_mask srq_attr_mask,
5659 				  struct ib_udata *udata)
5660 {
5661 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5662 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5663 	struct hns_roce_srq_context *srq_context;
5664 	struct hns_roce_srq_context *srqc_mask;
5665 	struct hns_roce_cmd_mailbox *mailbox;
5666 	int ret;
5667 
5668 	/* Resizing SRQs is not supported yet */
5669 	if (srq_attr_mask & IB_SRQ_MAX_WR)
5670 		return -EINVAL;
5671 
5672 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5673 		if (srq_attr->srq_limit > srq->wqe_cnt)
5674 			return -EINVAL;
5675 
5676 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5677 		if (IS_ERR(mailbox))
5678 			return PTR_ERR(mailbox);
5679 
5680 		srq_context = mailbox->buf;
5681 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5682 
5683 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5684 
5685 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5686 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5687 
5688 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5689 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5690 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5691 		if (ret) {
5692 			ibdev_err(&hr_dev->ib_dev,
5693 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5694 				  ret);
5695 			return ret;
5696 		}
5697 	}
5698 
5699 	return 0;
5700 }
5701 
5702 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5703 {
5704 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5705 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5706 	struct hns_roce_srq_context *srq_context;
5707 	struct hns_roce_cmd_mailbox *mailbox;
5708 	int ret;
5709 
5710 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5711 	if (IS_ERR(mailbox))
5712 		return PTR_ERR(mailbox);
5713 
5714 	srq_context = mailbox->buf;
5715 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5716 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5717 	if (ret) {
5718 		ibdev_err(&hr_dev->ib_dev,
5719 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5720 			  ret);
5721 		goto out;
5722 	}
5723 
5724 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5725 	attr->max_wr = srq->wqe_cnt;
5726 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5727 
5728 out:
5729 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5730 	return ret;
5731 }
5732 
5733 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5734 {
5735 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5736 	struct hns_roce_v2_cq_context *cq_context;
5737 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5738 	struct hns_roce_v2_cq_context *cqc_mask;
5739 	struct hns_roce_cmd_mailbox *mailbox;
5740 	int ret;
5741 
5742 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5743 	if (IS_ERR(mailbox))
5744 		return PTR_ERR(mailbox);
5745 
5746 	cq_context = mailbox->buf;
5747 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5748 
5749 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5750 
5751 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5752 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5753 
5754 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5755 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5756 			dev_info(hr_dev->dev,
5757 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5758 				 cq_period);
5759 			cq_period = HNS_ROCE_MAX_CQ_PERIOD;
5760 		}
5761 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5762 	}
5763 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5764 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5765 
5766 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5767 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5768 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5769 	if (ret)
5770 		ibdev_err(&hr_dev->ib_dev,
5771 			  "failed to process cmd when modifying CQ, ret = %d.\n",
5772 			  ret);
5773 
5774 	return ret;
5775 }
5776 
5777 static void hns_roce_irq_work_handle(struct work_struct *work)
5778 {
5779 	struct hns_roce_work *irq_work =
5780 				container_of(work, struct hns_roce_work, work);
5781 	struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
5782 
5783 	switch (irq_work->event_type) {
5784 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5785 		ibdev_info(ibdev, "Path migrated succeeded.\n");
5786 		break;
5787 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5788 		ibdev_warn(ibdev, "Path migration failed.\n");
5789 		break;
5790 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
5791 		break;
5792 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5793 		ibdev_warn(ibdev, "Send queue drained.\n");
5794 		break;
5795 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5796 		ibdev_err(ibdev, "Local work queue 0x%x catast error, sub_event type is: %d\n",
5797 			  irq_work->queue_num, irq_work->sub_type);
5798 		break;
5799 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5800 		ibdev_err(ibdev, "Invalid request local work queue 0x%x error.\n",
5801 			  irq_work->queue_num);
5802 		break;
5803 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5804 		ibdev_err(ibdev, "Local access violation work queue 0x%x error, sub_event type is: %d\n",
5805 			  irq_work->queue_num, irq_work->sub_type);
5806 		break;
5807 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5808 		ibdev_warn(ibdev, "SRQ limit reach.\n");
5809 		break;
5810 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5811 		ibdev_warn(ibdev, "SRQ last wqe reach.\n");
5812 		break;
5813 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
5814 		ibdev_err(ibdev, "SRQ catas error.\n");
5815 		break;
5816 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
5817 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
5818 		break;
5819 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
5820 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
5821 		break;
5822 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
5823 		ibdev_warn(ibdev, "DB overflow.\n");
5824 		break;
5825 	case HNS_ROCE_EVENT_TYPE_FLR:
5826 		ibdev_warn(ibdev, "Function level reset.\n");
5827 		break;
5828 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
5829 		ibdev_err(ibdev, "xrc domain violation error.\n");
5830 		break;
5831 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
5832 		ibdev_err(ibdev, "invalid xrceth error.\n");
5833 		break;
5834 	default:
5835 		break;
5836 	}
5837 
5838 	kfree(irq_work);
5839 }
5840 
5841 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
5842 				      struct hns_roce_eq *eq, u32 queue_num)
5843 {
5844 	struct hns_roce_work *irq_work;
5845 
5846 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
5847 	if (!irq_work)
5848 		return;
5849 
5850 	INIT_WORK(&(irq_work->work), hns_roce_irq_work_handle);
5851 	irq_work->hr_dev = hr_dev;
5852 	irq_work->event_type = eq->event_type;
5853 	irq_work->sub_type = eq->sub_type;
5854 	irq_work->queue_num = queue_num;
5855 	queue_work(hr_dev->irq_workq, &(irq_work->work));
5856 }
5857 
5858 static void update_eq_db(struct hns_roce_eq *eq)
5859 {
5860 	struct hns_roce_dev *hr_dev = eq->hr_dev;
5861 	struct hns_roce_v2_db eq_db = {};
5862 
5863 	if (eq->type_flag == HNS_ROCE_AEQ) {
5864 		hr_reg_write(&eq_db, EQ_DB_CMD,
5865 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
5866 			     HNS_ROCE_EQ_DB_CMD_AEQ :
5867 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
5868 	} else {
5869 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
5870 
5871 		hr_reg_write(&eq_db, EQ_DB_CMD,
5872 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
5873 			     HNS_ROCE_EQ_DB_CMD_CEQ :
5874 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
5875 	}
5876 
5877 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
5878 
5879 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
5880 }
5881 
5882 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
5883 {
5884 	struct hns_roce_aeqe *aeqe;
5885 
5886 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
5887 				   (eq->cons_index & (eq->entries - 1)) *
5888 				   eq->eqe_size);
5889 
5890 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
5891 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
5892 }
5893 
5894 static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
5895 				       struct hns_roce_eq *eq)
5896 {
5897 	struct device *dev = hr_dev->dev;
5898 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
5899 	irqreturn_t aeqe_found = IRQ_NONE;
5900 	int event_type;
5901 	u32 queue_num;
5902 	int sub_type;
5903 
5904 	while (aeqe) {
5905 		/* Make sure we read AEQ entry after we have checked the
5906 		 * ownership bit
5907 		 */
5908 		dma_rmb();
5909 
5910 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
5911 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
5912 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
5913 
5914 		switch (event_type) {
5915 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5916 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5917 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
5918 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5919 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5920 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5921 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5922 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5923 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
5924 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
5925 			hns_roce_qp_event(hr_dev, queue_num, event_type);
5926 			break;
5927 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5928 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
5929 			hns_roce_srq_event(hr_dev, queue_num, event_type);
5930 			break;
5931 		case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
5932 		case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
5933 			hns_roce_cq_event(hr_dev, queue_num, event_type);
5934 			break;
5935 		case HNS_ROCE_EVENT_TYPE_MB:
5936 			hns_roce_cmd_event(hr_dev,
5937 					le16_to_cpu(aeqe->event.cmd.token),
5938 					aeqe->event.cmd.status,
5939 					le64_to_cpu(aeqe->event.cmd.out_param));
5940 			break;
5941 		case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
5942 		case HNS_ROCE_EVENT_TYPE_FLR:
5943 			break;
5944 		default:
5945 			dev_err(dev, "Unhandled event %d on EQ %d at idx %u.\n",
5946 				event_type, eq->eqn, eq->cons_index);
5947 			break;
5948 		}
5949 
5950 		eq->event_type = event_type;
5951 		eq->sub_type = sub_type;
5952 		++eq->cons_index;
5953 		aeqe_found = IRQ_HANDLED;
5954 
5955 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
5956 
5957 		aeqe = next_aeqe_sw_v2(eq);
5958 	}
5959 
5960 	update_eq_db(eq);
5961 
5962 	return IRQ_RETVAL(aeqe_found);
5963 }
5964 
5965 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
5966 {
5967 	struct hns_roce_ceqe *ceqe;
5968 
5969 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
5970 				   (eq->cons_index & (eq->entries - 1)) *
5971 				   eq->eqe_size);
5972 
5973 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
5974 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
5975 }
5976 
5977 static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_dev *hr_dev,
5978 				       struct hns_roce_eq *eq)
5979 {
5980 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
5981 	irqreturn_t ceqe_found = IRQ_NONE;
5982 	u32 cqn;
5983 
5984 	while (ceqe) {
5985 		/* Make sure we read CEQ entry after we have checked the
5986 		 * ownership bit
5987 		 */
5988 		dma_rmb();
5989 
5990 		cqn = hr_reg_read(ceqe, CEQE_CQN);
5991 
5992 		hns_roce_cq_completion(hr_dev, cqn);
5993 
5994 		++eq->cons_index;
5995 		ceqe_found = IRQ_HANDLED;
5996 
5997 		ceqe = next_ceqe_sw_v2(eq);
5998 	}
5999 
6000 	update_eq_db(eq);
6001 
6002 	return IRQ_RETVAL(ceqe_found);
6003 }
6004 
6005 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
6006 {
6007 	struct hns_roce_eq *eq = eq_ptr;
6008 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6009 	irqreturn_t int_work;
6010 
6011 	if (eq->type_flag == HNS_ROCE_CEQ)
6012 		/* Completion event interrupt */
6013 		int_work = hns_roce_v2_ceq_int(hr_dev, eq);
6014 	else
6015 		/* Asychronous event interrupt */
6016 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
6017 
6018 	return IRQ_RETVAL(int_work);
6019 }
6020 
6021 static irqreturn_t abnormal_interrupt_basic(struct hns_roce_dev *hr_dev,
6022 					    u32 int_st)
6023 {
6024 	struct pci_dev *pdev = hr_dev->pci_dev;
6025 	struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
6026 	const struct hnae3_ae_ops *ops = ae_dev->ops;
6027 	irqreturn_t int_work = IRQ_NONE;
6028 	u32 int_en;
6029 
6030 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
6031 
6032 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
6033 		dev_err(hr_dev->dev, "AEQ overflow!\n");
6034 
6035 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG,
6036 			   1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S);
6037 
6038 		/* Set reset level for reset_event() */
6039 		if (ops->set_default_reset_request)
6040 			ops->set_default_reset_request(ae_dev,
6041 						       HNAE3_FUNC_RESET);
6042 		if (ops->reset_event)
6043 			ops->reset_event(pdev, NULL);
6044 
6045 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6046 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6047 
6048 		int_work = IRQ_HANDLED;
6049 	} else {
6050 		dev_err(hr_dev->dev, "there is no basic abn irq found.\n");
6051 	}
6052 
6053 	return IRQ_RETVAL(int_work);
6054 }
6055 
6056 static int fmea_ram_ecc_query(struct hns_roce_dev *hr_dev,
6057 			       struct fmea_ram_ecc *ecc_info)
6058 {
6059 	struct hns_roce_cmq_desc desc;
6060 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6061 	int ret;
6062 
6063 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_QUERY_RAM_ECC, true);
6064 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6065 	if (ret)
6066 		return ret;
6067 
6068 	ecc_info->is_ecc_err = hr_reg_read(req, QUERY_RAM_ECC_1BIT_ERR);
6069 	ecc_info->res_type = hr_reg_read(req, QUERY_RAM_ECC_RES_TYPE);
6070 	ecc_info->index = hr_reg_read(req, QUERY_RAM_ECC_TAG);
6071 
6072 	return 0;
6073 }
6074 
6075 static int fmea_recover_gmv(struct hns_roce_dev *hr_dev, u32 idx)
6076 {
6077 	struct hns_roce_cmq_desc desc;
6078 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6079 	u32 addr_upper;
6080 	u32 addr_low;
6081 	int ret;
6082 
6083 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, true);
6084 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6085 
6086 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6087 	if (ret) {
6088 		dev_err(hr_dev->dev,
6089 			"failed to execute cmd to read gmv, ret = %d.\n", ret);
6090 		return ret;
6091 	}
6092 
6093 	addr_low =  hr_reg_read(req, CFG_GMV_BT_BA_L);
6094 	addr_upper = hr_reg_read(req, CFG_GMV_BT_BA_H);
6095 
6096 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
6097 	hr_reg_write(req, CFG_GMV_BT_BA_L, addr_low);
6098 	hr_reg_write(req, CFG_GMV_BT_BA_H, addr_upper);
6099 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6100 
6101 	return hns_roce_cmq_send(hr_dev, &desc, 1);
6102 }
6103 
6104 static u64 fmea_get_ram_res_addr(u32 res_type, __le64 *data)
6105 {
6106 	if (res_type == ECC_RESOURCE_QPC_TIMER ||
6107 	    res_type == ECC_RESOURCE_CQC_TIMER ||
6108 	    res_type == ECC_RESOURCE_SCCC)
6109 		return le64_to_cpu(*data);
6110 
6111 	return le64_to_cpu(*data) << PAGE_SHIFT;
6112 }
6113 
6114 static int fmea_recover_others(struct hns_roce_dev *hr_dev, u32 res_type,
6115 			       u32 index)
6116 {
6117 	u8 write_bt0_op = fmea_ram_res[res_type].write_bt0_op;
6118 	u8 read_bt0_op = fmea_ram_res[res_type].read_bt0_op;
6119 	struct hns_roce_cmd_mailbox *mailbox;
6120 	u64 addr;
6121 	int ret;
6122 
6123 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6124 	if (IS_ERR(mailbox))
6125 		return PTR_ERR(mailbox);
6126 
6127 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, read_bt0_op, index);
6128 	if (ret) {
6129 		dev_err(hr_dev->dev,
6130 			"failed to execute cmd to read fmea ram, ret = %d.\n",
6131 			ret);
6132 		goto out;
6133 	}
6134 
6135 	addr = fmea_get_ram_res_addr(res_type, mailbox->buf);
6136 
6137 	ret = hns_roce_cmd_mbox(hr_dev, addr, 0, write_bt0_op, index);
6138 	if (ret)
6139 		dev_err(hr_dev->dev,
6140 			"failed to execute cmd to write fmea ram, ret = %d.\n",
6141 			ret);
6142 
6143 out:
6144 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6145 	return ret;
6146 }
6147 
6148 static void fmea_ram_ecc_recover(struct hns_roce_dev *hr_dev,
6149 				 struct fmea_ram_ecc *ecc_info)
6150 {
6151 	u32 res_type = ecc_info->res_type;
6152 	u32 index = ecc_info->index;
6153 	int ret;
6154 
6155 	BUILD_BUG_ON(ARRAY_SIZE(fmea_ram_res) != ECC_RESOURCE_COUNT);
6156 
6157 	if (res_type >= ECC_RESOURCE_COUNT) {
6158 		dev_err(hr_dev->dev, "unsupported fmea ram ecc type %u.\n",
6159 			res_type);
6160 		return;
6161 	}
6162 
6163 	if (res_type == ECC_RESOURCE_GMV)
6164 		ret = fmea_recover_gmv(hr_dev, index);
6165 	else
6166 		ret = fmea_recover_others(hr_dev, res_type, index);
6167 	if (ret)
6168 		dev_err(hr_dev->dev,
6169 			"failed to recover %s, index = %u, ret = %d.\n",
6170 			fmea_ram_res[res_type].name, index, ret);
6171 }
6172 
6173 static void fmea_ram_ecc_work(struct work_struct *ecc_work)
6174 {
6175 	struct hns_roce_dev *hr_dev =
6176 		container_of(ecc_work, struct hns_roce_dev, ecc_work);
6177 	struct fmea_ram_ecc ecc_info = {};
6178 
6179 	if (fmea_ram_ecc_query(hr_dev, &ecc_info)) {
6180 		dev_err(hr_dev->dev, "failed to query fmea ram ecc.\n");
6181 		return;
6182 	}
6183 
6184 	if (!ecc_info.is_ecc_err) {
6185 		dev_err(hr_dev->dev, "there is no fmea ram ecc err found.\n");
6186 		return;
6187 	}
6188 
6189 	fmea_ram_ecc_recover(hr_dev, &ecc_info);
6190 }
6191 
6192 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
6193 {
6194 	struct hns_roce_dev *hr_dev = dev_id;
6195 	irqreturn_t int_work = IRQ_NONE;
6196 	u32 int_st;
6197 
6198 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
6199 
6200 	if (int_st) {
6201 		int_work = abnormal_interrupt_basic(hr_dev, int_st);
6202 	} else if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
6203 		queue_work(hr_dev->irq_workq, &hr_dev->ecc_work);
6204 		int_work = IRQ_HANDLED;
6205 	} else {
6206 		dev_err(hr_dev->dev, "there is no abnormal irq found.\n");
6207 	}
6208 
6209 	return IRQ_RETVAL(int_work);
6210 }
6211 
6212 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6213 					int eq_num, u32 enable_flag)
6214 {
6215 	int i;
6216 
6217 	for (i = 0; i < eq_num; i++)
6218 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6219 			   i * EQ_REG_OFFSET, enable_flag);
6220 
6221 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6222 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6223 }
6224 
6225 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev, u32 eqn)
6226 {
6227 	struct device *dev = hr_dev->dev;
6228 	int ret;
6229 	u8 cmd;
6230 
6231 	if (eqn < hr_dev->caps.num_comp_vectors)
6232 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6233 	else
6234 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6235 
6236 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6237 	if (ret)
6238 		dev_err(dev, "[mailbox cmd] destroy eqc(%u) failed.\n", eqn);
6239 }
6240 
6241 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6242 {
6243 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6244 }
6245 
6246 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6247 {
6248 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6249 	eq->cons_index = 0;
6250 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6251 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6252 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6253 	eq->shift = ilog2((unsigned int)eq->entries);
6254 }
6255 
6256 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6257 		      void *mb_buf)
6258 {
6259 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6260 	struct hns_roce_eq_context *eqc;
6261 	u64 bt_ba = 0;
6262 	int count;
6263 
6264 	eqc = mb_buf;
6265 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6266 
6267 	init_eq_config(hr_dev, eq);
6268 
6269 	/* if not multi-hop, eqe buffer only use one trunk */
6270 	count = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba, MTT_MIN_COUNT,
6271 				  &bt_ba);
6272 	if (count < 1) {
6273 		dev_err(hr_dev->dev, "failed to find EQE mtr\n");
6274 		return -ENOBUFS;
6275 	}
6276 
6277 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6278 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6279 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6280 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6281 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6282 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6283 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6284 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6285 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6286 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6287 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6288 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6289 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6290 
6291 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6292 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6293 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6294 				 eq->eq_period);
6295 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6296 		}
6297 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6298 	}
6299 
6300 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6301 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6302 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6303 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6304 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6305 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6306 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6307 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6308 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6309 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6310 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6311 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6312 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6313 
6314 	return 0;
6315 }
6316 
6317 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6318 {
6319 	struct hns_roce_buf_attr buf_attr = {};
6320 	int err;
6321 
6322 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6323 		eq->hop_num = 0;
6324 	else
6325 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6326 
6327 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6328 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6329 	buf_attr.region[0].hopnum = eq->hop_num;
6330 	buf_attr.region_count = 1;
6331 
6332 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6333 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6334 				  0);
6335 	if (err)
6336 		dev_err(hr_dev->dev, "Failed to alloc EQE mtr, err %d\n", err);
6337 
6338 	return err;
6339 }
6340 
6341 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6342 				 struct hns_roce_eq *eq, u8 eq_cmd)
6343 {
6344 	struct hns_roce_cmd_mailbox *mailbox;
6345 	int ret;
6346 
6347 	/* Allocate mailbox memory */
6348 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6349 	if (IS_ERR(mailbox))
6350 		return PTR_ERR(mailbox);
6351 
6352 	ret = alloc_eq_buf(hr_dev, eq);
6353 	if (ret)
6354 		goto free_cmd_mbox;
6355 
6356 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6357 	if (ret)
6358 		goto err_cmd_mbox;
6359 
6360 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6361 	if (ret) {
6362 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6363 		goto err_cmd_mbox;
6364 	}
6365 
6366 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6367 
6368 	return 0;
6369 
6370 err_cmd_mbox:
6371 	free_eq_buf(hr_dev, eq);
6372 
6373 free_cmd_mbox:
6374 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6375 
6376 	return ret;
6377 }
6378 
6379 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6380 				  int comp_num, int aeq_num, int other_num)
6381 {
6382 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6383 	int i, j;
6384 	int ret;
6385 
6386 	for (i = 0; i < irq_num; i++) {
6387 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6388 					       GFP_KERNEL);
6389 		if (!hr_dev->irq_names[i]) {
6390 			ret = -ENOMEM;
6391 			goto err_kzalloc_failed;
6392 		}
6393 	}
6394 
6395 	/* irq contains: abnormal + AEQ + CEQ */
6396 	for (j = 0; j < other_num; j++)
6397 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6398 			 "hns-abn-%d", j);
6399 
6400 	for (j = other_num; j < (other_num + aeq_num); j++)
6401 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6402 			 "hns-aeq-%d", j - other_num);
6403 
6404 	for (j = (other_num + aeq_num); j < irq_num; j++)
6405 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6406 			 "hns-ceq-%d", j - other_num - aeq_num);
6407 
6408 	for (j = 0; j < irq_num; j++) {
6409 		if (j < other_num)
6410 			ret = request_irq(hr_dev->irq[j],
6411 					  hns_roce_v2_msix_interrupt_abn,
6412 					  0, hr_dev->irq_names[j], hr_dev);
6413 
6414 		else if (j < (other_num + comp_num))
6415 			ret = request_irq(eq_table->eq[j - other_num].irq,
6416 					  hns_roce_v2_msix_interrupt_eq,
6417 					  0, hr_dev->irq_names[j + aeq_num],
6418 					  &eq_table->eq[j - other_num]);
6419 		else
6420 			ret = request_irq(eq_table->eq[j - other_num].irq,
6421 					  hns_roce_v2_msix_interrupt_eq,
6422 					  0, hr_dev->irq_names[j - comp_num],
6423 					  &eq_table->eq[j - other_num]);
6424 		if (ret) {
6425 			dev_err(hr_dev->dev, "Request irq error!\n");
6426 			goto err_request_failed;
6427 		}
6428 	}
6429 
6430 	return 0;
6431 
6432 err_request_failed:
6433 	for (j -= 1; j >= 0; j--)
6434 		if (j < other_num)
6435 			free_irq(hr_dev->irq[j], hr_dev);
6436 		else
6437 			free_irq(eq_table->eq[j - other_num].irq,
6438 				 &eq_table->eq[j - other_num]);
6439 
6440 err_kzalloc_failed:
6441 	for (i -= 1; i >= 0; i--)
6442 		kfree(hr_dev->irq_names[i]);
6443 
6444 	return ret;
6445 }
6446 
6447 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6448 {
6449 	int irq_num;
6450 	int eq_num;
6451 	int i;
6452 
6453 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6454 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6455 
6456 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6457 		free_irq(hr_dev->irq[i], hr_dev);
6458 
6459 	for (i = 0; i < eq_num; i++)
6460 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6461 
6462 	for (i = 0; i < irq_num; i++)
6463 		kfree(hr_dev->irq_names[i]);
6464 }
6465 
6466 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6467 {
6468 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6469 	struct device *dev = hr_dev->dev;
6470 	struct hns_roce_eq *eq;
6471 	int other_num;
6472 	int comp_num;
6473 	int aeq_num;
6474 	int irq_num;
6475 	int eq_num;
6476 	u8 eq_cmd;
6477 	int ret;
6478 	int i;
6479 
6480 	other_num = hr_dev->caps.num_other_vectors;
6481 	comp_num = hr_dev->caps.num_comp_vectors;
6482 	aeq_num = hr_dev->caps.num_aeq_vectors;
6483 
6484 	eq_num = comp_num + aeq_num;
6485 	irq_num = eq_num + other_num;
6486 
6487 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6488 	if (!eq_table->eq)
6489 		return -ENOMEM;
6490 
6491 	/* create eq */
6492 	for (i = 0; i < eq_num; i++) {
6493 		eq = &eq_table->eq[i];
6494 		eq->hr_dev = hr_dev;
6495 		eq->eqn = i;
6496 		if (i < comp_num) {
6497 			/* CEQ */
6498 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6499 			eq->type_flag = HNS_ROCE_CEQ;
6500 			eq->entries = hr_dev->caps.ceqe_depth;
6501 			eq->eqe_size = hr_dev->caps.ceqe_size;
6502 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6503 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6504 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6505 		} else {
6506 			/* AEQ */
6507 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6508 			eq->type_flag = HNS_ROCE_AEQ;
6509 			eq->entries = hr_dev->caps.aeqe_depth;
6510 			eq->eqe_size = hr_dev->caps.aeqe_size;
6511 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6512 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6513 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6514 		}
6515 
6516 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6517 		if (ret) {
6518 			dev_err(dev, "failed to create eq.\n");
6519 			goto err_create_eq_fail;
6520 		}
6521 	}
6522 
6523 	INIT_WORK(&hr_dev->ecc_work, fmea_ram_ecc_work);
6524 
6525 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6526 	if (!hr_dev->irq_workq) {
6527 		dev_err(dev, "failed to create irq workqueue.\n");
6528 		ret = -ENOMEM;
6529 		goto err_create_eq_fail;
6530 	}
6531 
6532 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6533 				     other_num);
6534 	if (ret) {
6535 		dev_err(dev, "failed to request irq.\n");
6536 		goto err_request_irq_fail;
6537 	}
6538 
6539 	/* enable irq */
6540 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6541 
6542 	return 0;
6543 
6544 err_request_irq_fail:
6545 	destroy_workqueue(hr_dev->irq_workq);
6546 
6547 err_create_eq_fail:
6548 	for (i -= 1; i >= 0; i--)
6549 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6550 	kfree(eq_table->eq);
6551 
6552 	return ret;
6553 }
6554 
6555 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6556 {
6557 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6558 	int eq_num;
6559 	int i;
6560 
6561 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6562 
6563 	/* Disable irq */
6564 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6565 
6566 	__hns_roce_free_irq(hr_dev);
6567 	destroy_workqueue(hr_dev->irq_workq);
6568 
6569 	for (i = 0; i < eq_num; i++) {
6570 		hns_roce_v2_destroy_eqc(hr_dev, i);
6571 
6572 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6573 	}
6574 
6575 	kfree(eq_table->eq);
6576 }
6577 
6578 static const struct hns_roce_dfx_hw hns_roce_dfx_hw_v2 = {
6579 	.query_cqc_info = hns_roce_v2_query_cqc_info,
6580 };
6581 
6582 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6583 	.destroy_qp = hns_roce_v2_destroy_qp,
6584 	.modify_cq = hns_roce_v2_modify_cq,
6585 	.poll_cq = hns_roce_v2_poll_cq,
6586 	.post_recv = hns_roce_v2_post_recv,
6587 	.post_send = hns_roce_v2_post_send,
6588 	.query_qp = hns_roce_v2_query_qp,
6589 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6590 };
6591 
6592 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6593 	.modify_srq = hns_roce_v2_modify_srq,
6594 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6595 	.query_srq = hns_roce_v2_query_srq,
6596 };
6597 
6598 static const struct hns_roce_hw hns_roce_hw_v2 = {
6599 	.cmq_init = hns_roce_v2_cmq_init,
6600 	.cmq_exit = hns_roce_v2_cmq_exit,
6601 	.hw_profile = hns_roce_v2_profile,
6602 	.hw_init = hns_roce_v2_init,
6603 	.hw_exit = hns_roce_v2_exit,
6604 	.post_mbox = v2_post_mbox,
6605 	.poll_mbox_done = v2_poll_mbox_done,
6606 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6607 	.set_gid = hns_roce_v2_set_gid,
6608 	.set_mac = hns_roce_v2_set_mac,
6609 	.write_mtpt = hns_roce_v2_write_mtpt,
6610 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6611 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6612 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6613 	.write_cqc = hns_roce_v2_write_cqc,
6614 	.set_hem = hns_roce_v2_set_hem,
6615 	.clear_hem = hns_roce_v2_clear_hem,
6616 	.modify_qp = hns_roce_v2_modify_qp,
6617 	.dereg_mr = hns_roce_v2_dereg_mr,
6618 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6619 	.init_eq = hns_roce_v2_init_eq_table,
6620 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6621 	.write_srqc = hns_roce_v2_write_srqc,
6622 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6623 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6624 };
6625 
6626 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6627 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6628 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6629 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6630 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6631 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6632 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6633 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6634 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6635 	/* required last entry */
6636 	{0, }
6637 };
6638 
6639 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6640 
6641 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6642 				  struct hnae3_handle *handle)
6643 {
6644 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6645 	const struct pci_device_id *id;
6646 	int i;
6647 
6648 	hr_dev->pci_dev = handle->pdev;
6649 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6650 	hr_dev->is_vf = id->driver_data;
6651 	hr_dev->dev = &handle->pdev->dev;
6652 	hr_dev->hw = &hns_roce_hw_v2;
6653 	hr_dev->dfx = &hns_roce_dfx_hw_v2;
6654 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6655 	hr_dev->odb_offset = hr_dev->sdb_offset;
6656 
6657 	/* Get info from NIC driver. */
6658 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6659 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6660 	hr_dev->caps.num_ports = 1;
6661 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6662 	hr_dev->iboe.phy_port[0] = 0;
6663 
6664 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6665 			    hr_dev->iboe.netdevs[0]->dev_addr);
6666 
6667 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6668 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6669 						i + handle->rinfo.base_vector);
6670 
6671 	/* cmd issue mode: 0 is poll, 1 is event */
6672 	hr_dev->cmd_mod = 1;
6673 	hr_dev->loop_idc = 0;
6674 
6675 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6676 	priv->handle = handle;
6677 }
6678 
6679 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6680 {
6681 	struct hns_roce_dev *hr_dev;
6682 	int ret;
6683 
6684 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6685 	if (!hr_dev)
6686 		return -ENOMEM;
6687 
6688 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6689 	if (!hr_dev->priv) {
6690 		ret = -ENOMEM;
6691 		goto error_failed_kzalloc;
6692 	}
6693 
6694 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6695 
6696 	ret = hns_roce_init(hr_dev);
6697 	if (ret) {
6698 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6699 		goto error_failed_cfg;
6700 	}
6701 
6702 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6703 		ret = free_mr_init(hr_dev);
6704 		if (ret) {
6705 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6706 			goto error_failed_roce_init;
6707 		}
6708 	}
6709 
6710 	handle->priv = hr_dev;
6711 
6712 	return 0;
6713 
6714 error_failed_roce_init:
6715 	hns_roce_exit(hr_dev);
6716 
6717 error_failed_cfg:
6718 	kfree(hr_dev->priv);
6719 
6720 error_failed_kzalloc:
6721 	ib_dealloc_device(&hr_dev->ib_dev);
6722 
6723 	return ret;
6724 }
6725 
6726 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6727 					   bool reset)
6728 {
6729 	struct hns_roce_dev *hr_dev = handle->priv;
6730 
6731 	if (!hr_dev)
6732 		return;
6733 
6734 	handle->priv = NULL;
6735 
6736 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
6737 	hns_roce_handle_device_err(hr_dev);
6738 
6739 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
6740 		free_mr_exit(hr_dev);
6741 
6742 	hns_roce_exit(hr_dev);
6743 	kfree(hr_dev->priv);
6744 	ib_dealloc_device(&hr_dev->ib_dev);
6745 }
6746 
6747 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6748 {
6749 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
6750 	const struct pci_device_id *id;
6751 	struct device *dev = &handle->pdev->dev;
6752 	int ret;
6753 
6754 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
6755 
6756 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
6757 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6758 		goto reset_chk_err;
6759 	}
6760 
6761 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
6762 	if (!id)
6763 		return 0;
6764 
6765 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
6766 		return 0;
6767 
6768 	ret = __hns_roce_hw_v2_init_instance(handle);
6769 	if (ret) {
6770 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6771 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
6772 		if (ops->ae_dev_resetting(handle) ||
6773 		    ops->get_hw_reset_stat(handle))
6774 			goto reset_chk_err;
6775 		else
6776 			return ret;
6777 	}
6778 
6779 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
6780 
6781 	return 0;
6782 
6783 reset_chk_err:
6784 	dev_err(dev, "Device is busy in resetting state.\n"
6785 		     "please retry later.\n");
6786 
6787 	return -EBUSY;
6788 }
6789 
6790 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6791 					   bool reset)
6792 {
6793 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
6794 		return;
6795 
6796 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
6797 
6798 	__hns_roce_hw_v2_uninit_instance(handle, reset);
6799 
6800 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6801 }
6802 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
6803 {
6804 	struct hns_roce_dev *hr_dev;
6805 
6806 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
6807 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6808 		return 0;
6809 	}
6810 
6811 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
6812 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6813 
6814 	hr_dev = handle->priv;
6815 	if (!hr_dev)
6816 		return 0;
6817 
6818 	hr_dev->active = false;
6819 	hr_dev->dis_db = true;
6820 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
6821 
6822 	return 0;
6823 }
6824 
6825 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
6826 {
6827 	struct device *dev = &handle->pdev->dev;
6828 	int ret;
6829 
6830 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
6831 			       &handle->rinfo.state)) {
6832 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
6833 		return 0;
6834 	}
6835 
6836 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
6837 
6838 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
6839 	ret = __hns_roce_hw_v2_init_instance(handle);
6840 	if (ret) {
6841 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
6842 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
6843 		 * failed, we should inform NIC driver.
6844 		 */
6845 		handle->priv = NULL;
6846 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
6847 	} else {
6848 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
6849 		dev_info(dev, "Reset done, RoCE client reinit finished.\n");
6850 	}
6851 
6852 	return ret;
6853 }
6854 
6855 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
6856 {
6857 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
6858 		return 0;
6859 
6860 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
6861 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
6862 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
6863 	__hns_roce_hw_v2_uninit_instance(handle, false);
6864 
6865 	return 0;
6866 }
6867 
6868 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
6869 				       enum hnae3_reset_notify_type type)
6870 {
6871 	int ret = 0;
6872 
6873 	switch (type) {
6874 	case HNAE3_DOWN_CLIENT:
6875 		ret = hns_roce_hw_v2_reset_notify_down(handle);
6876 		break;
6877 	case HNAE3_INIT_CLIENT:
6878 		ret = hns_roce_hw_v2_reset_notify_init(handle);
6879 		break;
6880 	case HNAE3_UNINIT_CLIENT:
6881 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
6882 		break;
6883 	default:
6884 		break;
6885 	}
6886 
6887 	return ret;
6888 }
6889 
6890 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
6891 	.init_instance = hns_roce_hw_v2_init_instance,
6892 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
6893 	.reset_notify = hns_roce_hw_v2_reset_notify,
6894 };
6895 
6896 static struct hnae3_client hns_roce_hw_v2_client = {
6897 	.name = "hns_roce_hw_v2",
6898 	.type = HNAE3_CLIENT_ROCE,
6899 	.ops = &hns_roce_hw_v2_ops,
6900 };
6901 
6902 static int __init hns_roce_hw_v2_init(void)
6903 {
6904 	return hnae3_register_client(&hns_roce_hw_v2_client);
6905 }
6906 
6907 static void __exit hns_roce_hw_v2_exit(void)
6908 {
6909 	hnae3_unregister_client(&hns_roce_hw_v2_client);
6910 }
6911 
6912 module_init(hns_roce_hw_v2_init);
6913 module_exit(hns_roce_hw_v2_exit);
6914 
6915 MODULE_LICENSE("Dual BSD/GPL");
6916 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
6917 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
6918 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
6919 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
6920