xref: /linux/drivers/infiniband/hw/hns/hns_roce_hw_v2.c (revision 001821b0e79716c4e17c71d8e053a23599a7a508)
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(MASKED_ATOMIC_CMP_AND_SWP,	ATOM_MSK_CMP_AND_SWAP),
122 	HR_OPC_MAP(MASKED_ATOMIC_FETCH_AND_ADD,	ATOM_MSK_FETCH_AND_ADD),
123 	HR_OPC_MAP(REG_MR,			FAST_REG_PMR),
124 };
125 
126 static u32 to_hr_opcode(u32 ib_opcode)
127 {
128 	if (ib_opcode >= ARRAY_SIZE(hns_roce_op_code))
129 		return HNS_ROCE_V2_WQE_OP_MASK;
130 
131 	return hns_roce_op_code[ib_opcode] ? hns_roce_op_code[ib_opcode] - 1 :
132 					     HNS_ROCE_V2_WQE_OP_MASK;
133 }
134 
135 static void set_frmr_seg(struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
136 			 const struct ib_reg_wr *wr)
137 {
138 	struct hns_roce_wqe_frmr_seg *fseg =
139 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
140 	struct hns_roce_mr *mr = to_hr_mr(wr->mr);
141 	u64 pbl_ba;
142 
143 	/* use ib_access_flags */
144 	hr_reg_write_bool(fseg, FRMR_BIND_EN, wr->access & IB_ACCESS_MW_BIND);
145 	hr_reg_write_bool(fseg, FRMR_ATOMIC,
146 			  wr->access & IB_ACCESS_REMOTE_ATOMIC);
147 	hr_reg_write_bool(fseg, FRMR_RR, wr->access & IB_ACCESS_REMOTE_READ);
148 	hr_reg_write_bool(fseg, FRMR_RW, wr->access & IB_ACCESS_REMOTE_WRITE);
149 	hr_reg_write_bool(fseg, FRMR_LW, wr->access & IB_ACCESS_LOCAL_WRITE);
150 
151 	/* Data structure reuse may lead to confusion */
152 	pbl_ba = mr->pbl_mtr.hem_cfg.root_ba;
153 	rc_sq_wqe->msg_len = cpu_to_le32(lower_32_bits(pbl_ba));
154 	rc_sq_wqe->inv_key = cpu_to_le32(upper_32_bits(pbl_ba));
155 
156 	rc_sq_wqe->byte_16 = cpu_to_le32(wr->mr->length & 0xffffffff);
157 	rc_sq_wqe->byte_20 = cpu_to_le32(wr->mr->length >> 32);
158 	rc_sq_wqe->rkey = cpu_to_le32(wr->key);
159 	rc_sq_wqe->va = cpu_to_le64(wr->mr->iova);
160 
161 	hr_reg_write(fseg, FRMR_PBL_SIZE, mr->npages);
162 	hr_reg_write(fseg, FRMR_PBL_BUF_PG_SZ,
163 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
164 	hr_reg_clear(fseg, FRMR_BLK_MODE);
165 }
166 
167 static void set_atomic_seg(const struct ib_send_wr *wr,
168 			   struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
169 			   unsigned int valid_num_sge)
170 {
171 	struct hns_roce_v2_wqe_data_seg *dseg =
172 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
173 	struct hns_roce_wqe_atomic_seg *aseg =
174 		(void *)dseg + sizeof(struct hns_roce_v2_wqe_data_seg);
175 
176 	set_data_seg_v2(dseg, wr->sg_list);
177 
178 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
179 		aseg->fetchadd_swap_data = cpu_to_le64(atomic_wr(wr)->swap);
180 		aseg->cmp_data = cpu_to_le64(atomic_wr(wr)->compare_add);
181 	} else {
182 		aseg->fetchadd_swap_data =
183 			cpu_to_le64(atomic_wr(wr)->compare_add);
184 		aseg->cmp_data = 0;
185 	}
186 
187 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
188 }
189 
190 static int fill_ext_sge_inl_data(struct hns_roce_qp *qp,
191 				 const struct ib_send_wr *wr,
192 				 unsigned int *sge_idx, u32 msg_len)
193 {
194 	struct ib_device *ibdev = &(to_hr_dev(qp->ibqp.device))->ib_dev;
195 	unsigned int left_len_in_pg;
196 	unsigned int idx = *sge_idx;
197 	unsigned int i = 0;
198 	unsigned int len;
199 	void *addr;
200 	void *dseg;
201 
202 	if (msg_len > qp->sq.ext_sge_cnt * HNS_ROCE_SGE_SIZE) {
203 		ibdev_err(ibdev,
204 			  "no enough extended sge space for inline data.\n");
205 		return -EINVAL;
206 	}
207 
208 	dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
209 	left_len_in_pg = hr_hw_page_align((uintptr_t)dseg) - (uintptr_t)dseg;
210 	len = wr->sg_list[0].length;
211 	addr = (void *)(unsigned long)(wr->sg_list[0].addr);
212 
213 	/* When copying data to extended sge space, the left length in page may
214 	 * not long enough for current user's sge. So the data should be
215 	 * splited into several parts, one in the first page, and the others in
216 	 * the subsequent pages.
217 	 */
218 	while (1) {
219 		if (len <= left_len_in_pg) {
220 			memcpy(dseg, addr, len);
221 
222 			idx += len / HNS_ROCE_SGE_SIZE;
223 
224 			i++;
225 			if (i >= wr->num_sge)
226 				break;
227 
228 			left_len_in_pg -= len;
229 			len = wr->sg_list[i].length;
230 			addr = (void *)(unsigned long)(wr->sg_list[i].addr);
231 			dseg += len;
232 		} else {
233 			memcpy(dseg, addr, left_len_in_pg);
234 
235 			len -= left_len_in_pg;
236 			addr += left_len_in_pg;
237 			idx += left_len_in_pg / HNS_ROCE_SGE_SIZE;
238 			dseg = hns_roce_get_extend_sge(qp,
239 						idx & (qp->sge.sge_cnt - 1));
240 			left_len_in_pg = 1 << HNS_HW_PAGE_SHIFT;
241 		}
242 	}
243 
244 	*sge_idx = idx;
245 
246 	return 0;
247 }
248 
249 static void set_extend_sge(struct hns_roce_qp *qp, struct ib_sge *sge,
250 			   unsigned int *sge_ind, unsigned int cnt)
251 {
252 	struct hns_roce_v2_wqe_data_seg *dseg;
253 	unsigned int idx = *sge_ind;
254 
255 	while (cnt > 0) {
256 		dseg = hns_roce_get_extend_sge(qp, idx & (qp->sge.sge_cnt - 1));
257 		if (likely(sge->length)) {
258 			set_data_seg_v2(dseg, sge);
259 			idx++;
260 			cnt--;
261 		}
262 		sge++;
263 	}
264 
265 	*sge_ind = idx;
266 }
267 
268 static bool check_inl_data_len(struct hns_roce_qp *qp, unsigned int len)
269 {
270 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
271 	int mtu = ib_mtu_enum_to_int(qp->path_mtu);
272 
273 	if (mtu < 0 || len > qp->max_inline_data || len > mtu) {
274 		ibdev_err(&hr_dev->ib_dev,
275 			  "invalid length of data, data len = %u, max inline len = %u, path mtu = %d.\n",
276 			  len, qp->max_inline_data, mtu);
277 		return false;
278 	}
279 
280 	return true;
281 }
282 
283 static int set_rc_inl(struct hns_roce_qp *qp, const struct ib_send_wr *wr,
284 		      struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
285 		      unsigned int *sge_idx)
286 {
287 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
288 	u32 msg_len = le32_to_cpu(rc_sq_wqe->msg_len);
289 	struct ib_device *ibdev = &hr_dev->ib_dev;
290 	unsigned int curr_idx = *sge_idx;
291 	void *dseg = rc_sq_wqe;
292 	unsigned int i;
293 	int ret;
294 
295 	if (unlikely(wr->opcode == IB_WR_RDMA_READ)) {
296 		ibdev_err(ibdev, "invalid inline parameters!\n");
297 		return -EINVAL;
298 	}
299 
300 	if (!check_inl_data_len(qp, msg_len))
301 		return -EINVAL;
302 
303 	dseg += sizeof(struct hns_roce_v2_rc_send_wqe);
304 
305 	if (msg_len <= HNS_ROCE_V2_MAX_RC_INL_INN_SZ) {
306 		hr_reg_clear(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
307 
308 		for (i = 0; i < wr->num_sge; i++) {
309 			memcpy(dseg, ((void *)wr->sg_list[i].addr),
310 			       wr->sg_list[i].length);
311 			dseg += wr->sg_list[i].length;
312 		}
313 	} else {
314 		hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_INL_TYPE);
315 
316 		ret = fill_ext_sge_inl_data(qp, wr, &curr_idx, msg_len);
317 		if (ret)
318 			return ret;
319 
320 		hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, curr_idx - *sge_idx);
321 	}
322 
323 	*sge_idx = curr_idx;
324 
325 	return 0;
326 }
327 
328 static int set_rwqe_data_seg(struct ib_qp *ibqp, const struct ib_send_wr *wr,
329 			     struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
330 			     unsigned int *sge_ind,
331 			     unsigned int valid_num_sge)
332 {
333 	struct hns_roce_v2_wqe_data_seg *dseg =
334 		(void *)rc_sq_wqe + sizeof(struct hns_roce_v2_rc_send_wqe);
335 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
336 	int j = 0;
337 	int i;
338 
339 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_MSG_START_SGE_IDX,
340 		     (*sge_ind) & (qp->sge.sge_cnt - 1));
341 
342 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_INLINE,
343 		     !!(wr->send_flags & IB_SEND_INLINE));
344 	if (wr->send_flags & IB_SEND_INLINE)
345 		return set_rc_inl(qp, wr, rc_sq_wqe, sge_ind);
346 
347 	if (valid_num_sge <= HNS_ROCE_SGE_IN_WQE) {
348 		for (i = 0; i < wr->num_sge; i++) {
349 			if (likely(wr->sg_list[i].length)) {
350 				set_data_seg_v2(dseg, wr->sg_list + i);
351 				dseg++;
352 			}
353 		}
354 	} else {
355 		for (i = 0; i < wr->num_sge && j < HNS_ROCE_SGE_IN_WQE; i++) {
356 			if (likely(wr->sg_list[i].length)) {
357 				set_data_seg_v2(dseg, wr->sg_list + i);
358 				dseg++;
359 				j++;
360 			}
361 		}
362 
363 		set_extend_sge(qp, wr->sg_list + i, sge_ind,
364 			       valid_num_sge - HNS_ROCE_SGE_IN_WQE);
365 	}
366 
367 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SGE_NUM, valid_num_sge);
368 
369 	return 0;
370 }
371 
372 static int check_send_valid(struct hns_roce_dev *hr_dev,
373 			    struct hns_roce_qp *hr_qp)
374 {
375 	struct ib_device *ibdev = &hr_dev->ib_dev;
376 
377 	if (unlikely(hr_qp->state == IB_QPS_RESET ||
378 		     hr_qp->state == IB_QPS_INIT ||
379 		     hr_qp->state == IB_QPS_RTR)) {
380 		ibdev_err(ibdev, "failed to post WQE, QP state %u!\n",
381 			  hr_qp->state);
382 		return -EINVAL;
383 	} else if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN)) {
384 		ibdev_err(ibdev, "failed to post WQE, dev state %d!\n",
385 			  hr_dev->state);
386 		return -EIO;
387 	}
388 
389 	return 0;
390 }
391 
392 static unsigned int calc_wr_sge_num(const struct ib_send_wr *wr,
393 				    unsigned int *sge_len)
394 {
395 	unsigned int valid_num = 0;
396 	unsigned int len = 0;
397 	int i;
398 
399 	for (i = 0; i < wr->num_sge; i++) {
400 		if (likely(wr->sg_list[i].length)) {
401 			len += wr->sg_list[i].length;
402 			valid_num++;
403 		}
404 	}
405 
406 	*sge_len = len;
407 	return valid_num;
408 }
409 
410 static __le32 get_immtdata(const struct ib_send_wr *wr)
411 {
412 	switch (wr->opcode) {
413 	case IB_WR_SEND_WITH_IMM:
414 	case IB_WR_RDMA_WRITE_WITH_IMM:
415 		return cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
416 	default:
417 		return 0;
418 	}
419 }
420 
421 static int set_ud_opcode(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
422 			 const struct ib_send_wr *wr)
423 {
424 	u32 ib_op = wr->opcode;
425 
426 	if (ib_op != IB_WR_SEND && ib_op != IB_WR_SEND_WITH_IMM)
427 		return -EINVAL;
428 
429 	ud_sq_wqe->immtdata = get_immtdata(wr);
430 
431 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
432 
433 	return 0;
434 }
435 
436 static int fill_ud_av(struct hns_roce_v2_ud_send_wqe *ud_sq_wqe,
437 		      struct hns_roce_ah *ah)
438 {
439 	struct ib_device *ib_dev = ah->ibah.device;
440 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev);
441 
442 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_UDPSPN, ah->av.udp_sport);
443 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_HOPLIMIT, ah->av.hop_limit);
444 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_TCLASS, ah->av.tclass);
445 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_FLOW_LABEL, ah->av.flowlabel);
446 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SL, ah->av.sl);
447 
448 	ud_sq_wqe->sgid_index = ah->av.gid_index;
449 
450 	memcpy(ud_sq_wqe->dmac, ah->av.mac, ETH_ALEN);
451 	memcpy(ud_sq_wqe->dgid, ah->av.dgid, GID_LEN_V2);
452 
453 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
454 		return 0;
455 
456 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN_EN, ah->av.vlan_en);
457 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_VLAN, ah->av.vlan_id);
458 
459 	return 0;
460 }
461 
462 static inline int set_ud_wqe(struct hns_roce_qp *qp,
463 			     const struct ib_send_wr *wr,
464 			     void *wqe, unsigned int *sge_idx,
465 			     unsigned int owner_bit)
466 {
467 	struct hns_roce_ah *ah = to_hr_ah(ud_wr(wr)->ah);
468 	struct hns_roce_v2_ud_send_wqe *ud_sq_wqe = wqe;
469 	unsigned int curr_idx = *sge_idx;
470 	unsigned int valid_num_sge;
471 	u32 msg_len = 0;
472 	int ret;
473 
474 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
475 
476 	ret = set_ud_opcode(ud_sq_wqe, wr);
477 	if (WARN_ON(ret))
478 		return ret;
479 
480 	ud_sq_wqe->msg_len = cpu_to_le32(msg_len);
481 
482 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_CQE,
483 		     !!(wr->send_flags & IB_SEND_SIGNALED));
484 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SE,
485 		     !!(wr->send_flags & IB_SEND_SOLICITED));
486 
487 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_PD, to_hr_pd(qp->ibqp.pd)->pdn);
488 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_SGE_NUM, valid_num_sge);
489 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_MSG_START_SGE_IDX,
490 		     curr_idx & (qp->sge.sge_cnt - 1));
491 
492 	ud_sq_wqe->qkey = cpu_to_le32(ud_wr(wr)->remote_qkey & 0x80000000 ?
493 			  qp->qkey : ud_wr(wr)->remote_qkey);
494 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_DQPN, ud_wr(wr)->remote_qpn);
495 
496 	ret = fill_ud_av(ud_sq_wqe, ah);
497 	if (ret)
498 		return ret;
499 
500 	qp->sl = to_hr_ah(ud_wr(wr)->ah)->av.sl;
501 
502 	set_extend_sge(qp, wr->sg_list, &curr_idx, valid_num_sge);
503 
504 	/*
505 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
506 	 * including new WQEs waiting for the doorbell to update the PI again.
507 	 * Therefore, the owner bit of WQE MUST be updated after all fields
508 	 * and extSGEs have been written into DDR instead of cache.
509 	 */
510 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
511 		dma_wmb();
512 
513 	*sge_idx = curr_idx;
514 	hr_reg_write(ud_sq_wqe, UD_SEND_WQE_OWNER, owner_bit);
515 
516 	return 0;
517 }
518 
519 static int set_rc_opcode(struct hns_roce_dev *hr_dev,
520 			 struct hns_roce_v2_rc_send_wqe *rc_sq_wqe,
521 			 const struct ib_send_wr *wr)
522 {
523 	u32 ib_op = wr->opcode;
524 	int ret = 0;
525 
526 	rc_sq_wqe->immtdata = get_immtdata(wr);
527 
528 	switch (ib_op) {
529 	case IB_WR_RDMA_READ:
530 	case IB_WR_RDMA_WRITE:
531 	case IB_WR_RDMA_WRITE_WITH_IMM:
532 		rc_sq_wqe->rkey = cpu_to_le32(rdma_wr(wr)->rkey);
533 		rc_sq_wqe->va = cpu_to_le64(rdma_wr(wr)->remote_addr);
534 		break;
535 	case IB_WR_SEND:
536 	case IB_WR_SEND_WITH_IMM:
537 		break;
538 	case IB_WR_ATOMIC_CMP_AND_SWP:
539 	case IB_WR_ATOMIC_FETCH_AND_ADD:
540 		rc_sq_wqe->rkey = cpu_to_le32(atomic_wr(wr)->rkey);
541 		rc_sq_wqe->va = cpu_to_le64(atomic_wr(wr)->remote_addr);
542 		break;
543 	case IB_WR_REG_MR:
544 		if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
545 			set_frmr_seg(rc_sq_wqe, reg_wr(wr));
546 		else
547 			ret = -EOPNOTSUPP;
548 		break;
549 	case IB_WR_SEND_WITH_INV:
550 		rc_sq_wqe->inv_key = cpu_to_le32(wr->ex.invalidate_rkey);
551 		break;
552 	default:
553 		ret = -EINVAL;
554 	}
555 
556 	if (unlikely(ret))
557 		return ret;
558 
559 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OPCODE, to_hr_opcode(ib_op));
560 
561 	return ret;
562 }
563 
564 static inline int set_rc_wqe(struct hns_roce_qp *qp,
565 			     const struct ib_send_wr *wr,
566 			     void *wqe, unsigned int *sge_idx,
567 			     unsigned int owner_bit)
568 {
569 	struct hns_roce_dev *hr_dev = to_hr_dev(qp->ibqp.device);
570 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
571 	unsigned int curr_idx = *sge_idx;
572 	unsigned int valid_num_sge;
573 	u32 msg_len = 0;
574 	int ret;
575 
576 	valid_num_sge = calc_wr_sge_num(wr, &msg_len);
577 
578 	rc_sq_wqe->msg_len = cpu_to_le32(msg_len);
579 
580 	ret = set_rc_opcode(hr_dev, rc_sq_wqe, wr);
581 	if (WARN_ON(ret))
582 		return ret;
583 
584 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_FENCE,
585 		     (wr->send_flags & IB_SEND_FENCE) ? 1 : 0);
586 
587 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_SE,
588 		     (wr->send_flags & IB_SEND_SOLICITED) ? 1 : 0);
589 
590 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_CQE,
591 		     (wr->send_flags & IB_SEND_SIGNALED) ? 1 : 0);
592 
593 	if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
594 	    wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
595 		set_atomic_seg(wr, rc_sq_wqe, valid_num_sge);
596 	else if (wr->opcode != IB_WR_REG_MR)
597 		ret = set_rwqe_data_seg(&qp->ibqp, wr, rc_sq_wqe,
598 					&curr_idx, valid_num_sge);
599 
600 	/*
601 	 * The pipeline can sequentially post all valid WQEs into WQ buffer,
602 	 * including new WQEs waiting for the doorbell to update the PI again.
603 	 * Therefore, the owner bit of WQE MUST be updated after all fields
604 	 * and extSGEs have been written into DDR instead of cache.
605 	 */
606 	if (qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
607 		dma_wmb();
608 
609 	*sge_idx = curr_idx;
610 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_OWNER, owner_bit);
611 
612 	return ret;
613 }
614 
615 static inline void update_sq_db(struct hns_roce_dev *hr_dev,
616 				struct hns_roce_qp *qp)
617 {
618 	if (unlikely(qp->state == IB_QPS_ERR)) {
619 		flush_cqe(hr_dev, qp);
620 	} else {
621 		struct hns_roce_v2_db sq_db = {};
622 
623 		hr_reg_write(&sq_db, DB_TAG, qp->qpn);
624 		hr_reg_write(&sq_db, DB_CMD, HNS_ROCE_V2_SQ_DB);
625 		hr_reg_write(&sq_db, DB_PI, qp->sq.head);
626 		hr_reg_write(&sq_db, DB_SL, qp->sl);
627 
628 		hns_roce_write64(hr_dev, (__le32 *)&sq_db, qp->sq.db_reg);
629 	}
630 }
631 
632 static inline void update_rq_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 		if (likely(qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)) {
639 			*qp->rdb.db_record =
640 					qp->rq.head & V2_DB_PRODUCER_IDX_M;
641 		} else {
642 			struct hns_roce_v2_db rq_db = {};
643 
644 			hr_reg_write(&rq_db, DB_TAG, qp->qpn);
645 			hr_reg_write(&rq_db, DB_CMD, HNS_ROCE_V2_RQ_DB);
646 			hr_reg_write(&rq_db, DB_PI, qp->rq.head);
647 
648 			hns_roce_write64(hr_dev, (__le32 *)&rq_db,
649 					 qp->rq.db_reg);
650 		}
651 	}
652 }
653 
654 static void hns_roce_write512(struct hns_roce_dev *hr_dev, u64 *val,
655 			      u64 __iomem *dest)
656 {
657 #define HNS_ROCE_WRITE_TIMES 8
658 	struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
659 	struct hnae3_handle *handle = priv->handle;
660 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
661 	int i;
662 
663 	if (!hr_dev->dis_db && !ops->get_hw_reset_stat(handle))
664 		for (i = 0; i < HNS_ROCE_WRITE_TIMES; i++)
665 			writeq_relaxed(*(val + i), dest + i);
666 }
667 
668 static void write_dwqe(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
669 		       void *wqe)
670 {
671 #define HNS_ROCE_SL_SHIFT 2
672 	struct hns_roce_v2_rc_send_wqe *rc_sq_wqe = wqe;
673 
674 	/* All kinds of DirectWQE have the same header field layout */
675 	hr_reg_enable(rc_sq_wqe, RC_SEND_WQE_FLAG);
676 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_L, qp->sl);
677 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_DB_SL_H,
678 		     qp->sl >> HNS_ROCE_SL_SHIFT);
679 	hr_reg_write(rc_sq_wqe, RC_SEND_WQE_WQE_INDEX, qp->sq.head);
680 
681 	hns_roce_write512(hr_dev, wqe, qp->sq.db_reg);
682 }
683 
684 static int hns_roce_v2_post_send(struct ib_qp *ibqp,
685 				 const struct ib_send_wr *wr,
686 				 const struct ib_send_wr **bad_wr)
687 {
688 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
689 	struct ib_device *ibdev = &hr_dev->ib_dev;
690 	struct hns_roce_qp *qp = to_hr_qp(ibqp);
691 	unsigned long flags = 0;
692 	unsigned int owner_bit;
693 	unsigned int sge_idx;
694 	unsigned int wqe_idx;
695 	void *wqe = NULL;
696 	u32 nreq;
697 	int ret;
698 
699 	spin_lock_irqsave(&qp->sq.lock, flags);
700 
701 	ret = check_send_valid(hr_dev, qp);
702 	if (unlikely(ret)) {
703 		*bad_wr = wr;
704 		nreq = 0;
705 		goto out;
706 	}
707 
708 	sge_idx = qp->next_sge;
709 
710 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
711 		if (hns_roce_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
712 			ret = -ENOMEM;
713 			*bad_wr = wr;
714 			goto out;
715 		}
716 
717 		wqe_idx = (qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1);
718 
719 		if (unlikely(wr->num_sge > qp->sq.max_gs)) {
720 			ibdev_err(ibdev, "num_sge = %d > qp->sq.max_gs = %u.\n",
721 				  wr->num_sge, qp->sq.max_gs);
722 			ret = -EINVAL;
723 			*bad_wr = wr;
724 			goto out;
725 		}
726 
727 		wqe = hns_roce_get_send_wqe(qp, wqe_idx);
728 		qp->sq.wrid[wqe_idx] = wr->wr_id;
729 		owner_bit =
730 		       ~(((qp->sq.head + nreq) >> ilog2(qp->sq.wqe_cnt)) & 0x1);
731 
732 		/* Corresponding to the QP type, wqe process separately */
733 		if (ibqp->qp_type == IB_QPT_RC)
734 			ret = set_rc_wqe(qp, wr, wqe, &sge_idx, owner_bit);
735 		else
736 			ret = set_ud_wqe(qp, wr, wqe, &sge_idx, owner_bit);
737 
738 		if (unlikely(ret)) {
739 			*bad_wr = wr;
740 			goto out;
741 		}
742 	}
743 
744 out:
745 	if (likely(nreq)) {
746 		qp->sq.head += nreq;
747 		qp->next_sge = sge_idx;
748 
749 		if (nreq == 1 && !ret &&
750 		    (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
751 			write_dwqe(hr_dev, qp, wqe);
752 		else
753 			update_sq_db(hr_dev, qp);
754 	}
755 
756 	spin_unlock_irqrestore(&qp->sq.lock, flags);
757 
758 	return ret;
759 }
760 
761 static int check_recv_valid(struct hns_roce_dev *hr_dev,
762 			    struct hns_roce_qp *hr_qp)
763 {
764 	if (unlikely(hr_dev->state >= HNS_ROCE_DEVICE_STATE_RST_DOWN))
765 		return -EIO;
766 
767 	if (hr_qp->state == IB_QPS_RESET)
768 		return -EINVAL;
769 
770 	return 0;
771 }
772 
773 static void fill_recv_sge_to_wqe(const struct ib_recv_wr *wr, void *wqe,
774 				 u32 max_sge, bool rsv)
775 {
776 	struct hns_roce_v2_wqe_data_seg *dseg = wqe;
777 	u32 i, cnt;
778 
779 	for (i = 0, cnt = 0; i < wr->num_sge; i++) {
780 		/* Skip zero-length sge */
781 		if (!wr->sg_list[i].length)
782 			continue;
783 		set_data_seg_v2(dseg + cnt, wr->sg_list + i);
784 		cnt++;
785 	}
786 
787 	/* Fill a reserved sge to make hw stop reading remaining segments */
788 	if (rsv) {
789 		dseg[cnt].lkey = cpu_to_le32(HNS_ROCE_INVALID_LKEY);
790 		dseg[cnt].addr = 0;
791 		dseg[cnt].len = cpu_to_le32(HNS_ROCE_INVALID_SGE_LENGTH);
792 	} else {
793 		/* Clear remaining segments to make ROCEE ignore sges */
794 		if (cnt < max_sge)
795 			memset(dseg + cnt, 0,
796 			       (max_sge - cnt) * HNS_ROCE_SGE_SIZE);
797 	}
798 }
799 
800 static void fill_rq_wqe(struct hns_roce_qp *hr_qp, const struct ib_recv_wr *wr,
801 			u32 wqe_idx, u32 max_sge)
802 {
803 	void *wqe = NULL;
804 
805 	wqe = hns_roce_get_recv_wqe(hr_qp, wqe_idx);
806 	fill_recv_sge_to_wqe(wr, wqe, max_sge, hr_qp->rq.rsv_sge);
807 }
808 
809 static int hns_roce_v2_post_recv(struct ib_qp *ibqp,
810 				 const struct ib_recv_wr *wr,
811 				 const struct ib_recv_wr **bad_wr)
812 {
813 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
814 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
815 	struct ib_device *ibdev = &hr_dev->ib_dev;
816 	u32 wqe_idx, nreq, max_sge;
817 	unsigned long flags;
818 	int ret;
819 
820 	spin_lock_irqsave(&hr_qp->rq.lock, flags);
821 
822 	ret = check_recv_valid(hr_dev, hr_qp);
823 	if (unlikely(ret)) {
824 		*bad_wr = wr;
825 		nreq = 0;
826 		goto out;
827 	}
828 
829 	max_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
830 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
831 		if (unlikely(hns_roce_wq_overflow(&hr_qp->rq, nreq,
832 						  hr_qp->ibqp.recv_cq))) {
833 			ret = -ENOMEM;
834 			*bad_wr = wr;
835 			goto out;
836 		}
837 
838 		if (unlikely(wr->num_sge > max_sge)) {
839 			ibdev_err(ibdev, "num_sge = %d >= max_sge = %u.\n",
840 				  wr->num_sge, max_sge);
841 			ret = -EINVAL;
842 			*bad_wr = wr;
843 			goto out;
844 		}
845 
846 		wqe_idx = (hr_qp->rq.head + nreq) & (hr_qp->rq.wqe_cnt - 1);
847 		fill_rq_wqe(hr_qp, wr, wqe_idx, max_sge);
848 		hr_qp->rq.wrid[wqe_idx] = wr->wr_id;
849 	}
850 
851 out:
852 	if (likely(nreq)) {
853 		hr_qp->rq.head += nreq;
854 
855 		update_rq_db(hr_dev, hr_qp);
856 	}
857 	spin_unlock_irqrestore(&hr_qp->rq.lock, flags);
858 
859 	return ret;
860 }
861 
862 static void *get_srq_wqe_buf(struct hns_roce_srq *srq, u32 n)
863 {
864 	return hns_roce_buf_offset(srq->buf_mtr.kmem, n << srq->wqe_shift);
865 }
866 
867 static void *get_idx_buf(struct hns_roce_idx_que *idx_que, u32 n)
868 {
869 	return hns_roce_buf_offset(idx_que->mtr.kmem,
870 				   n << idx_que->entry_shift);
871 }
872 
873 static void hns_roce_free_srq_wqe(struct hns_roce_srq *srq, u32 wqe_index)
874 {
875 	/* always called with interrupts disabled. */
876 	spin_lock(&srq->lock);
877 
878 	bitmap_clear(srq->idx_que.bitmap, wqe_index, 1);
879 	srq->idx_que.tail++;
880 
881 	spin_unlock(&srq->lock);
882 }
883 
884 static int hns_roce_srqwq_overflow(struct hns_roce_srq *srq)
885 {
886 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
887 
888 	return idx_que->head - idx_que->tail >= srq->wqe_cnt;
889 }
890 
891 static int check_post_srq_valid(struct hns_roce_srq *srq, u32 max_sge,
892 				const struct ib_recv_wr *wr)
893 {
894 	struct ib_device *ib_dev = srq->ibsrq.device;
895 
896 	if (unlikely(wr->num_sge > max_sge)) {
897 		ibdev_err(ib_dev,
898 			  "failed to check sge, wr->num_sge = %d, max_sge = %u.\n",
899 			  wr->num_sge, max_sge);
900 		return -EINVAL;
901 	}
902 
903 	if (unlikely(hns_roce_srqwq_overflow(srq))) {
904 		ibdev_err(ib_dev,
905 			  "failed to check srqwq status, srqwq is full.\n");
906 		return -ENOMEM;
907 	}
908 
909 	return 0;
910 }
911 
912 static int get_srq_wqe_idx(struct hns_roce_srq *srq, u32 *wqe_idx)
913 {
914 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
915 	u32 pos;
916 
917 	pos = find_first_zero_bit(idx_que->bitmap, srq->wqe_cnt);
918 	if (unlikely(pos == srq->wqe_cnt))
919 		return -ENOSPC;
920 
921 	bitmap_set(idx_que->bitmap, pos, 1);
922 	*wqe_idx = pos;
923 	return 0;
924 }
925 
926 static void fill_wqe_idx(struct hns_roce_srq *srq, unsigned int wqe_idx)
927 {
928 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
929 	unsigned int head;
930 	__le32 *buf;
931 
932 	head = idx_que->head & (srq->wqe_cnt - 1);
933 
934 	buf = get_idx_buf(idx_que, head);
935 	*buf = cpu_to_le32(wqe_idx);
936 
937 	idx_que->head++;
938 }
939 
940 static void update_srq_db(struct hns_roce_srq *srq)
941 {
942 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
943 	struct hns_roce_v2_db db;
944 
945 	hr_reg_write(&db, DB_TAG, srq->srqn);
946 	hr_reg_write(&db, DB_CMD, HNS_ROCE_V2_SRQ_DB);
947 	hr_reg_write(&db, DB_PI, srq->idx_que.head);
948 
949 	hns_roce_write64(hr_dev, (__le32 *)&db, srq->db_reg);
950 }
951 
952 static int hns_roce_v2_post_srq_recv(struct ib_srq *ibsrq,
953 				     const struct ib_recv_wr *wr,
954 				     const struct ib_recv_wr **bad_wr)
955 {
956 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
957 	unsigned long flags;
958 	int ret = 0;
959 	u32 max_sge;
960 	u32 wqe_idx;
961 	void *wqe;
962 	u32 nreq;
963 
964 	spin_lock_irqsave(&srq->lock, flags);
965 
966 	max_sge = srq->max_gs - srq->rsv_sge;
967 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
968 		ret = check_post_srq_valid(srq, max_sge, wr);
969 		if (ret) {
970 			*bad_wr = wr;
971 			break;
972 		}
973 
974 		ret = get_srq_wqe_idx(srq, &wqe_idx);
975 		if (unlikely(ret)) {
976 			*bad_wr = wr;
977 			break;
978 		}
979 
980 		wqe = get_srq_wqe_buf(srq, wqe_idx);
981 		fill_recv_sge_to_wqe(wr, wqe, max_sge, srq->rsv_sge);
982 		fill_wqe_idx(srq, wqe_idx);
983 		srq->wrid[wqe_idx] = wr->wr_id;
984 	}
985 
986 	if (likely(nreq)) {
987 		if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB)
988 			*srq->rdb.db_record = srq->idx_que.head &
989 					      V2_DB_PRODUCER_IDX_M;
990 		else
991 			update_srq_db(srq);
992 	}
993 
994 	spin_unlock_irqrestore(&srq->lock, flags);
995 
996 	return ret;
997 }
998 
999 static u32 hns_roce_v2_cmd_hw_reseted(struct hns_roce_dev *hr_dev,
1000 				      unsigned long instance_stage,
1001 				      unsigned long reset_stage)
1002 {
1003 	/* When hardware reset has been completed once or more, we should stop
1004 	 * sending mailbox&cmq&doorbell to hardware. If now in .init_instance()
1005 	 * function, we should exit with error. If now at HNAE3_INIT_CLIENT
1006 	 * stage of soft reset process, we should exit with error, and then
1007 	 * HNAE3_INIT_CLIENT related process can rollback the operation like
1008 	 * notifing hardware to free resources, HNAE3_INIT_CLIENT related
1009 	 * process will exit with error to notify NIC driver to reschedule soft
1010 	 * reset process once again.
1011 	 */
1012 	hr_dev->is_reset = true;
1013 	hr_dev->dis_db = true;
1014 
1015 	if (reset_stage == HNS_ROCE_STATE_RST_INIT ||
1016 	    instance_stage == HNS_ROCE_STATE_INIT)
1017 		return CMD_RST_PRC_EBUSY;
1018 
1019 	return CMD_RST_PRC_SUCCESS;
1020 }
1021 
1022 static u32 hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
1023 					unsigned long instance_stage,
1024 					unsigned long reset_stage)
1025 {
1026 #define HW_RESET_TIMEOUT_US 1000000
1027 #define HW_RESET_SLEEP_US 1000
1028 
1029 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1030 	struct hnae3_handle *handle = priv->handle;
1031 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1032 	unsigned long val;
1033 	int ret;
1034 
1035 	/* When hardware reset is detected, we should stop sending mailbox&cmq&
1036 	 * doorbell to hardware. If now in .init_instance() function, we should
1037 	 * exit with error. If now at HNAE3_INIT_CLIENT stage of soft reset
1038 	 * process, we should exit with error, and then HNAE3_INIT_CLIENT
1039 	 * related process can rollback the operation like notifing hardware to
1040 	 * free resources, HNAE3_INIT_CLIENT related process will exit with
1041 	 * error to notify NIC driver to reschedule soft reset process once
1042 	 * again.
1043 	 */
1044 	hr_dev->dis_db = true;
1045 
1046 	ret = read_poll_timeout(ops->ae_dev_reset_cnt, val,
1047 				val > hr_dev->reset_cnt, HW_RESET_SLEEP_US,
1048 				HW_RESET_TIMEOUT_US, false, handle);
1049 	if (!ret)
1050 		hr_dev->is_reset = true;
1051 
1052 	if (!hr_dev->is_reset || reset_stage == HNS_ROCE_STATE_RST_INIT ||
1053 	    instance_stage == HNS_ROCE_STATE_INIT)
1054 		return CMD_RST_PRC_EBUSY;
1055 
1056 	return CMD_RST_PRC_SUCCESS;
1057 }
1058 
1059 static u32 hns_roce_v2_cmd_sw_resetting(struct hns_roce_dev *hr_dev)
1060 {
1061 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1062 	struct hnae3_handle *handle = priv->handle;
1063 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1064 
1065 	/* When software reset is detected at .init_instance() function, we
1066 	 * should stop sending mailbox&cmq&doorbell to hardware, and exit
1067 	 * with error.
1068 	 */
1069 	hr_dev->dis_db = true;
1070 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt)
1071 		hr_dev->is_reset = true;
1072 
1073 	return CMD_RST_PRC_EBUSY;
1074 }
1075 
1076 static u32 check_aedev_reset_status(struct hns_roce_dev *hr_dev,
1077 				    struct hnae3_handle *handle)
1078 {
1079 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1080 	unsigned long instance_stage; /* the current instance stage */
1081 	unsigned long reset_stage; /* the current reset stage */
1082 	unsigned long reset_cnt;
1083 	bool sw_resetting;
1084 	bool hw_resetting;
1085 
1086 	/* Get information about reset from NIC driver or RoCE driver itself,
1087 	 * the meaning of the following variables from NIC driver are described
1088 	 * as below:
1089 	 * reset_cnt -- The count value of completed hardware reset.
1090 	 * hw_resetting -- Whether hardware device is resetting now.
1091 	 * sw_resetting -- Whether NIC's software reset process is running now.
1092 	 */
1093 	instance_stage = handle->rinfo.instance_state;
1094 	reset_stage = handle->rinfo.reset_state;
1095 	reset_cnt = ops->ae_dev_reset_cnt(handle);
1096 	if (reset_cnt != hr_dev->reset_cnt)
1097 		return hns_roce_v2_cmd_hw_reseted(hr_dev, instance_stage,
1098 						  reset_stage);
1099 
1100 	hw_resetting = ops->get_cmdq_stat(handle);
1101 	if (hw_resetting)
1102 		return hns_roce_v2_cmd_hw_resetting(hr_dev, instance_stage,
1103 						    reset_stage);
1104 
1105 	sw_resetting = ops->ae_dev_resetting(handle);
1106 	if (sw_resetting && instance_stage == HNS_ROCE_STATE_INIT)
1107 		return hns_roce_v2_cmd_sw_resetting(hr_dev);
1108 
1109 	return CMD_RST_PRC_OTHERS;
1110 }
1111 
1112 static bool check_device_is_in_reset(struct hns_roce_dev *hr_dev)
1113 {
1114 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1115 	struct hnae3_handle *handle = priv->handle;
1116 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1117 
1118 	if (hr_dev->reset_cnt != ops->ae_dev_reset_cnt(handle))
1119 		return true;
1120 
1121 	if (ops->get_hw_reset_stat(handle))
1122 		return true;
1123 
1124 	if (ops->ae_dev_resetting(handle))
1125 		return true;
1126 
1127 	return false;
1128 }
1129 
1130 static bool v2_chk_mbox_is_avail(struct hns_roce_dev *hr_dev, bool *busy)
1131 {
1132 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1133 	u32 status;
1134 
1135 	if (hr_dev->is_reset)
1136 		status = CMD_RST_PRC_SUCCESS;
1137 	else
1138 		status = check_aedev_reset_status(hr_dev, priv->handle);
1139 
1140 	*busy = (status == CMD_RST_PRC_EBUSY);
1141 
1142 	return status == CMD_RST_PRC_OTHERS;
1143 }
1144 
1145 static int hns_roce_alloc_cmq_desc(struct hns_roce_dev *hr_dev,
1146 				   struct hns_roce_v2_cmq_ring *ring)
1147 {
1148 	int size = ring->desc_num * sizeof(struct hns_roce_cmq_desc);
1149 
1150 	ring->desc = dma_alloc_coherent(hr_dev->dev, size,
1151 					&ring->desc_dma_addr, GFP_KERNEL);
1152 	if (!ring->desc)
1153 		return -ENOMEM;
1154 
1155 	return 0;
1156 }
1157 
1158 static void hns_roce_free_cmq_desc(struct hns_roce_dev *hr_dev,
1159 				   struct hns_roce_v2_cmq_ring *ring)
1160 {
1161 	dma_free_coherent(hr_dev->dev,
1162 			  ring->desc_num * sizeof(struct hns_roce_cmq_desc),
1163 			  ring->desc, ring->desc_dma_addr);
1164 
1165 	ring->desc_dma_addr = 0;
1166 }
1167 
1168 static int init_csq(struct hns_roce_dev *hr_dev,
1169 		    struct hns_roce_v2_cmq_ring *csq)
1170 {
1171 	dma_addr_t dma;
1172 	int ret;
1173 
1174 	csq->desc_num = CMD_CSQ_DESC_NUM;
1175 	spin_lock_init(&csq->lock);
1176 	csq->flag = TYPE_CSQ;
1177 	csq->head = 0;
1178 
1179 	ret = hns_roce_alloc_cmq_desc(hr_dev, csq);
1180 	if (ret)
1181 		return ret;
1182 
1183 	dma = csq->desc_dma_addr;
1184 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_L_REG, lower_32_bits(dma));
1185 	roce_write(hr_dev, ROCEE_TX_CMQ_BASEADDR_H_REG, upper_32_bits(dma));
1186 	roce_write(hr_dev, ROCEE_TX_CMQ_DEPTH_REG,
1187 		   (u32)csq->desc_num >> HNS_ROCE_CMQ_DESC_NUM_S);
1188 
1189 	/* Make sure to write CI first and then PI */
1190 	roce_write(hr_dev, ROCEE_TX_CMQ_CI_REG, 0);
1191 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, 0);
1192 
1193 	return 0;
1194 }
1195 
1196 static int hns_roce_v2_cmq_init(struct hns_roce_dev *hr_dev)
1197 {
1198 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1199 	int ret;
1200 
1201 	priv->cmq.tx_timeout = HNS_ROCE_CMQ_TX_TIMEOUT;
1202 
1203 	ret = init_csq(hr_dev, &priv->cmq.csq);
1204 	if (ret)
1205 		dev_err(hr_dev->dev, "failed to init CSQ, ret = %d.\n", ret);
1206 
1207 	return ret;
1208 }
1209 
1210 static void hns_roce_v2_cmq_exit(struct hns_roce_dev *hr_dev)
1211 {
1212 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1213 
1214 	hns_roce_free_cmq_desc(hr_dev, &priv->cmq.csq);
1215 }
1216 
1217 static void hns_roce_cmq_setup_basic_desc(struct hns_roce_cmq_desc *desc,
1218 					  enum hns_roce_opcode_type opcode,
1219 					  bool is_read)
1220 {
1221 	memset((void *)desc, 0, sizeof(struct hns_roce_cmq_desc));
1222 	desc->opcode = cpu_to_le16(opcode);
1223 	desc->flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1224 	if (is_read)
1225 		desc->flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_WR);
1226 	else
1227 		desc->flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1228 }
1229 
1230 static int hns_roce_cmq_csq_done(struct hns_roce_dev *hr_dev)
1231 {
1232 	u32 tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1233 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1234 
1235 	return tail == priv->cmq.csq.head;
1236 }
1237 
1238 static void update_cmdq_status(struct hns_roce_dev *hr_dev)
1239 {
1240 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1241 	struct hnae3_handle *handle = priv->handle;
1242 
1243 	if (handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
1244 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT)
1245 		hr_dev->cmd.state = HNS_ROCE_CMDQ_STATE_FATAL_ERR;
1246 }
1247 
1248 static int hns_roce_cmd_err_convert_errno(u16 desc_ret)
1249 {
1250 	struct hns_roce_cmd_errcode errcode_table[] = {
1251 		{CMD_EXEC_SUCCESS, 0},
1252 		{CMD_NO_AUTH, -EPERM},
1253 		{CMD_NOT_EXIST, -EOPNOTSUPP},
1254 		{CMD_CRQ_FULL, -EXFULL},
1255 		{CMD_NEXT_ERR, -ENOSR},
1256 		{CMD_NOT_EXEC, -ENOTBLK},
1257 		{CMD_PARA_ERR, -EINVAL},
1258 		{CMD_RESULT_ERR, -ERANGE},
1259 		{CMD_TIMEOUT, -ETIME},
1260 		{CMD_HILINK_ERR, -ENOLINK},
1261 		{CMD_INFO_ILLEGAL, -ENXIO},
1262 		{CMD_INVALID, -EBADR},
1263 	};
1264 	u16 i;
1265 
1266 	for (i = 0; i < ARRAY_SIZE(errcode_table); i++)
1267 		if (desc_ret == errcode_table[i].return_status)
1268 			return errcode_table[i].errno;
1269 	return -EIO;
1270 }
1271 
1272 static int __hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1273 			       struct hns_roce_cmq_desc *desc, int num)
1274 {
1275 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1276 	struct hns_roce_v2_cmq_ring *csq = &priv->cmq.csq;
1277 	u32 timeout = 0;
1278 	u16 desc_ret;
1279 	u32 tail;
1280 	int ret;
1281 	int i;
1282 
1283 	spin_lock_bh(&csq->lock);
1284 
1285 	tail = csq->head;
1286 
1287 	for (i = 0; i < num; i++) {
1288 		csq->desc[csq->head++] = desc[i];
1289 		if (csq->head == csq->desc_num)
1290 			csq->head = 0;
1291 	}
1292 
1293 	/* Write to hardware */
1294 	roce_write(hr_dev, ROCEE_TX_CMQ_PI_REG, csq->head);
1295 
1296 	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_CNT]);
1297 
1298 	do {
1299 		if (hns_roce_cmq_csq_done(hr_dev))
1300 			break;
1301 		udelay(1);
1302 	} while (++timeout < priv->cmq.tx_timeout);
1303 
1304 	if (hns_roce_cmq_csq_done(hr_dev)) {
1305 		ret = 0;
1306 		for (i = 0; i < num; i++) {
1307 			/* check the result of hardware write back */
1308 			desc[i] = csq->desc[tail++];
1309 			if (tail == csq->desc_num)
1310 				tail = 0;
1311 
1312 			desc_ret = le16_to_cpu(desc[i].retval);
1313 			if (likely(desc_ret == CMD_EXEC_SUCCESS))
1314 				continue;
1315 
1316 			dev_err_ratelimited(hr_dev->dev,
1317 					    "Cmdq IO error, opcode = 0x%x, return = 0x%x.\n",
1318 					    desc->opcode, desc_ret);
1319 			ret = hns_roce_cmd_err_convert_errno(desc_ret);
1320 		}
1321 	} else {
1322 		/* FW/HW reset or incorrect number of desc */
1323 		tail = roce_read(hr_dev, ROCEE_TX_CMQ_CI_REG);
1324 		dev_warn(hr_dev->dev, "CMDQ move tail from %u to %u.\n",
1325 			 csq->head, tail);
1326 		csq->head = tail;
1327 
1328 		update_cmdq_status(hr_dev);
1329 
1330 		ret = -EAGAIN;
1331 	}
1332 
1333 	spin_unlock_bh(&csq->lock);
1334 
1335 	if (ret)
1336 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CMDS_ERR_CNT]);
1337 
1338 	return ret;
1339 }
1340 
1341 static int hns_roce_cmq_send(struct hns_roce_dev *hr_dev,
1342 			     struct hns_roce_cmq_desc *desc, int num)
1343 {
1344 	bool busy;
1345 	int ret;
1346 
1347 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1348 		return -EIO;
1349 
1350 	if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1351 		return busy ? -EBUSY : 0;
1352 
1353 	ret = __hns_roce_cmq_send(hr_dev, desc, num);
1354 	if (ret) {
1355 		if (!v2_chk_mbox_is_avail(hr_dev, &busy))
1356 			return busy ? -EBUSY : 0;
1357 	}
1358 
1359 	return ret;
1360 }
1361 
1362 static int config_hem_ba_to_hw(struct hns_roce_dev *hr_dev,
1363 			       dma_addr_t base_addr, u8 cmd, unsigned long tag)
1364 {
1365 	struct hns_roce_cmd_mailbox *mbox;
1366 	int ret;
1367 
1368 	mbox = hns_roce_alloc_cmd_mailbox(hr_dev);
1369 	if (IS_ERR(mbox))
1370 		return PTR_ERR(mbox);
1371 
1372 	ret = hns_roce_cmd_mbox(hr_dev, base_addr, mbox->dma, cmd, tag);
1373 	hns_roce_free_cmd_mailbox(hr_dev, mbox);
1374 	return ret;
1375 }
1376 
1377 static int hns_roce_cmq_query_hw_info(struct hns_roce_dev *hr_dev)
1378 {
1379 	struct hns_roce_query_version *resp;
1380 	struct hns_roce_cmq_desc desc;
1381 	int ret;
1382 
1383 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_HW_VER, true);
1384 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1385 	if (ret)
1386 		return ret;
1387 
1388 	resp = (struct hns_roce_query_version *)desc.data;
1389 	hr_dev->hw_rev = le16_to_cpu(resp->rocee_hw_version);
1390 	hr_dev->vendor_id = hr_dev->pci_dev->vendor;
1391 
1392 	return 0;
1393 }
1394 
1395 static void func_clr_hw_resetting_state(struct hns_roce_dev *hr_dev,
1396 					struct hnae3_handle *handle)
1397 {
1398 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1399 	unsigned long end;
1400 
1401 	hr_dev->dis_db = true;
1402 
1403 	dev_warn(hr_dev->dev,
1404 		 "func clear is pending, device in resetting state.\n");
1405 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1406 	while (end) {
1407 		if (!ops->get_hw_reset_stat(handle)) {
1408 			hr_dev->is_reset = true;
1409 			dev_info(hr_dev->dev,
1410 				 "func clear success after reset.\n");
1411 			return;
1412 		}
1413 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1414 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1415 	}
1416 
1417 	dev_warn(hr_dev->dev, "func clear failed.\n");
1418 }
1419 
1420 static void func_clr_sw_resetting_state(struct hns_roce_dev *hr_dev,
1421 					struct hnae3_handle *handle)
1422 {
1423 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1424 	unsigned long end;
1425 
1426 	hr_dev->dis_db = true;
1427 
1428 	dev_warn(hr_dev->dev,
1429 		 "func clear is pending, device in resetting state.\n");
1430 	end = HNS_ROCE_V2_HW_RST_TIMEOUT;
1431 	while (end) {
1432 		if (ops->ae_dev_reset_cnt(handle) !=
1433 		    hr_dev->reset_cnt) {
1434 			hr_dev->is_reset = true;
1435 			dev_info(hr_dev->dev,
1436 				 "func clear success after sw reset\n");
1437 			return;
1438 		}
1439 		msleep(HNS_ROCE_V2_HW_RST_COMPLETION_WAIT);
1440 		end -= HNS_ROCE_V2_HW_RST_COMPLETION_WAIT;
1441 	}
1442 
1443 	dev_warn(hr_dev->dev, "func clear failed because of unfinished sw reset\n");
1444 }
1445 
1446 static void hns_roce_func_clr_rst_proc(struct hns_roce_dev *hr_dev, int retval,
1447 				       int flag)
1448 {
1449 	struct hns_roce_v2_priv *priv = hr_dev->priv;
1450 	struct hnae3_handle *handle = priv->handle;
1451 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
1452 
1453 	if (ops->ae_dev_reset_cnt(handle) != hr_dev->reset_cnt) {
1454 		hr_dev->dis_db = true;
1455 		hr_dev->is_reset = true;
1456 		dev_info(hr_dev->dev, "func clear success after reset.\n");
1457 		return;
1458 	}
1459 
1460 	if (ops->get_hw_reset_stat(handle)) {
1461 		func_clr_hw_resetting_state(hr_dev, handle);
1462 		return;
1463 	}
1464 
1465 	if (ops->ae_dev_resetting(handle) &&
1466 	    handle->rinfo.instance_state == HNS_ROCE_STATE_INIT) {
1467 		func_clr_sw_resetting_state(hr_dev, handle);
1468 		return;
1469 	}
1470 
1471 	if (retval && !flag)
1472 		dev_warn(hr_dev->dev,
1473 			 "func clear read failed, ret = %d.\n", retval);
1474 
1475 	dev_warn(hr_dev->dev, "func clear failed.\n");
1476 }
1477 
1478 static void __hns_roce_function_clear(struct hns_roce_dev *hr_dev, int vf_id)
1479 {
1480 	bool fclr_write_fail_flag = false;
1481 	struct hns_roce_func_clear *resp;
1482 	struct hns_roce_cmq_desc desc;
1483 	unsigned long end;
1484 	int ret = 0;
1485 
1486 	if (check_device_is_in_reset(hr_dev))
1487 		goto out;
1488 
1489 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR, false);
1490 	resp = (struct hns_roce_func_clear *)desc.data;
1491 	resp->rst_funcid_en = cpu_to_le32(vf_id);
1492 
1493 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1494 	if (ret) {
1495 		fclr_write_fail_flag = true;
1496 		dev_err(hr_dev->dev, "func clear write failed, ret = %d.\n",
1497 			 ret);
1498 		goto out;
1499 	}
1500 
1501 	msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_INTERVAL);
1502 	end = HNS_ROCE_V2_FUNC_CLEAR_TIMEOUT_MSECS;
1503 	while (end) {
1504 		if (check_device_is_in_reset(hr_dev))
1505 			goto out;
1506 		msleep(HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT);
1507 		end -= HNS_ROCE_V2_READ_FUNC_CLEAR_FLAG_FAIL_WAIT;
1508 
1509 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_FUNC_CLEAR,
1510 					      true);
1511 
1512 		resp->rst_funcid_en = cpu_to_le32(vf_id);
1513 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1514 		if (ret)
1515 			continue;
1516 
1517 		if (hr_reg_read(resp, FUNC_CLEAR_RST_FUN_DONE)) {
1518 			if (vf_id == 0)
1519 				hr_dev->is_reset = true;
1520 			return;
1521 		}
1522 	}
1523 
1524 out:
1525 	hns_roce_func_clr_rst_proc(hr_dev, ret, fclr_write_fail_flag);
1526 }
1527 
1528 static int hns_roce_free_vf_resource(struct hns_roce_dev *hr_dev, int vf_id)
1529 {
1530 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1531 	struct hns_roce_cmq_desc desc[2];
1532 	struct hns_roce_cmq_req *req_a;
1533 
1534 	req_a = (struct hns_roce_cmq_req *)desc[0].data;
1535 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1536 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1537 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1538 	hr_reg_write(req_a, FUNC_RES_A_VF_ID, vf_id);
1539 
1540 	return hns_roce_cmq_send(hr_dev, desc, 2);
1541 }
1542 
1543 static void hns_roce_function_clear(struct hns_roce_dev *hr_dev)
1544 {
1545 	int ret;
1546 	int i;
1547 
1548 	if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
1549 		return;
1550 
1551 	for (i = hr_dev->func_num - 1; i >= 0; i--) {
1552 		__hns_roce_function_clear(hr_dev, i);
1553 
1554 		if (i == 0)
1555 			continue;
1556 
1557 		ret = hns_roce_free_vf_resource(hr_dev, i);
1558 		if (ret)
1559 			ibdev_err(&hr_dev->ib_dev,
1560 				  "failed to free vf resource, vf_id = %d, ret = %d.\n",
1561 				  i, ret);
1562 	}
1563 }
1564 
1565 static int hns_roce_clear_extdb_list_info(struct hns_roce_dev *hr_dev)
1566 {
1567 	struct hns_roce_cmq_desc desc;
1568 	int ret;
1569 
1570 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLEAR_EXTDB_LIST_INFO,
1571 				      false);
1572 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1573 	if (ret)
1574 		ibdev_err(&hr_dev->ib_dev,
1575 			  "failed to clear extended doorbell info, ret = %d.\n",
1576 			  ret);
1577 
1578 	return ret;
1579 }
1580 
1581 static int hns_roce_query_fw_ver(struct hns_roce_dev *hr_dev)
1582 {
1583 	struct hns_roce_query_fw_info *resp;
1584 	struct hns_roce_cmq_desc desc;
1585 	int ret;
1586 
1587 	hns_roce_cmq_setup_basic_desc(&desc, HNS_QUERY_FW_VER, true);
1588 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1589 	if (ret)
1590 		return ret;
1591 
1592 	resp = (struct hns_roce_query_fw_info *)desc.data;
1593 	hr_dev->caps.fw_ver = (u64)(le32_to_cpu(resp->fw_ver));
1594 
1595 	return 0;
1596 }
1597 
1598 static int hns_roce_query_func_info(struct hns_roce_dev *hr_dev)
1599 {
1600 	struct hns_roce_cmq_desc desc;
1601 	int ret;
1602 
1603 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
1604 		hr_dev->func_num = 1;
1605 		return 0;
1606 	}
1607 
1608 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_FUNC_INFO,
1609 				      true);
1610 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1611 	if (ret) {
1612 		hr_dev->func_num = 1;
1613 		return ret;
1614 	}
1615 
1616 	hr_dev->func_num = le32_to_cpu(desc.func_info.own_func_num);
1617 	hr_dev->cong_algo_tmpl_id = le32_to_cpu(desc.func_info.own_mac_id);
1618 
1619 	return 0;
1620 }
1621 
1622 static int hns_roce_hw_v2_query_counter(struct hns_roce_dev *hr_dev,
1623 					u64 *stats, u32 port, int *num_counters)
1624 {
1625 #define CNT_PER_DESC 3
1626 	struct hns_roce_cmq_desc *desc;
1627 	int bd_idx, cnt_idx;
1628 	__le64 *cnt_data;
1629 	int desc_num;
1630 	int ret;
1631 	int i;
1632 
1633 	if (port > hr_dev->caps.num_ports)
1634 		return -EINVAL;
1635 
1636 	desc_num = DIV_ROUND_UP(HNS_ROCE_HW_CNT_TOTAL, CNT_PER_DESC);
1637 	desc = kcalloc(desc_num, sizeof(*desc), GFP_KERNEL);
1638 	if (!desc)
1639 		return -ENOMEM;
1640 
1641 	for (i = 0; i < desc_num; i++) {
1642 		hns_roce_cmq_setup_basic_desc(&desc[i],
1643 					      HNS_ROCE_OPC_QUERY_COUNTER, true);
1644 		if (i != desc_num - 1)
1645 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1646 	}
1647 
1648 	ret = hns_roce_cmq_send(hr_dev, desc, desc_num);
1649 	if (ret) {
1650 		ibdev_err(&hr_dev->ib_dev,
1651 			  "failed to get counter, ret = %d.\n", ret);
1652 		goto err_out;
1653 	}
1654 
1655 	for (i = 0; i < HNS_ROCE_HW_CNT_TOTAL && i < *num_counters; i++) {
1656 		bd_idx = i / CNT_PER_DESC;
1657 		if (!(desc[bd_idx].flag & HNS_ROCE_CMD_FLAG_NEXT) &&
1658 		    bd_idx != HNS_ROCE_HW_CNT_TOTAL / CNT_PER_DESC)
1659 			break;
1660 
1661 		cnt_data = (__le64 *)&desc[bd_idx].data[0];
1662 		cnt_idx = i % CNT_PER_DESC;
1663 		stats[i] = le64_to_cpu(cnt_data[cnt_idx]);
1664 	}
1665 	*num_counters = i;
1666 
1667 err_out:
1668 	kfree(desc);
1669 	return ret;
1670 }
1671 
1672 static int hns_roce_config_global_param(struct hns_roce_dev *hr_dev)
1673 {
1674 	struct hns_roce_cmq_desc desc;
1675 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1676 	u32 clock_cycles_of_1us;
1677 
1678 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GLOBAL_PARAM,
1679 				      false);
1680 
1681 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
1682 		clock_cycles_of_1us = HNS_ROCE_1NS_CFG;
1683 	else
1684 		clock_cycles_of_1us = HNS_ROCE_1US_CFG;
1685 
1686 	hr_reg_write(req, CFG_GLOBAL_PARAM_1US_CYCLES, clock_cycles_of_1us);
1687 	hr_reg_write(req, CFG_GLOBAL_PARAM_UDP_PORT, ROCE_V2_UDP_DPORT);
1688 
1689 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1690 }
1691 
1692 static int load_func_res_caps(struct hns_roce_dev *hr_dev, bool is_vf)
1693 {
1694 	struct hns_roce_cmq_desc desc[2];
1695 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1696 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1697 	struct hns_roce_caps *caps = &hr_dev->caps;
1698 	enum hns_roce_opcode_type opcode;
1699 	u32 func_num;
1700 	int ret;
1701 
1702 	if (is_vf) {
1703 		opcode = HNS_ROCE_OPC_QUERY_VF_RES;
1704 		func_num = 1;
1705 	} else {
1706 		opcode = HNS_ROCE_OPC_QUERY_PF_RES;
1707 		func_num = hr_dev->func_num;
1708 	}
1709 
1710 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, true);
1711 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1712 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, true);
1713 
1714 	ret = hns_roce_cmq_send(hr_dev, desc, 2);
1715 	if (ret)
1716 		return ret;
1717 
1718 	caps->qpc_bt_num = hr_reg_read(r_a, FUNC_RES_A_QPC_BT_NUM) / func_num;
1719 	caps->srqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_SRQC_BT_NUM) / func_num;
1720 	caps->cqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_CQC_BT_NUM) / func_num;
1721 	caps->mpt_bt_num = hr_reg_read(r_a, FUNC_RES_A_MPT_BT_NUM) / func_num;
1722 	caps->eqc_bt_num = hr_reg_read(r_a, FUNC_RES_A_EQC_BT_NUM) / func_num;
1723 	caps->smac_bt_num = hr_reg_read(r_b, FUNC_RES_B_SMAC_NUM) / func_num;
1724 	caps->sgid_bt_num = hr_reg_read(r_b, FUNC_RES_B_SGID_NUM) / func_num;
1725 	caps->sccc_bt_num = hr_reg_read(r_b, FUNC_RES_B_SCCC_BT_NUM) / func_num;
1726 
1727 	if (is_vf) {
1728 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_V_QID_NUM) / func_num;
1729 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_V_GMV_BT_NUM) /
1730 					       func_num;
1731 	} else {
1732 		caps->sl_num = hr_reg_read(r_b, FUNC_RES_B_QID_NUM) / func_num;
1733 		caps->gmv_bt_num = hr_reg_read(r_b, FUNC_RES_B_GMV_BT_NUM) /
1734 					       func_num;
1735 	}
1736 
1737 	return 0;
1738 }
1739 
1740 static int load_pf_timer_res_caps(struct hns_roce_dev *hr_dev)
1741 {
1742 	struct hns_roce_cmq_desc desc;
1743 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1744 	struct hns_roce_caps *caps = &hr_dev->caps;
1745 	int ret;
1746 
1747 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_PF_TIMER_RES,
1748 				      true);
1749 
1750 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1751 	if (ret)
1752 		return ret;
1753 
1754 	caps->qpc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_QPC_ITEM_NUM);
1755 	caps->cqc_timer_bt_num = hr_reg_read(req, PF_TIMER_RES_CQC_ITEM_NUM);
1756 
1757 	return 0;
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 = load_func_res_caps(hr_dev, false);
1766 	if (ret) {
1767 		dev_err(dev, "failed to load pf res caps, ret = %d.\n", ret);
1768 		return ret;
1769 	}
1770 
1771 	ret = load_pf_timer_res_caps(hr_dev);
1772 	if (ret)
1773 		dev_err(dev, "failed to load pf timer resource, ret = %d.\n",
1774 			ret);
1775 
1776 	return ret;
1777 }
1778 
1779 static int hns_roce_query_vf_resource(struct hns_roce_dev *hr_dev)
1780 {
1781 	struct device *dev = hr_dev->dev;
1782 	int ret;
1783 
1784 	ret = load_func_res_caps(hr_dev, true);
1785 	if (ret)
1786 		dev_err(dev, "failed to load vf res caps, ret = %d.\n", ret);
1787 
1788 	return ret;
1789 }
1790 
1791 static int __hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev,
1792 					  u32 vf_id)
1793 {
1794 	struct hns_roce_vf_switch *swt;
1795 	struct hns_roce_cmq_desc desc;
1796 	int ret;
1797 
1798 	swt = (struct hns_roce_vf_switch *)desc.data;
1799 	hns_roce_cmq_setup_basic_desc(&desc, HNS_SWITCH_PARAMETER_CFG, true);
1800 	swt->rocee_sel |= cpu_to_le32(HNS_ICL_SWITCH_CMD_ROCEE_SEL);
1801 	hr_reg_write(swt, VF_SWITCH_VF_ID, vf_id);
1802 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
1803 	if (ret)
1804 		return ret;
1805 
1806 	desc.flag = cpu_to_le16(HNS_ROCE_CMD_FLAG_IN);
1807 	desc.flag &= cpu_to_le16(~HNS_ROCE_CMD_FLAG_WR);
1808 	hr_reg_enable(swt, VF_SWITCH_ALW_LPBK);
1809 	hr_reg_clear(swt, VF_SWITCH_ALW_LCL_LPBK);
1810 	hr_reg_enable(swt, VF_SWITCH_ALW_DST_OVRD);
1811 
1812 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1813 }
1814 
1815 static int hns_roce_set_vf_switch_param(struct hns_roce_dev *hr_dev)
1816 {
1817 	u32 vf_id;
1818 	int ret;
1819 
1820 	for (vf_id = 0; vf_id < hr_dev->func_num; vf_id++) {
1821 		ret = __hns_roce_set_vf_switch_param(hr_dev, vf_id);
1822 		if (ret)
1823 			return ret;
1824 	}
1825 	return 0;
1826 }
1827 
1828 static int config_vf_hem_resource(struct hns_roce_dev *hr_dev, int vf_id)
1829 {
1830 	struct hns_roce_cmq_desc desc[2];
1831 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
1832 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
1833 	enum hns_roce_opcode_type opcode = HNS_ROCE_OPC_ALLOC_VF_RES;
1834 	struct hns_roce_caps *caps = &hr_dev->caps;
1835 
1836 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
1837 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
1838 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
1839 
1840 	hr_reg_write(r_a, FUNC_RES_A_VF_ID, vf_id);
1841 
1842 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_NUM, caps->qpc_bt_num);
1843 	hr_reg_write(r_a, FUNC_RES_A_QPC_BT_IDX, vf_id * caps->qpc_bt_num);
1844 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_NUM, caps->srqc_bt_num);
1845 	hr_reg_write(r_a, FUNC_RES_A_SRQC_BT_IDX, vf_id * caps->srqc_bt_num);
1846 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_NUM, caps->cqc_bt_num);
1847 	hr_reg_write(r_a, FUNC_RES_A_CQC_BT_IDX, vf_id * caps->cqc_bt_num);
1848 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_NUM, caps->mpt_bt_num);
1849 	hr_reg_write(r_a, FUNC_RES_A_MPT_BT_IDX, vf_id * caps->mpt_bt_num);
1850 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_NUM, caps->eqc_bt_num);
1851 	hr_reg_write(r_a, FUNC_RES_A_EQC_BT_IDX, vf_id * caps->eqc_bt_num);
1852 	hr_reg_write(r_b, FUNC_RES_V_QID_NUM, caps->sl_num);
1853 	hr_reg_write(r_b, FUNC_RES_B_QID_IDX, vf_id * caps->sl_num);
1854 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_NUM, caps->sccc_bt_num);
1855 	hr_reg_write(r_b, FUNC_RES_B_SCCC_BT_IDX, vf_id * caps->sccc_bt_num);
1856 
1857 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
1858 		hr_reg_write(r_b, FUNC_RES_V_GMV_BT_NUM, caps->gmv_bt_num);
1859 		hr_reg_write(r_b, FUNC_RES_B_GMV_BT_IDX,
1860 			     vf_id * caps->gmv_bt_num);
1861 	} else {
1862 		hr_reg_write(r_b, FUNC_RES_B_SGID_NUM, caps->sgid_bt_num);
1863 		hr_reg_write(r_b, FUNC_RES_B_SGID_IDX,
1864 			     vf_id * caps->sgid_bt_num);
1865 		hr_reg_write(r_b, FUNC_RES_B_SMAC_NUM, caps->smac_bt_num);
1866 		hr_reg_write(r_b, FUNC_RES_B_SMAC_IDX,
1867 			     vf_id * caps->smac_bt_num);
1868 	}
1869 
1870 	return hns_roce_cmq_send(hr_dev, desc, 2);
1871 }
1872 
1873 static int hns_roce_alloc_vf_resource(struct hns_roce_dev *hr_dev)
1874 {
1875 	u32 func_num = max_t(u32, 1, hr_dev->func_num);
1876 	u32 vf_id;
1877 	int ret;
1878 
1879 	for (vf_id = 0; vf_id < func_num; vf_id++) {
1880 		ret = config_vf_hem_resource(hr_dev, vf_id);
1881 		if (ret) {
1882 			dev_err(hr_dev->dev,
1883 				"failed to config vf-%u hem res, ret = %d.\n",
1884 				vf_id, ret);
1885 			return ret;
1886 		}
1887 	}
1888 
1889 	return 0;
1890 }
1891 
1892 static int hns_roce_v2_set_bt(struct hns_roce_dev *hr_dev)
1893 {
1894 	struct hns_roce_cmq_desc desc;
1895 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
1896 	struct hns_roce_caps *caps = &hr_dev->caps;
1897 
1898 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_BT_ATTR, false);
1899 
1900 	hr_reg_write(req, CFG_BT_ATTR_QPC_BA_PGSZ,
1901 		     caps->qpc_ba_pg_sz + PG_SHIFT_OFFSET);
1902 	hr_reg_write(req, CFG_BT_ATTR_QPC_BUF_PGSZ,
1903 		     caps->qpc_buf_pg_sz + PG_SHIFT_OFFSET);
1904 	hr_reg_write(req, CFG_BT_ATTR_QPC_HOPNUM,
1905 		     to_hr_hem_hopnum(caps->qpc_hop_num, caps->num_qps));
1906 
1907 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BA_PGSZ,
1908 		     caps->srqc_ba_pg_sz + PG_SHIFT_OFFSET);
1909 	hr_reg_write(req, CFG_BT_ATTR_SRQC_BUF_PGSZ,
1910 		     caps->srqc_buf_pg_sz + PG_SHIFT_OFFSET);
1911 	hr_reg_write(req, CFG_BT_ATTR_SRQC_HOPNUM,
1912 		     to_hr_hem_hopnum(caps->srqc_hop_num, caps->num_srqs));
1913 
1914 	hr_reg_write(req, CFG_BT_ATTR_CQC_BA_PGSZ,
1915 		     caps->cqc_ba_pg_sz + PG_SHIFT_OFFSET);
1916 	hr_reg_write(req, CFG_BT_ATTR_CQC_BUF_PGSZ,
1917 		     caps->cqc_buf_pg_sz + PG_SHIFT_OFFSET);
1918 	hr_reg_write(req, CFG_BT_ATTR_CQC_HOPNUM,
1919 		     to_hr_hem_hopnum(caps->cqc_hop_num, caps->num_cqs));
1920 
1921 	hr_reg_write(req, CFG_BT_ATTR_MPT_BA_PGSZ,
1922 		     caps->mpt_ba_pg_sz + PG_SHIFT_OFFSET);
1923 	hr_reg_write(req, CFG_BT_ATTR_MPT_BUF_PGSZ,
1924 		     caps->mpt_buf_pg_sz + PG_SHIFT_OFFSET);
1925 	hr_reg_write(req, CFG_BT_ATTR_MPT_HOPNUM,
1926 		     to_hr_hem_hopnum(caps->mpt_hop_num, caps->num_mtpts));
1927 
1928 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BA_PGSZ,
1929 		     caps->sccc_ba_pg_sz + PG_SHIFT_OFFSET);
1930 	hr_reg_write(req, CFG_BT_ATTR_SCCC_BUF_PGSZ,
1931 		     caps->sccc_buf_pg_sz + PG_SHIFT_OFFSET);
1932 	hr_reg_write(req, CFG_BT_ATTR_SCCC_HOPNUM,
1933 		     to_hr_hem_hopnum(caps->sccc_hop_num, caps->num_qps));
1934 
1935 	return hns_roce_cmq_send(hr_dev, &desc, 1);
1936 }
1937 
1938 static void calc_pg_sz(u32 obj_num, u32 obj_size, u32 hop_num, u32 ctx_bt_num,
1939 		       u32 *buf_page_size, u32 *bt_page_size, u32 hem_type)
1940 {
1941 	u64 obj_per_chunk;
1942 	u64 bt_chunk_size = PAGE_SIZE;
1943 	u64 buf_chunk_size = PAGE_SIZE;
1944 	u64 obj_per_chunk_default = buf_chunk_size / obj_size;
1945 
1946 	*buf_page_size = 0;
1947 	*bt_page_size = 0;
1948 
1949 	switch (hop_num) {
1950 	case 3:
1951 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1952 				(bt_chunk_size / BA_BYTE_LEN) *
1953 				(bt_chunk_size / BA_BYTE_LEN) *
1954 				 obj_per_chunk_default;
1955 		break;
1956 	case 2:
1957 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1958 				(bt_chunk_size / BA_BYTE_LEN) *
1959 				 obj_per_chunk_default;
1960 		break;
1961 	case 1:
1962 		obj_per_chunk = ctx_bt_num * (bt_chunk_size / BA_BYTE_LEN) *
1963 				obj_per_chunk_default;
1964 		break;
1965 	case HNS_ROCE_HOP_NUM_0:
1966 		obj_per_chunk = ctx_bt_num * obj_per_chunk_default;
1967 		break;
1968 	default:
1969 		pr_err("table %u not support hop_num = %u!\n", hem_type,
1970 		       hop_num);
1971 		return;
1972 	}
1973 
1974 	if (hem_type >= HEM_TYPE_MTT)
1975 		*bt_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
1976 	else
1977 		*buf_page_size = ilog2(DIV_ROUND_UP(obj_num, obj_per_chunk));
1978 }
1979 
1980 static void set_hem_page_size(struct hns_roce_dev *hr_dev)
1981 {
1982 	struct hns_roce_caps *caps = &hr_dev->caps;
1983 
1984 	/* EQ */
1985 	caps->eqe_ba_pg_sz = 0;
1986 	caps->eqe_buf_pg_sz = 0;
1987 
1988 	/* Link Table */
1989 	caps->llm_buf_pg_sz = 0;
1990 
1991 	/* MR */
1992 	caps->mpt_ba_pg_sz = 0;
1993 	caps->mpt_buf_pg_sz = 0;
1994 	caps->pbl_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_16K;
1995 	caps->pbl_buf_pg_sz = 0;
1996 	calc_pg_sz(caps->num_mtpts, caps->mtpt_entry_sz, caps->mpt_hop_num,
1997 		   caps->mpt_bt_num, &caps->mpt_buf_pg_sz, &caps->mpt_ba_pg_sz,
1998 		   HEM_TYPE_MTPT);
1999 
2000 	/* QP */
2001 	caps->qpc_ba_pg_sz = 0;
2002 	caps->qpc_buf_pg_sz = 0;
2003 	caps->qpc_timer_ba_pg_sz = 0;
2004 	caps->qpc_timer_buf_pg_sz = 0;
2005 	caps->sccc_ba_pg_sz = 0;
2006 	caps->sccc_buf_pg_sz = 0;
2007 	caps->mtt_ba_pg_sz = 0;
2008 	caps->mtt_buf_pg_sz = 0;
2009 	calc_pg_sz(caps->num_qps, caps->qpc_sz, caps->qpc_hop_num,
2010 		   caps->qpc_bt_num, &caps->qpc_buf_pg_sz, &caps->qpc_ba_pg_sz,
2011 		   HEM_TYPE_QPC);
2012 
2013 	if (caps->flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
2014 		calc_pg_sz(caps->num_qps, caps->sccc_sz, caps->sccc_hop_num,
2015 			   caps->sccc_bt_num, &caps->sccc_buf_pg_sz,
2016 			   &caps->sccc_ba_pg_sz, HEM_TYPE_SCCC);
2017 
2018 	/* CQ */
2019 	caps->cqc_ba_pg_sz = 0;
2020 	caps->cqc_buf_pg_sz = 0;
2021 	caps->cqc_timer_ba_pg_sz = 0;
2022 	caps->cqc_timer_buf_pg_sz = 0;
2023 	caps->cqe_ba_pg_sz = HNS_ROCE_BA_PG_SZ_SUPPORTED_256K;
2024 	caps->cqe_buf_pg_sz = 0;
2025 	calc_pg_sz(caps->num_cqs, caps->cqc_entry_sz, caps->cqc_hop_num,
2026 		   caps->cqc_bt_num, &caps->cqc_buf_pg_sz, &caps->cqc_ba_pg_sz,
2027 		   HEM_TYPE_CQC);
2028 	calc_pg_sz(caps->max_cqes, caps->cqe_sz, caps->cqe_hop_num,
2029 		   1, &caps->cqe_buf_pg_sz, &caps->cqe_ba_pg_sz, HEM_TYPE_CQE);
2030 
2031 	/* SRQ */
2032 	if (caps->flags & HNS_ROCE_CAP_FLAG_SRQ) {
2033 		caps->srqc_ba_pg_sz = 0;
2034 		caps->srqc_buf_pg_sz = 0;
2035 		caps->srqwqe_ba_pg_sz = 0;
2036 		caps->srqwqe_buf_pg_sz = 0;
2037 		caps->idx_ba_pg_sz = 0;
2038 		caps->idx_buf_pg_sz = 0;
2039 		calc_pg_sz(caps->num_srqs, caps->srqc_entry_sz,
2040 			   caps->srqc_hop_num, caps->srqc_bt_num,
2041 			   &caps->srqc_buf_pg_sz, &caps->srqc_ba_pg_sz,
2042 			   HEM_TYPE_SRQC);
2043 		calc_pg_sz(caps->num_srqwqe_segs, caps->mtt_entry_sz,
2044 			   caps->srqwqe_hop_num, 1, &caps->srqwqe_buf_pg_sz,
2045 			   &caps->srqwqe_ba_pg_sz, HEM_TYPE_SRQWQE);
2046 		calc_pg_sz(caps->num_idx_segs, caps->idx_entry_sz,
2047 			   caps->idx_hop_num, 1, &caps->idx_buf_pg_sz,
2048 			   &caps->idx_ba_pg_sz, HEM_TYPE_IDX);
2049 	}
2050 
2051 	/* GMV */
2052 	caps->gmv_ba_pg_sz = 0;
2053 	caps->gmv_buf_pg_sz = 0;
2054 }
2055 
2056 /* Apply all loaded caps before setting to hardware */
2057 static void apply_func_caps(struct hns_roce_dev *hr_dev)
2058 {
2059 #define MAX_GID_TBL_LEN 256
2060 	struct hns_roce_caps *caps = &hr_dev->caps;
2061 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2062 
2063 	/* The following configurations don't need to be got from firmware. */
2064 	caps->qpc_timer_entry_sz = HNS_ROCE_V2_QPC_TIMER_ENTRY_SZ;
2065 	caps->cqc_timer_entry_sz = HNS_ROCE_V2_CQC_TIMER_ENTRY_SZ;
2066 	caps->mtt_entry_sz = HNS_ROCE_V2_MTT_ENTRY_SZ;
2067 
2068 	caps->pbl_hop_num = HNS_ROCE_PBL_HOP_NUM;
2069 	caps->qpc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2070 	caps->cqc_timer_hop_num = HNS_ROCE_HOP_NUM_0;
2071 
2072 	caps->num_srqwqe_segs = HNS_ROCE_V2_MAX_SRQWQE_SEGS;
2073 	caps->num_idx_segs = HNS_ROCE_V2_MAX_IDX_SEGS;
2074 
2075 	if (!caps->num_comp_vectors)
2076 		caps->num_comp_vectors =
2077 			min_t(u32, caps->eqc_bt_num - HNS_ROCE_V2_AEQE_VEC_NUM,
2078 				(u32)priv->handle->rinfo.num_vectors -
2079 		(HNS_ROCE_V2_AEQE_VEC_NUM + HNS_ROCE_V2_ABNORMAL_VEC_NUM));
2080 
2081 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
2082 		caps->eqe_hop_num = HNS_ROCE_V3_EQE_HOP_NUM;
2083 		caps->ceqe_size = HNS_ROCE_V3_EQE_SIZE;
2084 		caps->aeqe_size = HNS_ROCE_V3_EQE_SIZE;
2085 
2086 		/* The following configurations will be overwritten */
2087 		caps->qpc_sz = HNS_ROCE_V3_QPC_SZ;
2088 		caps->cqe_sz = HNS_ROCE_V3_CQE_SIZE;
2089 		caps->sccc_sz = HNS_ROCE_V3_SCCC_SZ;
2090 
2091 		/* The following configurations are not got from firmware */
2092 		caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;
2093 
2094 		caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0;
2095 
2096 		/* It's meaningless to support excessively large gid_table_len,
2097 		 * as the type of sgid_index in kernel struct ib_global_route
2098 		 * and userspace struct ibv_global_route are u8/uint8_t (0-255).
2099 		 */
2100 		caps->gid_table_len[0] = min_t(u32, MAX_GID_TBL_LEN,
2101 					 caps->gmv_bt_num *
2102 					 (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz));
2103 
2104 		caps->gmv_entry_num = caps->gmv_bt_num * (HNS_HW_PAGE_SIZE /
2105 							  caps->gmv_entry_sz);
2106 	} else {
2107 		u32 func_num = max_t(u32, 1, hr_dev->func_num);
2108 
2109 		caps->eqe_hop_num = HNS_ROCE_V2_EQE_HOP_NUM;
2110 		caps->ceqe_size = HNS_ROCE_CEQE_SIZE;
2111 		caps->aeqe_size = HNS_ROCE_AEQE_SIZE;
2112 		caps->gid_table_len[0] /= func_num;
2113 	}
2114 
2115 	if (hr_dev->is_vf) {
2116 		caps->default_aeq_arm_st = 0x3;
2117 		caps->default_ceq_arm_st = 0x3;
2118 		caps->default_ceq_max_cnt = 0x1;
2119 		caps->default_ceq_period = 0x10;
2120 		caps->default_aeq_max_cnt = 0x1;
2121 		caps->default_aeq_period = 0x10;
2122 	}
2123 
2124 	set_hem_page_size(hr_dev);
2125 }
2126 
2127 static int hns_roce_query_caps(struct hns_roce_dev *hr_dev)
2128 {
2129 	struct hns_roce_cmq_desc desc[HNS_ROCE_QUERY_PF_CAPS_CMD_NUM];
2130 	struct hns_roce_caps *caps = &hr_dev->caps;
2131 	struct hns_roce_query_pf_caps_a *resp_a;
2132 	struct hns_roce_query_pf_caps_b *resp_b;
2133 	struct hns_roce_query_pf_caps_c *resp_c;
2134 	struct hns_roce_query_pf_caps_d *resp_d;
2135 	struct hns_roce_query_pf_caps_e *resp_e;
2136 	enum hns_roce_opcode_type cmd;
2137 	int ctx_hop_num;
2138 	int pbl_hop_num;
2139 	int ret;
2140 	int i;
2141 
2142 	cmd = hr_dev->is_vf ? HNS_ROCE_OPC_QUERY_VF_CAPS_NUM :
2143 	      HNS_ROCE_OPC_QUERY_PF_CAPS_NUM;
2144 
2145 	for (i = 0; i < HNS_ROCE_QUERY_PF_CAPS_CMD_NUM; i++) {
2146 		hns_roce_cmq_setup_basic_desc(&desc[i], cmd, true);
2147 		if (i < (HNS_ROCE_QUERY_PF_CAPS_CMD_NUM - 1))
2148 			desc[i].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2149 		else
2150 			desc[i].flag &= ~cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2151 	}
2152 
2153 	ret = hns_roce_cmq_send(hr_dev, desc, HNS_ROCE_QUERY_PF_CAPS_CMD_NUM);
2154 	if (ret)
2155 		return ret;
2156 
2157 	resp_a = (struct hns_roce_query_pf_caps_a *)desc[0].data;
2158 	resp_b = (struct hns_roce_query_pf_caps_b *)desc[1].data;
2159 	resp_c = (struct hns_roce_query_pf_caps_c *)desc[2].data;
2160 	resp_d = (struct hns_roce_query_pf_caps_d *)desc[3].data;
2161 	resp_e = (struct hns_roce_query_pf_caps_e *)desc[4].data;
2162 
2163 	caps->local_ca_ack_delay = resp_a->local_ca_ack_delay;
2164 	caps->max_sq_sg = le16_to_cpu(resp_a->max_sq_sg);
2165 	caps->max_sq_inline = le16_to_cpu(resp_a->max_sq_inline);
2166 	caps->max_rq_sg = le16_to_cpu(resp_a->max_rq_sg);
2167 	caps->max_rq_sg = roundup_pow_of_two(caps->max_rq_sg);
2168 	caps->max_srq_sges = le16_to_cpu(resp_a->max_srq_sges);
2169 	caps->max_srq_sges = roundup_pow_of_two(caps->max_srq_sges);
2170 	caps->num_aeq_vectors = resp_a->num_aeq_vectors;
2171 	caps->num_other_vectors = resp_a->num_other_vectors;
2172 	caps->max_sq_desc_sz = resp_a->max_sq_desc_sz;
2173 	caps->max_rq_desc_sz = resp_a->max_rq_desc_sz;
2174 
2175 	caps->mtpt_entry_sz = resp_b->mtpt_entry_sz;
2176 	caps->irrl_entry_sz = resp_b->irrl_entry_sz;
2177 	caps->trrl_entry_sz = resp_b->trrl_entry_sz;
2178 	caps->cqc_entry_sz = resp_b->cqc_entry_sz;
2179 	caps->srqc_entry_sz = resp_b->srqc_entry_sz;
2180 	caps->idx_entry_sz = resp_b->idx_entry_sz;
2181 	caps->sccc_sz = resp_b->sccc_sz;
2182 	caps->max_mtu = resp_b->max_mtu;
2183 	caps->min_cqes = resp_b->min_cqes;
2184 	caps->min_wqes = resp_b->min_wqes;
2185 	caps->page_size_cap = le32_to_cpu(resp_b->page_size_cap);
2186 	caps->pkey_table_len[0] = resp_b->pkey_table_len;
2187 	caps->phy_num_uars = resp_b->phy_num_uars;
2188 	ctx_hop_num = resp_b->ctx_hop_num;
2189 	pbl_hop_num = resp_b->pbl_hop_num;
2190 
2191 	caps->num_pds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_PDS);
2192 
2193 	caps->flags = hr_reg_read(resp_c, PF_CAPS_C_CAP_FLAGS);
2194 	caps->flags |= le16_to_cpu(resp_d->cap_flags_ex) <<
2195 		       HNS_ROCE_CAP_FLAGS_EX_SHIFT;
2196 
2197 	caps->num_cqs = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_CQS);
2198 	caps->gid_table_len[0] = hr_reg_read(resp_c, PF_CAPS_C_MAX_GID);
2199 	caps->max_cqes = 1 << hr_reg_read(resp_c, PF_CAPS_C_CQ_DEPTH);
2200 	caps->num_xrcds = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_XRCDS);
2201 	caps->num_mtpts = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_MRWS);
2202 	caps->num_qps = 1 << hr_reg_read(resp_c, PF_CAPS_C_NUM_QPS);
2203 	caps->max_qp_init_rdma = hr_reg_read(resp_c, PF_CAPS_C_MAX_ORD);
2204 	caps->max_qp_dest_rdma = caps->max_qp_init_rdma;
2205 	caps->max_wqes = 1 << le16_to_cpu(resp_c->sq_depth);
2206 
2207 	caps->num_srqs = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_SRQS);
2208 	caps->cong_cap = hr_reg_read(resp_d, PF_CAPS_D_CONG_CAP);
2209 	caps->max_srq_wrs = 1 << le16_to_cpu(resp_d->srq_depth);
2210 	caps->ceqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_CEQ_DEPTH);
2211 	caps->num_comp_vectors = hr_reg_read(resp_d, PF_CAPS_D_NUM_CEQS);
2212 	caps->aeqe_depth = 1 << hr_reg_read(resp_d, PF_CAPS_D_AEQ_DEPTH);
2213 	caps->default_cong_type = hr_reg_read(resp_d, PF_CAPS_D_DEFAULT_ALG);
2214 	caps->reserved_pds = hr_reg_read(resp_d, PF_CAPS_D_RSV_PDS);
2215 	caps->num_uars = 1 << hr_reg_read(resp_d, PF_CAPS_D_NUM_UARS);
2216 	caps->reserved_qps = hr_reg_read(resp_d, PF_CAPS_D_RSV_QPS);
2217 	caps->reserved_uars = hr_reg_read(resp_d, PF_CAPS_D_RSV_UARS);
2218 
2219 	caps->reserved_mrws = hr_reg_read(resp_e, PF_CAPS_E_RSV_MRWS);
2220 	caps->chunk_sz = 1 << hr_reg_read(resp_e, PF_CAPS_E_CHUNK_SIZE_SHIFT);
2221 	caps->reserved_cqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_CQS);
2222 	caps->reserved_xrcds = hr_reg_read(resp_e, PF_CAPS_E_RSV_XRCDS);
2223 	caps->reserved_srqs = hr_reg_read(resp_e, PF_CAPS_E_RSV_SRQS);
2224 	caps->reserved_lkey = hr_reg_read(resp_e, PF_CAPS_E_RSV_LKEYS);
2225 
2226 	caps->qpc_hop_num = ctx_hop_num;
2227 	caps->sccc_hop_num = ctx_hop_num;
2228 	caps->srqc_hop_num = ctx_hop_num;
2229 	caps->cqc_hop_num = ctx_hop_num;
2230 	caps->mpt_hop_num = ctx_hop_num;
2231 	caps->mtt_hop_num = pbl_hop_num;
2232 	caps->cqe_hop_num = pbl_hop_num;
2233 	caps->srqwqe_hop_num = pbl_hop_num;
2234 	caps->idx_hop_num = pbl_hop_num;
2235 	caps->wqe_sq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_SQWQE_HOP_NUM);
2236 	caps->wqe_sge_hop_num = hr_reg_read(resp_d, PF_CAPS_D_EX_SGE_HOP_NUM);
2237 	caps->wqe_rq_hop_num = hr_reg_read(resp_d, PF_CAPS_D_RQWQE_HOP_NUM);
2238 
2239 	if (!(caps->page_size_cap & PAGE_SIZE))
2240 		caps->page_size_cap = HNS_ROCE_V2_PAGE_SIZE_SUPPORTED;
2241 
2242 	if (!hr_dev->is_vf) {
2243 		caps->cqe_sz = resp_a->cqe_sz;
2244 		caps->qpc_sz = le16_to_cpu(resp_b->qpc_sz);
2245 		caps->default_aeq_arm_st =
2246 				hr_reg_read(resp_d, PF_CAPS_D_AEQ_ARM_ST);
2247 		caps->default_ceq_arm_st =
2248 				hr_reg_read(resp_d, PF_CAPS_D_CEQ_ARM_ST);
2249 		caps->default_ceq_max_cnt = le16_to_cpu(resp_e->ceq_max_cnt);
2250 		caps->default_ceq_period = le16_to_cpu(resp_e->ceq_period);
2251 		caps->default_aeq_max_cnt = le16_to_cpu(resp_e->aeq_max_cnt);
2252 		caps->default_aeq_period = le16_to_cpu(resp_e->aeq_period);
2253 	}
2254 
2255 	return 0;
2256 }
2257 
2258 static int config_hem_entry_size(struct hns_roce_dev *hr_dev, u32 type, u32 val)
2259 {
2260 	struct hns_roce_cmq_desc desc;
2261 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
2262 
2263 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_ENTRY_SIZE,
2264 				      false);
2265 
2266 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_TYPE, type);
2267 	hr_reg_write(req, CFG_HEM_ENTRY_SIZE_VALUE, val);
2268 
2269 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2270 }
2271 
2272 static int hns_roce_config_entry_size(struct hns_roce_dev *hr_dev)
2273 {
2274 	struct hns_roce_caps *caps = &hr_dev->caps;
2275 	int ret;
2276 
2277 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
2278 		return 0;
2279 
2280 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_QPC_SIZE,
2281 				    caps->qpc_sz);
2282 	if (ret) {
2283 		dev_err(hr_dev->dev, "failed to cfg qpc sz, ret = %d.\n", ret);
2284 		return ret;
2285 	}
2286 
2287 	ret = config_hem_entry_size(hr_dev, HNS_ROCE_CFG_SCCC_SIZE,
2288 				    caps->sccc_sz);
2289 	if (ret)
2290 		dev_err(hr_dev->dev, "failed to cfg sccc sz, ret = %d.\n", ret);
2291 
2292 	return ret;
2293 }
2294 
2295 static int hns_roce_v2_vf_profile(struct hns_roce_dev *hr_dev)
2296 {
2297 	struct device *dev = hr_dev->dev;
2298 	int ret;
2299 
2300 	hr_dev->func_num = 1;
2301 
2302 	ret = hns_roce_query_caps(hr_dev);
2303 	if (ret) {
2304 		dev_err(dev, "failed to query VF caps, ret = %d.\n", ret);
2305 		return ret;
2306 	}
2307 
2308 	ret = hns_roce_query_vf_resource(hr_dev);
2309 	if (ret) {
2310 		dev_err(dev, "failed to query VF resource, ret = %d.\n", ret);
2311 		return ret;
2312 	}
2313 
2314 	apply_func_caps(hr_dev);
2315 
2316 	ret = hns_roce_v2_set_bt(hr_dev);
2317 	if (ret)
2318 		dev_err(dev, "failed to config VF BA table, ret = %d.\n", ret);
2319 
2320 	return ret;
2321 }
2322 
2323 static int hns_roce_v2_pf_profile(struct hns_roce_dev *hr_dev)
2324 {
2325 	struct device *dev = hr_dev->dev;
2326 	int ret;
2327 
2328 	ret = hns_roce_query_func_info(hr_dev);
2329 	if (ret) {
2330 		dev_err(dev, "failed to query func info, ret = %d.\n", ret);
2331 		return ret;
2332 	}
2333 
2334 	ret = hns_roce_config_global_param(hr_dev);
2335 	if (ret) {
2336 		dev_err(dev, "failed to config global param, ret = %d.\n", ret);
2337 		return ret;
2338 	}
2339 
2340 	ret = hns_roce_set_vf_switch_param(hr_dev);
2341 	if (ret) {
2342 		dev_err(dev, "failed to set switch param, ret = %d.\n", ret);
2343 		return ret;
2344 	}
2345 
2346 	ret = hns_roce_query_caps(hr_dev);
2347 	if (ret) {
2348 		dev_err(dev, "failed to query PF caps, ret = %d.\n", ret);
2349 		return ret;
2350 	}
2351 
2352 	ret = hns_roce_query_pf_resource(hr_dev);
2353 	if (ret) {
2354 		dev_err(dev, "failed to query pf resource, ret = %d.\n", ret);
2355 		return ret;
2356 	}
2357 
2358 	apply_func_caps(hr_dev);
2359 
2360 	ret = hns_roce_alloc_vf_resource(hr_dev);
2361 	if (ret) {
2362 		dev_err(dev, "failed to alloc vf resource, ret = %d.\n", ret);
2363 		return ret;
2364 	}
2365 
2366 	ret = hns_roce_v2_set_bt(hr_dev);
2367 	if (ret) {
2368 		dev_err(dev, "failed to config BA table, ret = %d.\n", ret);
2369 		return ret;
2370 	}
2371 
2372 	/* Configure the size of QPC, SCCC, etc. */
2373 	return hns_roce_config_entry_size(hr_dev);
2374 }
2375 
2376 static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
2377 {
2378 	struct device *dev = hr_dev->dev;
2379 	int ret;
2380 
2381 	ret = hns_roce_cmq_query_hw_info(hr_dev);
2382 	if (ret) {
2383 		dev_err(dev, "failed to query hardware info, ret = %d.\n", ret);
2384 		return ret;
2385 	}
2386 
2387 	ret = hns_roce_query_fw_ver(hr_dev);
2388 	if (ret) {
2389 		dev_err(dev, "failed to query firmware info, ret = %d.\n", ret);
2390 		return ret;
2391 	}
2392 
2393 	hr_dev->vendor_part_id = hr_dev->pci_dev->device;
2394 	hr_dev->sys_image_guid = be64_to_cpu(hr_dev->ib_dev.node_guid);
2395 
2396 	if (hr_dev->is_vf)
2397 		return hns_roce_v2_vf_profile(hr_dev);
2398 	else
2399 		return hns_roce_v2_pf_profile(hr_dev);
2400 }
2401 
2402 static void config_llm_table(struct hns_roce_buf *data_buf, void *cfg_buf)
2403 {
2404 	u32 i, next_ptr, page_num;
2405 	__le64 *entry = cfg_buf;
2406 	dma_addr_t addr;
2407 	u64 val;
2408 
2409 	page_num = data_buf->npages;
2410 	for (i = 0; i < page_num; i++) {
2411 		addr = hns_roce_buf_page(data_buf, i);
2412 		if (i == (page_num - 1))
2413 			next_ptr = 0;
2414 		else
2415 			next_ptr = i + 1;
2416 
2417 		val = HNS_ROCE_EXT_LLM_ENTRY(addr, (u64)next_ptr);
2418 		entry[i] = cpu_to_le64(val);
2419 	}
2420 }
2421 
2422 static int set_llm_cfg_to_hw(struct hns_roce_dev *hr_dev,
2423 			     struct hns_roce_link_table *table)
2424 {
2425 	struct hns_roce_cmq_desc desc[2];
2426 	struct hns_roce_cmq_req *r_a = (struct hns_roce_cmq_req *)desc[0].data;
2427 	struct hns_roce_cmq_req *r_b = (struct hns_roce_cmq_req *)desc[1].data;
2428 	struct hns_roce_buf *buf = table->buf;
2429 	enum hns_roce_opcode_type opcode;
2430 	dma_addr_t addr;
2431 
2432 	opcode = HNS_ROCE_OPC_CFG_EXT_LLM;
2433 	hns_roce_cmq_setup_basic_desc(&desc[0], opcode, false);
2434 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
2435 	hns_roce_cmq_setup_basic_desc(&desc[1], opcode, false);
2436 
2437 	hr_reg_write(r_a, CFG_LLM_A_BA_L, lower_32_bits(table->table.map));
2438 	hr_reg_write(r_a, CFG_LLM_A_BA_H, upper_32_bits(table->table.map));
2439 	hr_reg_write(r_a, CFG_LLM_A_DEPTH, buf->npages);
2440 	hr_reg_write(r_a, CFG_LLM_A_PGSZ, to_hr_hw_page_shift(buf->page_shift));
2441 	hr_reg_enable(r_a, CFG_LLM_A_INIT_EN);
2442 
2443 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, 0));
2444 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_L, lower_32_bits(addr));
2445 	hr_reg_write(r_a, CFG_LLM_A_HEAD_BA_H, upper_32_bits(addr));
2446 	hr_reg_write(r_a, CFG_LLM_A_HEAD_NXTPTR, 1);
2447 	hr_reg_write(r_a, CFG_LLM_A_HEAD_PTR, 0);
2448 
2449 	addr = to_hr_hw_page_addr(hns_roce_buf_page(buf, buf->npages - 1));
2450 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_L, lower_32_bits(addr));
2451 	hr_reg_write(r_b, CFG_LLM_B_TAIL_BA_H, upper_32_bits(addr));
2452 	hr_reg_write(r_b, CFG_LLM_B_TAIL_PTR, buf->npages - 1);
2453 
2454 	return hns_roce_cmq_send(hr_dev, desc, 2);
2455 }
2456 
2457 static struct hns_roce_link_table *
2458 alloc_link_table_buf(struct hns_roce_dev *hr_dev)
2459 {
2460 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2461 	struct hns_roce_link_table *link_tbl;
2462 	u32 pg_shift, size, min_size;
2463 
2464 	link_tbl = &priv->ext_llm;
2465 	pg_shift = hr_dev->caps.llm_buf_pg_sz + PAGE_SHIFT;
2466 	size = hr_dev->caps.num_qps * HNS_ROCE_V2_EXT_LLM_ENTRY_SZ;
2467 	min_size = HNS_ROCE_EXT_LLM_MIN_PAGES(hr_dev->caps.sl_num) << pg_shift;
2468 
2469 	/* Alloc data table */
2470 	size = max(size, min_size);
2471 	link_tbl->buf = hns_roce_buf_alloc(hr_dev, size, pg_shift, 0);
2472 	if (IS_ERR(link_tbl->buf))
2473 		return ERR_PTR(-ENOMEM);
2474 
2475 	/* Alloc config table */
2476 	size = link_tbl->buf->npages * sizeof(u64);
2477 	link_tbl->table.buf = dma_alloc_coherent(hr_dev->dev, size,
2478 						 &link_tbl->table.map,
2479 						 GFP_KERNEL);
2480 	if (!link_tbl->table.buf) {
2481 		hns_roce_buf_free(hr_dev, link_tbl->buf);
2482 		return ERR_PTR(-ENOMEM);
2483 	}
2484 
2485 	return link_tbl;
2486 }
2487 
2488 static void free_link_table_buf(struct hns_roce_dev *hr_dev,
2489 				struct hns_roce_link_table *tbl)
2490 {
2491 	if (tbl->buf) {
2492 		u32 size = tbl->buf->npages * sizeof(u64);
2493 
2494 		dma_free_coherent(hr_dev->dev, size, tbl->table.buf,
2495 				  tbl->table.map);
2496 	}
2497 
2498 	hns_roce_buf_free(hr_dev, tbl->buf);
2499 }
2500 
2501 static int hns_roce_init_link_table(struct hns_roce_dev *hr_dev)
2502 {
2503 	struct hns_roce_link_table *link_tbl;
2504 	int ret;
2505 
2506 	link_tbl = alloc_link_table_buf(hr_dev);
2507 	if (IS_ERR(link_tbl))
2508 		return -ENOMEM;
2509 
2510 	if (WARN_ON(link_tbl->buf->npages > HNS_ROCE_V2_EXT_LLM_MAX_DEPTH)) {
2511 		ret = -EINVAL;
2512 		goto err_alloc;
2513 	}
2514 
2515 	config_llm_table(link_tbl->buf, link_tbl->table.buf);
2516 	ret = set_llm_cfg_to_hw(hr_dev, link_tbl);
2517 	if (ret)
2518 		goto err_alloc;
2519 
2520 	return 0;
2521 
2522 err_alloc:
2523 	free_link_table_buf(hr_dev, link_tbl);
2524 	return ret;
2525 }
2526 
2527 static void hns_roce_free_link_table(struct hns_roce_dev *hr_dev)
2528 {
2529 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2530 
2531 	free_link_table_buf(hr_dev, &priv->ext_llm);
2532 }
2533 
2534 static void free_dip_list(struct hns_roce_dev *hr_dev)
2535 {
2536 	struct hns_roce_dip *hr_dip;
2537 	struct hns_roce_dip *tmp;
2538 	unsigned long flags;
2539 
2540 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
2541 
2542 	list_for_each_entry_safe(hr_dip, tmp, &hr_dev->dip_list, node) {
2543 		list_del(&hr_dip->node);
2544 		kfree(hr_dip);
2545 	}
2546 
2547 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
2548 }
2549 
2550 static struct ib_pd *free_mr_init_pd(struct hns_roce_dev *hr_dev)
2551 {
2552 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2553 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2554 	struct ib_device *ibdev = &hr_dev->ib_dev;
2555 	struct hns_roce_pd *hr_pd;
2556 	struct ib_pd *pd;
2557 
2558 	hr_pd = kzalloc(sizeof(*hr_pd), GFP_KERNEL);
2559 	if (ZERO_OR_NULL_PTR(hr_pd))
2560 		return NULL;
2561 	pd = &hr_pd->ibpd;
2562 	pd->device = ibdev;
2563 
2564 	if (hns_roce_alloc_pd(pd, NULL)) {
2565 		ibdev_err(ibdev, "failed to create pd for free mr.\n");
2566 		kfree(hr_pd);
2567 		return NULL;
2568 	}
2569 	free_mr->rsv_pd = to_hr_pd(pd);
2570 	free_mr->rsv_pd->ibpd.device = &hr_dev->ib_dev;
2571 	free_mr->rsv_pd->ibpd.uobject = NULL;
2572 	free_mr->rsv_pd->ibpd.__internal_mr = NULL;
2573 	atomic_set(&free_mr->rsv_pd->ibpd.usecnt, 0);
2574 
2575 	return pd;
2576 }
2577 
2578 static struct ib_cq *free_mr_init_cq(struct hns_roce_dev *hr_dev)
2579 {
2580 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2581 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2582 	struct ib_device *ibdev = &hr_dev->ib_dev;
2583 	struct ib_cq_init_attr cq_init_attr = {};
2584 	struct hns_roce_cq *hr_cq;
2585 	struct ib_cq *cq;
2586 
2587 	cq_init_attr.cqe = HNS_ROCE_FREE_MR_USED_CQE_NUM;
2588 
2589 	hr_cq = kzalloc(sizeof(*hr_cq), GFP_KERNEL);
2590 	if (ZERO_OR_NULL_PTR(hr_cq))
2591 		return NULL;
2592 
2593 	cq = &hr_cq->ib_cq;
2594 	cq->device = ibdev;
2595 
2596 	if (hns_roce_create_cq(cq, &cq_init_attr, NULL)) {
2597 		ibdev_err(ibdev, "failed to create cq for free mr.\n");
2598 		kfree(hr_cq);
2599 		return NULL;
2600 	}
2601 	free_mr->rsv_cq = to_hr_cq(cq);
2602 	free_mr->rsv_cq->ib_cq.device = &hr_dev->ib_dev;
2603 	free_mr->rsv_cq->ib_cq.uobject = NULL;
2604 	free_mr->rsv_cq->ib_cq.comp_handler = NULL;
2605 	free_mr->rsv_cq->ib_cq.event_handler = NULL;
2606 	free_mr->rsv_cq->ib_cq.cq_context = NULL;
2607 	atomic_set(&free_mr->rsv_cq->ib_cq.usecnt, 0);
2608 
2609 	return cq;
2610 }
2611 
2612 static int free_mr_init_qp(struct hns_roce_dev *hr_dev, struct ib_cq *cq,
2613 			   struct ib_qp_init_attr *init_attr, int i)
2614 {
2615 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2616 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2617 	struct ib_device *ibdev = &hr_dev->ib_dev;
2618 	struct hns_roce_qp *hr_qp;
2619 	struct ib_qp *qp;
2620 	int ret;
2621 
2622 	hr_qp = kzalloc(sizeof(*hr_qp), GFP_KERNEL);
2623 	if (ZERO_OR_NULL_PTR(hr_qp))
2624 		return -ENOMEM;
2625 
2626 	qp = &hr_qp->ibqp;
2627 	qp->device = ibdev;
2628 
2629 	ret = hns_roce_create_qp(qp, init_attr, NULL);
2630 	if (ret) {
2631 		ibdev_err(ibdev, "failed to create qp for free mr.\n");
2632 		kfree(hr_qp);
2633 		return ret;
2634 	}
2635 
2636 	free_mr->rsv_qp[i] = hr_qp;
2637 	free_mr->rsv_qp[i]->ibqp.recv_cq = cq;
2638 	free_mr->rsv_qp[i]->ibqp.send_cq = cq;
2639 
2640 	return 0;
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 	struct ib_qp *qp;
2648 	int i;
2649 
2650 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2651 		if (free_mr->rsv_qp[i]) {
2652 			qp = &free_mr->rsv_qp[i]->ibqp;
2653 			hns_roce_v2_destroy_qp(qp, NULL);
2654 			kfree(free_mr->rsv_qp[i]);
2655 			free_mr->rsv_qp[i] = NULL;
2656 		}
2657 	}
2658 
2659 	if (free_mr->rsv_cq) {
2660 		hns_roce_destroy_cq(&free_mr->rsv_cq->ib_cq, NULL);
2661 		kfree(free_mr->rsv_cq);
2662 		free_mr->rsv_cq = NULL;
2663 	}
2664 
2665 	if (free_mr->rsv_pd) {
2666 		hns_roce_dealloc_pd(&free_mr->rsv_pd->ibpd, NULL);
2667 		kfree(free_mr->rsv_pd);
2668 		free_mr->rsv_pd = NULL;
2669 	}
2670 
2671 	mutex_destroy(&free_mr->mutex);
2672 }
2673 
2674 static int free_mr_alloc_res(struct hns_roce_dev *hr_dev)
2675 {
2676 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2677 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2678 	struct ib_qp_init_attr qp_init_attr = {};
2679 	struct ib_pd *pd;
2680 	struct ib_cq *cq;
2681 	int ret;
2682 	int i;
2683 
2684 	pd = free_mr_init_pd(hr_dev);
2685 	if (!pd)
2686 		return -ENOMEM;
2687 
2688 	cq = free_mr_init_cq(hr_dev);
2689 	if (!cq) {
2690 		ret = -ENOMEM;
2691 		goto create_failed_cq;
2692 	}
2693 
2694 	qp_init_attr.qp_type = IB_QPT_RC;
2695 	qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2696 	qp_init_attr.send_cq = cq;
2697 	qp_init_attr.recv_cq = cq;
2698 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2699 		qp_init_attr.cap.max_send_wr = HNS_ROCE_FREE_MR_USED_SQWQE_NUM;
2700 		qp_init_attr.cap.max_send_sge = HNS_ROCE_FREE_MR_USED_SQSGE_NUM;
2701 		qp_init_attr.cap.max_recv_wr = HNS_ROCE_FREE_MR_USED_RQWQE_NUM;
2702 		qp_init_attr.cap.max_recv_sge = HNS_ROCE_FREE_MR_USED_RQSGE_NUM;
2703 
2704 		ret = free_mr_init_qp(hr_dev, cq, &qp_init_attr, i);
2705 		if (ret)
2706 			goto create_failed_qp;
2707 	}
2708 
2709 	return 0;
2710 
2711 create_failed_qp:
2712 	for (i--; i >= 0; i--) {
2713 		hns_roce_v2_destroy_qp(&free_mr->rsv_qp[i]->ibqp, NULL);
2714 		kfree(free_mr->rsv_qp[i]);
2715 	}
2716 	hns_roce_destroy_cq(cq, NULL);
2717 	kfree(cq);
2718 
2719 create_failed_cq:
2720 	hns_roce_dealloc_pd(pd, NULL);
2721 	kfree(pd);
2722 
2723 	return ret;
2724 }
2725 
2726 static int free_mr_modify_rsv_qp(struct hns_roce_dev *hr_dev,
2727 				 struct ib_qp_attr *attr, int sl_num)
2728 {
2729 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2730 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2731 	struct ib_device *ibdev = &hr_dev->ib_dev;
2732 	struct hns_roce_qp *hr_qp;
2733 	int loopback;
2734 	int mask;
2735 	int ret;
2736 
2737 	hr_qp = to_hr_qp(&free_mr->rsv_qp[sl_num]->ibqp);
2738 	hr_qp->free_mr_en = 1;
2739 	hr_qp->ibqp.device = ibdev;
2740 	hr_qp->ibqp.qp_type = IB_QPT_RC;
2741 
2742 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_ACCESS_FLAGS;
2743 	attr->qp_state = IB_QPS_INIT;
2744 	attr->port_num = 1;
2745 	attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
2746 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2747 				    IB_QPS_INIT, NULL);
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 = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_INIT,
2769 				    IB_QPS_RTR, NULL);
2770 	hr_dev->loop_idc = loopback;
2771 	if (ret) {
2772 		ibdev_err(ibdev, "failed to modify qp to rtr, ret = %d.\n",
2773 			  ret);
2774 		return ret;
2775 	}
2776 
2777 	mask = IB_QP_STATE | IB_QP_SQ_PSN | IB_QP_RETRY_CNT | IB_QP_TIMEOUT |
2778 	       IB_QP_RNR_RETRY | IB_QP_MAX_QP_RD_ATOMIC;
2779 	attr->qp_state = IB_QPS_RTS;
2780 	attr->sq_psn = HNS_ROCE_FREE_MR_USED_PSN;
2781 	attr->retry_cnt = HNS_ROCE_FREE_MR_USED_QP_RETRY_CNT;
2782 	attr->timeout = HNS_ROCE_FREE_MR_USED_QP_TIMEOUT;
2783 	ret = hr_dev->hw->modify_qp(&hr_qp->ibqp, attr, mask, IB_QPS_RTR,
2784 				    IB_QPS_RTS, NULL);
2785 	if (ret)
2786 		ibdev_err(ibdev, "failed to modify qp to rts, ret = %d.\n",
2787 			  ret);
2788 
2789 	return ret;
2790 }
2791 
2792 static int free_mr_modify_qp(struct hns_roce_dev *hr_dev)
2793 {
2794 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2795 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2796 	struct ib_qp_attr attr = {};
2797 	int ret;
2798 	int i;
2799 
2800 	rdma_ah_set_grh(&attr.ah_attr, NULL, 0, 0, 1, 0);
2801 	rdma_ah_set_static_rate(&attr.ah_attr, 3);
2802 	rdma_ah_set_port_num(&attr.ah_attr, 1);
2803 
2804 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
2805 		ret = free_mr_modify_rsv_qp(hr_dev, &attr, i);
2806 		if (ret)
2807 			return ret;
2808 	}
2809 
2810 	return 0;
2811 }
2812 
2813 static int free_mr_init(struct hns_roce_dev *hr_dev)
2814 {
2815 	struct hns_roce_v2_priv *priv = hr_dev->priv;
2816 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
2817 	int ret;
2818 
2819 	mutex_init(&free_mr->mutex);
2820 
2821 	ret = free_mr_alloc_res(hr_dev);
2822 	if (ret) {
2823 		mutex_destroy(&free_mr->mutex);
2824 		return ret;
2825 	}
2826 
2827 	ret = free_mr_modify_qp(hr_dev);
2828 	if (ret)
2829 		goto err_modify_qp;
2830 
2831 	return 0;
2832 
2833 err_modify_qp:
2834 	free_mr_exit(hr_dev);
2835 
2836 	return ret;
2837 }
2838 
2839 static int get_hem_table(struct hns_roce_dev *hr_dev)
2840 {
2841 	unsigned int qpc_count;
2842 	unsigned int cqc_count;
2843 	unsigned int gmv_count;
2844 	int ret;
2845 	int i;
2846 
2847 	/* Alloc memory for source address table buffer space chunk */
2848 	for (gmv_count = 0; gmv_count < hr_dev->caps.gmv_entry_num;
2849 	     gmv_count++) {
2850 		ret = hns_roce_table_get(hr_dev, &hr_dev->gmv_table, gmv_count);
2851 		if (ret)
2852 			goto err_gmv_failed;
2853 	}
2854 
2855 	if (hr_dev->is_vf)
2856 		return 0;
2857 
2858 	/* Alloc memory for QPC Timer buffer space chunk */
2859 	for (qpc_count = 0; qpc_count < hr_dev->caps.qpc_timer_bt_num;
2860 	     qpc_count++) {
2861 		ret = hns_roce_table_get(hr_dev, &hr_dev->qpc_timer_table,
2862 					 qpc_count);
2863 		if (ret) {
2864 			dev_err(hr_dev->dev, "QPC Timer get failed\n");
2865 			goto err_qpc_timer_failed;
2866 		}
2867 	}
2868 
2869 	/* Alloc memory for CQC Timer buffer space chunk */
2870 	for (cqc_count = 0; cqc_count < hr_dev->caps.cqc_timer_bt_num;
2871 	     cqc_count++) {
2872 		ret = hns_roce_table_get(hr_dev, &hr_dev->cqc_timer_table,
2873 					 cqc_count);
2874 		if (ret) {
2875 			dev_err(hr_dev->dev, "CQC Timer get failed\n");
2876 			goto err_cqc_timer_failed;
2877 		}
2878 	}
2879 
2880 	return 0;
2881 
2882 err_cqc_timer_failed:
2883 	for (i = 0; i < cqc_count; i++)
2884 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2885 
2886 err_qpc_timer_failed:
2887 	for (i = 0; i < qpc_count; i++)
2888 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2889 
2890 err_gmv_failed:
2891 	for (i = 0; i < gmv_count; i++)
2892 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2893 
2894 	return ret;
2895 }
2896 
2897 static void put_hem_table(struct hns_roce_dev *hr_dev)
2898 {
2899 	int i;
2900 
2901 	for (i = 0; i < hr_dev->caps.gmv_entry_num; i++)
2902 		hns_roce_table_put(hr_dev, &hr_dev->gmv_table, i);
2903 
2904 	if (hr_dev->is_vf)
2905 		return;
2906 
2907 	for (i = 0; i < hr_dev->caps.qpc_timer_bt_num; i++)
2908 		hns_roce_table_put(hr_dev, &hr_dev->qpc_timer_table, i);
2909 
2910 	for (i = 0; i < hr_dev->caps.cqc_timer_bt_num; i++)
2911 		hns_roce_table_put(hr_dev, &hr_dev->cqc_timer_table, i);
2912 }
2913 
2914 static int hns_roce_v2_init(struct hns_roce_dev *hr_dev)
2915 {
2916 	int ret;
2917 
2918 	/* The hns ROCEE requires the extdb info to be cleared before using */
2919 	ret = hns_roce_clear_extdb_list_info(hr_dev);
2920 	if (ret)
2921 		return ret;
2922 
2923 	ret = get_hem_table(hr_dev);
2924 	if (ret)
2925 		return ret;
2926 
2927 	if (hr_dev->is_vf)
2928 		return 0;
2929 
2930 	ret = hns_roce_init_link_table(hr_dev);
2931 	if (ret) {
2932 		dev_err(hr_dev->dev, "failed to init llm, ret = %d.\n", ret);
2933 		goto err_llm_init_failed;
2934 	}
2935 
2936 	return 0;
2937 
2938 err_llm_init_failed:
2939 	put_hem_table(hr_dev);
2940 
2941 	return ret;
2942 }
2943 
2944 static void hns_roce_v2_exit(struct hns_roce_dev *hr_dev)
2945 {
2946 	hns_roce_function_clear(hr_dev);
2947 
2948 	if (!hr_dev->is_vf)
2949 		hns_roce_free_link_table(hr_dev);
2950 
2951 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP09)
2952 		free_dip_list(hr_dev);
2953 }
2954 
2955 static int hns_roce_mbox_post(struct hns_roce_dev *hr_dev,
2956 			      struct hns_roce_mbox_msg *mbox_msg)
2957 {
2958 	struct hns_roce_cmq_desc desc;
2959 	struct hns_roce_post_mbox *mb = (struct hns_roce_post_mbox *)desc.data;
2960 
2961 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_POST_MB, false);
2962 
2963 	mb->in_param_l = cpu_to_le32(mbox_msg->in_param);
2964 	mb->in_param_h = cpu_to_le32(mbox_msg->in_param >> 32);
2965 	mb->out_param_l = cpu_to_le32(mbox_msg->out_param);
2966 	mb->out_param_h = cpu_to_le32(mbox_msg->out_param >> 32);
2967 	mb->cmd_tag = cpu_to_le32(mbox_msg->tag << 8 | mbox_msg->cmd);
2968 	mb->token_event_en = cpu_to_le32(mbox_msg->event_en << 16 |
2969 					 mbox_msg->token);
2970 
2971 	return hns_roce_cmq_send(hr_dev, &desc, 1);
2972 }
2973 
2974 static int v2_wait_mbox_complete(struct hns_roce_dev *hr_dev, u32 timeout,
2975 				 u8 *complete_status)
2976 {
2977 	struct hns_roce_mbox_status *mb_st;
2978 	struct hns_roce_cmq_desc desc;
2979 	unsigned long end;
2980 	int ret = -EBUSY;
2981 	u32 status;
2982 	bool busy;
2983 
2984 	mb_st = (struct hns_roce_mbox_status *)desc.data;
2985 	end = msecs_to_jiffies(timeout) + jiffies;
2986 	while (v2_chk_mbox_is_avail(hr_dev, &busy)) {
2987 		if (hr_dev->cmd.state == HNS_ROCE_CMDQ_STATE_FATAL_ERR)
2988 			return -EIO;
2989 
2990 		status = 0;
2991 		hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_QUERY_MB_ST,
2992 					      true);
2993 		ret = __hns_roce_cmq_send(hr_dev, &desc, 1);
2994 		if (!ret) {
2995 			status = le32_to_cpu(mb_st->mb_status_hw_run);
2996 			/* No pending message exists in ROCEE mbox. */
2997 			if (!(status & MB_ST_HW_RUN_M))
2998 				break;
2999 		} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3000 			break;
3001 		}
3002 
3003 		if (time_after(jiffies, end)) {
3004 			dev_err_ratelimited(hr_dev->dev,
3005 					    "failed to wait mbox status 0x%x\n",
3006 					    status);
3007 			return -ETIMEDOUT;
3008 		}
3009 
3010 		cond_resched();
3011 		ret = -EBUSY;
3012 	}
3013 
3014 	if (!ret) {
3015 		*complete_status = (u8)(status & MB_ST_COMPLETE_M);
3016 	} else if (!v2_chk_mbox_is_avail(hr_dev, &busy)) {
3017 		/* Ignore all errors if the mbox is unavailable. */
3018 		ret = 0;
3019 		*complete_status = MB_ST_COMPLETE_M;
3020 	}
3021 
3022 	return ret;
3023 }
3024 
3025 static int v2_post_mbox(struct hns_roce_dev *hr_dev,
3026 			struct hns_roce_mbox_msg *mbox_msg)
3027 {
3028 	u8 status = 0;
3029 	int ret;
3030 
3031 	/* Waiting for the mbox to be idle */
3032 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_V2_GO_BIT_TIMEOUT_MSECS,
3033 				    &status);
3034 	if (unlikely(ret)) {
3035 		dev_err_ratelimited(hr_dev->dev,
3036 				    "failed to check post mbox status = 0x%x, ret = %d.\n",
3037 				    status, ret);
3038 		return ret;
3039 	}
3040 
3041 	/* Post new message to mbox */
3042 	ret = hns_roce_mbox_post(hr_dev, mbox_msg);
3043 	if (ret)
3044 		dev_err_ratelimited(hr_dev->dev,
3045 				    "failed to post mailbox, ret = %d.\n", ret);
3046 
3047 	return ret;
3048 }
3049 
3050 static int v2_poll_mbox_done(struct hns_roce_dev *hr_dev)
3051 {
3052 	u8 status = 0;
3053 	int ret;
3054 
3055 	ret = v2_wait_mbox_complete(hr_dev, HNS_ROCE_CMD_TIMEOUT_MSECS,
3056 				    &status);
3057 	if (!ret) {
3058 		if (status != MB_ST_COMPLETE_SUCC)
3059 			return -EBUSY;
3060 	} else {
3061 		dev_err_ratelimited(hr_dev->dev,
3062 				    "failed to check mbox status = 0x%x, ret = %d.\n",
3063 				    status, ret);
3064 	}
3065 
3066 	return ret;
3067 }
3068 
3069 static void copy_gid(void *dest, const union ib_gid *gid)
3070 {
3071 #define GID_SIZE 4
3072 	const union ib_gid *src = gid;
3073 	__le32 (*p)[GID_SIZE] = dest;
3074 	int i;
3075 
3076 	if (!gid)
3077 		src = &zgid;
3078 
3079 	for (i = 0; i < GID_SIZE; i++)
3080 		(*p)[i] = cpu_to_le32(*(u32 *)&src->raw[i * sizeof(u32)]);
3081 }
3082 
3083 static int config_sgid_table(struct hns_roce_dev *hr_dev,
3084 			     int gid_index, const union ib_gid *gid,
3085 			     enum hns_roce_sgid_type sgid_type)
3086 {
3087 	struct hns_roce_cmq_desc desc;
3088 	struct hns_roce_cfg_sgid_tb *sgid_tb =
3089 				    (struct hns_roce_cfg_sgid_tb *)desc.data;
3090 
3091 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SGID_TB, false);
3092 
3093 	hr_reg_write(sgid_tb, CFG_SGID_TB_TABLE_IDX, gid_index);
3094 	hr_reg_write(sgid_tb, CFG_SGID_TB_VF_SGID_TYPE, sgid_type);
3095 
3096 	copy_gid(&sgid_tb->vf_sgid_l, gid);
3097 
3098 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3099 }
3100 
3101 static int config_gmv_table(struct hns_roce_dev *hr_dev,
3102 			    int gid_index, const union ib_gid *gid,
3103 			    enum hns_roce_sgid_type sgid_type,
3104 			    const struct ib_gid_attr *attr)
3105 {
3106 	struct hns_roce_cmq_desc desc[2];
3107 	struct hns_roce_cfg_gmv_tb_a *tb_a =
3108 				(struct hns_roce_cfg_gmv_tb_a *)desc[0].data;
3109 	struct hns_roce_cfg_gmv_tb_b *tb_b =
3110 				(struct hns_roce_cfg_gmv_tb_b *)desc[1].data;
3111 
3112 	u16 vlan_id = VLAN_CFI_MASK;
3113 	u8 mac[ETH_ALEN] = {};
3114 	int ret;
3115 
3116 	if (gid) {
3117 		ret = rdma_read_gid_l2_fields(attr, &vlan_id, mac);
3118 		if (ret)
3119 			return ret;
3120 	}
3121 
3122 	hns_roce_cmq_setup_basic_desc(&desc[0], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3123 	desc[0].flag |= cpu_to_le16(HNS_ROCE_CMD_FLAG_NEXT);
3124 
3125 	hns_roce_cmq_setup_basic_desc(&desc[1], HNS_ROCE_OPC_CFG_GMV_TBL, false);
3126 
3127 	copy_gid(&tb_a->vf_sgid_l, gid);
3128 
3129 	hr_reg_write(tb_a, GMV_TB_A_VF_SGID_TYPE, sgid_type);
3130 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_EN, vlan_id < VLAN_CFI_MASK);
3131 	hr_reg_write(tb_a, GMV_TB_A_VF_VLAN_ID, vlan_id);
3132 
3133 	tb_b->vf_smac_l = cpu_to_le32(*(u32 *)mac);
3134 
3135 	hr_reg_write(tb_b, GMV_TB_B_SMAC_H, *(u16 *)&mac[4]);
3136 	hr_reg_write(tb_b, GMV_TB_B_SGID_IDX, gid_index);
3137 
3138 	return hns_roce_cmq_send(hr_dev, desc, 2);
3139 }
3140 
3141 static int hns_roce_v2_set_gid(struct hns_roce_dev *hr_dev, int gid_index,
3142 			       const union ib_gid *gid,
3143 			       const struct ib_gid_attr *attr)
3144 {
3145 	enum hns_roce_sgid_type sgid_type = GID_TYPE_FLAG_ROCE_V1;
3146 	int ret;
3147 
3148 	if (gid) {
3149 		if (attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3150 			if (ipv6_addr_v4mapped((void *)gid))
3151 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV4;
3152 			else
3153 				sgid_type = GID_TYPE_FLAG_ROCE_V2_IPV6;
3154 		} else if (attr->gid_type == IB_GID_TYPE_ROCE) {
3155 			sgid_type = GID_TYPE_FLAG_ROCE_V1;
3156 		}
3157 	}
3158 
3159 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
3160 		ret = config_gmv_table(hr_dev, gid_index, gid, sgid_type, attr);
3161 	else
3162 		ret = config_sgid_table(hr_dev, gid_index, gid, sgid_type);
3163 
3164 	if (ret)
3165 		ibdev_err(&hr_dev->ib_dev, "failed to set gid, ret = %d!\n",
3166 			  ret);
3167 
3168 	return ret;
3169 }
3170 
3171 static int hns_roce_v2_set_mac(struct hns_roce_dev *hr_dev, u8 phy_port,
3172 			       const u8 *addr)
3173 {
3174 	struct hns_roce_cmq_desc desc;
3175 	struct hns_roce_cfg_smac_tb *smac_tb =
3176 				    (struct hns_roce_cfg_smac_tb *)desc.data;
3177 	u16 reg_smac_h;
3178 	u32 reg_smac_l;
3179 
3180 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_SMAC_TB, false);
3181 
3182 	reg_smac_l = *(u32 *)(&addr[0]);
3183 	reg_smac_h = *(u16 *)(&addr[4]);
3184 
3185 	hr_reg_write(smac_tb, CFG_SMAC_TB_IDX, phy_port);
3186 	hr_reg_write(smac_tb, CFG_SMAC_TB_VF_SMAC_H, reg_smac_h);
3187 	smac_tb->vf_smac_l = cpu_to_le32(reg_smac_l);
3188 
3189 	return hns_roce_cmq_send(hr_dev, &desc, 1);
3190 }
3191 
3192 static int set_mtpt_pbl(struct hns_roce_dev *hr_dev,
3193 			struct hns_roce_v2_mpt_entry *mpt_entry,
3194 			struct hns_roce_mr *mr)
3195 {
3196 	u64 pages[HNS_ROCE_V2_MAX_INNER_MTPT_NUM] = { 0 };
3197 	struct ib_device *ibdev = &hr_dev->ib_dev;
3198 	dma_addr_t pbl_ba;
3199 	int ret;
3200 	int i;
3201 
3202 	ret = hns_roce_mtr_find(hr_dev, &mr->pbl_mtr, 0, pages,
3203 				min_t(int, ARRAY_SIZE(pages), mr->npages));
3204 	if (ret) {
3205 		ibdev_err(ibdev, "failed to find PBL mtr, ret = %d.\n", ret);
3206 		return ret;
3207 	}
3208 
3209 	/* Aligned to the hardware address access unit */
3210 	for (i = 0; i < ARRAY_SIZE(pages); i++)
3211 		pages[i] >>= MPT_PBL_BUF_ADDR_S;
3212 
3213 	pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3214 
3215 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3216 	mpt_entry->pbl_ba_l = cpu_to_le32(pbl_ba >> MPT_PBL_BA_ADDR_S);
3217 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3218 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3219 
3220 	mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
3221 	hr_reg_write(mpt_entry, MPT_PA0_H, upper_32_bits(pages[0]));
3222 
3223 	mpt_entry->pa1_l = cpu_to_le32(lower_32_bits(pages[1]));
3224 	hr_reg_write(mpt_entry, MPT_PA1_H, upper_32_bits(pages[1]));
3225 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3226 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3227 
3228 	return 0;
3229 }
3230 
3231 static int hns_roce_v2_write_mtpt(struct hns_roce_dev *hr_dev,
3232 				  void *mb_buf, struct hns_roce_mr *mr)
3233 {
3234 	struct hns_roce_v2_mpt_entry *mpt_entry;
3235 
3236 	mpt_entry = mb_buf;
3237 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3238 
3239 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3240 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3241 
3242 	hr_reg_write_bool(mpt_entry, MPT_BIND_EN,
3243 			  mr->access & IB_ACCESS_MW_BIND);
3244 	hr_reg_write_bool(mpt_entry, MPT_ATOMIC_EN,
3245 			  mr->access & IB_ACCESS_REMOTE_ATOMIC);
3246 	hr_reg_write_bool(mpt_entry, MPT_RR_EN,
3247 			  mr->access & IB_ACCESS_REMOTE_READ);
3248 	hr_reg_write_bool(mpt_entry, MPT_RW_EN,
3249 			  mr->access & IB_ACCESS_REMOTE_WRITE);
3250 	hr_reg_write_bool(mpt_entry, MPT_LW_EN,
3251 			  mr->access & IB_ACCESS_LOCAL_WRITE);
3252 
3253 	mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3254 	mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3255 	mpt_entry->lkey = cpu_to_le32(mr->key);
3256 	mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3257 	mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3258 
3259 	if (mr->type != MR_TYPE_MR)
3260 		hr_reg_enable(mpt_entry, MPT_PA);
3261 
3262 	if (mr->type == MR_TYPE_DMA)
3263 		return 0;
3264 
3265 	if (mr->pbl_hop_num != HNS_ROCE_HOP_NUM_0)
3266 		hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, mr->pbl_hop_num);
3267 
3268 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3269 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3270 	hr_reg_enable(mpt_entry, MPT_INNER_PA_VLD);
3271 
3272 	return set_mtpt_pbl(hr_dev, mpt_entry, mr);
3273 }
3274 
3275 static int hns_roce_v2_rereg_write_mtpt(struct hns_roce_dev *hr_dev,
3276 					struct hns_roce_mr *mr, int flags,
3277 					void *mb_buf)
3278 {
3279 	struct hns_roce_v2_mpt_entry *mpt_entry = mb_buf;
3280 	u32 mr_access_flags = mr->access;
3281 	int ret = 0;
3282 
3283 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_VALID);
3284 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3285 
3286 	if (flags & IB_MR_REREG_ACCESS) {
3287 		hr_reg_write(mpt_entry, MPT_BIND_EN,
3288 			     (mr_access_flags & IB_ACCESS_MW_BIND ? 1 : 0));
3289 		hr_reg_write(mpt_entry, MPT_ATOMIC_EN,
3290 			     mr_access_flags & IB_ACCESS_REMOTE_ATOMIC ? 1 : 0);
3291 		hr_reg_write(mpt_entry, MPT_RR_EN,
3292 			     mr_access_flags & IB_ACCESS_REMOTE_READ ? 1 : 0);
3293 		hr_reg_write(mpt_entry, MPT_RW_EN,
3294 			     mr_access_flags & IB_ACCESS_REMOTE_WRITE ? 1 : 0);
3295 		hr_reg_write(mpt_entry, MPT_LW_EN,
3296 			     mr_access_flags & IB_ACCESS_LOCAL_WRITE ? 1 : 0);
3297 	}
3298 
3299 	if (flags & IB_MR_REREG_TRANS) {
3300 		mpt_entry->va_l = cpu_to_le32(lower_32_bits(mr->iova));
3301 		mpt_entry->va_h = cpu_to_le32(upper_32_bits(mr->iova));
3302 		mpt_entry->len_l = cpu_to_le32(lower_32_bits(mr->size));
3303 		mpt_entry->len_h = cpu_to_le32(upper_32_bits(mr->size));
3304 
3305 		ret = set_mtpt_pbl(hr_dev, mpt_entry, mr);
3306 	}
3307 
3308 	return ret;
3309 }
3310 
3311 static int hns_roce_v2_frmr_write_mtpt(void *mb_buf, struct hns_roce_mr *mr)
3312 {
3313 	dma_addr_t pbl_ba = hns_roce_get_mtr_ba(&mr->pbl_mtr);
3314 	struct hns_roce_v2_mpt_entry *mpt_entry;
3315 
3316 	mpt_entry = mb_buf;
3317 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3318 
3319 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3320 	hr_reg_write(mpt_entry, MPT_PD, mr->pd);
3321 
3322 	hr_reg_enable(mpt_entry, MPT_RA_EN);
3323 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3324 
3325 	hr_reg_enable(mpt_entry, MPT_FRE);
3326 	hr_reg_clear(mpt_entry, MPT_MR_MW);
3327 	hr_reg_enable(mpt_entry, MPT_BPD);
3328 	hr_reg_clear(mpt_entry, MPT_PA);
3329 
3330 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM, 1);
3331 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3332 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.ba_pg_shift));
3333 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3334 		     to_hr_hw_page_shift(mr->pbl_mtr.hem_cfg.buf_pg_shift));
3335 
3336 	mpt_entry->pbl_size = cpu_to_le32(mr->npages);
3337 
3338 	mpt_entry->pbl_ba_l = cpu_to_le32(lower_32_bits(pbl_ba >>
3339 							MPT_PBL_BA_ADDR_S));
3340 	hr_reg_write(mpt_entry, MPT_PBL_BA_H,
3341 		     upper_32_bits(pbl_ba >> MPT_PBL_BA_ADDR_S));
3342 
3343 	return 0;
3344 }
3345 
3346 static int hns_roce_v2_mw_write_mtpt(void *mb_buf, struct hns_roce_mw *mw)
3347 {
3348 	struct hns_roce_v2_mpt_entry *mpt_entry;
3349 
3350 	mpt_entry = mb_buf;
3351 	memset(mpt_entry, 0, sizeof(*mpt_entry));
3352 
3353 	hr_reg_write(mpt_entry, MPT_ST, V2_MPT_ST_FREE);
3354 	hr_reg_write(mpt_entry, MPT_PD, mw->pdn);
3355 
3356 	hr_reg_enable(mpt_entry, MPT_R_INV_EN);
3357 	hr_reg_enable(mpt_entry, MPT_LW_EN);
3358 
3359 	hr_reg_enable(mpt_entry, MPT_MR_MW);
3360 	hr_reg_enable(mpt_entry, MPT_BPD);
3361 	hr_reg_clear(mpt_entry, MPT_PA);
3362 	hr_reg_write(mpt_entry, MPT_BQP,
3363 		     mw->ibmw.type == IB_MW_TYPE_1 ? 0 : 1);
3364 
3365 	mpt_entry->lkey = cpu_to_le32(mw->rkey);
3366 
3367 	hr_reg_write(mpt_entry, MPT_PBL_HOP_NUM,
3368 		     mw->pbl_hop_num == HNS_ROCE_HOP_NUM_0 ? 0 :
3369 							     mw->pbl_hop_num);
3370 	hr_reg_write(mpt_entry, MPT_PBL_BA_PG_SZ,
3371 		     mw->pbl_ba_pg_sz + PG_SHIFT_OFFSET);
3372 	hr_reg_write(mpt_entry, MPT_PBL_BUF_PG_SZ,
3373 		     mw->pbl_buf_pg_sz + PG_SHIFT_OFFSET);
3374 
3375 	return 0;
3376 }
3377 
3378 static int free_mr_post_send_lp_wqe(struct hns_roce_qp *hr_qp)
3379 {
3380 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device);
3381 	struct ib_device *ibdev = &hr_dev->ib_dev;
3382 	const struct ib_send_wr *bad_wr;
3383 	struct ib_rdma_wr rdma_wr = {};
3384 	struct ib_send_wr *send_wr;
3385 	int ret;
3386 
3387 	send_wr = &rdma_wr.wr;
3388 	send_wr->opcode = IB_WR_RDMA_WRITE;
3389 
3390 	ret = hns_roce_v2_post_send(&hr_qp->ibqp, send_wr, &bad_wr);
3391 	if (ret) {
3392 		ibdev_err(ibdev, "failed to post wqe for free mr, ret = %d.\n",
3393 			  ret);
3394 		return ret;
3395 	}
3396 
3397 	return 0;
3398 }
3399 
3400 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3401 			       struct ib_wc *wc);
3402 
3403 static void free_mr_send_cmd_to_hw(struct hns_roce_dev *hr_dev)
3404 {
3405 	struct hns_roce_v2_priv *priv = hr_dev->priv;
3406 	struct hns_roce_v2_free_mr *free_mr = &priv->free_mr;
3407 	struct ib_wc wc[ARRAY_SIZE(free_mr->rsv_qp)];
3408 	struct ib_device *ibdev = &hr_dev->ib_dev;
3409 	struct hns_roce_qp *hr_qp;
3410 	unsigned long end;
3411 	int cqe_cnt = 0;
3412 	int npolled;
3413 	int ret;
3414 	int i;
3415 
3416 	/*
3417 	 * If the device initialization is not complete or in the uninstall
3418 	 * process, then there is no need to execute free mr.
3419 	 */
3420 	if (priv->handle->rinfo.reset_state == HNS_ROCE_STATE_RST_INIT ||
3421 	    priv->handle->rinfo.instance_state == HNS_ROCE_STATE_INIT ||
3422 	    hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT)
3423 		return;
3424 
3425 	mutex_lock(&free_mr->mutex);
3426 
3427 	for (i = 0; i < ARRAY_SIZE(free_mr->rsv_qp); i++) {
3428 		hr_qp = free_mr->rsv_qp[i];
3429 
3430 		ret = free_mr_post_send_lp_wqe(hr_qp);
3431 		if (ret) {
3432 			ibdev_err(ibdev,
3433 				  "failed to send wqe (qp:0x%lx) for free mr, ret = %d.\n",
3434 				  hr_qp->qpn, ret);
3435 			break;
3436 		}
3437 
3438 		cqe_cnt++;
3439 	}
3440 
3441 	end = msecs_to_jiffies(HNS_ROCE_V2_FREE_MR_TIMEOUT) + jiffies;
3442 	while (cqe_cnt) {
3443 		npolled = hns_roce_v2_poll_cq(&free_mr->rsv_cq->ib_cq, cqe_cnt, wc);
3444 		if (npolled < 0) {
3445 			ibdev_err(ibdev,
3446 				  "failed to poll cqe for free mr, remain %d cqe.\n",
3447 				  cqe_cnt);
3448 			goto out;
3449 		}
3450 
3451 		if (time_after(jiffies, end)) {
3452 			ibdev_err(ibdev,
3453 				  "failed to poll cqe for free mr and timeout, remain %d cqe.\n",
3454 				  cqe_cnt);
3455 			goto out;
3456 		}
3457 		cqe_cnt -= npolled;
3458 	}
3459 
3460 out:
3461 	mutex_unlock(&free_mr->mutex);
3462 }
3463 
3464 static void hns_roce_v2_dereg_mr(struct hns_roce_dev *hr_dev)
3465 {
3466 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
3467 		free_mr_send_cmd_to_hw(hr_dev);
3468 }
3469 
3470 static void *get_cqe_v2(struct hns_roce_cq *hr_cq, int n)
3471 {
3472 	return hns_roce_buf_offset(hr_cq->mtr.kmem, n * hr_cq->cqe_size);
3473 }
3474 
3475 static void *get_sw_cqe_v2(struct hns_roce_cq *hr_cq, unsigned int n)
3476 {
3477 	struct hns_roce_v2_cqe *cqe = get_cqe_v2(hr_cq, n & hr_cq->ib_cq.cqe);
3478 
3479 	/* Get cqe when Owner bit is Conversely with the MSB of cons_idx */
3480 	return (hr_reg_read(cqe, CQE_OWNER) ^ !!(n & hr_cq->cq_depth)) ? cqe :
3481 									 NULL;
3482 }
3483 
3484 static inline void update_cq_db(struct hns_roce_dev *hr_dev,
3485 				struct hns_roce_cq *hr_cq)
3486 {
3487 	if (likely(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) {
3488 		*hr_cq->set_ci_db = hr_cq->cons_index & V2_CQ_DB_CONS_IDX_M;
3489 	} else {
3490 		struct hns_roce_v2_db cq_db = {};
3491 
3492 		hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3493 		hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB);
3494 		hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3495 		hr_reg_write(&cq_db, DB_CQ_CMD_SN, 1);
3496 
3497 		hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3498 	}
3499 }
3500 
3501 static void __hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3502 				   struct hns_roce_srq *srq)
3503 {
3504 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3505 	struct hns_roce_v2_cqe *cqe, *dest;
3506 	u32 prod_index;
3507 	int nfreed = 0;
3508 	int wqe_index;
3509 	u8 owner_bit;
3510 
3511 	for (prod_index = hr_cq->cons_index; get_sw_cqe_v2(hr_cq, prod_index);
3512 	     ++prod_index) {
3513 		if (prod_index > hr_cq->cons_index + hr_cq->ib_cq.cqe)
3514 			break;
3515 	}
3516 
3517 	/*
3518 	 * Now backwards through the CQ, removing CQ entries
3519 	 * that match our QP by overwriting them with next entries.
3520 	 */
3521 	while ((int) --prod_index - (int) hr_cq->cons_index >= 0) {
3522 		cqe = get_cqe_v2(hr_cq, prod_index & hr_cq->ib_cq.cqe);
3523 		if (hr_reg_read(cqe, CQE_LCL_QPN) == qpn) {
3524 			if (srq && hr_reg_read(cqe, CQE_S_R)) {
3525 				wqe_index = hr_reg_read(cqe, CQE_WQE_IDX);
3526 				hns_roce_free_srq_wqe(srq, wqe_index);
3527 			}
3528 			++nfreed;
3529 		} else if (nfreed) {
3530 			dest = get_cqe_v2(hr_cq, (prod_index + nfreed) &
3531 					  hr_cq->ib_cq.cqe);
3532 			owner_bit = hr_reg_read(dest, CQE_OWNER);
3533 			memcpy(dest, cqe, hr_cq->cqe_size);
3534 			hr_reg_write(dest, CQE_OWNER, owner_bit);
3535 		}
3536 	}
3537 
3538 	if (nfreed) {
3539 		hr_cq->cons_index += nfreed;
3540 		update_cq_db(hr_dev, hr_cq);
3541 	}
3542 }
3543 
3544 static void hns_roce_v2_cq_clean(struct hns_roce_cq *hr_cq, u32 qpn,
3545 				 struct hns_roce_srq *srq)
3546 {
3547 	spin_lock_irq(&hr_cq->lock);
3548 	__hns_roce_v2_cq_clean(hr_cq, qpn, srq);
3549 	spin_unlock_irq(&hr_cq->lock);
3550 }
3551 
3552 static void hns_roce_v2_write_cqc(struct hns_roce_dev *hr_dev,
3553 				  struct hns_roce_cq *hr_cq, void *mb_buf,
3554 				  u64 *mtts, dma_addr_t dma_handle)
3555 {
3556 	struct hns_roce_v2_cq_context *cq_context;
3557 
3558 	cq_context = mb_buf;
3559 	memset(cq_context, 0, sizeof(*cq_context));
3560 
3561 	hr_reg_write(cq_context, CQC_CQ_ST, V2_CQ_STATE_VALID);
3562 	hr_reg_write(cq_context, CQC_ARM_ST, NO_ARMED);
3563 	hr_reg_write(cq_context, CQC_SHIFT, ilog2(hr_cq->cq_depth));
3564 	hr_reg_write(cq_context, CQC_CEQN, hr_cq->vector);
3565 	hr_reg_write(cq_context, CQC_CQN, hr_cq->cqn);
3566 
3567 	if (hr_cq->cqe_size == HNS_ROCE_V3_CQE_SIZE)
3568 		hr_reg_write(cq_context, CQC_CQE_SIZE, CQE_SIZE_64B);
3569 
3570 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
3571 		hr_reg_enable(cq_context, CQC_STASH);
3572 
3573 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_L,
3574 		     to_hr_hw_page_addr(mtts[0]));
3575 	hr_reg_write(cq_context, CQC_CQE_CUR_BLK_ADDR_H,
3576 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
3577 	hr_reg_write(cq_context, CQC_CQE_HOP_NUM, hr_dev->caps.cqe_hop_num ==
3578 		     HNS_ROCE_HOP_NUM_0 ? 0 : hr_dev->caps.cqe_hop_num);
3579 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_L,
3580 		     to_hr_hw_page_addr(mtts[1]));
3581 	hr_reg_write(cq_context, CQC_CQE_NEX_BLK_ADDR_H,
3582 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
3583 	hr_reg_write(cq_context, CQC_CQE_BAR_PG_SZ,
3584 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.ba_pg_shift));
3585 	hr_reg_write(cq_context, CQC_CQE_BUF_PG_SZ,
3586 		     to_hr_hw_page_shift(hr_cq->mtr.hem_cfg.buf_pg_shift));
3587 	hr_reg_write(cq_context, CQC_CQE_BA_L, dma_handle >> CQC_CQE_BA_L_S);
3588 	hr_reg_write(cq_context, CQC_CQE_BA_H, dma_handle >> CQC_CQE_BA_H_S);
3589 	hr_reg_write_bool(cq_context, CQC_DB_RECORD_EN,
3590 			  hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB);
3591 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_L,
3592 		     ((u32)hr_cq->db.dma) >> 1);
3593 	hr_reg_write(cq_context, CQC_CQE_DB_RECORD_ADDR_H,
3594 		     hr_cq->db.dma >> CQC_CQE_DB_RECORD_ADDR_H_S);
3595 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT,
3596 		     HNS_ROCE_V2_CQ_DEFAULT_BURST_NUM);
3597 	hr_reg_write(cq_context, CQC_CQ_PERIOD,
3598 		     HNS_ROCE_V2_CQ_DEFAULT_INTERVAL);
3599 }
3600 
3601 static int hns_roce_v2_req_notify_cq(struct ib_cq *ibcq,
3602 				     enum ib_cq_notify_flags flags)
3603 {
3604 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3605 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3606 	struct hns_roce_v2_db cq_db = {};
3607 	u32 notify_flag;
3608 
3609 	/*
3610 	 * flags = 0, then notify_flag : next
3611 	 * flags = 1, then notify flag : solocited
3612 	 */
3613 	notify_flag = (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
3614 		      V2_CQ_DB_REQ_NOT : V2_CQ_DB_REQ_NOT_SOL;
3615 
3616 	hr_reg_write(&cq_db, DB_TAG, hr_cq->cqn);
3617 	hr_reg_write(&cq_db, DB_CMD, HNS_ROCE_V2_CQ_DB_NOTIFY);
3618 	hr_reg_write(&cq_db, DB_CQ_CI, hr_cq->cons_index);
3619 	hr_reg_write(&cq_db, DB_CQ_CMD_SN, hr_cq->arm_sn);
3620 	hr_reg_write(&cq_db, DB_CQ_NOTIFY, notify_flag);
3621 
3622 	hns_roce_write64(hr_dev, (__le32 *)&cq_db, hr_cq->db_reg);
3623 
3624 	return 0;
3625 }
3626 
3627 static int sw_comp(struct hns_roce_qp *hr_qp, struct hns_roce_wq *wq,
3628 		   int num_entries, struct ib_wc *wc)
3629 {
3630 	unsigned int left;
3631 	int npolled = 0;
3632 
3633 	left = wq->head - wq->tail;
3634 	if (left == 0)
3635 		return 0;
3636 
3637 	left = min_t(unsigned int, (unsigned int)num_entries, left);
3638 	while (npolled < left) {
3639 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3640 		wc->status = IB_WC_WR_FLUSH_ERR;
3641 		wc->vendor_err = 0;
3642 		wc->qp = &hr_qp->ibqp;
3643 
3644 		wq->tail++;
3645 		wc++;
3646 		npolled++;
3647 	}
3648 
3649 	return npolled;
3650 }
3651 
3652 static int hns_roce_v2_sw_poll_cq(struct hns_roce_cq *hr_cq, int num_entries,
3653 				  struct ib_wc *wc)
3654 {
3655 	struct hns_roce_qp *hr_qp;
3656 	int npolled = 0;
3657 
3658 	list_for_each_entry(hr_qp, &hr_cq->sq_list, sq_node) {
3659 		npolled += sw_comp(hr_qp, &hr_qp->sq,
3660 				   num_entries - npolled, wc + npolled);
3661 		if (npolled >= num_entries)
3662 			goto out;
3663 	}
3664 
3665 	list_for_each_entry(hr_qp, &hr_cq->rq_list, rq_node) {
3666 		npolled += sw_comp(hr_qp, &hr_qp->rq,
3667 				   num_entries - npolled, wc + npolled);
3668 		if (npolled >= num_entries)
3669 			goto out;
3670 	}
3671 
3672 out:
3673 	return npolled;
3674 }
3675 
3676 static void get_cqe_status(struct hns_roce_dev *hr_dev, struct hns_roce_qp *qp,
3677 			   struct hns_roce_cq *cq, struct hns_roce_v2_cqe *cqe,
3678 			   struct ib_wc *wc)
3679 {
3680 	static const struct {
3681 		u32 cqe_status;
3682 		enum ib_wc_status wc_status;
3683 	} map[] = {
3684 		{ HNS_ROCE_CQE_V2_SUCCESS, IB_WC_SUCCESS },
3685 		{ HNS_ROCE_CQE_V2_LOCAL_LENGTH_ERR, IB_WC_LOC_LEN_ERR },
3686 		{ HNS_ROCE_CQE_V2_LOCAL_QP_OP_ERR, IB_WC_LOC_QP_OP_ERR },
3687 		{ HNS_ROCE_CQE_V2_LOCAL_PROT_ERR, IB_WC_LOC_PROT_ERR },
3688 		{ HNS_ROCE_CQE_V2_WR_FLUSH_ERR, IB_WC_WR_FLUSH_ERR },
3689 		{ HNS_ROCE_CQE_V2_MW_BIND_ERR, IB_WC_MW_BIND_ERR },
3690 		{ HNS_ROCE_CQE_V2_BAD_RESP_ERR, IB_WC_BAD_RESP_ERR },
3691 		{ HNS_ROCE_CQE_V2_LOCAL_ACCESS_ERR, IB_WC_LOC_ACCESS_ERR },
3692 		{ HNS_ROCE_CQE_V2_REMOTE_INVAL_REQ_ERR, IB_WC_REM_INV_REQ_ERR },
3693 		{ HNS_ROCE_CQE_V2_REMOTE_ACCESS_ERR, IB_WC_REM_ACCESS_ERR },
3694 		{ HNS_ROCE_CQE_V2_REMOTE_OP_ERR, IB_WC_REM_OP_ERR },
3695 		{ HNS_ROCE_CQE_V2_TRANSPORT_RETRY_EXC_ERR,
3696 		  IB_WC_RETRY_EXC_ERR },
3697 		{ HNS_ROCE_CQE_V2_RNR_RETRY_EXC_ERR, IB_WC_RNR_RETRY_EXC_ERR },
3698 		{ HNS_ROCE_CQE_V2_REMOTE_ABORT_ERR, IB_WC_REM_ABORT_ERR },
3699 		{ HNS_ROCE_CQE_V2_GENERAL_ERR, IB_WC_GENERAL_ERR}
3700 	};
3701 
3702 	u32 cqe_status = hr_reg_read(cqe, CQE_STATUS);
3703 	int i;
3704 
3705 	wc->status = IB_WC_GENERAL_ERR;
3706 	for (i = 0; i < ARRAY_SIZE(map); i++)
3707 		if (cqe_status == map[i].cqe_status) {
3708 			wc->status = map[i].wc_status;
3709 			break;
3710 		}
3711 
3712 	if (likely(wc->status == IB_WC_SUCCESS ||
3713 		   wc->status == IB_WC_WR_FLUSH_ERR))
3714 		return;
3715 
3716 	ibdev_err_ratelimited(&hr_dev->ib_dev, "error cqe status 0x%x:\n",
3717 			      cqe_status);
3718 	print_hex_dump(KERN_DEBUG, "", DUMP_PREFIX_NONE, 16, 4, cqe,
3719 		       cq->cqe_size, false);
3720 	wc->vendor_err = hr_reg_read(cqe, CQE_SUB_STATUS);
3721 
3722 	/*
3723 	 * For hns ROCEE, GENERAL_ERR is an error type that is not defined in
3724 	 * the standard protocol, the driver must ignore it and needn't to set
3725 	 * the QP to an error state.
3726 	 */
3727 	if (cqe_status == HNS_ROCE_CQE_V2_GENERAL_ERR)
3728 		return;
3729 
3730 	flush_cqe(hr_dev, qp);
3731 }
3732 
3733 static int get_cur_qp(struct hns_roce_cq *hr_cq, struct hns_roce_v2_cqe *cqe,
3734 		      struct hns_roce_qp **cur_qp)
3735 {
3736 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3737 	struct hns_roce_qp *hr_qp = *cur_qp;
3738 	u32 qpn;
3739 
3740 	qpn = hr_reg_read(cqe, CQE_LCL_QPN);
3741 
3742 	if (!hr_qp || qpn != hr_qp->qpn) {
3743 		hr_qp = __hns_roce_qp_lookup(hr_dev, qpn);
3744 		if (unlikely(!hr_qp)) {
3745 			ibdev_err(&hr_dev->ib_dev,
3746 				  "CQ %06lx with entry for unknown QPN %06x\n",
3747 				  hr_cq->cqn, qpn);
3748 			return -EINVAL;
3749 		}
3750 		*cur_qp = hr_qp;
3751 	}
3752 
3753 	return 0;
3754 }
3755 
3756 /*
3757  * mapped-value = 1 + real-value
3758  * The ib wc opcode's real value is start from 0, In order to distinguish
3759  * between initialized and uninitialized map values, we plus 1 to the actual
3760  * value when defining the mapping, so that the validity can be identified by
3761  * checking whether the mapped value is greater than 0.
3762  */
3763 #define HR_WC_OP_MAP(hr_key, ib_key) \
3764 		[HNS_ROCE_V2_WQE_OP_ ## hr_key] = 1 + IB_WC_ ## ib_key
3765 
3766 static const u32 wc_send_op_map[] = {
3767 	HR_WC_OP_MAP(SEND,			SEND),
3768 	HR_WC_OP_MAP(SEND_WITH_INV,		SEND),
3769 	HR_WC_OP_MAP(SEND_WITH_IMM,		SEND),
3770 	HR_WC_OP_MAP(RDMA_READ,			RDMA_READ),
3771 	HR_WC_OP_MAP(RDMA_WRITE,		RDMA_WRITE),
3772 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,	RDMA_WRITE),
3773 	HR_WC_OP_MAP(ATOM_CMP_AND_SWAP,		COMP_SWAP),
3774 	HR_WC_OP_MAP(ATOM_FETCH_AND_ADD,	FETCH_ADD),
3775 	HR_WC_OP_MAP(ATOM_MSK_CMP_AND_SWAP,	MASKED_COMP_SWAP),
3776 	HR_WC_OP_MAP(ATOM_MSK_FETCH_AND_ADD,	MASKED_FETCH_ADD),
3777 	HR_WC_OP_MAP(FAST_REG_PMR,		REG_MR),
3778 	HR_WC_OP_MAP(BIND_MW,			REG_MR),
3779 };
3780 
3781 static int to_ib_wc_send_op(u32 hr_opcode)
3782 {
3783 	if (hr_opcode >= ARRAY_SIZE(wc_send_op_map))
3784 		return -EINVAL;
3785 
3786 	return wc_send_op_map[hr_opcode] ? wc_send_op_map[hr_opcode] - 1 :
3787 					   -EINVAL;
3788 }
3789 
3790 static const u32 wc_recv_op_map[] = {
3791 	HR_WC_OP_MAP(RDMA_WRITE_WITH_IMM,		WITH_IMM),
3792 	HR_WC_OP_MAP(SEND,				RECV),
3793 	HR_WC_OP_MAP(SEND_WITH_IMM,			WITH_IMM),
3794 	HR_WC_OP_MAP(SEND_WITH_INV,			RECV),
3795 };
3796 
3797 static int to_ib_wc_recv_op(u32 hr_opcode)
3798 {
3799 	if (hr_opcode >= ARRAY_SIZE(wc_recv_op_map))
3800 		return -EINVAL;
3801 
3802 	return wc_recv_op_map[hr_opcode] ? wc_recv_op_map[hr_opcode] - 1 :
3803 					   -EINVAL;
3804 }
3805 
3806 static void fill_send_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3807 {
3808 	u32 hr_opcode;
3809 	int ib_opcode;
3810 
3811 	wc->wc_flags = 0;
3812 
3813 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3814 	switch (hr_opcode) {
3815 	case HNS_ROCE_V2_WQE_OP_RDMA_READ:
3816 		wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3817 		break;
3818 	case HNS_ROCE_V2_WQE_OP_SEND_WITH_IMM:
3819 	case HNS_ROCE_V2_WQE_OP_RDMA_WRITE_WITH_IMM:
3820 		wc->wc_flags |= IB_WC_WITH_IMM;
3821 		break;
3822 	case HNS_ROCE_V2_WQE_OP_ATOM_CMP_AND_SWAP:
3823 	case HNS_ROCE_V2_WQE_OP_ATOM_FETCH_AND_ADD:
3824 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_CMP_AND_SWAP:
3825 	case HNS_ROCE_V2_WQE_OP_ATOM_MSK_FETCH_AND_ADD:
3826 		wc->byte_len  = 8;
3827 		break;
3828 	default:
3829 		break;
3830 	}
3831 
3832 	ib_opcode = to_ib_wc_send_op(hr_opcode);
3833 	if (ib_opcode < 0)
3834 		wc->status = IB_WC_GENERAL_ERR;
3835 	else
3836 		wc->opcode = ib_opcode;
3837 }
3838 
3839 static int fill_recv_wc(struct ib_wc *wc, struct hns_roce_v2_cqe *cqe)
3840 {
3841 	u32 hr_opcode;
3842 	int ib_opcode;
3843 
3844 	wc->byte_len = le32_to_cpu(cqe->byte_cnt);
3845 
3846 	hr_opcode = hr_reg_read(cqe, CQE_OPCODE);
3847 	switch (hr_opcode) {
3848 	case HNS_ROCE_V2_OPCODE_RDMA_WRITE_IMM:
3849 	case HNS_ROCE_V2_OPCODE_SEND_WITH_IMM:
3850 		wc->wc_flags = IB_WC_WITH_IMM;
3851 		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->immtdata));
3852 		break;
3853 	case HNS_ROCE_V2_OPCODE_SEND_WITH_INV:
3854 		wc->wc_flags = IB_WC_WITH_INVALIDATE;
3855 		wc->ex.invalidate_rkey = le32_to_cpu(cqe->rkey);
3856 		break;
3857 	default:
3858 		wc->wc_flags = 0;
3859 	}
3860 
3861 	ib_opcode = to_ib_wc_recv_op(hr_opcode);
3862 	if (ib_opcode < 0)
3863 		wc->status = IB_WC_GENERAL_ERR;
3864 	else
3865 		wc->opcode = ib_opcode;
3866 
3867 	wc->sl = hr_reg_read(cqe, CQE_SL);
3868 	wc->src_qp = hr_reg_read(cqe, CQE_RMT_QPN);
3869 	wc->slid = 0;
3870 	wc->wc_flags |= hr_reg_read(cqe, CQE_GRH) ? IB_WC_GRH : 0;
3871 	wc->port_num = hr_reg_read(cqe, CQE_PORTN);
3872 	wc->pkey_index = 0;
3873 
3874 	if (hr_reg_read(cqe, CQE_VID_VLD)) {
3875 		wc->vlan_id = hr_reg_read(cqe, CQE_VID);
3876 		wc->wc_flags |= IB_WC_WITH_VLAN;
3877 	} else {
3878 		wc->vlan_id = 0xffff;
3879 	}
3880 
3881 	wc->network_hdr_type = hr_reg_read(cqe, CQE_PORT_TYPE);
3882 
3883 	return 0;
3884 }
3885 
3886 static int hns_roce_v2_poll_one(struct hns_roce_cq *hr_cq,
3887 				struct hns_roce_qp **cur_qp, struct ib_wc *wc)
3888 {
3889 	struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device);
3890 	struct hns_roce_qp *qp = *cur_qp;
3891 	struct hns_roce_srq *srq = NULL;
3892 	struct hns_roce_v2_cqe *cqe;
3893 	struct hns_roce_wq *wq;
3894 	int is_send;
3895 	u16 wqe_idx;
3896 	int ret;
3897 
3898 	cqe = get_sw_cqe_v2(hr_cq, hr_cq->cons_index);
3899 	if (!cqe)
3900 		return -EAGAIN;
3901 
3902 	++hr_cq->cons_index;
3903 	/* Memory barrier */
3904 	rmb();
3905 
3906 	ret = get_cur_qp(hr_cq, cqe, &qp);
3907 	if (ret)
3908 		return ret;
3909 
3910 	wc->qp = &qp->ibqp;
3911 	wc->vendor_err = 0;
3912 
3913 	wqe_idx = hr_reg_read(cqe, CQE_WQE_IDX);
3914 
3915 	is_send = !hr_reg_read(cqe, CQE_S_R);
3916 	if (is_send) {
3917 		wq = &qp->sq;
3918 
3919 		/* If sg_signal_bit is set, tail pointer will be updated to
3920 		 * the WQE corresponding to the current CQE.
3921 		 */
3922 		if (qp->sq_signal_bits)
3923 			wq->tail += (wqe_idx - (u16)wq->tail) &
3924 				    (wq->wqe_cnt - 1);
3925 
3926 		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3927 		++wq->tail;
3928 
3929 		fill_send_wc(wc, cqe);
3930 	} else {
3931 		if (qp->ibqp.srq) {
3932 			srq = to_hr_srq(qp->ibqp.srq);
3933 			wc->wr_id = srq->wrid[wqe_idx];
3934 			hns_roce_free_srq_wqe(srq, wqe_idx);
3935 		} else {
3936 			wq = &qp->rq;
3937 			wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
3938 			++wq->tail;
3939 		}
3940 
3941 		ret = fill_recv_wc(wc, cqe);
3942 	}
3943 
3944 	get_cqe_status(hr_dev, qp, hr_cq, cqe, wc);
3945 	if (unlikely(wc->status != IB_WC_SUCCESS))
3946 		return 0;
3947 
3948 	return ret;
3949 }
3950 
3951 static int hns_roce_v2_poll_cq(struct ib_cq *ibcq, int num_entries,
3952 			       struct ib_wc *wc)
3953 {
3954 	struct hns_roce_dev *hr_dev = to_hr_dev(ibcq->device);
3955 	struct hns_roce_cq *hr_cq = to_hr_cq(ibcq);
3956 	struct hns_roce_qp *cur_qp = NULL;
3957 	unsigned long flags;
3958 	int npolled;
3959 
3960 	spin_lock_irqsave(&hr_cq->lock, flags);
3961 
3962 	/*
3963 	 * When the device starts to reset, the state is RST_DOWN. At this time,
3964 	 * there may still be some valid CQEs in the hardware that are not
3965 	 * polled. Therefore, it is not allowed to switch to the software mode
3966 	 * immediately. When the state changes to UNINIT, CQE no longer exists
3967 	 * in the hardware, and then switch to software mode.
3968 	 */
3969 	if (hr_dev->state == HNS_ROCE_DEVICE_STATE_UNINIT) {
3970 		npolled = hns_roce_v2_sw_poll_cq(hr_cq, num_entries, wc);
3971 		goto out;
3972 	}
3973 
3974 	for (npolled = 0; npolled < num_entries; ++npolled) {
3975 		if (hns_roce_v2_poll_one(hr_cq, &cur_qp, wc + npolled))
3976 			break;
3977 	}
3978 
3979 	if (npolled)
3980 		update_cq_db(hr_dev, hr_cq);
3981 
3982 out:
3983 	spin_unlock_irqrestore(&hr_cq->lock, flags);
3984 
3985 	return npolled;
3986 }
3987 
3988 static int get_op_for_set_hem(struct hns_roce_dev *hr_dev, u32 type,
3989 			      u32 step_idx, u8 *mbox_cmd)
3990 {
3991 	u8 cmd;
3992 
3993 	switch (type) {
3994 	case HEM_TYPE_QPC:
3995 		cmd = HNS_ROCE_CMD_WRITE_QPC_BT0;
3996 		break;
3997 	case HEM_TYPE_MTPT:
3998 		cmd = HNS_ROCE_CMD_WRITE_MPT_BT0;
3999 		break;
4000 	case HEM_TYPE_CQC:
4001 		cmd = HNS_ROCE_CMD_WRITE_CQC_BT0;
4002 		break;
4003 	case HEM_TYPE_SRQC:
4004 		cmd = HNS_ROCE_CMD_WRITE_SRQC_BT0;
4005 		break;
4006 	case HEM_TYPE_SCCC:
4007 		cmd = HNS_ROCE_CMD_WRITE_SCCC_BT0;
4008 		break;
4009 	case HEM_TYPE_QPC_TIMER:
4010 		cmd = HNS_ROCE_CMD_WRITE_QPC_TIMER_BT0;
4011 		break;
4012 	case HEM_TYPE_CQC_TIMER:
4013 		cmd = HNS_ROCE_CMD_WRITE_CQC_TIMER_BT0;
4014 		break;
4015 	default:
4016 		dev_warn(hr_dev->dev, "failed to check hem type %u.\n", type);
4017 		return -EINVAL;
4018 	}
4019 
4020 	*mbox_cmd = cmd + step_idx;
4021 
4022 	return 0;
4023 }
4024 
4025 static int config_gmv_ba_to_hw(struct hns_roce_dev *hr_dev, unsigned long obj,
4026 			       dma_addr_t base_addr)
4027 {
4028 	struct hns_roce_cmq_desc desc;
4029 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
4030 	u32 idx = obj / (HNS_HW_PAGE_SIZE / hr_dev->caps.gmv_entry_sz);
4031 	u64 addr = to_hr_hw_page_addr(base_addr);
4032 
4033 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
4034 
4035 	hr_reg_write(req, CFG_GMV_BT_BA_L, lower_32_bits(addr));
4036 	hr_reg_write(req, CFG_GMV_BT_BA_H, upper_32_bits(addr));
4037 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
4038 
4039 	return hns_roce_cmq_send(hr_dev, &desc, 1);
4040 }
4041 
4042 static int set_hem_to_hw(struct hns_roce_dev *hr_dev, int obj,
4043 			 dma_addr_t base_addr, u32 hem_type, u32 step_idx)
4044 {
4045 	int ret;
4046 	u8 cmd;
4047 
4048 	if (unlikely(hem_type == HEM_TYPE_GMV))
4049 		return config_gmv_ba_to_hw(hr_dev, obj, base_addr);
4050 
4051 	if (unlikely(hem_type == HEM_TYPE_SCCC && step_idx))
4052 		return 0;
4053 
4054 	ret = get_op_for_set_hem(hr_dev, hem_type, step_idx, &cmd);
4055 	if (ret < 0)
4056 		return ret;
4057 
4058 	return config_hem_ba_to_hw(hr_dev, base_addr, cmd, obj);
4059 }
4060 
4061 static int hns_roce_v2_set_hem(struct hns_roce_dev *hr_dev,
4062 			       struct hns_roce_hem_table *table, int obj,
4063 			       u32 step_idx)
4064 {
4065 	struct hns_roce_hem_mhop mhop;
4066 	struct hns_roce_hem *hem;
4067 	unsigned long mhop_obj = obj;
4068 	int i, j, k;
4069 	int ret = 0;
4070 	u64 hem_idx = 0;
4071 	u64 l1_idx = 0;
4072 	u64 bt_ba = 0;
4073 	u32 chunk_ba_num;
4074 	u32 hop_num;
4075 
4076 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4077 		return 0;
4078 
4079 	hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
4080 	i = mhop.l0_idx;
4081 	j = mhop.l1_idx;
4082 	k = mhop.l2_idx;
4083 	hop_num = mhop.hop_num;
4084 	chunk_ba_num = mhop.bt_chunk_size / 8;
4085 
4086 	if (hop_num == 2) {
4087 		hem_idx = i * chunk_ba_num * chunk_ba_num + j * chunk_ba_num +
4088 			  k;
4089 		l1_idx = i * chunk_ba_num + j;
4090 	} else if (hop_num == 1) {
4091 		hem_idx = i * chunk_ba_num + j;
4092 	} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
4093 		hem_idx = i;
4094 	}
4095 
4096 	if (table->type == HEM_TYPE_SCCC)
4097 		obj = mhop.l0_idx;
4098 
4099 	if (check_whether_last_step(hop_num, step_idx)) {
4100 		hem = table->hem[hem_idx];
4101 
4102 		ret = set_hem_to_hw(hr_dev, obj, hem->dma, table->type, step_idx);
4103 	} else {
4104 		if (step_idx == 0)
4105 			bt_ba = table->bt_l0_dma_addr[i];
4106 		else if (step_idx == 1 && hop_num == 2)
4107 			bt_ba = table->bt_l1_dma_addr[l1_idx];
4108 
4109 		ret = set_hem_to_hw(hr_dev, obj, bt_ba, table->type, step_idx);
4110 	}
4111 
4112 	return ret;
4113 }
4114 
4115 static int hns_roce_v2_clear_hem(struct hns_roce_dev *hr_dev,
4116 				 struct hns_roce_hem_table *table,
4117 				 int tag, u32 step_idx)
4118 {
4119 	struct hns_roce_cmd_mailbox *mailbox;
4120 	struct device *dev = hr_dev->dev;
4121 	u8 cmd = 0xff;
4122 	int ret;
4123 
4124 	if (!hns_roce_check_whether_mhop(hr_dev, table->type))
4125 		return 0;
4126 
4127 	switch (table->type) {
4128 	case HEM_TYPE_QPC:
4129 		cmd = HNS_ROCE_CMD_DESTROY_QPC_BT0;
4130 		break;
4131 	case HEM_TYPE_MTPT:
4132 		cmd = HNS_ROCE_CMD_DESTROY_MPT_BT0;
4133 		break;
4134 	case HEM_TYPE_CQC:
4135 		cmd = HNS_ROCE_CMD_DESTROY_CQC_BT0;
4136 		break;
4137 	case HEM_TYPE_SRQC:
4138 		cmd = HNS_ROCE_CMD_DESTROY_SRQC_BT0;
4139 		break;
4140 	case HEM_TYPE_SCCC:
4141 	case HEM_TYPE_QPC_TIMER:
4142 	case HEM_TYPE_CQC_TIMER:
4143 	case HEM_TYPE_GMV:
4144 		return 0;
4145 	default:
4146 		dev_warn(dev, "table %u not to be destroyed by mailbox!\n",
4147 			 table->type);
4148 		return 0;
4149 	}
4150 
4151 	cmd += step_idx;
4152 
4153 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4154 	if (IS_ERR(mailbox))
4155 		return PTR_ERR(mailbox);
4156 
4157 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, cmd, tag);
4158 
4159 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4160 	return ret;
4161 }
4162 
4163 static int hns_roce_v2_qp_modify(struct hns_roce_dev *hr_dev,
4164 				 struct hns_roce_v2_qp_context *context,
4165 				 struct hns_roce_v2_qp_context *qpc_mask,
4166 				 struct hns_roce_qp *hr_qp)
4167 {
4168 	struct hns_roce_cmd_mailbox *mailbox;
4169 	int qpc_size;
4170 	int ret;
4171 
4172 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
4173 	if (IS_ERR(mailbox))
4174 		return PTR_ERR(mailbox);
4175 
4176 	/* The qpc size of HIP08 is only 256B, which is half of HIP09 */
4177 	qpc_size = hr_dev->caps.qpc_sz;
4178 	memcpy(mailbox->buf, context, qpc_size);
4179 	memcpy(mailbox->buf + qpc_size, qpc_mask, qpc_size);
4180 
4181 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
4182 				HNS_ROCE_CMD_MODIFY_QPC, hr_qp->qpn);
4183 
4184 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
4185 
4186 	return ret;
4187 }
4188 
4189 static void set_access_flags(struct hns_roce_qp *hr_qp,
4190 			     struct hns_roce_v2_qp_context *context,
4191 			     struct hns_roce_v2_qp_context *qpc_mask,
4192 			     const struct ib_qp_attr *attr, int attr_mask)
4193 {
4194 	u8 dest_rd_atomic;
4195 	u32 access_flags;
4196 
4197 	dest_rd_atomic = (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) ?
4198 			 attr->max_dest_rd_atomic : hr_qp->resp_depth;
4199 
4200 	access_flags = (attr_mask & IB_QP_ACCESS_FLAGS) ?
4201 		       attr->qp_access_flags : hr_qp->atomic_rd_en;
4202 
4203 	if (!dest_rd_atomic)
4204 		access_flags &= IB_ACCESS_REMOTE_WRITE;
4205 
4206 	hr_reg_write_bool(context, QPC_RRE,
4207 			  access_flags & IB_ACCESS_REMOTE_READ);
4208 	hr_reg_clear(qpc_mask, QPC_RRE);
4209 
4210 	hr_reg_write_bool(context, QPC_RWE,
4211 			  access_flags & IB_ACCESS_REMOTE_WRITE);
4212 	hr_reg_clear(qpc_mask, QPC_RWE);
4213 
4214 	hr_reg_write_bool(context, QPC_ATE,
4215 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4216 	hr_reg_clear(qpc_mask, QPC_ATE);
4217 	hr_reg_write_bool(context, QPC_EXT_ATE,
4218 			  access_flags & IB_ACCESS_REMOTE_ATOMIC);
4219 	hr_reg_clear(qpc_mask, QPC_EXT_ATE);
4220 }
4221 
4222 static void set_qpc_wqe_cnt(struct hns_roce_qp *hr_qp,
4223 			    struct hns_roce_v2_qp_context *context)
4224 {
4225 	hr_reg_write(context, QPC_SGE_SHIFT,
4226 		     to_hr_hem_entries_shift(hr_qp->sge.sge_cnt,
4227 					     hr_qp->sge.sge_shift));
4228 
4229 	hr_reg_write(context, QPC_SQ_SHIFT, ilog2(hr_qp->sq.wqe_cnt));
4230 
4231 	hr_reg_write(context, QPC_RQ_SHIFT, ilog2(hr_qp->rq.wqe_cnt));
4232 }
4233 
4234 static inline int get_cqn(struct ib_cq *ib_cq)
4235 {
4236 	return ib_cq ? to_hr_cq(ib_cq)->cqn : 0;
4237 }
4238 
4239 static inline int get_pdn(struct ib_pd *ib_pd)
4240 {
4241 	return ib_pd ? to_hr_pd(ib_pd)->pdn : 0;
4242 }
4243 
4244 static void modify_qp_reset_to_init(struct ib_qp *ibqp,
4245 				    struct hns_roce_v2_qp_context *context,
4246 				    struct hns_roce_v2_qp_context *qpc_mask)
4247 {
4248 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4249 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4250 
4251 	/*
4252 	 * In v2 engine, software pass context and context mask to hardware
4253 	 * when modifying qp. If software need modify some fields in context,
4254 	 * we should set all bits of the relevant fields in context mask to
4255 	 * 0 at the same time, else set them to 0x1.
4256 	 */
4257 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4258 
4259 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4260 
4261 	hr_reg_write(context, QPC_RQWS, ilog2(hr_qp->rq.max_gs));
4262 
4263 	set_qpc_wqe_cnt(hr_qp, context);
4264 
4265 	/* No VLAN need to set 0xFFF */
4266 	hr_reg_write(context, QPC_VLAN_ID, 0xfff);
4267 
4268 	if (ibqp->qp_type == IB_QPT_XRC_TGT) {
4269 		context->qkey_xrcd = cpu_to_le32(hr_qp->xrcdn);
4270 
4271 		hr_reg_enable(context, QPC_XRC_QP_TYPE);
4272 	}
4273 
4274 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
4275 		hr_reg_enable(context, QPC_RQ_RECORD_EN);
4276 
4277 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_OWNER_DB)
4278 		hr_reg_enable(context, QPC_OWNER_MODE);
4279 
4280 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_L,
4281 		     lower_32_bits(hr_qp->rdb.dma) >> 1);
4282 	hr_reg_write(context, QPC_RQ_DB_RECORD_ADDR_H,
4283 		     upper_32_bits(hr_qp->rdb.dma));
4284 
4285 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4286 
4287 	if (ibqp->srq) {
4288 		hr_reg_enable(context, QPC_SRQ_EN);
4289 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4290 	}
4291 
4292 	hr_reg_enable(context, QPC_FRE);
4293 
4294 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4295 
4296 	if (hr_dev->caps.qpc_sz < HNS_ROCE_V3_QPC_SZ)
4297 		return;
4298 
4299 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_STASH)
4300 		hr_reg_enable(&context->ext, QPCEX_STASH);
4301 }
4302 
4303 static void modify_qp_init_to_init(struct ib_qp *ibqp,
4304 				   struct hns_roce_v2_qp_context *context,
4305 				   struct hns_roce_v2_qp_context *qpc_mask)
4306 {
4307 	/*
4308 	 * In v2 engine, software pass context and context mask to hardware
4309 	 * when modifying qp. If software need modify some fields in context,
4310 	 * we should set all bits of the relevant fields in context mask to
4311 	 * 0 at the same time, else set them to 0x1.
4312 	 */
4313 	hr_reg_write(context, QPC_TST, to_hr_qp_type(ibqp->qp_type));
4314 	hr_reg_clear(qpc_mask, QPC_TST);
4315 
4316 	hr_reg_write(context, QPC_PD, get_pdn(ibqp->pd));
4317 	hr_reg_clear(qpc_mask, QPC_PD);
4318 
4319 	hr_reg_write(context, QPC_RX_CQN, get_cqn(ibqp->recv_cq));
4320 	hr_reg_clear(qpc_mask, QPC_RX_CQN);
4321 
4322 	hr_reg_write(context, QPC_TX_CQN, get_cqn(ibqp->send_cq));
4323 	hr_reg_clear(qpc_mask, QPC_TX_CQN);
4324 
4325 	if (ibqp->srq) {
4326 		hr_reg_enable(context, QPC_SRQ_EN);
4327 		hr_reg_clear(qpc_mask, QPC_SRQ_EN);
4328 		hr_reg_write(context, QPC_SRQN, to_hr_srq(ibqp->srq)->srqn);
4329 		hr_reg_clear(qpc_mask, QPC_SRQN);
4330 	}
4331 }
4332 
4333 static int config_qp_rq_buf(struct hns_roce_dev *hr_dev,
4334 			    struct hns_roce_qp *hr_qp,
4335 			    struct hns_roce_v2_qp_context *context,
4336 			    struct hns_roce_v2_qp_context *qpc_mask)
4337 {
4338 	u64 mtts[MTT_MIN_COUNT] = { 0 };
4339 	u64 wqe_sge_ba;
4340 	int ret;
4341 
4342 	/* Search qp buf's mtts */
4343 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->rq.offset, mtts,
4344 				MTT_MIN_COUNT);
4345 	if (hr_qp->rq.wqe_cnt && ret) {
4346 		ibdev_err(&hr_dev->ib_dev,
4347 			  "failed to find QP(0x%lx) RQ WQE buf, ret = %d.\n",
4348 			  hr_qp->qpn, ret);
4349 		return ret;
4350 	}
4351 
4352 	wqe_sge_ba = hns_roce_get_mtr_ba(&hr_qp->mtr);
4353 
4354 	context->wqe_sge_ba = cpu_to_le32(wqe_sge_ba >> 3);
4355 	qpc_mask->wqe_sge_ba = 0;
4356 
4357 	/*
4358 	 * In v2 engine, software pass context and context mask to hardware
4359 	 * when modifying qp. If software need modify some fields in context,
4360 	 * we should set all bits of the relevant fields in context mask to
4361 	 * 0 at the same time, else set them to 0x1.
4362 	 */
4363 	hr_reg_write(context, QPC_WQE_SGE_BA_H, wqe_sge_ba >> (32 + 3));
4364 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_H);
4365 
4366 	hr_reg_write(context, QPC_SQ_HOP_NUM,
4367 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sq_hop_num,
4368 				      hr_qp->sq.wqe_cnt));
4369 	hr_reg_clear(qpc_mask, QPC_SQ_HOP_NUM);
4370 
4371 	hr_reg_write(context, QPC_SGE_HOP_NUM,
4372 		     to_hr_hem_hopnum(hr_dev->caps.wqe_sge_hop_num,
4373 				      hr_qp->sge.sge_cnt));
4374 	hr_reg_clear(qpc_mask, QPC_SGE_HOP_NUM);
4375 
4376 	hr_reg_write(context, QPC_RQ_HOP_NUM,
4377 		     to_hr_hem_hopnum(hr_dev->caps.wqe_rq_hop_num,
4378 				      hr_qp->rq.wqe_cnt));
4379 
4380 	hr_reg_clear(qpc_mask, QPC_RQ_HOP_NUM);
4381 
4382 	hr_reg_write(context, QPC_WQE_SGE_BA_PG_SZ,
4383 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.ba_pg_shift));
4384 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BA_PG_SZ);
4385 
4386 	hr_reg_write(context, QPC_WQE_SGE_BUF_PG_SZ,
4387 		     to_hr_hw_page_shift(hr_qp->mtr.hem_cfg.buf_pg_shift));
4388 	hr_reg_clear(qpc_mask, QPC_WQE_SGE_BUF_PG_SZ);
4389 
4390 	context->rq_cur_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[0]));
4391 	qpc_mask->rq_cur_blk_addr = 0;
4392 
4393 	hr_reg_write(context, QPC_RQ_CUR_BLK_ADDR_H,
4394 		     upper_32_bits(to_hr_hw_page_addr(mtts[0])));
4395 	hr_reg_clear(qpc_mask, QPC_RQ_CUR_BLK_ADDR_H);
4396 
4397 	context->rq_nxt_blk_addr = cpu_to_le32(to_hr_hw_page_addr(mtts[1]));
4398 	qpc_mask->rq_nxt_blk_addr = 0;
4399 
4400 	hr_reg_write(context, QPC_RQ_NXT_BLK_ADDR_H,
4401 		     upper_32_bits(to_hr_hw_page_addr(mtts[1])));
4402 	hr_reg_clear(qpc_mask, QPC_RQ_NXT_BLK_ADDR_H);
4403 
4404 	return 0;
4405 }
4406 
4407 static int config_qp_sq_buf(struct hns_roce_dev *hr_dev,
4408 			    struct hns_roce_qp *hr_qp,
4409 			    struct hns_roce_v2_qp_context *context,
4410 			    struct hns_roce_v2_qp_context *qpc_mask)
4411 {
4412 	struct ib_device *ibdev = &hr_dev->ib_dev;
4413 	u64 sge_cur_blk = 0;
4414 	u64 sq_cur_blk = 0;
4415 	int ret;
4416 
4417 	/* search qp buf's mtts */
4418 	ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr, hr_qp->sq.offset,
4419 				&sq_cur_blk, 1);
4420 	if (ret) {
4421 		ibdev_err(ibdev, "failed to find QP(0x%lx) SQ WQE buf, ret = %d.\n",
4422 			  hr_qp->qpn, ret);
4423 		return ret;
4424 	}
4425 	if (hr_qp->sge.sge_cnt > 0) {
4426 		ret = hns_roce_mtr_find(hr_dev, &hr_qp->mtr,
4427 					hr_qp->sge.offset, &sge_cur_blk, 1);
4428 		if (ret) {
4429 			ibdev_err(ibdev, "failed to find QP(0x%lx) SGE buf, ret = %d.\n",
4430 				  hr_qp->qpn, ret);
4431 			return ret;
4432 		}
4433 	}
4434 
4435 	/*
4436 	 * In v2 engine, software pass context and context mask to hardware
4437 	 * when modifying qp. If software need modify some fields in context,
4438 	 * we should set all bits of the relevant fields in context mask to
4439 	 * 0 at the same time, else set them to 0x1.
4440 	 */
4441 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_L,
4442 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4443 	hr_reg_write(context, QPC_SQ_CUR_BLK_ADDR_H,
4444 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4445 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_L);
4446 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_BLK_ADDR_H);
4447 
4448 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_L,
4449 		     lower_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4450 	hr_reg_write(context, QPC_SQ_CUR_SGE_BLK_ADDR_H,
4451 		     upper_32_bits(to_hr_hw_page_addr(sge_cur_blk)));
4452 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_L);
4453 	hr_reg_clear(qpc_mask, QPC_SQ_CUR_SGE_BLK_ADDR_H);
4454 
4455 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_L,
4456 		     lower_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4457 	hr_reg_write(context, QPC_RX_SQ_CUR_BLK_ADDR_H,
4458 		     upper_32_bits(to_hr_hw_page_addr(sq_cur_blk)));
4459 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_L);
4460 	hr_reg_clear(qpc_mask, QPC_RX_SQ_CUR_BLK_ADDR_H);
4461 
4462 	return 0;
4463 }
4464 
4465 static inline enum ib_mtu get_mtu(struct ib_qp *ibqp,
4466 				  const struct ib_qp_attr *attr)
4467 {
4468 	if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_UD)
4469 		return IB_MTU_4096;
4470 
4471 	return attr->path_mtu;
4472 }
4473 
4474 static int modify_qp_init_to_rtr(struct ib_qp *ibqp,
4475 				 const struct ib_qp_attr *attr, int attr_mask,
4476 				 struct hns_roce_v2_qp_context *context,
4477 				 struct hns_roce_v2_qp_context *qpc_mask,
4478 				 struct ib_udata *udata)
4479 {
4480 	struct hns_roce_ucontext *uctx = rdma_udata_to_drv_context(udata,
4481 					  struct hns_roce_ucontext, ibucontext);
4482 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4483 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4484 	struct ib_device *ibdev = &hr_dev->ib_dev;
4485 	dma_addr_t trrl_ba;
4486 	dma_addr_t irrl_ba;
4487 	enum ib_mtu ib_mtu;
4488 	const u8 *smac;
4489 	u8 lp_pktn_ini;
4490 	u64 *mtts;
4491 	u8 *dmac;
4492 	u32 port;
4493 	int mtu;
4494 	int ret;
4495 
4496 	ret = config_qp_rq_buf(hr_dev, hr_qp, context, qpc_mask);
4497 	if (ret) {
4498 		ibdev_err(ibdev, "failed to config rq buf, ret = %d.\n", ret);
4499 		return ret;
4500 	}
4501 
4502 	/* Search IRRL's mtts */
4503 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.irrl_table,
4504 				   hr_qp->qpn, &irrl_ba);
4505 	if (!mtts) {
4506 		ibdev_err(ibdev, "failed to find qp irrl_table.\n");
4507 		return -EINVAL;
4508 	}
4509 
4510 	/* Search TRRL's mtts */
4511 	mtts = hns_roce_table_find(hr_dev, &hr_dev->qp_table.trrl_table,
4512 				   hr_qp->qpn, &trrl_ba);
4513 	if (!mtts) {
4514 		ibdev_err(ibdev, "failed to find qp trrl_table.\n");
4515 		return -EINVAL;
4516 	}
4517 
4518 	if (attr_mask & IB_QP_ALT_PATH) {
4519 		ibdev_err(ibdev, "INIT2RTR attr_mask (0x%x) error.\n",
4520 			  attr_mask);
4521 		return -EINVAL;
4522 	}
4523 
4524 	hr_reg_write(context, QPC_TRRL_BA_L, trrl_ba >> QPC_TRRL_BA_L_S);
4525 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_L);
4526 	context->trrl_ba = cpu_to_le32(trrl_ba >> QPC_TRRL_BA_M_S);
4527 	qpc_mask->trrl_ba = 0;
4528 	hr_reg_write(context, QPC_TRRL_BA_H, trrl_ba >> QPC_TRRL_BA_H_S);
4529 	hr_reg_clear(qpc_mask, QPC_TRRL_BA_H);
4530 
4531 	context->irrl_ba = cpu_to_le32(irrl_ba >> QPC_IRRL_BA_L_S);
4532 	qpc_mask->irrl_ba = 0;
4533 	hr_reg_write(context, QPC_IRRL_BA_H, irrl_ba >> QPC_IRRL_BA_H_S);
4534 	hr_reg_clear(qpc_mask, QPC_IRRL_BA_H);
4535 
4536 	hr_reg_enable(context, QPC_RMT_E2E);
4537 	hr_reg_clear(qpc_mask, QPC_RMT_E2E);
4538 
4539 	hr_reg_write(context, QPC_SIG_TYPE, hr_qp->sq_signal_bits);
4540 	hr_reg_clear(qpc_mask, QPC_SIG_TYPE);
4541 
4542 	port = (attr_mask & IB_QP_PORT) ? (attr->port_num - 1) : hr_qp->port;
4543 
4544 	smac = (const u8 *)hr_dev->dev_addr[port];
4545 	dmac = (u8 *)attr->ah_attr.roce.dmac;
4546 	/* when dmac equals smac or loop_idc is 1, it should loopback */
4547 	if (ether_addr_equal_unaligned(dmac, smac) ||
4548 	    hr_dev->loop_idc == 0x1) {
4549 		hr_reg_write(context, QPC_LBI, hr_dev->loop_idc);
4550 		hr_reg_clear(qpc_mask, QPC_LBI);
4551 	}
4552 
4553 	if (attr_mask & IB_QP_DEST_QPN) {
4554 		hr_reg_write(context, QPC_DQPN, attr->dest_qp_num);
4555 		hr_reg_clear(qpc_mask, QPC_DQPN);
4556 	}
4557 
4558 	memcpy(&context->dmac, dmac, sizeof(u32));
4559 	hr_reg_write(context, QPC_DMAC_H, *((u16 *)(&dmac[4])));
4560 	qpc_mask->dmac = 0;
4561 	hr_reg_clear(qpc_mask, QPC_DMAC_H);
4562 
4563 	ib_mtu = get_mtu(ibqp, attr);
4564 	hr_qp->path_mtu = ib_mtu;
4565 
4566 	mtu = ib_mtu_enum_to_int(ib_mtu);
4567 	if (WARN_ON(mtu <= 0))
4568 		return -EINVAL;
4569 #define MIN_LP_MSG_LEN 1024
4570 	/* mtu * (2 ^ lp_pktn_ini) should be in the range of 1024 to mtu */
4571 	lp_pktn_ini = ilog2(max(mtu, MIN_LP_MSG_LEN) / mtu);
4572 
4573 	if (attr_mask & IB_QP_PATH_MTU) {
4574 		hr_reg_write(context, QPC_MTU, ib_mtu);
4575 		hr_reg_clear(qpc_mask, QPC_MTU);
4576 	}
4577 
4578 	hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
4579 	hr_reg_clear(qpc_mask, QPC_LP_PKTN_INI);
4580 
4581 	/* ACK_REQ_FREQ should be larger than or equal to LP_PKTN_INI */
4582 	hr_reg_write(context, QPC_ACK_REQ_FREQ, lp_pktn_ini);
4583 	hr_reg_clear(qpc_mask, QPC_ACK_REQ_FREQ);
4584 
4585 	hr_reg_clear(qpc_mask, QPC_RX_REQ_PSN_ERR);
4586 	hr_reg_clear(qpc_mask, QPC_RX_REQ_MSN);
4587 	hr_reg_clear(qpc_mask, QPC_RX_REQ_LAST_OPTYPE);
4588 
4589 	context->rq_rnr_timer = 0;
4590 	qpc_mask->rq_rnr_timer = 0;
4591 
4592 	hr_reg_clear(qpc_mask, QPC_TRRL_HEAD_MAX);
4593 	hr_reg_clear(qpc_mask, QPC_TRRL_TAIL_MAX);
4594 
4595 #define MAX_LP_SGEN 3
4596 	/* rocee send 2^lp_sgen_ini segs every time */
4597 	hr_reg_write(context, QPC_LP_SGEN_INI, MAX_LP_SGEN);
4598 	hr_reg_clear(qpc_mask, QPC_LP_SGEN_INI);
4599 
4600 	if (udata && ibqp->qp_type == IB_QPT_RC &&
4601 	    (uctx->config & HNS_ROCE_RQ_INLINE_FLAGS)) {
4602 		hr_reg_write_bool(context, QPC_RQIE,
4603 				  hr_dev->caps.flags &
4604 				  HNS_ROCE_CAP_FLAG_RQ_INLINE);
4605 		hr_reg_clear(qpc_mask, QPC_RQIE);
4606 	}
4607 
4608 	if (udata &&
4609 	    (ibqp->qp_type == IB_QPT_RC || ibqp->qp_type == IB_QPT_XRC_TGT) &&
4610 	    (uctx->config & HNS_ROCE_CQE_INLINE_FLAGS)) {
4611 		hr_reg_write_bool(context, QPC_CQEIE,
4612 				  hr_dev->caps.flags &
4613 				  HNS_ROCE_CAP_FLAG_CQE_INLINE);
4614 		hr_reg_clear(qpc_mask, QPC_CQEIE);
4615 
4616 		hr_reg_write(context, QPC_CQEIS, 0);
4617 		hr_reg_clear(qpc_mask, QPC_CQEIS);
4618 	}
4619 
4620 	return 0;
4621 }
4622 
4623 static int modify_qp_rtr_to_rts(struct ib_qp *ibqp, int attr_mask,
4624 				struct hns_roce_v2_qp_context *context,
4625 				struct hns_roce_v2_qp_context *qpc_mask)
4626 {
4627 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4628 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4629 	struct ib_device *ibdev = &hr_dev->ib_dev;
4630 	int ret;
4631 
4632 	/* Not support alternate path and path migration */
4633 	if (attr_mask & (IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE)) {
4634 		ibdev_err(ibdev, "RTR2RTS attr_mask (0x%x)error\n", attr_mask);
4635 		return -EINVAL;
4636 	}
4637 
4638 	ret = config_qp_sq_buf(hr_dev, hr_qp, context, qpc_mask);
4639 	if (ret) {
4640 		ibdev_err(ibdev, "failed to config sq buf, ret = %d.\n", ret);
4641 		return ret;
4642 	}
4643 
4644 	/*
4645 	 * Set some fields in context to zero, Because the default values
4646 	 * of all fields in context are zero, we need not set them to 0 again.
4647 	 * but we should set the relevant fields of context mask to 0.
4648 	 */
4649 	hr_reg_clear(qpc_mask, QPC_IRRL_SGE_IDX);
4650 
4651 	hr_reg_clear(qpc_mask, QPC_RX_ACK_MSN);
4652 
4653 	hr_reg_clear(qpc_mask, QPC_ACK_LAST_OPTYPE);
4654 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN_VLD);
4655 	hr_reg_clear(qpc_mask, QPC_IRRL_PSN);
4656 
4657 	hr_reg_clear(qpc_mask, QPC_IRRL_TAIL_REAL);
4658 
4659 	hr_reg_clear(qpc_mask, QPC_RETRY_MSG_MSN);
4660 
4661 	hr_reg_clear(qpc_mask, QPC_RNR_RETRY_FLAG);
4662 
4663 	hr_reg_clear(qpc_mask, QPC_CHECK_FLG);
4664 
4665 	hr_reg_clear(qpc_mask, QPC_V2_IRRL_HEAD);
4666 
4667 	return 0;
4668 }
4669 
4670 static int get_dip_ctx_idx(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4671 			   u32 *dip_idx)
4672 {
4673 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4674 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4675 	u32 *spare_idx = hr_dev->qp_table.idx_table.spare_idx;
4676 	u32 *head =  &hr_dev->qp_table.idx_table.head;
4677 	u32 *tail =  &hr_dev->qp_table.idx_table.tail;
4678 	struct hns_roce_dip *hr_dip;
4679 	unsigned long flags;
4680 	int ret = 0;
4681 
4682 	spin_lock_irqsave(&hr_dev->dip_list_lock, flags);
4683 
4684 	spare_idx[*tail] = ibqp->qp_num;
4685 	*tail = (*tail == hr_dev->caps.num_qps - 1) ? 0 : (*tail + 1);
4686 
4687 	list_for_each_entry(hr_dip, &hr_dev->dip_list, node) {
4688 		if (!memcmp(grh->dgid.raw, hr_dip->dgid, GID_LEN_V2)) {
4689 			*dip_idx = hr_dip->dip_idx;
4690 			goto out;
4691 		}
4692 	}
4693 
4694 	/* If no dgid is found, a new dip and a mapping between dgid and
4695 	 * dip_idx will be created.
4696 	 */
4697 	hr_dip = kzalloc(sizeof(*hr_dip), GFP_ATOMIC);
4698 	if (!hr_dip) {
4699 		ret = -ENOMEM;
4700 		goto out;
4701 	}
4702 
4703 	memcpy(hr_dip->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4704 	hr_dip->dip_idx = *dip_idx = spare_idx[*head];
4705 	*head = (*head == hr_dev->caps.num_qps - 1) ? 0 : (*head + 1);
4706 	list_add_tail(&hr_dip->node, &hr_dev->dip_list);
4707 
4708 out:
4709 	spin_unlock_irqrestore(&hr_dev->dip_list_lock, flags);
4710 	return ret;
4711 }
4712 
4713 enum {
4714 	CONG_DCQCN,
4715 	CONG_WINDOW,
4716 };
4717 
4718 enum {
4719 	UNSUPPORT_CONG_LEVEL,
4720 	SUPPORT_CONG_LEVEL,
4721 };
4722 
4723 enum {
4724 	CONG_LDCP,
4725 	CONG_HC3,
4726 };
4727 
4728 enum {
4729 	DIP_INVALID,
4730 	DIP_VALID,
4731 };
4732 
4733 enum {
4734 	WND_LIMIT,
4735 	WND_UNLIMIT,
4736 };
4737 
4738 static int check_cong_type(struct ib_qp *ibqp,
4739 			   struct hns_roce_congestion_algorithm *cong_alg)
4740 {
4741 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4742 
4743 	/* different congestion types match different configurations */
4744 	switch (hr_qp->cong_type) {
4745 	case CONG_TYPE_DCQCN:
4746 		cong_alg->alg_sel = CONG_DCQCN;
4747 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4748 		cong_alg->dip_vld = DIP_INVALID;
4749 		cong_alg->wnd_mode_sel = WND_LIMIT;
4750 		break;
4751 	case CONG_TYPE_LDCP:
4752 		cong_alg->alg_sel = CONG_WINDOW;
4753 		cong_alg->alg_sub_sel = CONG_LDCP;
4754 		cong_alg->dip_vld = DIP_INVALID;
4755 		cong_alg->wnd_mode_sel = WND_UNLIMIT;
4756 		break;
4757 	case CONG_TYPE_HC3:
4758 		cong_alg->alg_sel = CONG_WINDOW;
4759 		cong_alg->alg_sub_sel = CONG_HC3;
4760 		cong_alg->dip_vld = DIP_INVALID;
4761 		cong_alg->wnd_mode_sel = WND_LIMIT;
4762 		break;
4763 	case CONG_TYPE_DIP:
4764 		cong_alg->alg_sel = CONG_DCQCN;
4765 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4766 		cong_alg->dip_vld = DIP_VALID;
4767 		cong_alg->wnd_mode_sel = WND_LIMIT;
4768 		break;
4769 	default:
4770 		hr_qp->cong_type = CONG_TYPE_DCQCN;
4771 		cong_alg->alg_sel = CONG_DCQCN;
4772 		cong_alg->alg_sub_sel = UNSUPPORT_CONG_LEVEL;
4773 		cong_alg->dip_vld = DIP_INVALID;
4774 		cong_alg->wnd_mode_sel = WND_LIMIT;
4775 		break;
4776 	}
4777 
4778 	return 0;
4779 }
4780 
4781 static int fill_cong_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,
4782 			   struct hns_roce_v2_qp_context *context,
4783 			   struct hns_roce_v2_qp_context *qpc_mask)
4784 {
4785 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4786 	struct hns_roce_congestion_algorithm cong_field;
4787 	struct ib_device *ibdev = ibqp->device;
4788 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
4789 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4790 	u32 dip_idx = 0;
4791 	int ret;
4792 
4793 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ||
4794 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE)
4795 		return 0;
4796 
4797 	ret = check_cong_type(ibqp, &cong_field);
4798 	if (ret)
4799 		return ret;
4800 
4801 	hr_reg_write(context, QPC_CONG_ALGO_TMPL_ID, hr_dev->cong_algo_tmpl_id +
4802 		     hr_qp->cong_type * HNS_ROCE_CONG_SIZE);
4803 	hr_reg_clear(qpc_mask, QPC_CONG_ALGO_TMPL_ID);
4804 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SEL, cong_field.alg_sel);
4805 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SEL);
4806 	hr_reg_write(&context->ext, QPCEX_CONG_ALG_SUB_SEL,
4807 		     cong_field.alg_sub_sel);
4808 	hr_reg_clear(&qpc_mask->ext, QPCEX_CONG_ALG_SUB_SEL);
4809 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX_VLD, cong_field.dip_vld);
4810 	hr_reg_clear(&qpc_mask->ext, QPCEX_DIP_CTX_IDX_VLD);
4811 	hr_reg_write(&context->ext, QPCEX_SQ_RQ_NOT_FORBID_EN,
4812 		     cong_field.wnd_mode_sel);
4813 	hr_reg_clear(&qpc_mask->ext, QPCEX_SQ_RQ_NOT_FORBID_EN);
4814 
4815 	/* if dip is disabled, there is no need to set dip idx */
4816 	if (cong_field.dip_vld == 0)
4817 		return 0;
4818 
4819 	ret = get_dip_ctx_idx(ibqp, attr, &dip_idx);
4820 	if (ret) {
4821 		ibdev_err(ibdev, "failed to fill cong field, ret = %d.\n", ret);
4822 		return ret;
4823 	}
4824 
4825 	hr_reg_write(&context->ext, QPCEX_DIP_CTX_IDX, dip_idx);
4826 	hr_reg_write(&qpc_mask->ext, QPCEX_DIP_CTX_IDX, 0);
4827 
4828 	return 0;
4829 }
4830 
4831 static int hns_roce_hw_v2_get_dscp(struct hns_roce_dev *hr_dev, u8 dscp,
4832 				   u8 *tc_mode, u8 *priority)
4833 {
4834 	struct hns_roce_v2_priv *priv = hr_dev->priv;
4835 	struct hnae3_handle *handle = priv->handle;
4836 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
4837 
4838 	if (!ops->get_dscp_prio)
4839 		return -EOPNOTSUPP;
4840 
4841 	return ops->get_dscp_prio(handle, dscp, tc_mode, priority);
4842 }
4843 
4844 bool check_sl_valid(struct hns_roce_dev *hr_dev, u8 sl)
4845 {
4846 	u32 max_sl;
4847 
4848 	max_sl = min_t(u32, MAX_SERVICE_LEVEL, hr_dev->caps.sl_num - 1);
4849 	if (unlikely(sl > max_sl)) {
4850 		ibdev_err_ratelimited(&hr_dev->ib_dev,
4851 				      "failed to set SL(%u). Shouldn't be larger than %u.\n",
4852 				      sl, max_sl);
4853 		return false;
4854 	}
4855 
4856 	return true;
4857 }
4858 
4859 static int hns_roce_set_sl(struct ib_qp *ibqp,
4860 			   const struct ib_qp_attr *attr,
4861 			   struct hns_roce_v2_qp_context *context,
4862 			   struct hns_roce_v2_qp_context *qpc_mask)
4863 {
4864 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4865 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4866 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4867 	struct ib_device *ibdev = &hr_dev->ib_dev;
4868 	int ret;
4869 
4870 	ret = hns_roce_hw_v2_get_dscp(hr_dev, get_tclass(&attr->ah_attr.grh),
4871 				      &hr_qp->tc_mode, &hr_qp->priority);
4872 	if (ret && ret != -EOPNOTSUPP &&
4873 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4874 		ibdev_err_ratelimited(ibdev,
4875 				      "failed to get dscp, ret = %d.\n", ret);
4876 		return ret;
4877 	}
4878 
4879 	if (hr_qp->tc_mode == HNAE3_TC_MAP_MODE_DSCP &&
4880 	    grh->sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4881 		hr_qp->sl = hr_qp->priority;
4882 	else
4883 		hr_qp->sl = rdma_ah_get_sl(&attr->ah_attr);
4884 
4885 	if (!check_sl_valid(hr_dev, hr_qp->sl))
4886 		return -EINVAL;
4887 
4888 	hr_reg_write(context, QPC_SL, hr_qp->sl);
4889 	hr_reg_clear(qpc_mask, QPC_SL);
4890 
4891 	return 0;
4892 }
4893 
4894 static int hns_roce_v2_set_path(struct ib_qp *ibqp,
4895 				const struct ib_qp_attr *attr,
4896 				int attr_mask,
4897 				struct hns_roce_v2_qp_context *context,
4898 				struct hns_roce_v2_qp_context *qpc_mask)
4899 {
4900 	const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
4901 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
4902 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
4903 	struct ib_device *ibdev = &hr_dev->ib_dev;
4904 	const struct ib_gid_attr *gid_attr = NULL;
4905 	u8 sl = rdma_ah_get_sl(&attr->ah_attr);
4906 	int is_roce_protocol;
4907 	u16 vlan_id = 0xffff;
4908 	bool is_udp = false;
4909 	u8 ib_port;
4910 	u8 hr_port;
4911 	int ret;
4912 
4913 	/*
4914 	 * If free_mr_en of qp is set, it means that this qp comes from
4915 	 * free mr. This qp will perform the loopback operation.
4916 	 * In the loopback scenario, only sl needs to be set.
4917 	 */
4918 	if (hr_qp->free_mr_en) {
4919 		if (!check_sl_valid(hr_dev, sl))
4920 			return -EINVAL;
4921 		hr_reg_write(context, QPC_SL, sl);
4922 		hr_reg_clear(qpc_mask, QPC_SL);
4923 		hr_qp->sl = sl;
4924 		return 0;
4925 	}
4926 
4927 	ib_port = (attr_mask & IB_QP_PORT) ? attr->port_num : hr_qp->port + 1;
4928 	hr_port = ib_port - 1;
4929 	is_roce_protocol = rdma_cap_eth_ah(&hr_dev->ib_dev, ib_port) &&
4930 			   rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH;
4931 
4932 	if (is_roce_protocol) {
4933 		gid_attr = attr->ah_attr.grh.sgid_attr;
4934 		ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
4935 		if (ret)
4936 			return ret;
4937 
4938 		is_udp = (gid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP);
4939 	}
4940 
4941 	/* Only HIP08 needs to set the vlan_en bits in QPC */
4942 	if (vlan_id < VLAN_N_VID &&
4943 	    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
4944 		hr_reg_enable(context, QPC_RQ_VLAN_EN);
4945 		hr_reg_clear(qpc_mask, QPC_RQ_VLAN_EN);
4946 		hr_reg_enable(context, QPC_SQ_VLAN_EN);
4947 		hr_reg_clear(qpc_mask, QPC_SQ_VLAN_EN);
4948 	}
4949 
4950 	hr_reg_write(context, QPC_VLAN_ID, vlan_id);
4951 	hr_reg_clear(qpc_mask, QPC_VLAN_ID);
4952 
4953 	if (grh->sgid_index >= hr_dev->caps.gid_table_len[hr_port]) {
4954 		ibdev_err(ibdev, "sgid_index(%u) too large. max is %d\n",
4955 			  grh->sgid_index, hr_dev->caps.gid_table_len[hr_port]);
4956 		return -EINVAL;
4957 	}
4958 
4959 	if (attr->ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE) {
4960 		ibdev_err(ibdev, "ah attr is not RDMA roce type\n");
4961 		return -EINVAL;
4962 	}
4963 
4964 	hr_reg_write(context, QPC_UDPSPN,
4965 		     is_udp ? rdma_get_udp_sport(grh->flow_label, ibqp->qp_num,
4966 						 attr->dest_qp_num) :
4967 				    0);
4968 
4969 	hr_reg_clear(qpc_mask, QPC_UDPSPN);
4970 
4971 	hr_reg_write(context, QPC_GMV_IDX, grh->sgid_index);
4972 
4973 	hr_reg_clear(qpc_mask, QPC_GMV_IDX);
4974 
4975 	hr_reg_write(context, QPC_HOPLIMIT, grh->hop_limit);
4976 	hr_reg_clear(qpc_mask, QPC_HOPLIMIT);
4977 
4978 	ret = fill_cong_field(ibqp, attr, context, qpc_mask);
4979 	if (ret)
4980 		return ret;
4981 
4982 	hr_reg_write(context, QPC_TC, get_tclass(&attr->ah_attr.grh));
4983 	hr_reg_clear(qpc_mask, QPC_TC);
4984 
4985 	hr_reg_write(context, QPC_FL, grh->flow_label);
4986 	hr_reg_clear(qpc_mask, QPC_FL);
4987 	memcpy(context->dgid, grh->dgid.raw, sizeof(grh->dgid.raw));
4988 	memset(qpc_mask->dgid, 0, sizeof(grh->dgid.raw));
4989 
4990 	return  hns_roce_set_sl(ibqp, attr, context, qpc_mask);
4991 }
4992 
4993 static bool check_qp_state(enum ib_qp_state cur_state,
4994 			   enum ib_qp_state new_state)
4995 {
4996 	static const bool sm[][IB_QPS_ERR + 1] = {
4997 		[IB_QPS_RESET] = { [IB_QPS_RESET] = true,
4998 				   [IB_QPS_INIT] = true },
4999 		[IB_QPS_INIT] = { [IB_QPS_RESET] = true,
5000 				  [IB_QPS_INIT] = true,
5001 				  [IB_QPS_RTR] = true,
5002 				  [IB_QPS_ERR] = true },
5003 		[IB_QPS_RTR] = { [IB_QPS_RESET] = true,
5004 				 [IB_QPS_RTS] = true,
5005 				 [IB_QPS_ERR] = true },
5006 		[IB_QPS_RTS] = { [IB_QPS_RESET] = true,
5007 				 [IB_QPS_RTS] = true,
5008 				 [IB_QPS_ERR] = true },
5009 		[IB_QPS_SQD] = {},
5010 		[IB_QPS_SQE] = {},
5011 		[IB_QPS_ERR] = { [IB_QPS_RESET] = true,
5012 				 [IB_QPS_ERR] = true }
5013 	};
5014 
5015 	return sm[cur_state][new_state];
5016 }
5017 
5018 static int hns_roce_v2_set_abs_fields(struct ib_qp *ibqp,
5019 				      const struct ib_qp_attr *attr,
5020 				      int attr_mask,
5021 				      enum ib_qp_state cur_state,
5022 				      enum ib_qp_state new_state,
5023 				      struct hns_roce_v2_qp_context *context,
5024 				      struct hns_roce_v2_qp_context *qpc_mask,
5025 				      struct ib_udata *udata)
5026 {
5027 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5028 	int ret = 0;
5029 
5030 	if (!check_qp_state(cur_state, new_state)) {
5031 		ibdev_err(&hr_dev->ib_dev, "Illegal state for QP!\n");
5032 		return -EINVAL;
5033 	}
5034 
5035 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
5036 		memset(qpc_mask, 0, hr_dev->caps.qpc_sz);
5037 		modify_qp_reset_to_init(ibqp, context, qpc_mask);
5038 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_INIT) {
5039 		modify_qp_init_to_init(ibqp, context, qpc_mask);
5040 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
5041 		ret = modify_qp_init_to_rtr(ibqp, attr, attr_mask, context,
5042 					    qpc_mask, udata);
5043 	} else if (cur_state == IB_QPS_RTR && new_state == IB_QPS_RTS) {
5044 		ret = modify_qp_rtr_to_rts(ibqp, attr_mask, context, qpc_mask);
5045 	}
5046 
5047 	return ret;
5048 }
5049 
5050 static bool check_qp_timeout_cfg_range(struct hns_roce_dev *hr_dev, u8 *timeout)
5051 {
5052 #define QP_ACK_TIMEOUT_MAX_HIP08 20
5053 #define QP_ACK_TIMEOUT_MAX 31
5054 
5055 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5056 		if (*timeout > QP_ACK_TIMEOUT_MAX_HIP08) {
5057 			ibdev_warn(&hr_dev->ib_dev,
5058 				   "local ACK timeout shall be 0 to 20.\n");
5059 			return false;
5060 		}
5061 		*timeout += HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5062 	} else if (hr_dev->pci_dev->revision > PCI_REVISION_ID_HIP08) {
5063 		if (*timeout > QP_ACK_TIMEOUT_MAX) {
5064 			ibdev_warn(&hr_dev->ib_dev,
5065 				   "local ACK timeout shall be 0 to 31.\n");
5066 			return false;
5067 		}
5068 	}
5069 
5070 	return true;
5071 }
5072 
5073 static int hns_roce_v2_set_opt_fields(struct ib_qp *ibqp,
5074 				      const struct ib_qp_attr *attr,
5075 				      int attr_mask,
5076 				      struct hns_roce_v2_qp_context *context,
5077 				      struct hns_roce_v2_qp_context *qpc_mask)
5078 {
5079 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5080 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5081 	int ret = 0;
5082 	u8 timeout;
5083 
5084 	if (attr_mask & IB_QP_AV) {
5085 		ret = hns_roce_v2_set_path(ibqp, attr, attr_mask, context,
5086 					   qpc_mask);
5087 		if (ret)
5088 			return ret;
5089 	}
5090 
5091 	if (attr_mask & IB_QP_TIMEOUT) {
5092 		timeout = attr->timeout;
5093 		if (check_qp_timeout_cfg_range(hr_dev, &timeout)) {
5094 			hr_reg_write(context, QPC_AT, timeout);
5095 			hr_reg_clear(qpc_mask, QPC_AT);
5096 		}
5097 	}
5098 
5099 	if (attr_mask & IB_QP_RETRY_CNT) {
5100 		hr_reg_write(context, QPC_RETRY_NUM_INIT, attr->retry_cnt);
5101 		hr_reg_clear(qpc_mask, QPC_RETRY_NUM_INIT);
5102 
5103 		hr_reg_write(context, QPC_RETRY_CNT, attr->retry_cnt);
5104 		hr_reg_clear(qpc_mask, QPC_RETRY_CNT);
5105 	}
5106 
5107 	if (attr_mask & IB_QP_RNR_RETRY) {
5108 		hr_reg_write(context, QPC_RNR_NUM_INIT, attr->rnr_retry);
5109 		hr_reg_clear(qpc_mask, QPC_RNR_NUM_INIT);
5110 
5111 		hr_reg_write(context, QPC_RNR_CNT, attr->rnr_retry);
5112 		hr_reg_clear(qpc_mask, QPC_RNR_CNT);
5113 	}
5114 
5115 	if (attr_mask & IB_QP_SQ_PSN) {
5116 		hr_reg_write(context, QPC_SQ_CUR_PSN, attr->sq_psn);
5117 		hr_reg_clear(qpc_mask, QPC_SQ_CUR_PSN);
5118 
5119 		hr_reg_write(context, QPC_SQ_MAX_PSN, attr->sq_psn);
5120 		hr_reg_clear(qpc_mask, QPC_SQ_MAX_PSN);
5121 
5122 		hr_reg_write(context, QPC_RETRY_MSG_PSN_L, attr->sq_psn);
5123 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_L);
5124 
5125 		hr_reg_write(context, QPC_RETRY_MSG_PSN_H,
5126 			     attr->sq_psn >> RETRY_MSG_PSN_SHIFT);
5127 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_PSN_H);
5128 
5129 		hr_reg_write(context, QPC_RETRY_MSG_FPKT_PSN, attr->sq_psn);
5130 		hr_reg_clear(qpc_mask, QPC_RETRY_MSG_FPKT_PSN);
5131 
5132 		hr_reg_write(context, QPC_RX_ACK_EPSN, attr->sq_psn);
5133 		hr_reg_clear(qpc_mask, QPC_RX_ACK_EPSN);
5134 	}
5135 
5136 	if ((attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) &&
5137 	     attr->max_dest_rd_atomic) {
5138 		hr_reg_write(context, QPC_RR_MAX,
5139 			     fls(attr->max_dest_rd_atomic - 1));
5140 		hr_reg_clear(qpc_mask, QPC_RR_MAX);
5141 	}
5142 
5143 	if ((attr_mask & IB_QP_MAX_QP_RD_ATOMIC) && attr->max_rd_atomic) {
5144 		hr_reg_write(context, QPC_SR_MAX, fls(attr->max_rd_atomic - 1));
5145 		hr_reg_clear(qpc_mask, QPC_SR_MAX);
5146 	}
5147 
5148 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
5149 		set_access_flags(hr_qp, context, qpc_mask, attr, attr_mask);
5150 
5151 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
5152 		hr_reg_write(context, QPC_MIN_RNR_TIME,
5153 			    hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08 ?
5154 			    HNS_ROCE_RNR_TIMER_10NS : attr->min_rnr_timer);
5155 		hr_reg_clear(qpc_mask, QPC_MIN_RNR_TIME);
5156 	}
5157 
5158 	if (attr_mask & IB_QP_RQ_PSN) {
5159 		hr_reg_write(context, QPC_RX_REQ_EPSN, attr->rq_psn);
5160 		hr_reg_clear(qpc_mask, QPC_RX_REQ_EPSN);
5161 
5162 		hr_reg_write(context, QPC_RAQ_PSN, attr->rq_psn - 1);
5163 		hr_reg_clear(qpc_mask, QPC_RAQ_PSN);
5164 	}
5165 
5166 	if (attr_mask & IB_QP_QKEY) {
5167 		context->qkey_xrcd = cpu_to_le32(attr->qkey);
5168 		qpc_mask->qkey_xrcd = 0;
5169 		hr_qp->qkey = attr->qkey;
5170 	}
5171 
5172 	return ret;
5173 }
5174 
5175 static void hns_roce_v2_record_opt_fields(struct ib_qp *ibqp,
5176 					  const struct ib_qp_attr *attr,
5177 					  int attr_mask)
5178 {
5179 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5180 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5181 
5182 	if (attr_mask & IB_QP_ACCESS_FLAGS)
5183 		hr_qp->atomic_rd_en = attr->qp_access_flags;
5184 
5185 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
5186 		hr_qp->resp_depth = attr->max_dest_rd_atomic;
5187 	if (attr_mask & IB_QP_PORT) {
5188 		hr_qp->port = attr->port_num - 1;
5189 		hr_qp->phy_port = hr_dev->iboe.phy_port[hr_qp->port];
5190 	}
5191 }
5192 
5193 static void clear_qp(struct hns_roce_qp *hr_qp)
5194 {
5195 	struct ib_qp *ibqp = &hr_qp->ibqp;
5196 
5197 	if (ibqp->send_cq)
5198 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->send_cq),
5199 				     hr_qp->qpn, NULL);
5200 
5201 	if (ibqp->recv_cq  && ibqp->recv_cq != ibqp->send_cq)
5202 		hns_roce_v2_cq_clean(to_hr_cq(ibqp->recv_cq),
5203 				     hr_qp->qpn, ibqp->srq ?
5204 				     to_hr_srq(ibqp->srq) : NULL);
5205 
5206 	if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB)
5207 		*hr_qp->rdb.db_record = 0;
5208 
5209 	hr_qp->rq.head = 0;
5210 	hr_qp->rq.tail = 0;
5211 	hr_qp->sq.head = 0;
5212 	hr_qp->sq.tail = 0;
5213 	hr_qp->next_sge = 0;
5214 }
5215 
5216 static void v2_set_flushed_fields(struct ib_qp *ibqp,
5217 				  struct hns_roce_v2_qp_context *context,
5218 				  struct hns_roce_v2_qp_context *qpc_mask)
5219 {
5220 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5221 	unsigned long sq_flag = 0;
5222 	unsigned long rq_flag = 0;
5223 
5224 	if (ibqp->qp_type == IB_QPT_XRC_TGT)
5225 		return;
5226 
5227 	spin_lock_irqsave(&hr_qp->sq.lock, sq_flag);
5228 	hr_reg_write(context, QPC_SQ_PRODUCER_IDX, hr_qp->sq.head);
5229 	hr_reg_clear(qpc_mask, QPC_SQ_PRODUCER_IDX);
5230 	hr_qp->state = IB_QPS_ERR;
5231 	spin_unlock_irqrestore(&hr_qp->sq.lock, sq_flag);
5232 
5233 	if (ibqp->srq || ibqp->qp_type == IB_QPT_XRC_INI) /* no RQ */
5234 		return;
5235 
5236 	spin_lock_irqsave(&hr_qp->rq.lock, rq_flag);
5237 	hr_reg_write(context, QPC_RQ_PRODUCER_IDX, hr_qp->rq.head);
5238 	hr_reg_clear(qpc_mask, QPC_RQ_PRODUCER_IDX);
5239 	spin_unlock_irqrestore(&hr_qp->rq.lock, rq_flag);
5240 }
5241 
5242 static int hns_roce_v2_modify_qp(struct ib_qp *ibqp,
5243 				 const struct ib_qp_attr *attr,
5244 				 int attr_mask, enum ib_qp_state cur_state,
5245 				 enum ib_qp_state new_state, struct ib_udata *udata)
5246 {
5247 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5248 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5249 	struct hns_roce_v2_qp_context ctx[2];
5250 	struct hns_roce_v2_qp_context *context = ctx;
5251 	struct hns_roce_v2_qp_context *qpc_mask = ctx + 1;
5252 	struct ib_device *ibdev = &hr_dev->ib_dev;
5253 	int ret;
5254 
5255 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
5256 		return -EOPNOTSUPP;
5257 
5258 	/*
5259 	 * In v2 engine, software pass context and context mask to hardware
5260 	 * when modifying qp. If software need modify some fields in context,
5261 	 * we should set all bits of the relevant fields in context mask to
5262 	 * 0 at the same time, else set them to 0x1.
5263 	 */
5264 	memset(context, 0, hr_dev->caps.qpc_sz);
5265 	memset(qpc_mask, 0xff, hr_dev->caps.qpc_sz);
5266 
5267 	ret = hns_roce_v2_set_abs_fields(ibqp, attr, attr_mask, cur_state,
5268 					 new_state, context, qpc_mask, udata);
5269 	if (ret)
5270 		goto out;
5271 
5272 	/* When QP state is err, SQ and RQ WQE should be flushed */
5273 	if (new_state == IB_QPS_ERR)
5274 		v2_set_flushed_fields(ibqp, context, qpc_mask);
5275 
5276 	/* Configure the optional fields */
5277 	ret = hns_roce_v2_set_opt_fields(ibqp, attr, attr_mask, context,
5278 					 qpc_mask);
5279 	if (ret)
5280 		goto out;
5281 
5282 	hr_reg_write_bool(context, QPC_INV_CREDIT,
5283 			  to_hr_qp_type(hr_qp->ibqp.qp_type) == SERV_TYPE_XRC ||
5284 			  ibqp->srq);
5285 	hr_reg_clear(qpc_mask, QPC_INV_CREDIT);
5286 
5287 	/* Every status migrate must change state */
5288 	hr_reg_write(context, QPC_QP_ST, new_state);
5289 	hr_reg_clear(qpc_mask, QPC_QP_ST);
5290 
5291 	/* SW pass context to HW */
5292 	ret = hns_roce_v2_qp_modify(hr_dev, context, qpc_mask, hr_qp);
5293 	if (ret) {
5294 		ibdev_err(ibdev, "failed to modify QP, ret = %d.\n", ret);
5295 		goto out;
5296 	}
5297 
5298 	hr_qp->state = new_state;
5299 
5300 	hns_roce_v2_record_opt_fields(ibqp, attr, attr_mask);
5301 
5302 	if (new_state == IB_QPS_RESET && !ibqp->uobject)
5303 		clear_qp(hr_qp);
5304 
5305 out:
5306 	return ret;
5307 }
5308 
5309 static int to_ib_qp_st(enum hns_roce_v2_qp_state state)
5310 {
5311 	static const enum ib_qp_state map[] = {
5312 		[HNS_ROCE_QP_ST_RST] = IB_QPS_RESET,
5313 		[HNS_ROCE_QP_ST_INIT] = IB_QPS_INIT,
5314 		[HNS_ROCE_QP_ST_RTR] = IB_QPS_RTR,
5315 		[HNS_ROCE_QP_ST_RTS] = IB_QPS_RTS,
5316 		[HNS_ROCE_QP_ST_SQD] = IB_QPS_SQD,
5317 		[HNS_ROCE_QP_ST_SQER] = IB_QPS_SQE,
5318 		[HNS_ROCE_QP_ST_ERR] = IB_QPS_ERR,
5319 		[HNS_ROCE_QP_ST_SQ_DRAINING] = IB_QPS_SQD
5320 	};
5321 
5322 	return (state < ARRAY_SIZE(map)) ? map[state] : -1;
5323 }
5324 
5325 static int hns_roce_v2_query_qpc(struct hns_roce_dev *hr_dev, u32 qpn,
5326 				 void *buffer)
5327 {
5328 	struct hns_roce_cmd_mailbox *mailbox;
5329 	int ret;
5330 
5331 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5332 	if (IS_ERR(mailbox))
5333 		return PTR_ERR(mailbox);
5334 
5335 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_QPC,
5336 				qpn);
5337 	if (ret)
5338 		goto out;
5339 
5340 	memcpy(buffer, mailbox->buf, hr_dev->caps.qpc_sz);
5341 
5342 out:
5343 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5344 	return ret;
5345 }
5346 
5347 static int hns_roce_v2_query_srqc(struct hns_roce_dev *hr_dev, u32 srqn,
5348 				 void *buffer)
5349 {
5350 	struct hns_roce_srq_context *context;
5351 	struct hns_roce_cmd_mailbox *mailbox;
5352 	int ret;
5353 
5354 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5355 	if (IS_ERR(mailbox))
5356 		return PTR_ERR(mailbox);
5357 
5358 	context = mailbox->buf;
5359 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SRQC,
5360 				srqn);
5361 	if (ret)
5362 		goto out;
5363 
5364 	memcpy(buffer, context, sizeof(*context));
5365 
5366 out:
5367 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5368 	return ret;
5369 }
5370 
5371 static int hns_roce_v2_query_sccc(struct hns_roce_dev *hr_dev, u32 qpn,
5372 				  void *buffer)
5373 {
5374 	struct hns_roce_v2_scc_context *context;
5375 	struct hns_roce_cmd_mailbox *mailbox;
5376 	int ret;
5377 
5378 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5379 	if (IS_ERR(mailbox))
5380 		return PTR_ERR(mailbox);
5381 
5382 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_SCCC,
5383 				qpn);
5384 	if (ret)
5385 		goto out;
5386 
5387 	context = mailbox->buf;
5388 	memcpy(buffer, context, sizeof(*context));
5389 
5390 out:
5391 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5392 	return ret;
5393 }
5394 
5395 static u8 get_qp_timeout_attr(struct hns_roce_dev *hr_dev,
5396 			      struct hns_roce_v2_qp_context *context)
5397 {
5398 	u8 timeout;
5399 
5400 	timeout = (u8)hr_reg_read(context, QPC_AT);
5401 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
5402 		timeout -= HNS_ROCE_V2_QP_ACK_TIMEOUT_OFS_HIP08;
5403 
5404 	return timeout;
5405 }
5406 
5407 static int hns_roce_v2_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
5408 				int qp_attr_mask,
5409 				struct ib_qp_init_attr *qp_init_attr)
5410 {
5411 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5412 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5413 	struct hns_roce_v2_qp_context context = {};
5414 	struct ib_device *ibdev = &hr_dev->ib_dev;
5415 	int tmp_qp_state;
5416 	int state;
5417 	int ret;
5418 
5419 	memset(qp_attr, 0, sizeof(*qp_attr));
5420 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
5421 
5422 	mutex_lock(&hr_qp->mutex);
5423 
5424 	if (hr_qp->state == IB_QPS_RESET) {
5425 		qp_attr->qp_state = IB_QPS_RESET;
5426 		ret = 0;
5427 		goto done;
5428 	}
5429 
5430 	ret = hns_roce_v2_query_qpc(hr_dev, hr_qp->qpn, &context);
5431 	if (ret) {
5432 		ibdev_err(ibdev, "failed to query QPC, ret = %d.\n", ret);
5433 		ret = -EINVAL;
5434 		goto out;
5435 	}
5436 
5437 	state = hr_reg_read(&context, QPC_QP_ST);
5438 	tmp_qp_state = to_ib_qp_st((enum hns_roce_v2_qp_state)state);
5439 	if (tmp_qp_state == -1) {
5440 		ibdev_err(ibdev, "Illegal ib_qp_state\n");
5441 		ret = -EINVAL;
5442 		goto out;
5443 	}
5444 	hr_qp->state = (u8)tmp_qp_state;
5445 	qp_attr->qp_state = (enum ib_qp_state)hr_qp->state;
5446 	qp_attr->path_mtu = (enum ib_mtu)hr_reg_read(&context, QPC_MTU);
5447 	qp_attr->path_mig_state = IB_MIG_ARMED;
5448 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
5449 	if (hr_qp->ibqp.qp_type == IB_QPT_UD)
5450 		qp_attr->qkey = le32_to_cpu(context.qkey_xrcd);
5451 
5452 	qp_attr->rq_psn = hr_reg_read(&context, QPC_RX_REQ_EPSN);
5453 	qp_attr->sq_psn = (u32)hr_reg_read(&context, QPC_SQ_CUR_PSN);
5454 	qp_attr->dest_qp_num = hr_reg_read(&context, QPC_DQPN);
5455 	qp_attr->qp_access_flags =
5456 		((hr_reg_read(&context, QPC_RRE)) << V2_QP_RRE_S) |
5457 		((hr_reg_read(&context, QPC_RWE)) << V2_QP_RWE_S) |
5458 		((hr_reg_read(&context, QPC_ATE)) << V2_QP_ATE_S);
5459 
5460 	if (hr_qp->ibqp.qp_type == IB_QPT_RC ||
5461 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5462 	    hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) {
5463 		struct ib_global_route *grh =
5464 			rdma_ah_retrieve_grh(&qp_attr->ah_attr);
5465 
5466 		rdma_ah_set_sl(&qp_attr->ah_attr,
5467 			       hr_reg_read(&context, QPC_SL));
5468 		rdma_ah_set_port_num(&qp_attr->ah_attr, hr_qp->port + 1);
5469 		rdma_ah_set_ah_flags(&qp_attr->ah_attr, IB_AH_GRH);
5470 		grh->flow_label = hr_reg_read(&context, QPC_FL);
5471 		grh->sgid_index = hr_reg_read(&context, QPC_GMV_IDX);
5472 		grh->hop_limit = hr_reg_read(&context, QPC_HOPLIMIT);
5473 		grh->traffic_class = hr_reg_read(&context, QPC_TC);
5474 
5475 		memcpy(grh->dgid.raw, context.dgid, sizeof(grh->dgid.raw));
5476 	}
5477 
5478 	qp_attr->port_num = hr_qp->port + 1;
5479 	qp_attr->sq_draining = 0;
5480 	qp_attr->max_rd_atomic = 1 << hr_reg_read(&context, QPC_SR_MAX);
5481 	qp_attr->max_dest_rd_atomic = 1 << hr_reg_read(&context, QPC_RR_MAX);
5482 
5483 	qp_attr->min_rnr_timer = (u8)hr_reg_read(&context, QPC_MIN_RNR_TIME);
5484 	qp_attr->timeout = get_qp_timeout_attr(hr_dev, &context);
5485 	qp_attr->retry_cnt = hr_reg_read(&context, QPC_RETRY_NUM_INIT);
5486 	qp_attr->rnr_retry = hr_reg_read(&context, QPC_RNR_NUM_INIT);
5487 
5488 done:
5489 	qp_attr->cur_qp_state = qp_attr->qp_state;
5490 	qp_attr->cap.max_recv_wr = hr_qp->rq.wqe_cnt;
5491 	qp_attr->cap.max_recv_sge = hr_qp->rq.max_gs - hr_qp->rq.rsv_sge;
5492 	qp_attr->cap.max_inline_data = hr_qp->max_inline_data;
5493 
5494 	qp_attr->cap.max_send_wr = hr_qp->sq.wqe_cnt;
5495 	qp_attr->cap.max_send_sge = hr_qp->sq.max_gs;
5496 
5497 	qp_init_attr->qp_context = ibqp->qp_context;
5498 	qp_init_attr->qp_type = ibqp->qp_type;
5499 	qp_init_attr->recv_cq = ibqp->recv_cq;
5500 	qp_init_attr->send_cq = ibqp->send_cq;
5501 	qp_init_attr->srq = ibqp->srq;
5502 	qp_init_attr->cap = qp_attr->cap;
5503 	qp_init_attr->sq_sig_type = hr_qp->sq_signal_bits;
5504 
5505 out:
5506 	mutex_unlock(&hr_qp->mutex);
5507 	return ret;
5508 }
5509 
5510 static inline int modify_qp_is_ok(struct hns_roce_qp *hr_qp)
5511 {
5512 	return ((hr_qp->ibqp.qp_type == IB_QPT_RC ||
5513 		 hr_qp->ibqp.qp_type == IB_QPT_UD ||
5514 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_INI ||
5515 		 hr_qp->ibqp.qp_type == IB_QPT_XRC_TGT) &&
5516 		hr_qp->state != IB_QPS_RESET);
5517 }
5518 
5519 static int hns_roce_v2_destroy_qp_common(struct hns_roce_dev *hr_dev,
5520 					 struct hns_roce_qp *hr_qp,
5521 					 struct ib_udata *udata)
5522 {
5523 	struct ib_device *ibdev = &hr_dev->ib_dev;
5524 	struct hns_roce_cq *send_cq, *recv_cq;
5525 	unsigned long flags;
5526 	int ret = 0;
5527 
5528 	if (modify_qp_is_ok(hr_qp)) {
5529 		/* Modify qp to reset before destroying qp */
5530 		ret = hns_roce_v2_modify_qp(&hr_qp->ibqp, NULL, 0,
5531 					    hr_qp->state, IB_QPS_RESET, udata);
5532 		if (ret)
5533 			ibdev_err(ibdev,
5534 				  "failed to modify QP to RST, ret = %d.\n",
5535 				  ret);
5536 	}
5537 
5538 	send_cq = hr_qp->ibqp.send_cq ? to_hr_cq(hr_qp->ibqp.send_cq) : NULL;
5539 	recv_cq = hr_qp->ibqp.recv_cq ? to_hr_cq(hr_qp->ibqp.recv_cq) : NULL;
5540 
5541 	spin_lock_irqsave(&hr_dev->qp_list_lock, flags);
5542 	hns_roce_lock_cqs(send_cq, recv_cq);
5543 
5544 	if (!udata) {
5545 		if (recv_cq)
5546 			__hns_roce_v2_cq_clean(recv_cq, hr_qp->qpn,
5547 					       (hr_qp->ibqp.srq ?
5548 						to_hr_srq(hr_qp->ibqp.srq) :
5549 						NULL));
5550 
5551 		if (send_cq && send_cq != recv_cq)
5552 			__hns_roce_v2_cq_clean(send_cq, hr_qp->qpn, NULL);
5553 	}
5554 
5555 	hns_roce_qp_remove(hr_dev, hr_qp);
5556 
5557 	hns_roce_unlock_cqs(send_cq, recv_cq);
5558 	spin_unlock_irqrestore(&hr_dev->qp_list_lock, flags);
5559 
5560 	return ret;
5561 }
5562 
5563 int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
5564 {
5565 	struct hns_roce_dev *hr_dev = to_hr_dev(ibqp->device);
5566 	struct hns_roce_qp *hr_qp = to_hr_qp(ibqp);
5567 	int ret;
5568 
5569 	ret = hns_roce_v2_destroy_qp_common(hr_dev, hr_qp, udata);
5570 	if (ret)
5571 		ibdev_err(&hr_dev->ib_dev,
5572 			  "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n",
5573 			  hr_qp->qpn, ret);
5574 
5575 	hns_roce_qp_destroy(hr_dev, hr_qp, udata);
5576 
5577 	return 0;
5578 }
5579 
5580 static int hns_roce_v2_qp_flow_control_init(struct hns_roce_dev *hr_dev,
5581 					    struct hns_roce_qp *hr_qp)
5582 {
5583 	struct ib_device *ibdev = &hr_dev->ib_dev;
5584 	struct hns_roce_sccc_clr_done *resp;
5585 	struct hns_roce_sccc_clr *clr;
5586 	struct hns_roce_cmq_desc desc;
5587 	int ret, i;
5588 
5589 	if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
5590 		return 0;
5591 
5592 	mutex_lock(&hr_dev->qp_table.scc_mutex);
5593 
5594 	/* set scc ctx clear done flag */
5595 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_RESET_SCCC, false);
5596 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5597 	if (ret) {
5598 		ibdev_err(ibdev, "failed to reset SCC ctx, ret = %d.\n", ret);
5599 		goto out;
5600 	}
5601 
5602 	/* clear scc context */
5603 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CLR_SCCC, false);
5604 	clr = (struct hns_roce_sccc_clr *)desc.data;
5605 	clr->qpn = cpu_to_le32(hr_qp->qpn);
5606 	ret =  hns_roce_cmq_send(hr_dev, &desc, 1);
5607 	if (ret) {
5608 		ibdev_err(ibdev, "failed to clear SCC ctx, ret = %d.\n", ret);
5609 		goto out;
5610 	}
5611 
5612 	/* query scc context clear is done or not */
5613 	resp = (struct hns_roce_sccc_clr_done *)desc.data;
5614 	for (i = 0; i <= HNS_ROCE_CMQ_SCC_CLR_DONE_CNT; i++) {
5615 		hns_roce_cmq_setup_basic_desc(&desc,
5616 					      HNS_ROCE_OPC_QUERY_SCCC, true);
5617 		ret = hns_roce_cmq_send(hr_dev, &desc, 1);
5618 		if (ret) {
5619 			ibdev_err(ibdev, "failed to query clr cmq, ret = %d\n",
5620 				  ret);
5621 			goto out;
5622 		}
5623 
5624 		if (resp->clr_done)
5625 			goto out;
5626 
5627 		msleep(20);
5628 	}
5629 
5630 	ibdev_err(ibdev, "query SCC clr done flag overtime.\n");
5631 	ret = -ETIMEDOUT;
5632 
5633 out:
5634 	mutex_unlock(&hr_dev->qp_table.scc_mutex);
5635 	return ret;
5636 }
5637 
5638 #define DMA_IDX_SHIFT 3
5639 #define DMA_WQE_SHIFT 3
5640 
5641 static int hns_roce_v2_write_srqc_index_queue(struct hns_roce_srq *srq,
5642 					      struct hns_roce_srq_context *ctx)
5643 {
5644 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
5645 	struct ib_device *ibdev = srq->ibsrq.device;
5646 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5647 	u64 mtts_idx[MTT_MIN_COUNT] = {};
5648 	dma_addr_t dma_handle_idx;
5649 	int ret;
5650 
5651 	/* Get physical address of idx que buf */
5652 	ret = hns_roce_mtr_find(hr_dev, &idx_que->mtr, 0, mtts_idx,
5653 				ARRAY_SIZE(mtts_idx));
5654 	if (ret) {
5655 		ibdev_err(ibdev, "failed to find mtr for SRQ idx, ret = %d.\n",
5656 			  ret);
5657 		return ret;
5658 	}
5659 
5660 	dma_handle_idx = hns_roce_get_mtr_ba(&idx_que->mtr);
5661 
5662 	hr_reg_write(ctx, SRQC_IDX_HOP_NUM,
5663 		     to_hr_hem_hopnum(hr_dev->caps.idx_hop_num, srq->wqe_cnt));
5664 
5665 	hr_reg_write(ctx, SRQC_IDX_BT_BA_L, dma_handle_idx >> DMA_IDX_SHIFT);
5666 	hr_reg_write(ctx, SRQC_IDX_BT_BA_H,
5667 		     upper_32_bits(dma_handle_idx >> DMA_IDX_SHIFT));
5668 
5669 	hr_reg_write(ctx, SRQC_IDX_BA_PG_SZ,
5670 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.ba_pg_shift));
5671 	hr_reg_write(ctx, SRQC_IDX_BUF_PG_SZ,
5672 		     to_hr_hw_page_shift(idx_que->mtr.hem_cfg.buf_pg_shift));
5673 
5674 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_L,
5675 		     to_hr_hw_page_addr(mtts_idx[0]));
5676 	hr_reg_write(ctx, SRQC_IDX_CUR_BLK_ADDR_H,
5677 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[0])));
5678 
5679 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_L,
5680 		     to_hr_hw_page_addr(mtts_idx[1]));
5681 	hr_reg_write(ctx, SRQC_IDX_NXT_BLK_ADDR_H,
5682 		     upper_32_bits(to_hr_hw_page_addr(mtts_idx[1])));
5683 
5684 	return 0;
5685 }
5686 
5687 static int hns_roce_v2_write_srqc(struct hns_roce_srq *srq, void *mb_buf)
5688 {
5689 	struct ib_device *ibdev = srq->ibsrq.device;
5690 	struct hns_roce_dev *hr_dev = to_hr_dev(ibdev);
5691 	struct hns_roce_srq_context *ctx = mb_buf;
5692 	u64 mtts_wqe[MTT_MIN_COUNT] = {};
5693 	dma_addr_t dma_handle_wqe;
5694 	int ret;
5695 
5696 	memset(ctx, 0, sizeof(*ctx));
5697 
5698 	/* Get the physical address of srq buf */
5699 	ret = hns_roce_mtr_find(hr_dev, &srq->buf_mtr, 0, mtts_wqe,
5700 				ARRAY_SIZE(mtts_wqe));
5701 	if (ret) {
5702 		ibdev_err(ibdev, "failed to find mtr for SRQ WQE, ret = %d.\n",
5703 			  ret);
5704 		return ret;
5705 	}
5706 
5707 	dma_handle_wqe = hns_roce_get_mtr_ba(&srq->buf_mtr);
5708 
5709 	hr_reg_write(ctx, SRQC_SRQ_ST, 1);
5710 	hr_reg_write_bool(ctx, SRQC_SRQ_TYPE,
5711 			  srq->ibsrq.srq_type == IB_SRQT_XRC);
5712 	hr_reg_write(ctx, SRQC_PD, to_hr_pd(srq->ibsrq.pd)->pdn);
5713 	hr_reg_write(ctx, SRQC_SRQN, srq->srqn);
5714 	hr_reg_write(ctx, SRQC_XRCD, srq->xrcdn);
5715 	hr_reg_write(ctx, SRQC_XRC_CQN, srq->cqn);
5716 	hr_reg_write(ctx, SRQC_SHIFT, ilog2(srq->wqe_cnt));
5717 	hr_reg_write(ctx, SRQC_RQWS,
5718 		     srq->max_gs <= 0 ? 0 : fls(srq->max_gs - 1));
5719 
5720 	hr_reg_write(ctx, SRQC_WQE_HOP_NUM,
5721 		     to_hr_hem_hopnum(hr_dev->caps.srqwqe_hop_num,
5722 				      srq->wqe_cnt));
5723 
5724 	hr_reg_write(ctx, SRQC_WQE_BT_BA_L, dma_handle_wqe >> DMA_WQE_SHIFT);
5725 	hr_reg_write(ctx, SRQC_WQE_BT_BA_H,
5726 		     upper_32_bits(dma_handle_wqe >> DMA_WQE_SHIFT));
5727 
5728 	hr_reg_write(ctx, SRQC_WQE_BA_PG_SZ,
5729 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.ba_pg_shift));
5730 	hr_reg_write(ctx, SRQC_WQE_BUF_PG_SZ,
5731 		     to_hr_hw_page_shift(srq->buf_mtr.hem_cfg.buf_pg_shift));
5732 
5733 	if (srq->cap_flags & HNS_ROCE_SRQ_CAP_RECORD_DB) {
5734 		hr_reg_enable(ctx, SRQC_DB_RECORD_EN);
5735 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_L,
5736 			     lower_32_bits(srq->rdb.dma) >> 1);
5737 		hr_reg_write(ctx, SRQC_DB_RECORD_ADDR_H,
5738 			     upper_32_bits(srq->rdb.dma));
5739 	}
5740 
5741 	return hns_roce_v2_write_srqc_index_queue(srq, ctx);
5742 }
5743 
5744 static int hns_roce_v2_modify_srq(struct ib_srq *ibsrq,
5745 				  struct ib_srq_attr *srq_attr,
5746 				  enum ib_srq_attr_mask srq_attr_mask,
5747 				  struct ib_udata *udata)
5748 {
5749 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5750 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5751 	struct hns_roce_srq_context *srq_context;
5752 	struct hns_roce_srq_context *srqc_mask;
5753 	struct hns_roce_cmd_mailbox *mailbox;
5754 	int ret = 0;
5755 
5756 	/* Resizing SRQs is not supported yet */
5757 	if (srq_attr_mask & IB_SRQ_MAX_WR) {
5758 		ret = -EOPNOTSUPP;
5759 		goto out;
5760 	}
5761 
5762 	if (srq_attr_mask & IB_SRQ_LIMIT) {
5763 		if (srq_attr->srq_limit > srq->wqe_cnt) {
5764 			ret = -EINVAL;
5765 			goto out;
5766 		}
5767 
5768 		mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5769 		if (IS_ERR(mailbox)) {
5770 			ret = PTR_ERR(mailbox);
5771 			goto out;
5772 		}
5773 
5774 		srq_context = mailbox->buf;
5775 		srqc_mask = (struct hns_roce_srq_context *)mailbox->buf + 1;
5776 
5777 		memset(srqc_mask, 0xff, sizeof(*srqc_mask));
5778 
5779 		hr_reg_write(srq_context, SRQC_LIMIT_WL, srq_attr->srq_limit);
5780 		hr_reg_clear(srqc_mask, SRQC_LIMIT_WL);
5781 
5782 		ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5783 					HNS_ROCE_CMD_MODIFY_SRQC, srq->srqn);
5784 		hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5785 		if (ret)
5786 			ibdev_err(&hr_dev->ib_dev,
5787 				  "failed to handle cmd of modifying SRQ, ret = %d.\n",
5788 				  ret);
5789 	}
5790 
5791 out:
5792 	if (ret)
5793 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_SRQ_MODIFY_ERR_CNT]);
5794 
5795 	return ret;
5796 }
5797 
5798 static int hns_roce_v2_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
5799 {
5800 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
5801 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
5802 	struct hns_roce_srq_context *srq_context;
5803 	struct hns_roce_cmd_mailbox *mailbox;
5804 	int ret;
5805 
5806 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5807 	if (IS_ERR(mailbox))
5808 		return PTR_ERR(mailbox);
5809 
5810 	srq_context = mailbox->buf;
5811 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5812 				HNS_ROCE_CMD_QUERY_SRQC, srq->srqn);
5813 	if (ret) {
5814 		ibdev_err(&hr_dev->ib_dev,
5815 			  "failed to process cmd of querying SRQ, ret = %d.\n",
5816 			  ret);
5817 		goto out;
5818 	}
5819 
5820 	attr->srq_limit = hr_reg_read(srq_context, SRQC_LIMIT_WL);
5821 	attr->max_wr = srq->wqe_cnt;
5822 	attr->max_sge = srq->max_gs - srq->rsv_sge;
5823 
5824 out:
5825 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5826 	return ret;
5827 }
5828 
5829 static int hns_roce_v2_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
5830 {
5831 	struct hns_roce_dev *hr_dev = to_hr_dev(cq->device);
5832 	struct hns_roce_v2_cq_context *cq_context;
5833 	struct hns_roce_cq *hr_cq = to_hr_cq(cq);
5834 	struct hns_roce_v2_cq_context *cqc_mask;
5835 	struct hns_roce_cmd_mailbox *mailbox;
5836 	int ret;
5837 
5838 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5839 	ret = PTR_ERR_OR_ZERO(mailbox);
5840 	if (ret)
5841 		goto err_out;
5842 
5843 	cq_context = mailbox->buf;
5844 	cqc_mask = (struct hns_roce_v2_cq_context *)mailbox->buf + 1;
5845 
5846 	memset(cqc_mask, 0xff, sizeof(*cqc_mask));
5847 
5848 	hr_reg_write(cq_context, CQC_CQ_MAX_CNT, cq_count);
5849 	hr_reg_clear(cqc_mask, CQC_CQ_MAX_CNT);
5850 
5851 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
5852 		if (cq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
5853 			dev_info(hr_dev->dev,
5854 				 "cq_period(%u) reached the upper limit, adjusted to 65.\n",
5855 				 cq_period);
5856 			cq_period = HNS_ROCE_MAX_CQ_PERIOD_HIP08;
5857 		}
5858 		cq_period *= HNS_ROCE_CLOCK_ADJUST;
5859 	}
5860 	hr_reg_write(cq_context, CQC_CQ_PERIOD, cq_period);
5861 	hr_reg_clear(cqc_mask, CQC_CQ_PERIOD);
5862 
5863 	ret = hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0,
5864 				HNS_ROCE_CMD_MODIFY_CQC, hr_cq->cqn);
5865 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5866 	if (ret)
5867 		ibdev_err(&hr_dev->ib_dev,
5868 			  "failed to process cmd when modifying CQ, ret = %d.\n",
5869 			  ret);
5870 
5871 err_out:
5872 	if (ret)
5873 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CQ_MODIFY_ERR_CNT]);
5874 
5875 	return ret;
5876 }
5877 
5878 static int hns_roce_v2_query_cqc(struct hns_roce_dev *hr_dev, u32 cqn,
5879 				 void *buffer)
5880 {
5881 	struct hns_roce_v2_cq_context *context;
5882 	struct hns_roce_cmd_mailbox *mailbox;
5883 	int ret;
5884 
5885 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5886 	if (IS_ERR(mailbox))
5887 		return PTR_ERR(mailbox);
5888 
5889 	context = mailbox->buf;
5890 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma,
5891 				HNS_ROCE_CMD_QUERY_CQC, cqn);
5892 	if (ret) {
5893 		ibdev_err(&hr_dev->ib_dev,
5894 			  "failed to process cmd when querying CQ, ret = %d.\n",
5895 			  ret);
5896 		goto err_mailbox;
5897 	}
5898 
5899 	memcpy(buffer, context, sizeof(*context));
5900 
5901 err_mailbox:
5902 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5903 
5904 	return ret;
5905 }
5906 
5907 static int hns_roce_v2_query_mpt(struct hns_roce_dev *hr_dev, u32 key,
5908 				 void *buffer)
5909 {
5910 	struct hns_roce_v2_mpt_entry *context;
5911 	struct hns_roce_cmd_mailbox *mailbox;
5912 	int ret;
5913 
5914 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
5915 	if (IS_ERR(mailbox))
5916 		return PTR_ERR(mailbox);
5917 
5918 	context = mailbox->buf;
5919 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
5920 				key_to_hw_index(key));
5921 	if (ret) {
5922 		ibdev_err(&hr_dev->ib_dev,
5923 			  "failed to process cmd when querying MPT, ret = %d.\n",
5924 			  ret);
5925 		goto err_mailbox;
5926 	}
5927 
5928 	memcpy(buffer, context, sizeof(*context));
5929 
5930 err_mailbox:
5931 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
5932 
5933 	return ret;
5934 }
5935 
5936 static void hns_roce_irq_work_handle(struct work_struct *work)
5937 {
5938 	struct hns_roce_work *irq_work =
5939 				container_of(work, struct hns_roce_work, work);
5940 	struct ib_device *ibdev = &irq_work->hr_dev->ib_dev;
5941 
5942 	switch (irq_work->event_type) {
5943 	case HNS_ROCE_EVENT_TYPE_PATH_MIG:
5944 		ibdev_info(ibdev, "path migrated succeeded.\n");
5945 		break;
5946 	case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
5947 		ibdev_warn(ibdev, "path migration failed.\n");
5948 		break;
5949 	case HNS_ROCE_EVENT_TYPE_COMM_EST:
5950 		break;
5951 	case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
5952 		ibdev_dbg(ibdev, "send queue drained.\n");
5953 		break;
5954 	case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
5955 		ibdev_err(ibdev, "local work queue 0x%x catast error, sub_event type is: %d\n",
5956 			  irq_work->queue_num, irq_work->sub_type);
5957 		break;
5958 	case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
5959 		ibdev_err(ibdev, "invalid request local work queue 0x%x error.\n",
5960 			  irq_work->queue_num);
5961 		break;
5962 	case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
5963 		ibdev_err(ibdev, "local access violation work queue 0x%x error, sub_event type is: %d\n",
5964 			  irq_work->queue_num, irq_work->sub_type);
5965 		break;
5966 	case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
5967 		ibdev_dbg(ibdev, "SRQ limit reach.\n");
5968 		break;
5969 	case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
5970 		ibdev_dbg(ibdev, "SRQ last wqe reach.\n");
5971 		break;
5972 	case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
5973 		ibdev_err(ibdev, "SRQ catas error.\n");
5974 		break;
5975 	case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
5976 		ibdev_err(ibdev, "CQ 0x%x access err.\n", irq_work->queue_num);
5977 		break;
5978 	case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
5979 		ibdev_warn(ibdev, "CQ 0x%x overflow\n", irq_work->queue_num);
5980 		break;
5981 	case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
5982 		ibdev_warn(ibdev, "DB overflow.\n");
5983 		break;
5984 	case HNS_ROCE_EVENT_TYPE_FLR:
5985 		ibdev_warn(ibdev, "function level reset.\n");
5986 		break;
5987 	case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
5988 		ibdev_err(ibdev, "xrc domain violation error.\n");
5989 		break;
5990 	case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
5991 		ibdev_err(ibdev, "invalid xrceth error.\n");
5992 		break;
5993 	default:
5994 		break;
5995 	}
5996 
5997 	kfree(irq_work);
5998 }
5999 
6000 static void hns_roce_v2_init_irq_work(struct hns_roce_dev *hr_dev,
6001 				      struct hns_roce_eq *eq, u32 queue_num)
6002 {
6003 	struct hns_roce_work *irq_work;
6004 
6005 	irq_work = kzalloc(sizeof(struct hns_roce_work), GFP_ATOMIC);
6006 	if (!irq_work)
6007 		return;
6008 
6009 	INIT_WORK(&irq_work->work, hns_roce_irq_work_handle);
6010 	irq_work->hr_dev = hr_dev;
6011 	irq_work->event_type = eq->event_type;
6012 	irq_work->sub_type = eq->sub_type;
6013 	irq_work->queue_num = queue_num;
6014 	queue_work(hr_dev->irq_workq, &irq_work->work);
6015 }
6016 
6017 static void update_eq_db(struct hns_roce_eq *eq)
6018 {
6019 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6020 	struct hns_roce_v2_db eq_db = {};
6021 
6022 	if (eq->type_flag == HNS_ROCE_AEQ) {
6023 		hr_reg_write(&eq_db, EQ_DB_CMD,
6024 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6025 			     HNS_ROCE_EQ_DB_CMD_AEQ :
6026 			     HNS_ROCE_EQ_DB_CMD_AEQ_ARMED);
6027 	} else {
6028 		hr_reg_write(&eq_db, EQ_DB_TAG, eq->eqn);
6029 
6030 		hr_reg_write(&eq_db, EQ_DB_CMD,
6031 			     eq->arm_st == HNS_ROCE_V2_EQ_ALWAYS_ARMED ?
6032 			     HNS_ROCE_EQ_DB_CMD_CEQ :
6033 			     HNS_ROCE_EQ_DB_CMD_CEQ_ARMED);
6034 	}
6035 
6036 	hr_reg_write(&eq_db, EQ_DB_CI, eq->cons_index);
6037 
6038 	hns_roce_write64(hr_dev, (__le32 *)&eq_db, eq->db_reg);
6039 }
6040 
6041 static struct hns_roce_aeqe *next_aeqe_sw_v2(struct hns_roce_eq *eq)
6042 {
6043 	struct hns_roce_aeqe *aeqe;
6044 
6045 	aeqe = hns_roce_buf_offset(eq->mtr.kmem,
6046 				   (eq->cons_index & (eq->entries - 1)) *
6047 				   eq->eqe_size);
6048 
6049 	return (hr_reg_read(aeqe, AEQE_OWNER) ^
6050 		!!(eq->cons_index & eq->entries)) ? aeqe : NULL;
6051 }
6052 
6053 static irqreturn_t hns_roce_v2_aeq_int(struct hns_roce_dev *hr_dev,
6054 				       struct hns_roce_eq *eq)
6055 {
6056 	struct device *dev = hr_dev->dev;
6057 	struct hns_roce_aeqe *aeqe = next_aeqe_sw_v2(eq);
6058 	irqreturn_t aeqe_found = IRQ_NONE;
6059 	int event_type;
6060 	u32 queue_num;
6061 	int sub_type;
6062 
6063 	while (aeqe) {
6064 		/* Make sure we read AEQ entry after we have checked the
6065 		 * ownership bit
6066 		 */
6067 		dma_rmb();
6068 
6069 		event_type = hr_reg_read(aeqe, AEQE_EVENT_TYPE);
6070 		sub_type = hr_reg_read(aeqe, AEQE_SUB_TYPE);
6071 		queue_num = hr_reg_read(aeqe, AEQE_EVENT_QUEUE_NUM);
6072 
6073 		switch (event_type) {
6074 		case HNS_ROCE_EVENT_TYPE_PATH_MIG:
6075 		case HNS_ROCE_EVENT_TYPE_PATH_MIG_FAILED:
6076 		case HNS_ROCE_EVENT_TYPE_COMM_EST:
6077 		case HNS_ROCE_EVENT_TYPE_SQ_DRAINED:
6078 		case HNS_ROCE_EVENT_TYPE_WQ_CATAS_ERROR:
6079 		case HNS_ROCE_EVENT_TYPE_SRQ_LAST_WQE_REACH:
6080 		case HNS_ROCE_EVENT_TYPE_INV_REQ_LOCAL_WQ_ERROR:
6081 		case HNS_ROCE_EVENT_TYPE_LOCAL_WQ_ACCESS_ERROR:
6082 		case HNS_ROCE_EVENT_TYPE_XRCD_VIOLATION:
6083 		case HNS_ROCE_EVENT_TYPE_INVALID_XRCETH:
6084 			hns_roce_qp_event(hr_dev, queue_num, event_type);
6085 			break;
6086 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
6087 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
6088 			hns_roce_srq_event(hr_dev, queue_num, event_type);
6089 			break;
6090 		case HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR:
6091 		case HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW:
6092 			hns_roce_cq_event(hr_dev, queue_num, event_type);
6093 			break;
6094 		case HNS_ROCE_EVENT_TYPE_MB:
6095 			hns_roce_cmd_event(hr_dev,
6096 					le16_to_cpu(aeqe->event.cmd.token),
6097 					aeqe->event.cmd.status,
6098 					le64_to_cpu(aeqe->event.cmd.out_param));
6099 			break;
6100 		case HNS_ROCE_EVENT_TYPE_DB_OVERFLOW:
6101 		case HNS_ROCE_EVENT_TYPE_FLR:
6102 			break;
6103 		default:
6104 			dev_err(dev, "unhandled event %d on EQ %d at idx %u.\n",
6105 				event_type, eq->eqn, eq->cons_index);
6106 			break;
6107 		}
6108 
6109 		eq->event_type = event_type;
6110 		eq->sub_type = sub_type;
6111 		++eq->cons_index;
6112 		aeqe_found = IRQ_HANDLED;
6113 
6114 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_AEQE_CNT]);
6115 
6116 		hns_roce_v2_init_irq_work(hr_dev, eq, queue_num);
6117 
6118 		aeqe = next_aeqe_sw_v2(eq);
6119 	}
6120 
6121 	update_eq_db(eq);
6122 
6123 	return IRQ_RETVAL(aeqe_found);
6124 }
6125 
6126 static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq)
6127 {
6128 	struct hns_roce_ceqe *ceqe;
6129 
6130 	ceqe = hns_roce_buf_offset(eq->mtr.kmem,
6131 				   (eq->cons_index & (eq->entries - 1)) *
6132 				   eq->eqe_size);
6133 
6134 	return (hr_reg_read(ceqe, CEQE_OWNER) ^
6135 		!!(eq->cons_index & eq->entries)) ? ceqe : NULL;
6136 }
6137 
6138 static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_dev *hr_dev,
6139 				       struct hns_roce_eq *eq)
6140 {
6141 	struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq);
6142 	irqreturn_t ceqe_found = IRQ_NONE;
6143 	u32 cqn;
6144 
6145 	while (ceqe) {
6146 		/* Make sure we read CEQ entry after we have checked the
6147 		 * ownership bit
6148 		 */
6149 		dma_rmb();
6150 
6151 		cqn = hr_reg_read(ceqe, CEQE_CQN);
6152 
6153 		hns_roce_cq_completion(hr_dev, cqn);
6154 
6155 		++eq->cons_index;
6156 		ceqe_found = IRQ_HANDLED;
6157 		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]);
6158 
6159 		ceqe = next_ceqe_sw_v2(eq);
6160 	}
6161 
6162 	update_eq_db(eq);
6163 
6164 	return IRQ_RETVAL(ceqe_found);
6165 }
6166 
6167 static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
6168 {
6169 	struct hns_roce_eq *eq = eq_ptr;
6170 	struct hns_roce_dev *hr_dev = eq->hr_dev;
6171 	irqreturn_t int_work;
6172 
6173 	if (eq->type_flag == HNS_ROCE_CEQ)
6174 		/* Completion event interrupt */
6175 		int_work = hns_roce_v2_ceq_int(hr_dev, eq);
6176 	else
6177 		/* Asynchronous event interrupt */
6178 		int_work = hns_roce_v2_aeq_int(hr_dev, eq);
6179 
6180 	return IRQ_RETVAL(int_work);
6181 }
6182 
6183 static irqreturn_t abnormal_interrupt_basic(struct hns_roce_dev *hr_dev,
6184 					    u32 int_st)
6185 {
6186 	struct pci_dev *pdev = hr_dev->pci_dev;
6187 	struct hnae3_ae_dev *ae_dev = pci_get_drvdata(pdev);
6188 	const struct hnae3_ae_ops *ops = ae_dev->ops;
6189 	irqreturn_t int_work = IRQ_NONE;
6190 	u32 int_en;
6191 
6192 	int_en = roce_read(hr_dev, ROCEE_VF_ABN_INT_EN_REG);
6193 
6194 	if (int_st & BIT(HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S)) {
6195 		dev_err(hr_dev->dev, "AEQ overflow!\n");
6196 
6197 		roce_write(hr_dev, ROCEE_VF_ABN_INT_ST_REG,
6198 			   1 << HNS_ROCE_V2_VF_INT_ST_AEQ_OVERFLOW_S);
6199 
6200 		/* Set reset level for reset_event() */
6201 		if (ops->set_default_reset_request)
6202 			ops->set_default_reset_request(ae_dev,
6203 						       HNAE3_FUNC_RESET);
6204 		if (ops->reset_event)
6205 			ops->reset_event(pdev, NULL);
6206 
6207 		int_en |= 1 << HNS_ROCE_V2_VF_ABN_INT_EN_S;
6208 		roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, int_en);
6209 
6210 		int_work = IRQ_HANDLED;
6211 	} else {
6212 		dev_err(hr_dev->dev, "there is no basic abn irq found.\n");
6213 	}
6214 
6215 	return IRQ_RETVAL(int_work);
6216 }
6217 
6218 static int fmea_ram_ecc_query(struct hns_roce_dev *hr_dev,
6219 			       struct fmea_ram_ecc *ecc_info)
6220 {
6221 	struct hns_roce_cmq_desc desc;
6222 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6223 	int ret;
6224 
6225 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_QUERY_RAM_ECC, true);
6226 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6227 	if (ret)
6228 		return ret;
6229 
6230 	ecc_info->is_ecc_err = hr_reg_read(req, QUERY_RAM_ECC_1BIT_ERR);
6231 	ecc_info->res_type = hr_reg_read(req, QUERY_RAM_ECC_RES_TYPE);
6232 	ecc_info->index = hr_reg_read(req, QUERY_RAM_ECC_TAG);
6233 
6234 	return 0;
6235 }
6236 
6237 static int fmea_recover_gmv(struct hns_roce_dev *hr_dev, u32 idx)
6238 {
6239 	struct hns_roce_cmq_desc desc;
6240 	struct hns_roce_cmq_req *req = (struct hns_roce_cmq_req *)desc.data;
6241 	u32 addr_upper;
6242 	u32 addr_low;
6243 	int ret;
6244 
6245 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, true);
6246 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6247 
6248 	ret = hns_roce_cmq_send(hr_dev, &desc, 1);
6249 	if (ret) {
6250 		dev_err(hr_dev->dev,
6251 			"failed to execute cmd to read gmv, ret = %d.\n", ret);
6252 		return ret;
6253 	}
6254 
6255 	addr_low =  hr_reg_read(req, CFG_GMV_BT_BA_L);
6256 	addr_upper = hr_reg_read(req, CFG_GMV_BT_BA_H);
6257 
6258 	hns_roce_cmq_setup_basic_desc(&desc, HNS_ROCE_OPC_CFG_GMV_BT, false);
6259 	hr_reg_write(req, CFG_GMV_BT_BA_L, addr_low);
6260 	hr_reg_write(req, CFG_GMV_BT_BA_H, addr_upper);
6261 	hr_reg_write(req, CFG_GMV_BT_IDX, idx);
6262 
6263 	return hns_roce_cmq_send(hr_dev, &desc, 1);
6264 }
6265 
6266 static u64 fmea_get_ram_res_addr(u32 res_type, __le64 *data)
6267 {
6268 	if (res_type == ECC_RESOURCE_QPC_TIMER ||
6269 	    res_type == ECC_RESOURCE_CQC_TIMER ||
6270 	    res_type == ECC_RESOURCE_SCCC)
6271 		return le64_to_cpu(*data);
6272 
6273 	return le64_to_cpu(*data) << PAGE_SHIFT;
6274 }
6275 
6276 static int fmea_recover_others(struct hns_roce_dev *hr_dev, u32 res_type,
6277 			       u32 index)
6278 {
6279 	u8 write_bt0_op = fmea_ram_res[res_type].write_bt0_op;
6280 	u8 read_bt0_op = fmea_ram_res[res_type].read_bt0_op;
6281 	struct hns_roce_cmd_mailbox *mailbox;
6282 	u64 addr;
6283 	int ret;
6284 
6285 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6286 	if (IS_ERR(mailbox))
6287 		return PTR_ERR(mailbox);
6288 
6289 	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, read_bt0_op, index);
6290 	if (ret) {
6291 		dev_err(hr_dev->dev,
6292 			"failed to execute cmd to read fmea ram, ret = %d.\n",
6293 			ret);
6294 		goto out;
6295 	}
6296 
6297 	addr = fmea_get_ram_res_addr(res_type, mailbox->buf);
6298 
6299 	ret = hns_roce_cmd_mbox(hr_dev, addr, 0, write_bt0_op, index);
6300 	if (ret)
6301 		dev_err(hr_dev->dev,
6302 			"failed to execute cmd to write fmea ram, ret = %d.\n",
6303 			ret);
6304 
6305 out:
6306 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6307 	return ret;
6308 }
6309 
6310 static void fmea_ram_ecc_recover(struct hns_roce_dev *hr_dev,
6311 				 struct fmea_ram_ecc *ecc_info)
6312 {
6313 	u32 res_type = ecc_info->res_type;
6314 	u32 index = ecc_info->index;
6315 	int ret;
6316 
6317 	BUILD_BUG_ON(ARRAY_SIZE(fmea_ram_res) != ECC_RESOURCE_COUNT);
6318 
6319 	if (res_type >= ECC_RESOURCE_COUNT) {
6320 		dev_err(hr_dev->dev, "unsupported fmea ram ecc type %u.\n",
6321 			res_type);
6322 		return;
6323 	}
6324 
6325 	if (res_type == ECC_RESOURCE_GMV)
6326 		ret = fmea_recover_gmv(hr_dev, index);
6327 	else
6328 		ret = fmea_recover_others(hr_dev, res_type, index);
6329 	if (ret)
6330 		dev_err(hr_dev->dev,
6331 			"failed to recover %s, index = %u, ret = %d.\n",
6332 			fmea_ram_res[res_type].name, index, ret);
6333 }
6334 
6335 static void fmea_ram_ecc_work(struct work_struct *ecc_work)
6336 {
6337 	struct hns_roce_dev *hr_dev =
6338 		container_of(ecc_work, struct hns_roce_dev, ecc_work);
6339 	struct fmea_ram_ecc ecc_info = {};
6340 
6341 	if (fmea_ram_ecc_query(hr_dev, &ecc_info)) {
6342 		dev_err(hr_dev->dev, "failed to query fmea ram ecc.\n");
6343 		return;
6344 	}
6345 
6346 	if (!ecc_info.is_ecc_err) {
6347 		dev_err(hr_dev->dev, "there is no fmea ram ecc err found.\n");
6348 		return;
6349 	}
6350 
6351 	fmea_ram_ecc_recover(hr_dev, &ecc_info);
6352 }
6353 
6354 static irqreturn_t hns_roce_v2_msix_interrupt_abn(int irq, void *dev_id)
6355 {
6356 	struct hns_roce_dev *hr_dev = dev_id;
6357 	irqreturn_t int_work = IRQ_NONE;
6358 	u32 int_st;
6359 
6360 	int_st = roce_read(hr_dev, ROCEE_VF_ABN_INT_ST_REG);
6361 
6362 	if (int_st) {
6363 		int_work = abnormal_interrupt_basic(hr_dev, int_st);
6364 	} else if (hr_dev->pci_dev->revision >= PCI_REVISION_ID_HIP09) {
6365 		queue_work(hr_dev->irq_workq, &hr_dev->ecc_work);
6366 		int_work = IRQ_HANDLED;
6367 	} else {
6368 		dev_err(hr_dev->dev, "there is no abnormal irq found.\n");
6369 	}
6370 
6371 	return IRQ_RETVAL(int_work);
6372 }
6373 
6374 static void hns_roce_v2_int_mask_enable(struct hns_roce_dev *hr_dev,
6375 					int eq_num, u32 enable_flag)
6376 {
6377 	int i;
6378 
6379 	for (i = 0; i < eq_num; i++)
6380 		roce_write(hr_dev, ROCEE_VF_EVENT_INT_EN_REG +
6381 			   i * EQ_REG_OFFSET, enable_flag);
6382 
6383 	roce_write(hr_dev, ROCEE_VF_ABN_INT_EN_REG, enable_flag);
6384 	roce_write(hr_dev, ROCEE_VF_ABN_INT_CFG_REG, enable_flag);
6385 }
6386 
6387 static void hns_roce_v2_destroy_eqc(struct hns_roce_dev *hr_dev, u32 eqn)
6388 {
6389 	struct device *dev = hr_dev->dev;
6390 	int ret;
6391 	u8 cmd;
6392 
6393 	if (eqn < hr_dev->caps.num_comp_vectors)
6394 		cmd = HNS_ROCE_CMD_DESTROY_CEQC;
6395 	else
6396 		cmd = HNS_ROCE_CMD_DESTROY_AEQC;
6397 
6398 	ret = hns_roce_destroy_hw_ctx(hr_dev, cmd, eqn & HNS_ROCE_V2_EQN_M);
6399 	if (ret)
6400 		dev_err(dev, "[mailbox cmd] destroy eqc(%u) failed.\n", eqn);
6401 }
6402 
6403 static void free_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6404 {
6405 	hns_roce_mtr_destroy(hr_dev, &eq->mtr);
6406 }
6407 
6408 static void init_eq_config(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6409 {
6410 	eq->db_reg = hr_dev->reg_base + ROCEE_VF_EQ_DB_CFG0_REG;
6411 	eq->cons_index = 0;
6412 	eq->over_ignore = HNS_ROCE_V2_EQ_OVER_IGNORE_0;
6413 	eq->coalesce = HNS_ROCE_V2_EQ_COALESCE_0;
6414 	eq->arm_st = HNS_ROCE_V2_EQ_ALWAYS_ARMED;
6415 	eq->shift = ilog2((unsigned int)eq->entries);
6416 }
6417 
6418 static int config_eqc(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq,
6419 		      void *mb_buf)
6420 {
6421 	u64 eqe_ba[MTT_MIN_COUNT] = { 0 };
6422 	struct hns_roce_eq_context *eqc;
6423 	u64 bt_ba = 0;
6424 	int ret;
6425 
6426 	eqc = mb_buf;
6427 	memset(eqc, 0, sizeof(struct hns_roce_eq_context));
6428 
6429 	init_eq_config(hr_dev, eq);
6430 
6431 	/* if not multi-hop, eqe buffer only use one trunk */
6432 	ret = hns_roce_mtr_find(hr_dev, &eq->mtr, 0, eqe_ba,
6433 				ARRAY_SIZE(eqe_ba));
6434 	if (ret) {
6435 		dev_err(hr_dev->dev, "failed to find EQE mtr, ret = %d\n", ret);
6436 		return ret;
6437 	}
6438 
6439 	bt_ba = hns_roce_get_mtr_ba(&eq->mtr);
6440 
6441 	hr_reg_write(eqc, EQC_EQ_ST, HNS_ROCE_V2_EQ_STATE_VALID);
6442 	hr_reg_write(eqc, EQC_EQE_HOP_NUM, eq->hop_num);
6443 	hr_reg_write(eqc, EQC_OVER_IGNORE, eq->over_ignore);
6444 	hr_reg_write(eqc, EQC_COALESCE, eq->coalesce);
6445 	hr_reg_write(eqc, EQC_ARM_ST, eq->arm_st);
6446 	hr_reg_write(eqc, EQC_EQN, eq->eqn);
6447 	hr_reg_write(eqc, EQC_EQE_CNT, HNS_ROCE_EQ_INIT_EQE_CNT);
6448 	hr_reg_write(eqc, EQC_EQE_BA_PG_SZ,
6449 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.ba_pg_shift));
6450 	hr_reg_write(eqc, EQC_EQE_BUF_PG_SZ,
6451 		     to_hr_hw_page_shift(eq->mtr.hem_cfg.buf_pg_shift));
6452 	hr_reg_write(eqc, EQC_EQ_PROD_INDX, HNS_ROCE_EQ_INIT_PROD_IDX);
6453 	hr_reg_write(eqc, EQC_EQ_MAX_CNT, eq->eq_max_cnt);
6454 
6455 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6456 		if (eq->eq_period * HNS_ROCE_CLOCK_ADJUST > USHRT_MAX) {
6457 			dev_info(hr_dev->dev, "eq_period(%u) reached the upper limit, adjusted to 65.\n",
6458 				 eq->eq_period);
6459 			eq->eq_period = HNS_ROCE_MAX_EQ_PERIOD;
6460 		}
6461 		eq->eq_period *= HNS_ROCE_CLOCK_ADJUST;
6462 	}
6463 
6464 	hr_reg_write(eqc, EQC_EQ_PERIOD, eq->eq_period);
6465 	hr_reg_write(eqc, EQC_EQE_REPORT_TIMER, HNS_ROCE_EQ_INIT_REPORT_TIMER);
6466 	hr_reg_write(eqc, EQC_EQE_BA_L, bt_ba >> 3);
6467 	hr_reg_write(eqc, EQC_EQE_BA_H, bt_ba >> 35);
6468 	hr_reg_write(eqc, EQC_SHIFT, eq->shift);
6469 	hr_reg_write(eqc, EQC_MSI_INDX, HNS_ROCE_EQ_INIT_MSI_IDX);
6470 	hr_reg_write(eqc, EQC_CUR_EQE_BA_L, eqe_ba[0] >> 12);
6471 	hr_reg_write(eqc, EQC_CUR_EQE_BA_M, eqe_ba[0] >> 28);
6472 	hr_reg_write(eqc, EQC_CUR_EQE_BA_H, eqe_ba[0] >> 60);
6473 	hr_reg_write(eqc, EQC_EQ_CONS_INDX, HNS_ROCE_EQ_INIT_CONS_IDX);
6474 	hr_reg_write(eqc, EQC_NEX_EQE_BA_L, eqe_ba[1] >> 12);
6475 	hr_reg_write(eqc, EQC_NEX_EQE_BA_H, eqe_ba[1] >> 44);
6476 	hr_reg_write(eqc, EQC_EQE_SIZE, eq->eqe_size == HNS_ROCE_V3_EQE_SIZE);
6477 
6478 	return 0;
6479 }
6480 
6481 static int alloc_eq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_eq *eq)
6482 {
6483 	struct hns_roce_buf_attr buf_attr = {};
6484 	int err;
6485 
6486 	if (hr_dev->caps.eqe_hop_num == HNS_ROCE_HOP_NUM_0)
6487 		eq->hop_num = 0;
6488 	else
6489 		eq->hop_num = hr_dev->caps.eqe_hop_num;
6490 
6491 	buf_attr.page_shift = hr_dev->caps.eqe_buf_pg_sz + PAGE_SHIFT;
6492 	buf_attr.region[0].size = eq->entries * eq->eqe_size;
6493 	buf_attr.region[0].hopnum = eq->hop_num;
6494 	buf_attr.region_count = 1;
6495 
6496 	err = hns_roce_mtr_create(hr_dev, &eq->mtr, &buf_attr,
6497 				  hr_dev->caps.eqe_ba_pg_sz + PAGE_SHIFT, NULL,
6498 				  0);
6499 	if (err)
6500 		dev_err(hr_dev->dev, "failed to alloc EQE mtr, err %d\n", err);
6501 
6502 	return err;
6503 }
6504 
6505 static int hns_roce_v2_create_eq(struct hns_roce_dev *hr_dev,
6506 				 struct hns_roce_eq *eq, u8 eq_cmd)
6507 {
6508 	struct hns_roce_cmd_mailbox *mailbox;
6509 	int ret;
6510 
6511 	/* Allocate mailbox memory */
6512 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
6513 	if (IS_ERR(mailbox))
6514 		return PTR_ERR(mailbox);
6515 
6516 	ret = alloc_eq_buf(hr_dev, eq);
6517 	if (ret)
6518 		goto free_cmd_mbox;
6519 
6520 	ret = config_eqc(hr_dev, eq, mailbox->buf);
6521 	if (ret)
6522 		goto err_cmd_mbox;
6523 
6524 	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, eq_cmd, eq->eqn);
6525 	if (ret) {
6526 		dev_err(hr_dev->dev, "[mailbox cmd] create eqc failed.\n");
6527 		goto err_cmd_mbox;
6528 	}
6529 
6530 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6531 
6532 	return 0;
6533 
6534 err_cmd_mbox:
6535 	free_eq_buf(hr_dev, eq);
6536 
6537 free_cmd_mbox:
6538 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
6539 
6540 	return ret;
6541 }
6542 
6543 static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
6544 				  int comp_num, int aeq_num, int other_num)
6545 {
6546 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6547 	int i, j;
6548 	int ret;
6549 
6550 	for (i = 0; i < irq_num; i++) {
6551 		hr_dev->irq_names[i] = kzalloc(HNS_ROCE_INT_NAME_LEN,
6552 					       GFP_KERNEL);
6553 		if (!hr_dev->irq_names[i]) {
6554 			ret = -ENOMEM;
6555 			goto err_kzalloc_failed;
6556 		}
6557 	}
6558 
6559 	/* irq contains: abnormal + AEQ + CEQ */
6560 	for (j = 0; j < other_num; j++)
6561 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6562 			 "hns-%s-abn-%d", pci_name(hr_dev->pci_dev), j);
6563 
6564 	for (j = other_num; j < (other_num + aeq_num); j++)
6565 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6566 			 "hns-%s-aeq-%d", pci_name(hr_dev->pci_dev), j - other_num);
6567 
6568 	for (j = (other_num + aeq_num); j < irq_num; j++)
6569 		snprintf((char *)hr_dev->irq_names[j], HNS_ROCE_INT_NAME_LEN,
6570 			 "hns-%s-ceq-%d", pci_name(hr_dev->pci_dev),
6571 			 j - other_num - aeq_num);
6572 
6573 	for (j = 0; j < irq_num; j++) {
6574 		if (j < other_num)
6575 			ret = request_irq(hr_dev->irq[j],
6576 					  hns_roce_v2_msix_interrupt_abn,
6577 					  0, hr_dev->irq_names[j], hr_dev);
6578 
6579 		else if (j < (other_num + comp_num))
6580 			ret = request_irq(eq_table->eq[j - other_num].irq,
6581 					  hns_roce_v2_msix_interrupt_eq,
6582 					  0, hr_dev->irq_names[j + aeq_num],
6583 					  &eq_table->eq[j - other_num]);
6584 		else
6585 			ret = request_irq(eq_table->eq[j - other_num].irq,
6586 					  hns_roce_v2_msix_interrupt_eq,
6587 					  0, hr_dev->irq_names[j - comp_num],
6588 					  &eq_table->eq[j - other_num]);
6589 		if (ret) {
6590 			dev_err(hr_dev->dev, "request irq error!\n");
6591 			goto err_request_failed;
6592 		}
6593 	}
6594 
6595 	return 0;
6596 
6597 err_request_failed:
6598 	for (j -= 1; j >= 0; j--)
6599 		if (j < other_num)
6600 			free_irq(hr_dev->irq[j], hr_dev);
6601 		else
6602 			free_irq(eq_table->eq[j - other_num].irq,
6603 				 &eq_table->eq[j - other_num]);
6604 
6605 err_kzalloc_failed:
6606 	for (i -= 1; i >= 0; i--)
6607 		kfree(hr_dev->irq_names[i]);
6608 
6609 	return ret;
6610 }
6611 
6612 static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev)
6613 {
6614 	int irq_num;
6615 	int eq_num;
6616 	int i;
6617 
6618 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6619 	irq_num = eq_num + hr_dev->caps.num_other_vectors;
6620 
6621 	for (i = 0; i < hr_dev->caps.num_other_vectors; i++)
6622 		free_irq(hr_dev->irq[i], hr_dev);
6623 
6624 	for (i = 0; i < eq_num; i++)
6625 		free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]);
6626 
6627 	for (i = 0; i < irq_num; i++)
6628 		kfree(hr_dev->irq_names[i]);
6629 }
6630 
6631 static int hns_roce_v2_init_eq_table(struct hns_roce_dev *hr_dev)
6632 {
6633 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6634 	struct device *dev = hr_dev->dev;
6635 	struct hns_roce_eq *eq;
6636 	int other_num;
6637 	int comp_num;
6638 	int aeq_num;
6639 	int irq_num;
6640 	int eq_num;
6641 	u8 eq_cmd;
6642 	int ret;
6643 	int i;
6644 
6645 	other_num = hr_dev->caps.num_other_vectors;
6646 	comp_num = hr_dev->caps.num_comp_vectors;
6647 	aeq_num = hr_dev->caps.num_aeq_vectors;
6648 
6649 	eq_num = comp_num + aeq_num;
6650 	irq_num = eq_num + other_num;
6651 
6652 	eq_table->eq = kcalloc(eq_num, sizeof(*eq_table->eq), GFP_KERNEL);
6653 	if (!eq_table->eq)
6654 		return -ENOMEM;
6655 
6656 	/* create eq */
6657 	for (i = 0; i < eq_num; i++) {
6658 		eq = &eq_table->eq[i];
6659 		eq->hr_dev = hr_dev;
6660 		eq->eqn = i;
6661 		if (i < comp_num) {
6662 			/* CEQ */
6663 			eq_cmd = HNS_ROCE_CMD_CREATE_CEQC;
6664 			eq->type_flag = HNS_ROCE_CEQ;
6665 			eq->entries = hr_dev->caps.ceqe_depth;
6666 			eq->eqe_size = hr_dev->caps.ceqe_size;
6667 			eq->irq = hr_dev->irq[i + other_num + aeq_num];
6668 			eq->eq_max_cnt = HNS_ROCE_CEQ_DEFAULT_BURST_NUM;
6669 			eq->eq_period = HNS_ROCE_CEQ_DEFAULT_INTERVAL;
6670 		} else {
6671 			/* AEQ */
6672 			eq_cmd = HNS_ROCE_CMD_CREATE_AEQC;
6673 			eq->type_flag = HNS_ROCE_AEQ;
6674 			eq->entries = hr_dev->caps.aeqe_depth;
6675 			eq->eqe_size = hr_dev->caps.aeqe_size;
6676 			eq->irq = hr_dev->irq[i - comp_num + other_num];
6677 			eq->eq_max_cnt = HNS_ROCE_AEQ_DEFAULT_BURST_NUM;
6678 			eq->eq_period = HNS_ROCE_AEQ_DEFAULT_INTERVAL;
6679 		}
6680 
6681 		ret = hns_roce_v2_create_eq(hr_dev, eq, eq_cmd);
6682 		if (ret) {
6683 			dev_err(dev, "failed to create eq.\n");
6684 			goto err_create_eq_fail;
6685 		}
6686 	}
6687 
6688 	INIT_WORK(&hr_dev->ecc_work, fmea_ram_ecc_work);
6689 
6690 	hr_dev->irq_workq = alloc_ordered_workqueue("hns_roce_irq_workq", 0);
6691 	if (!hr_dev->irq_workq) {
6692 		dev_err(dev, "failed to create irq workqueue.\n");
6693 		ret = -ENOMEM;
6694 		goto err_create_eq_fail;
6695 	}
6696 
6697 	ret = __hns_roce_request_irq(hr_dev, irq_num, comp_num, aeq_num,
6698 				     other_num);
6699 	if (ret) {
6700 		dev_err(dev, "failed to request irq.\n");
6701 		goto err_request_irq_fail;
6702 	}
6703 
6704 	/* enable irq */
6705 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_ENABLE);
6706 
6707 	return 0;
6708 
6709 err_request_irq_fail:
6710 	destroy_workqueue(hr_dev->irq_workq);
6711 
6712 err_create_eq_fail:
6713 	for (i -= 1; i >= 0; i--)
6714 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6715 	kfree(eq_table->eq);
6716 
6717 	return ret;
6718 }
6719 
6720 static void hns_roce_v2_cleanup_eq_table(struct hns_roce_dev *hr_dev)
6721 {
6722 	struct hns_roce_eq_table *eq_table = &hr_dev->eq_table;
6723 	int eq_num;
6724 	int i;
6725 
6726 	eq_num = hr_dev->caps.num_comp_vectors + hr_dev->caps.num_aeq_vectors;
6727 
6728 	/* Disable irq */
6729 	hns_roce_v2_int_mask_enable(hr_dev, eq_num, EQ_DISABLE);
6730 
6731 	__hns_roce_free_irq(hr_dev);
6732 	destroy_workqueue(hr_dev->irq_workq);
6733 
6734 	for (i = 0; i < eq_num; i++) {
6735 		hns_roce_v2_destroy_eqc(hr_dev, i);
6736 
6737 		free_eq_buf(hr_dev, &eq_table->eq[i]);
6738 	}
6739 
6740 	kfree(eq_table->eq);
6741 }
6742 
6743 static const struct ib_device_ops hns_roce_v2_dev_ops = {
6744 	.destroy_qp = hns_roce_v2_destroy_qp,
6745 	.modify_cq = hns_roce_v2_modify_cq,
6746 	.poll_cq = hns_roce_v2_poll_cq,
6747 	.post_recv = hns_roce_v2_post_recv,
6748 	.post_send = hns_roce_v2_post_send,
6749 	.query_qp = hns_roce_v2_query_qp,
6750 	.req_notify_cq = hns_roce_v2_req_notify_cq,
6751 };
6752 
6753 static const struct ib_device_ops hns_roce_v2_dev_srq_ops = {
6754 	.modify_srq = hns_roce_v2_modify_srq,
6755 	.post_srq_recv = hns_roce_v2_post_srq_recv,
6756 	.query_srq = hns_roce_v2_query_srq,
6757 };
6758 
6759 static const struct hns_roce_hw hns_roce_hw_v2 = {
6760 	.cmq_init = hns_roce_v2_cmq_init,
6761 	.cmq_exit = hns_roce_v2_cmq_exit,
6762 	.hw_profile = hns_roce_v2_profile,
6763 	.hw_init = hns_roce_v2_init,
6764 	.hw_exit = hns_roce_v2_exit,
6765 	.post_mbox = v2_post_mbox,
6766 	.poll_mbox_done = v2_poll_mbox_done,
6767 	.chk_mbox_avail = v2_chk_mbox_is_avail,
6768 	.set_gid = hns_roce_v2_set_gid,
6769 	.set_mac = hns_roce_v2_set_mac,
6770 	.write_mtpt = hns_roce_v2_write_mtpt,
6771 	.rereg_write_mtpt = hns_roce_v2_rereg_write_mtpt,
6772 	.frmr_write_mtpt = hns_roce_v2_frmr_write_mtpt,
6773 	.mw_write_mtpt = hns_roce_v2_mw_write_mtpt,
6774 	.write_cqc = hns_roce_v2_write_cqc,
6775 	.set_hem = hns_roce_v2_set_hem,
6776 	.clear_hem = hns_roce_v2_clear_hem,
6777 	.modify_qp = hns_roce_v2_modify_qp,
6778 	.dereg_mr = hns_roce_v2_dereg_mr,
6779 	.qp_flow_control_init = hns_roce_v2_qp_flow_control_init,
6780 	.init_eq = hns_roce_v2_init_eq_table,
6781 	.cleanup_eq = hns_roce_v2_cleanup_eq_table,
6782 	.write_srqc = hns_roce_v2_write_srqc,
6783 	.query_cqc = hns_roce_v2_query_cqc,
6784 	.query_qpc = hns_roce_v2_query_qpc,
6785 	.query_mpt = hns_roce_v2_query_mpt,
6786 	.query_srqc = hns_roce_v2_query_srqc,
6787 	.query_sccc = hns_roce_v2_query_sccc,
6788 	.query_hw_counter = hns_roce_hw_v2_query_counter,
6789 	.get_dscp = hns_roce_hw_v2_get_dscp,
6790 	.hns_roce_dev_ops = &hns_roce_v2_dev_ops,
6791 	.hns_roce_dev_srq_ops = &hns_roce_v2_dev_srq_ops,
6792 };
6793 
6794 static const struct pci_device_id hns_roce_hw_v2_pci_tbl[] = {
6795 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA), 0},
6796 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_25GE_RDMA_MACSEC), 0},
6797 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA), 0},
6798 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_50GE_RDMA_MACSEC), 0},
6799 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_100G_RDMA_MACSEC), 0},
6800 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_200G_RDMA), 0},
6801 	{PCI_VDEVICE(HUAWEI, HNAE3_DEV_ID_RDMA_DCB_PFC_VF),
6802 	 HNAE3_DEV_SUPPORT_ROCE_DCB_BITS},
6803 	/* required last entry */
6804 	{0, }
6805 };
6806 
6807 MODULE_DEVICE_TABLE(pci, hns_roce_hw_v2_pci_tbl);
6808 
6809 static void hns_roce_hw_v2_get_cfg(struct hns_roce_dev *hr_dev,
6810 				  struct hnae3_handle *handle)
6811 {
6812 	struct hns_roce_v2_priv *priv = hr_dev->priv;
6813 	const struct pci_device_id *id;
6814 	int i;
6815 
6816 	hr_dev->pci_dev = handle->pdev;
6817 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, hr_dev->pci_dev);
6818 	hr_dev->is_vf = id->driver_data;
6819 	hr_dev->dev = &handle->pdev->dev;
6820 	hr_dev->hw = &hns_roce_hw_v2;
6821 	hr_dev->sdb_offset = ROCEE_DB_SQ_L_0_REG;
6822 	hr_dev->odb_offset = hr_dev->sdb_offset;
6823 
6824 	/* Get info from NIC driver. */
6825 	hr_dev->reg_base = handle->rinfo.roce_io_base;
6826 	hr_dev->mem_base = handle->rinfo.roce_mem_base;
6827 	hr_dev->caps.num_ports = 1;
6828 	hr_dev->iboe.netdevs[0] = handle->rinfo.netdev;
6829 	hr_dev->iboe.phy_port[0] = 0;
6830 
6831 	addrconf_addr_eui48((u8 *)&hr_dev->ib_dev.node_guid,
6832 			    hr_dev->iboe.netdevs[0]->dev_addr);
6833 
6834 	for (i = 0; i < handle->rinfo.num_vectors; i++)
6835 		hr_dev->irq[i] = pci_irq_vector(handle->pdev,
6836 						i + handle->rinfo.base_vector);
6837 
6838 	/* cmd issue mode: 0 is poll, 1 is event */
6839 	hr_dev->cmd_mod = 1;
6840 	hr_dev->loop_idc = 0;
6841 
6842 	hr_dev->reset_cnt = handle->ae_algo->ops->ae_dev_reset_cnt(handle);
6843 	priv->handle = handle;
6844 }
6845 
6846 static int __hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6847 {
6848 	struct hns_roce_dev *hr_dev;
6849 	int ret;
6850 
6851 	hr_dev = ib_alloc_device(hns_roce_dev, ib_dev);
6852 	if (!hr_dev)
6853 		return -ENOMEM;
6854 
6855 	hr_dev->priv = kzalloc(sizeof(struct hns_roce_v2_priv), GFP_KERNEL);
6856 	if (!hr_dev->priv) {
6857 		ret = -ENOMEM;
6858 		goto error_failed_kzalloc;
6859 	}
6860 
6861 	hns_roce_hw_v2_get_cfg(hr_dev, handle);
6862 
6863 	ret = hns_roce_init(hr_dev);
6864 	if (ret) {
6865 		dev_err(hr_dev->dev, "RoCE Engine init failed!\n");
6866 		goto error_failed_roce_init;
6867 	}
6868 
6869 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08) {
6870 		ret = free_mr_init(hr_dev);
6871 		if (ret) {
6872 			dev_err(hr_dev->dev, "failed to init free mr!\n");
6873 			goto error_failed_free_mr_init;
6874 		}
6875 	}
6876 
6877 	handle->priv = hr_dev;
6878 
6879 	return 0;
6880 
6881 error_failed_free_mr_init:
6882 	hns_roce_exit(hr_dev);
6883 
6884 error_failed_roce_init:
6885 	kfree(hr_dev->priv);
6886 
6887 error_failed_kzalloc:
6888 	ib_dealloc_device(&hr_dev->ib_dev);
6889 
6890 	return ret;
6891 }
6892 
6893 static void __hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6894 					   bool reset)
6895 {
6896 	struct hns_roce_dev *hr_dev = handle->priv;
6897 
6898 	if (!hr_dev)
6899 		return;
6900 
6901 	handle->priv = NULL;
6902 
6903 	hr_dev->state = HNS_ROCE_DEVICE_STATE_UNINIT;
6904 	hns_roce_handle_device_err(hr_dev);
6905 
6906 	if (hr_dev->pci_dev->revision == PCI_REVISION_ID_HIP08)
6907 		free_mr_exit(hr_dev);
6908 
6909 	hns_roce_exit(hr_dev);
6910 	kfree(hr_dev->priv);
6911 	ib_dealloc_device(&hr_dev->ib_dev);
6912 }
6913 
6914 static int hns_roce_hw_v2_init_instance(struct hnae3_handle *handle)
6915 {
6916 	const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
6917 	const struct pci_device_id *id;
6918 	struct device *dev = &handle->pdev->dev;
6919 	int ret;
6920 
6921 	handle->rinfo.instance_state = HNS_ROCE_STATE_INIT;
6922 
6923 	if (ops->ae_dev_resetting(handle) || ops->get_hw_reset_stat(handle)) {
6924 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6925 		goto reset_chk_err;
6926 	}
6927 
6928 	id = pci_match_id(hns_roce_hw_v2_pci_tbl, handle->pdev);
6929 	if (!id)
6930 		return 0;
6931 
6932 	if (id->driver_data && handle->pdev->revision == PCI_REVISION_ID_HIP08)
6933 		return 0;
6934 
6935 	ret = __hns_roce_hw_v2_init_instance(handle);
6936 	if (ret) {
6937 		handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6938 		dev_err(dev, "RoCE instance init failed! ret = %d\n", ret);
6939 		if (ops->ae_dev_resetting(handle) ||
6940 		    ops->get_hw_reset_stat(handle))
6941 			goto reset_chk_err;
6942 		else
6943 			return ret;
6944 	}
6945 
6946 	handle->rinfo.instance_state = HNS_ROCE_STATE_INITED;
6947 
6948 	return 0;
6949 
6950 reset_chk_err:
6951 	dev_err(dev, "Device is busy in resetting state.\n"
6952 		     "please retry later.\n");
6953 
6954 	return -EBUSY;
6955 }
6956 
6957 static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
6958 					   bool reset)
6959 {
6960 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED)
6961 		return;
6962 
6963 	handle->rinfo.instance_state = HNS_ROCE_STATE_UNINIT;
6964 
6965 	__hns_roce_hw_v2_uninit_instance(handle, reset);
6966 
6967 	handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
6968 }
6969 static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
6970 {
6971 	struct hns_roce_dev *hr_dev;
6972 
6973 	if (handle->rinfo.instance_state != HNS_ROCE_STATE_INITED) {
6974 		set_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6975 		return 0;
6976 	}
6977 
6978 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_DOWN;
6979 	clear_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state);
6980 
6981 	hr_dev = handle->priv;
6982 	if (!hr_dev)
6983 		return 0;
6984 
6985 	hr_dev->active = false;
6986 	hr_dev->dis_db = true;
6987 	hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
6988 
6989 	return 0;
6990 }
6991 
6992 static int hns_roce_hw_v2_reset_notify_init(struct hnae3_handle *handle)
6993 {
6994 	struct device *dev = &handle->pdev->dev;
6995 	int ret;
6996 
6997 	if (test_and_clear_bit(HNS_ROCE_RST_DIRECT_RETURN,
6998 			       &handle->rinfo.state)) {
6999 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7000 		return 0;
7001 	}
7002 
7003 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INIT;
7004 
7005 	dev_info(&handle->pdev->dev, "In reset process RoCE client reinit.\n");
7006 	ret = __hns_roce_hw_v2_init_instance(handle);
7007 	if (ret) {
7008 		/* when reset notify type is HNAE3_INIT_CLIENT In reset notify
7009 		 * callback function, RoCE Engine reinitialize. If RoCE reinit
7010 		 * failed, we should inform NIC driver.
7011 		 */
7012 		handle->priv = NULL;
7013 		dev_err(dev, "In reset process RoCE reinit failed %d.\n", ret);
7014 	} else {
7015 		handle->rinfo.reset_state = HNS_ROCE_STATE_RST_INITED;
7016 		dev_info(dev, "reset done, RoCE client reinit finished.\n");
7017 	}
7018 
7019 	return ret;
7020 }
7021 
7022 static int hns_roce_hw_v2_reset_notify_uninit(struct hnae3_handle *handle)
7023 {
7024 	if (test_bit(HNS_ROCE_RST_DIRECT_RETURN, &handle->rinfo.state))
7025 		return 0;
7026 
7027 	handle->rinfo.reset_state = HNS_ROCE_STATE_RST_UNINIT;
7028 	dev_info(&handle->pdev->dev, "In reset process RoCE client uninit.\n");
7029 	msleep(HNS_ROCE_V2_HW_RST_UNINT_DELAY);
7030 	__hns_roce_hw_v2_uninit_instance(handle, false);
7031 
7032 	return 0;
7033 }
7034 
7035 static int hns_roce_hw_v2_reset_notify(struct hnae3_handle *handle,
7036 				       enum hnae3_reset_notify_type type)
7037 {
7038 	int ret = 0;
7039 
7040 	switch (type) {
7041 	case HNAE3_DOWN_CLIENT:
7042 		ret = hns_roce_hw_v2_reset_notify_down(handle);
7043 		break;
7044 	case HNAE3_INIT_CLIENT:
7045 		ret = hns_roce_hw_v2_reset_notify_init(handle);
7046 		break;
7047 	case HNAE3_UNINIT_CLIENT:
7048 		ret = hns_roce_hw_v2_reset_notify_uninit(handle);
7049 		break;
7050 	default:
7051 		break;
7052 	}
7053 
7054 	return ret;
7055 }
7056 
7057 static const struct hnae3_client_ops hns_roce_hw_v2_ops = {
7058 	.init_instance = hns_roce_hw_v2_init_instance,
7059 	.uninit_instance = hns_roce_hw_v2_uninit_instance,
7060 	.reset_notify = hns_roce_hw_v2_reset_notify,
7061 };
7062 
7063 static struct hnae3_client hns_roce_hw_v2_client = {
7064 	.name = "hns_roce_hw_v2",
7065 	.type = HNAE3_CLIENT_ROCE,
7066 	.ops = &hns_roce_hw_v2_ops,
7067 };
7068 
7069 static int __init hns_roce_hw_v2_init(void)
7070 {
7071 	hns_roce_init_debugfs();
7072 	return hnae3_register_client(&hns_roce_hw_v2_client);
7073 }
7074 
7075 static void __exit hns_roce_hw_v2_exit(void)
7076 {
7077 	hnae3_unregister_client(&hns_roce_hw_v2_client);
7078 	hns_roce_cleanup_debugfs();
7079 }
7080 
7081 module_init(hns_roce_hw_v2_init);
7082 module_exit(hns_roce_hw_v2_exit);
7083 
7084 MODULE_LICENSE("Dual BSD/GPL");
7085 MODULE_AUTHOR("Wei Hu <xavier.huwei@huawei.com>");
7086 MODULE_AUTHOR("Lijun Ou <oulijun@huawei.com>");
7087 MODULE_AUTHOR("Shaobo Xu <xushaobo2@huawei.com>");
7088 MODULE_DESCRIPTION("Hisilicon Hip08 Family RoCE Driver");
7089